program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1.2"}})]
{
    func embed_128<ios18>(tensor<fp16, [1, 128]> attention_mask, tensor<fp16, [1, 128, 512]> inputs_embeds) {
            tensor<fp16, [12288, 512]> tm_ple_per_layer_model_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [12288, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1728))), scale = tensor<fp16, [12288, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6293248))))[name = string("tm_ple_per_layer_model_projection_weight_to_fp16_quantized")];
            tensor<fp16, [12288]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [12288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6317888)))];
            tensor<fp16, [1, 128, 12288]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = tm_ple_per_layer_model_projection_weight_to_fp16_quantized, x = inputs_embeds)[name = string("linear_0_cast_fp16")];
            fp16 var_2129_to_fp16 = const()[name = string("op_2129_to_fp16"), val = fp16(0x1.6ap-5)];
            tensor<fp16, [1, 128, 12288]> ple_1_cast_fp16 = mul(x = linear_0_cast_fp16, y = var_2129_to_fp16)[name = string("ple_1_cast_fp16")];
            tensor<int32, [4]> var_2135 = const()[name = string("op_2135"), val = tensor<int32, [4]>([1, -1, 24, 512])];
            tensor<fp16, [1, 128, 24, 512]> x_1_cast_fp16 = reshape(shape = var_2135, x = ple_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 128, 24, 512]> var_2137_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_2137_cast_fp16")];
            bool var_2140 = const()[name = string("op_2140"), val = bool(true)];
            tensor<int32, [1]> var_2141_axes_0 = const()[name = string("op_2141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 24, 1]> var_2141_cast_fp16 = reduce_max(axes = var_2141_axes_0, keep_dims = var_2140, x = var_2137_cast_fp16)[name = string("op_2141_cast_fp16")];
            fp16 var_2142_to_fp16 = const()[name = string("op_2142_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 24, 1]> s_1_cast_fp16 = maximum(x = var_2141_cast_fp16, y = var_2142_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 128, 24, 512]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 128, 24, 512]> var_2145_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [1]> var_2150_axes_0 = const()[name = string("op_2150_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2150_keep_dims_0 = const()[name = string("op_2150_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 24, 1]> var_2150_cast_fp16 = reduce_mean(axes = var_2150_axes_0, keep_dims = var_2150_keep_dims_0, x = var_2145_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 128, 24, 1]> var_2151_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 24, 1]> var_2152_cast_fp16 = inverse(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            fp16 var_2153_to_fp16 = const()[name = string("op_2153_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 24, 1]> var_2154_cast_fp16 = mul(x = var_2152_cast_fp16, y = var_2153_to_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 128, 24, 1]> var_2156_cast_fp16 = add(x = var_2150_cast_fp16, y = var_2154_cast_fp16)[name = string("op_2156_cast_fp16")];
            fp32 var_2157_epsilon_0 = const()[name = string("op_2157_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 24, 1]> var_2157_cast_fp16 = rsqrt(epsilon = var_2157_epsilon_0, x = var_2156_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<fp16, [1, 128, 24, 512]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_2157_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [512]> tm_ple_per_layer_projection_norm_weight_to_fp16 = const()[name = string("tm_ple_per_layer_projection_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6342528)))];
            tensor<fp16, [1, 128, 24, 512]> ple_3_cast_fp16 = mul(x = out_1_cast_fp16, y = tm_ple_per_layer_projection_norm_weight_to_fp16)[name = string("ple_3_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 128]> var_2162_cast_fp16 = sub(x = var_2160_to_fp16, y = attention_mask)[name = string("op_2162_cast_fp16")];
            fp16 var_2163_to_fp16 = const()[name = string("op_2163_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 128]> var_2164_cast_fp16 = mul(x = var_2162_cast_fp16, y = var_2163_to_fp16)[name = string("op_2164_cast_fp16")];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 128]> var_2171_cast_fp16 = expand_dims(axes = var_2171_axes_0, x = var_2164_cast_fp16)[name = string("op_2171_cast_fp16")];
            tensor<int32, [1]> var_2173_axes_0 = const()[name = string("op_2173_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 128]> var_2173_cast_fp16 = expand_dims(axes = var_2173_axes_0, x = var_2171_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2179_cast_fp16 = abs(x = inputs_embeds)[name = string("op_2179_cast_fp16")];
            bool var_2182 = const()[name = string("op_2182"), val = bool(true)];
            tensor<int32, [1]> var_2183_axes_0 = const()[name = string("op_2183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2183_cast_fp16 = reduce_max(axes = var_2183_axes_0, keep_dims = var_2182, x = var_2179_cast_fp16)[name = string("op_2183_cast_fp16")];
            fp16 var_2184_to_fp16 = const()[name = string("op_2184_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_3_cast_fp16 = maximum(x = var_2183_cast_fp16, y = var_2184_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_3_cast_fp16 = real_div(x = inputs_embeds, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2187_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> var_2192_axes_0 = const()[name = string("op_2192_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2192_keep_dims_0 = const()[name = string("op_2192_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2192_cast_fp16 = reduce_mean(axes = var_2192_axes_0, keep_dims = var_2192_keep_dims_0, x = var_2187_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2193_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp32 var_2194_epsilon_0 = const()[name = string("op_2194_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2194_cast_fp16 = inverse(epsilon = var_2194_epsilon_0, x = var_2193_cast_fp16)[name = string("op_2194_cast_fp16")];
            fp16 var_2195_to_fp16 = const()[name = string("op_2195_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2196_cast_fp16 = mul(x = var_2194_cast_fp16, y = var_2195_to_fp16)[name = string("op_2196_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2198_cast_fp16 = add(x = var_2192_cast_fp16, y = var_2196_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_2199_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6343616)))];
            tensor<fp16, [1, 128, 512]> input_1_cast_fp16 = mul(x = out_3_cast_fp16, y = tm_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6344704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6869056))))[name = string("tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6871168)))];
            tensor<fp16, [1, 128, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_1_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6873280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7135488))))[name = string("tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7136576)))];
            tensor<fp16, [1, 128, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_2221 = const()[name = string("op_2221"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_2222_cast_fp16 = reshape(shape = var_2221, x = linear_2_cast_fp16)[name = string("op_2222_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7137664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7399872))))[name = string("tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_2234_cast_fp16 = reshape(shape = var_2233, x = linear_3_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<int32, [4]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_2210_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 4, 128, 256]> var_2238_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_2238_cast_fp16")];
            bool var_2241 = const()[name = string("op_2241"), val = bool(true)];
            tensor<int32, [1]> var_2242_axes_0 = const()[name = string("op_2242_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_2242_cast_fp16 = reduce_max(axes = var_2242_axes_0, keep_dims = var_2241, x = var_2238_cast_fp16)[name = string("op_2242_cast_fp16")];
            fp16 var_2243_to_fp16 = const()[name = string("op_2243_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_5_cast_fp16 = maximum(x = var_2242_cast_fp16, y = var_2243_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_2246_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<int32, [1]> var_2251_axes_0 = const()[name = string("op_2251_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2251_keep_dims_0 = const()[name = string("op_2251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_2251_cast_fp16 = reduce_mean(axes = var_2251_axes_0, keep_dims = var_2251_keep_dims_0, x = var_2246_cast_fp16)[name = string("op_2251_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_2252_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_2252_cast_fp16")];
            fp32 var_2253_epsilon_0 = const()[name = string("op_2253_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_2253_cast_fp16 = inverse(epsilon = var_2253_epsilon_0, x = var_2252_cast_fp16)[name = string("op_2253_cast_fp16")];
            fp16 var_2254_to_fp16 = const()[name = string("op_2254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_2255_cast_fp16 = mul(x = var_2253_cast_fp16, y = var_2254_to_fp16)[name = string("op_2255_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_2257_cast_fp16 = add(x = var_2251_cast_fp16, y = var_2255_cast_fp16)[name = string("op_2257_cast_fp16")];
            fp32 var_2258_epsilon_0 = const()[name = string("op_2258_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_2258_cast_fp16 = rsqrt(epsilon = var_2258_epsilon_0, x = var_2257_cast_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_2258_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7400960)))];
            tensor<fp16, [1, 4, 128, 256]> q_1_cast_fp16 = mul(x = out_5_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_2222_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 2, 128, 256]> var_2261_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_2261_cast_fp16")];
            bool var_2264 = const()[name = string("op_2264"), val = bool(true)];
            tensor<int32, [1]> var_2265_axes_0 = const()[name = string("op_2265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_2265_cast_fp16 = reduce_max(axes = var_2265_axes_0, keep_dims = var_2264, x = var_2261_cast_fp16)[name = string("op_2265_cast_fp16")];
            fp16 var_2266_to_fp16 = const()[name = string("op_2266_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_7_cast_fp16 = maximum(x = var_2265_cast_fp16, y = var_2266_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2269_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> var_2274_axes_0 = const()[name = string("op_2274_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2274_keep_dims_0 = const()[name = string("op_2274_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_2274_cast_fp16 = reduce_mean(axes = var_2274_axes_0, keep_dims = var_2274_keep_dims_0, x = var_2269_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2275_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_2276_cast_fp16 = inverse(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            fp16 var_2277_to_fp16 = const()[name = string("op_2277_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_2278_cast_fp16 = mul(x = var_2276_cast_fp16, y = var_2277_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2280_cast_fp16 = add(x = var_2274_cast_fp16, y = var_2278_cast_fp16)[name = string("op_2280_cast_fp16")];
            fp32 var_2281_epsilon_0 = const()[name = string("op_2281_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_2281_cast_fp16 = rsqrt(epsilon = var_2281_epsilon_0, x = var_2280_cast_fp16)[name = string("op_2281_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_7_cast_fp16 = mul(x = y_7_cast_fp16, y = var_2281_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7401536)))];
            tensor<fp16, [1, 2, 128, 256]> k_1_cast_fp16 = mul(x = out_7_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = var_2234_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2, 128, 256]> var_2284_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_2284_cast_fp16")];
            bool var_2287 = const()[name = string("op_2287"), val = bool(true)];
            tensor<int32, [1]> var_2288_axes_0 = const()[name = string("op_2288_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_2288_cast_fp16 = reduce_max(axes = var_2288_axes_0, keep_dims = var_2287, x = var_2284_cast_fp16)[name = string("op_2288_cast_fp16")];
            fp16 var_2289_to_fp16 = const()[name = string("op_2289_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_9_cast_fp16 = maximum(x = var_2288_cast_fp16, y = var_2289_to_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_9_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2292_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_2292_cast_fp16")];
            tensor<int32, [1]> var_2297_axes_0 = const()[name = string("op_2297_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2297_keep_dims_0 = const()[name = string("op_2297_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_2297_cast_fp16 = reduce_mean(axes = var_2297_axes_0, keep_dims = var_2297_keep_dims_0, x = var_2292_cast_fp16)[name = string("op_2297_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2298_cast_fp16 = mul(x = s_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_2298_cast_fp16")];
            fp32 var_2299_epsilon_0 = const()[name = string("op_2299_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_2299_cast_fp16 = inverse(epsilon = var_2299_epsilon_0, x = var_2298_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 var_2300_to_fp16 = const()[name = string("op_2300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_2301_cast_fp16 = mul(x = var_2299_cast_fp16, y = var_2300_to_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2303_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2301_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp32 var_2304_epsilon_0 = const()[name = string("op_2304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_2304_cast_fp16 = rsqrt(epsilon = var_2304_epsilon_0, x = var_2303_cast_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_1_cast_fp16 = mul(x = y_9_cast_fp16, y = var_2304_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<fp16, [128, 256]> cos_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138762112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138794944))))[name = string("cos_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 128, 256]> var_2306_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2306_cast_fp16")];
            tensor<int32, [4]> var_2311_begin_0 = const()[name = string("op_2311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2311_end_0 = const()[name = string("op_2311_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_2311_end_mask_0 = const()[name = string("op_2311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_2311_cast_fp16 = slice_by_index(begin = var_2311_begin_0, end = var_2311_end_0, end_mask = var_2311_end_mask_0, x = q_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = q_1_cast_fp16)[name = string("op_2317_cast_fp16")];
            int32 var_2319 = const()[name = string("op_2319"), val = int32(-1)];
            bool var_2320_interleave_0 = const()[name = string("op_2320_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_2320_cast_fp16 = concat(axis = var_2319, interleave = var_2320_interleave_0, values = (var_2312_cast_fp16, var_2317_cast_fp16))[name = string("op_2320_cast_fp16")];
            tensor<fp16, [128, 256]> sin_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138795264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138828096))))[name = string("sin_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 128, 256]> var_2321_cast_fp16 = mul(x = var_2320_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_3_cast_fp16 = add(x = var_2306_cast_fp16, y = var_2321_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2324_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2329_begin_0 = const()[name = string("op_2329_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2329_end_0 = const()[name = string("op_2329_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_2329_end_mask_0 = const()[name = string("op_2329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, x = k_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_2330_cast_fp16 = mul(x = var_2329_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_2330_cast_fp16")];
            tensor<int32, [4]> var_2335_begin_0 = const()[name = string("op_2335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2335_end_0 = const()[name = string("op_2335_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_2335_end_mask_0 = const()[name = string("op_2335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_2335_cast_fp16 = slice_by_index(begin = var_2335_begin_0, end = var_2335_end_0, end_mask = var_2335_end_mask_0, x = k_1_cast_fp16)[name = string("op_2335_cast_fp16")];
            int32 var_2337 = const()[name = string("op_2337"), val = int32(-1)];
            bool var_2338_interleave_0 = const()[name = string("op_2338_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_2338_cast_fp16 = concat(axis = var_2337, interleave = var_2338_interleave_0, values = (var_2330_cast_fp16, var_2335_cast_fp16))[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2339_cast_fp16 = mul(x = var_2338_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2339_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_3_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2339_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = k_3_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [4, 1, 128, 256]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
            tensor<int32, [5]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_0_cast_fp16 = reshape(shape = concat_0, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [4, 1, 128, 256]> reshape_1_cast_fp16 = reshape(shape = concat_1, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = v_1_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [4, 1, 128, 256]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
            tensor<int32, [5]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_2_cast_fp16 = reshape(shape = concat_2, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [4, 1, 128, 256]> reshape_3_cast_fp16 = reshape(shape = concat_3, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2353_transpose_x_0 = const()[name = string("op_2353_transpose_x_0"), val = bool(false)];
            bool var_2353_transpose_y_0 = const()[name = string("op_2353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 4, 128, 128]> var_2353_cast_fp16 = matmul(transpose_x = var_2353_transpose_x_0, transpose_y = var_2353_transpose_y_0, x = q_3_cast_fp16, y = transpose_96_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_2355_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2356 = const()[name = string("op_2356"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_9_cast_fp16 = softmax(axis = var_2356, x = var_2355_cast_fp16)[name = string("w_9_cast_fp16")];
            bool var_2359_transpose_x_0 = const()[name = string("op_2359_transpose_x_0"), val = bool(false)];
            bool var_2359_transpose_y_0 = const()[name = string("op_2359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 4, 128, 256]> var_2359_cast_fp16 = matmul(transpose_x = var_2359_transpose_x_0, transpose_y = var_2359_transpose_y_0, x = w_9_cast_fp16, y = v_3_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<int32, [4]> var_2362_perm_0 = const()[name = string("op_2362_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2366 = const()[name = string("op_2366"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_2362_cast_fp16 = transpose(perm = var_2362_perm_0, x = var_2359_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 128, 1024]> input_3_cast_fp16 = reshape(shape = var_2366, x = var_2362_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7402112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7926464))))[name = string("tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2371_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_2371_cast_fp16")];
            bool var_2374 = const()[name = string("op_2374"), val = bool(true)];
            tensor<int32, [1]> var_2375_axes_0 = const()[name = string("op_2375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2375_cast_fp16 = reduce_max(axes = var_2375_axes_0, keep_dims = var_2374, x = var_2371_cast_fp16)[name = string("op_2375_cast_fp16")];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_11_cast_fp16 = maximum(x = var_2375_cast_fp16, y = var_2376_to_fp16)[name = string("s_11_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_11_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_11_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2379_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_2379_cast_fp16")];
            tensor<int32, [1]> var_2384_axes_0 = const()[name = string("op_2384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2384_keep_dims_0 = const()[name = string("op_2384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2384_cast_fp16 = reduce_mean(axes = var_2384_axes_0, keep_dims = var_2384_keep_dims_0, x = var_2379_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2385_cast_fp16 = mul(x = s_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2385_cast_fp16")];
            fp32 var_2386_epsilon_0 = const()[name = string("op_2386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2386_cast_fp16 = inverse(epsilon = var_2386_epsilon_0, x = var_2385_cast_fp16)[name = string("op_2386_cast_fp16")];
            fp16 var_2387_to_fp16 = const()[name = string("op_2387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2388_cast_fp16 = mul(x = var_2386_cast_fp16, y = var_2387_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2390_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2388_cast_fp16)[name = string("op_2390_cast_fp16")];
            fp32 var_2391_epsilon_0 = const()[name = string("op_2391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2391_cast_fp16 = rsqrt(epsilon = var_2391_epsilon_0, x = var_2390_cast_fp16)[name = string("op_2391_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_2391_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7927552)))];
            tensor<fp16, [1, 128, 512]> var_2393_cast_fp16 = mul(x = out_9_cast_fp16, y = tm_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_11_cast_fp16 = add(x = inputs_embeds, y = var_2393_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2396_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_2396_cast_fp16")];
            bool var_2399 = const()[name = string("op_2399"), val = bool(true)];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_2399, x = var_2396_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 var_2401_to_fp16 = const()[name = string("op_2401_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_13_cast_fp16 = maximum(x = var_2400_cast_fp16, y = var_2401_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = s_13_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2404_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [1]> var_2409_axes_0 = const()[name = string("op_2409_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2409_keep_dims_0 = const()[name = string("op_2409_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2409_cast_fp16 = reduce_mean(axes = var_2409_axes_0, keep_dims = var_2409_keep_dims_0, x = var_2404_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2410_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2411_cast_fp16 = inverse(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2413_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("op_2413_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2415_cast_fp16 = add(x = var_2409_cast_fp16, y = var_2413_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp32 var_2416_epsilon_0 = const()[name = string("op_2416_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2416_cast_fp16 = rsqrt(epsilon = var_2416_epsilon_0, x = var_2415_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_2416_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7928640)))];
            tensor<fp16, [1, 128, 512]> input_5_cast_fp16 = mul(x = out_11_cast_fp16, y = tm_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7929728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8978368))))[name = string("tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8982528)))];
            tensor<fp16, [1, 128, 2048]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_2423_mode_0 = const()[name = string("op_2423_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_2423_cast_fp16 = gelu(mode = var_2423_mode_0, x = linear_5_cast_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8986688))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10035328))))[name = string("tm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_7_cast_fp16 = mul(x = var_2423_cast_fp16, y = linear_6_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10039488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11088128))))[name = string("tm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2431_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_2431_cast_fp16")];
            bool var_2434 = const()[name = string("op_2434"), val = bool(true)];
            tensor<int32, [1]> var_2435_axes_0 = const()[name = string("op_2435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2435_cast_fp16 = reduce_max(axes = var_2435_axes_0, keep_dims = var_2434, x = var_2431_cast_fp16)[name = string("op_2435_cast_fp16")];
            fp16 var_2436_to_fp16 = const()[name = string("op_2436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_15_cast_fp16 = maximum(x = var_2435_cast_fp16, y = var_2436_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_15_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_15_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2439_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_2439_cast_fp16")];
            tensor<int32, [1]> var_2444_axes_0 = const()[name = string("op_2444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2444_keep_dims_0 = const()[name = string("op_2444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2444_cast_fp16 = reduce_mean(axes = var_2444_axes_0, keep_dims = var_2444_keep_dims_0, x = var_2439_cast_fp16)[name = string("op_2444_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2445_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_2445_cast_fp16")];
            fp32 var_2446_epsilon_0 = const()[name = string("op_2446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2446_cast_fp16 = inverse(epsilon = var_2446_epsilon_0, x = var_2445_cast_fp16)[name = string("op_2446_cast_fp16")];
            fp16 var_2447_to_fp16 = const()[name = string("op_2447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2448_cast_fp16 = mul(x = var_2446_cast_fp16, y = var_2447_to_fp16)[name = string("op_2448_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2450_cast_fp16 = add(x = var_2444_cast_fp16, y = var_2448_cast_fp16)[name = string("op_2450_cast_fp16")];
            fp32 var_2451_epsilon_0 = const()[name = string("op_2451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2451_cast_fp16 = rsqrt(epsilon = var_2451_epsilon_0, x = var_2450_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_2451_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11089216)))];
            tensor<fp16, [1, 128, 512]> var_2453_cast_fp16 = mul(x = out_13_cast_fp16, y = tm_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_9_cast_fp16 = add(x = x_11_cast_fp16, y = var_2453_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11090304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11352512))))[name = string("tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_8_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_8_cast_fp16")];
            string x_15_mode_0 = const()[name = string("x_15_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_15_cast_fp16 = gelu(mode = x_15_mode_0, x = linear_8_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([1, 128, 1, 512])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2473_squeeze_mask_0 = const()[name = string("op_2473_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, squeeze_mask = var_2473_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2473_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_11_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2473_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11353600))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11615808))))[name = string("tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2483_cast_fp16 = abs(x = linear_9_cast_fp16)[name = string("op_2483_cast_fp16")];
            bool var_2486 = const()[name = string("op_2486"), val = bool(true)];
            tensor<int32, [1]> var_2487_axes_0 = const()[name = string("op_2487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2487_cast_fp16 = reduce_max(axes = var_2487_axes_0, keep_dims = var_2486, x = var_2483_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp16 var_2488_to_fp16 = const()[name = string("op_2488_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_17_cast_fp16 = maximum(x = var_2487_cast_fp16, y = var_2488_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_17_cast_fp16 = real_div(x = linear_9_cast_fp16, y = s_17_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2491_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_2491_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2496_keep_dims_0 = const()[name = string("op_2496_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2496_cast_fp16 = reduce_mean(axes = var_2496_axes_0, keep_dims = var_2496_keep_dims_0, x = var_2491_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2497_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_2497_cast_fp16")];
            fp32 var_2498_epsilon_0 = const()[name = string("op_2498_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2498_cast_fp16 = inverse(epsilon = var_2498_epsilon_0, x = var_2497_cast_fp16)[name = string("op_2498_cast_fp16")];
            fp16 var_2499_to_fp16 = const()[name = string("op_2499_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2500_cast_fp16 = mul(x = var_2498_cast_fp16, y = var_2499_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2502_cast_fp16 = add(x = var_2496_cast_fp16, y = var_2500_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp32 var_2503_epsilon_0 = const()[name = string("op_2503_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2503_cast_fp16 = rsqrt(epsilon = var_2503_epsilon_0, x = var_2502_cast_fp16)[name = string("op_2503_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_2503_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11616896)))];
            tensor<fp16, [1, 128, 512]> var_2505_cast_fp16 = mul(x = out_15_cast_fp16, y = tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_1_cast_fp16 = add(x = input_9_cast_fp16, y = var_2505_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<fp16, [1]> tm_layers_0_layer_scalar_to_fp16 = const()[name = string("tm_layers_0_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9p-2])];
            tensor<fp16, [1, 128, 512]> x_19_cast_fp16 = mul(x = h_1_cast_fp16, y = tm_layers_0_layer_scalar_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2509_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool var_2512 = const()[name = string("op_2512"), val = bool(true)];
            tensor<int32, [1]> var_2513_axes_0 = const()[name = string("op_2513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2513_cast_fp16 = reduce_max(axes = var_2513_axes_0, keep_dims = var_2512, x = var_2509_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_19_cast_fp16 = maximum(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_19_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2517_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [1]> var_2522_axes_0 = const()[name = string("op_2522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2522_keep_dims_0 = const()[name = string("op_2522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2522_cast_fp16 = reduce_mean(axes = var_2522_axes_0, keep_dims = var_2522_keep_dims_0, x = var_2517_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2523_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_2523_cast_fp16")];
            fp32 var_2524_epsilon_0 = const()[name = string("op_2524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2524_cast_fp16 = inverse(epsilon = var_2524_epsilon_0, x = var_2523_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp16 var_2525_to_fp16 = const()[name = string("op_2525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2526_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_2525_to_fp16)[name = string("op_2526_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2528_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2526_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_2529_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11617984)))];
            tensor<fp16, [1, 128, 512]> input_13_cast_fp16 = mul(x = out_17_cast_fp16, y = tm_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11619072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12143424))))[name = string("tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_2540_cast_fp16 = reshape(shape = var_2539, x = linear_10_cast_fp16)[name = string("op_2540_cast_fp16")];
            tensor<int32, [4]> x_21_perm_0 = const()[name = string("x_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12145536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12407744))))[name = string("tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<int32, [4]> var_2551 = const()[name = string("op_2551"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_2552_cast_fp16 = reshape(shape = var_2551, x = linear_11_cast_fp16)[name = string("op_2552_cast_fp16")];
            tensor<int32, [4]> x_23_perm_0 = const()[name = string("x_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12408832))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671040))))[name = string("tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [4]> var_2563 = const()[name = string("op_2563"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_2564_cast_fp16 = reshape(shape = var_2563, x = linear_12_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = var_2540_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 4, 128, 256]> var_2568_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_2568_cast_fp16")];
            bool var_2571 = const()[name = string("op_2571"), val = bool(true)];
            tensor<int32, [1]> var_2572_axes_0 = const()[name = string("op_2572_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_2572_cast_fp16 = reduce_max(axes = var_2572_axes_0, keep_dims = var_2571, x = var_2568_cast_fp16)[name = string("op_2572_cast_fp16")];
            fp16 var_2573_to_fp16 = const()[name = string("op_2573_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_21_cast_fp16 = maximum(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_21_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_2576_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> var_2581_axes_0 = const()[name = string("op_2581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2581_keep_dims_0 = const()[name = string("op_2581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_2581_cast_fp16 = reduce_mean(axes = var_2581_axes_0, keep_dims = var_2581_keep_dims_0, x = var_2576_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_2582_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp32 var_2583_epsilon_0 = const()[name = string("op_2583_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_2583_cast_fp16 = inverse(epsilon = var_2583_epsilon_0, x = var_2582_cast_fp16)[name = string("op_2583_cast_fp16")];
            fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_2585_cast_fp16 = mul(x = var_2583_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_2587_cast_fp16 = add(x = var_2581_cast_fp16, y = var_2585_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_19_cast_fp16 = mul(x = y_21_cast_fp16, y = var_2588_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672128)))];
            tensor<fp16, [1, 4, 128, 256]> q_5_cast_fp16 = mul(x = out_19_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_23_cast_fp16 = transpose(perm = x_23_perm_0, x = var_2552_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2, 128, 256]> var_2591_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2594 = const()[name = string("op_2594"), val = bool(true)];
            tensor<int32, [1]> var_2595_axes_0 = const()[name = string("op_2595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_2595_cast_fp16 = reduce_max(axes = var_2595_axes_0, keep_dims = var_2594, x = var_2591_cast_fp16)[name = string("op_2595_cast_fp16")];
            fp16 var_2596_to_fp16 = const()[name = string("op_2596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_23_cast_fp16 = maximum(x = var_2595_cast_fp16, y = var_2596_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_23_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2599_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2599_cast_fp16")];
            tensor<int32, [1]> var_2604_axes_0 = const()[name = string("op_2604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2604_keep_dims_0 = const()[name = string("op_2604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_2604_cast_fp16 = reduce_mean(axes = var_2604_axes_0, keep_dims = var_2604_keep_dims_0, x = var_2599_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2605_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_2605_cast_fp16")];
            fp32 var_2606_epsilon_0 = const()[name = string("op_2606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_2606_cast_fp16 = inverse(epsilon = var_2606_epsilon_0, x = var_2605_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp16 var_2607_to_fp16 = const()[name = string("op_2607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_2608_cast_fp16 = mul(x = var_2606_cast_fp16, y = var_2607_to_fp16)[name = string("op_2608_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2610_cast_fp16 = add(x = var_2604_cast_fp16, y = var_2608_cast_fp16)[name = string("op_2610_cast_fp16")];
            fp32 var_2611_epsilon_0 = const()[name = string("op_2611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_2611_cast_fp16 = rsqrt(epsilon = var_2611_epsilon_0, x = var_2610_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_21_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2611_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672704)))];
            tensor<fp16, [1, 2, 128, 256]> k_7_cast_fp16 = mul(x = out_21_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_2564_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 2, 128, 256]> var_2614_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_2614_cast_fp16")];
            bool var_2617 = const()[name = string("op_2617"), val = bool(true)];
            tensor<int32, [1]> var_2618_axes_0 = const()[name = string("op_2618_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_2618_cast_fp16 = reduce_max(axes = var_2618_axes_0, keep_dims = var_2617, x = var_2614_cast_fp16)[name = string("op_2618_cast_fp16")];
            fp16 var_2619_to_fp16 = const()[name = string("op_2619_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_25_cast_fp16 = maximum(x = var_2618_cast_fp16, y = var_2619_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_25_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2622_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2622_cast_fp16")];
            tensor<int32, [1]> var_2627_axes_0 = const()[name = string("op_2627_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2627_keep_dims_0 = const()[name = string("op_2627_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_2627_cast_fp16 = reduce_mean(axes = var_2627_axes_0, keep_dims = var_2627_keep_dims_0, x = var_2622_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2628_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_2628_cast_fp16")];
            fp32 var_2629_epsilon_0 = const()[name = string("op_2629_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_2629_cast_fp16 = inverse(epsilon = var_2629_epsilon_0, x = var_2628_cast_fp16)[name = string("op_2629_cast_fp16")];
            fp16 var_2630_to_fp16 = const()[name = string("op_2630_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_2631_cast_fp16 = mul(x = var_2629_cast_fp16, y = var_2630_to_fp16)[name = string("op_2631_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2633_cast_fp16 = add(x = var_2627_cast_fp16, y = var_2631_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_2634_cast_fp16 = rsqrt(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_5_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2634_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_2636_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2636_cast_fp16")];
            tensor<int32, [4]> var_2641_begin_0 = const()[name = string("op_2641_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2641_end_0 = const()[name = string("op_2641_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_2641_end_mask_0 = const()[name = string("op_2641_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_2641_cast_fp16 = slice_by_index(begin = var_2641_begin_0, end = var_2641_end_0, end_mask = var_2641_end_mask_0, x = q_5_cast_fp16)[name = string("op_2641_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_2642_cast_fp16 = mul(x = var_2641_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_2642_cast_fp16")];
            tensor<int32, [4]> var_2647_begin_0 = const()[name = string("op_2647_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2647_end_0 = const()[name = string("op_2647_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_2647_end_mask_0 = const()[name = string("op_2647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_2647_cast_fp16 = slice_by_index(begin = var_2647_begin_0, end = var_2647_end_0, end_mask = var_2647_end_mask_0, x = q_5_cast_fp16)[name = string("op_2647_cast_fp16")];
            int32 var_2649 = const()[name = string("op_2649"), val = int32(-1)];
            bool var_2650_interleave_0 = const()[name = string("op_2650_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_2650_cast_fp16 = concat(axis = var_2649, interleave = var_2650_interleave_0, values = (var_2642_cast_fp16, var_2647_cast_fp16))[name = string("op_2650_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2651_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_7_cast_fp16 = add(x = var_2636_cast_fp16, y = var_2651_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2654_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = k_7_cast_fp16)[name = string("op_2659_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_2660_cast_fp16 = mul(x = var_2659_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2660_cast_fp16")];
            tensor<int32, [4]> var_2665_begin_0 = const()[name = string("op_2665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2665_end_0 = const()[name = string("op_2665_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_2665_end_mask_0 = const()[name = string("op_2665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_2665_cast_fp16 = slice_by_index(begin = var_2665_begin_0, end = var_2665_end_0, end_mask = var_2665_end_mask_0, x = k_7_cast_fp16)[name = string("op_2665_cast_fp16")];
            int32 var_2667 = const()[name = string("op_2667"), val = int32(-1)];
            bool var_2668_interleave_0 = const()[name = string("op_2668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_2668_cast_fp16 = concat(axis = var_2667, interleave = var_2668_interleave_0, values = (var_2660_cast_fp16, var_2665_cast_fp16))[name = string("op_2668_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2669_cast_fp16 = mul(x = var_2668_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_9_cast_fp16 = add(x = var_2654_cast_fp16, y = var_2669_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = k_9_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [4, 1, 128, 256]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_4_cast_fp16 = reshape(shape = concat_4, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [4, 1, 128, 256]> reshape_5_cast_fp16 = reshape(shape = concat_5, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = v_5_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [4, 1, 128, 256]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_6_cast_fp16 = reshape(shape = concat_6, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [4, 1, 128, 256]> reshape_7_cast_fp16 = reshape(shape = concat_7, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2683_transpose_x_0 = const()[name = string("op_2683_transpose_x_0"), val = bool(false)];
            bool var_2683_transpose_y_0 = const()[name = string("op_2683_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_97_cast_fp16 = transpose(perm = transpose_97_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 4, 128, 128]> var_2683_cast_fp16 = matmul(transpose_x = var_2683_transpose_x_0, transpose_y = var_2683_transpose_y_0, x = q_7_cast_fp16, y = transpose_97_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_2685_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2685_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_25_cast_fp16 = softmax(axis = var_2686, x = var_2685_cast_fp16)[name = string("w_25_cast_fp16")];
            bool var_2689_transpose_x_0 = const()[name = string("op_2689_transpose_x_0"), val = bool(false)];
            bool var_2689_transpose_y_0 = const()[name = string("op_2689_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 4, 128, 256]> var_2689_cast_fp16 = matmul(transpose_x = var_2689_transpose_x_0, transpose_y = var_2689_transpose_y_0, x = w_25_cast_fp16, y = v_7_cast_fp16)[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> var_2692_perm_0 = const()[name = string("op_2692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2696 = const()[name = string("op_2696"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_2692_cast_fp16 = transpose(perm = var_2692_perm_0, x = var_2689_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 128, 1024]> input_15_cast_fp16 = reshape(shape = var_2696, x = var_2692_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13197632))))[name = string("tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2701_cast_fp16 = abs(x = linear_13_cast_fp16)[name = string("op_2701_cast_fp16")];
            bool var_2704 = const()[name = string("op_2704"), val = bool(true)];
            tensor<int32, [1]> var_2705_axes_0 = const()[name = string("op_2705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2705_cast_fp16 = reduce_max(axes = var_2705_axes_0, keep_dims = var_2704, x = var_2701_cast_fp16)[name = string("op_2705_cast_fp16")];
            fp16 var_2706_to_fp16 = const()[name = string("op_2706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_27_cast_fp16 = maximum(x = var_2705_cast_fp16, y = var_2706_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_27_cast_fp16 = real_div(x = linear_13_cast_fp16, y = s_27_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2709_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2709_cast_fp16")];
            tensor<int32, [1]> var_2714_axes_0 = const()[name = string("op_2714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2714_keep_dims_0 = const()[name = string("op_2714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2714_cast_fp16 = reduce_mean(axes = var_2714_axes_0, keep_dims = var_2714_keep_dims_0, x = var_2709_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2715_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2715_cast_fp16")];
            fp32 var_2716_epsilon_0 = const()[name = string("op_2716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2716_cast_fp16 = inverse(epsilon = var_2716_epsilon_0, x = var_2715_cast_fp16)[name = string("op_2716_cast_fp16")];
            fp16 var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2718_cast_fp16 = mul(x = var_2716_cast_fp16, y = var_2717_to_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2720_cast_fp16 = add(x = var_2714_cast_fp16, y = var_2718_cast_fp16)[name = string("op_2720_cast_fp16")];
            fp32 var_2721_epsilon_0 = const()[name = string("op_2721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2721_cast_fp16 = rsqrt(epsilon = var_2721_epsilon_0, x = var_2720_cast_fp16)[name = string("op_2721_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2721_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13198720)))];
            tensor<fp16, [1, 128, 512]> var_2723_cast_fp16 = mul(x = out_23_cast_fp16, y = tm_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2723_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_29_cast_fp16 = add(x = x_19_cast_fp16, y = var_2723_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2726_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_2726_cast_fp16")];
            bool var_2729 = const()[name = string("op_2729"), val = bool(true)];
            tensor<int32, [1]> var_2730_axes_0 = const()[name = string("op_2730_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2730_cast_fp16 = reduce_max(axes = var_2730_axes_0, keep_dims = var_2729, x = var_2726_cast_fp16)[name = string("op_2730_cast_fp16")];
            fp16 var_2731_to_fp16 = const()[name = string("op_2731_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_29_cast_fp16 = maximum(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_29_cast_fp16 = real_div(x = x_29_cast_fp16, y = s_29_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2734_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2734_cast_fp16")];
            tensor<int32, [1]> var_2739_axes_0 = const()[name = string("op_2739_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2739_keep_dims_0 = const()[name = string("op_2739_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2739_cast_fp16 = reduce_mean(axes = var_2739_axes_0, keep_dims = var_2739_keep_dims_0, x = var_2734_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2740_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2740_cast_fp16")];
            fp32 var_2741_epsilon_0 = const()[name = string("op_2741_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2741_cast_fp16 = inverse(epsilon = var_2741_epsilon_0, x = var_2740_cast_fp16)[name = string("op_2741_cast_fp16")];
            fp16 var_2742_to_fp16 = const()[name = string("op_2742_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2743_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_2742_to_fp16)[name = string("op_2743_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2745_cast_fp16 = add(x = var_2739_cast_fp16, y = var_2743_cast_fp16)[name = string("op_2745_cast_fp16")];
            fp32 var_2746_epsilon_0 = const()[name = string("op_2746_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2746_cast_fp16 = rsqrt(epsilon = var_2746_epsilon_0, x = var_2745_cast_fp16)[name = string("op_2746_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2746_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13199808)))];
            tensor<fp16, [1, 128, 512]> input_17_cast_fp16 = mul(x = out_25_cast_fp16, y = tm_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13200896))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14249536))))[name = string("tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_14_cast_fp16")];
            string var_2753_mode_0 = const()[name = string("op_2753_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_2753_cast_fp16 = gelu(mode = var_2753_mode_0, x = linear_14_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14253696))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15302336))))[name = string("tm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_15_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_19_cast_fp16 = mul(x = var_2753_cast_fp16, y = linear_15_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15306496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16355136))))[name = string("tm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2761_cast_fp16 = abs(x = linear_16_cast_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2764 = const()[name = string("op_2764"), val = bool(true)];
            tensor<int32, [1]> var_2765_axes_0 = const()[name = string("op_2765_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2765_cast_fp16 = reduce_max(axes = var_2765_axes_0, keep_dims = var_2764, x = var_2761_cast_fp16)[name = string("op_2765_cast_fp16")];
            fp16 var_2766_to_fp16 = const()[name = string("op_2766_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_31_cast_fp16 = maximum(x = var_2765_cast_fp16, y = var_2766_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_31_cast_fp16 = real_div(x = linear_16_cast_fp16, y = s_31_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2769_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [1]> var_2774_axes_0 = const()[name = string("op_2774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2774_keep_dims_0 = const()[name = string("op_2774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2774_cast_fp16 = reduce_mean(axes = var_2774_axes_0, keep_dims = var_2774_keep_dims_0, x = var_2769_cast_fp16)[name = string("op_2774_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2775_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp32 var_2776_epsilon_0 = const()[name = string("op_2776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2776_cast_fp16 = inverse(epsilon = var_2776_epsilon_0, x = var_2775_cast_fp16)[name = string("op_2776_cast_fp16")];
            fp16 var_2777_to_fp16 = const()[name = string("op_2777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2778_cast_fp16 = mul(x = var_2776_cast_fp16, y = var_2777_to_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2780_cast_fp16 = add(x = var_2774_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2780_cast_fp16")];
            fp32 var_2781_epsilon_0 = const()[name = string("op_2781_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2781_cast_fp16 = rsqrt(epsilon = var_2781_epsilon_0, x = var_2780_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2781_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16356224)))];
            tensor<fp16, [1, 128, 512]> var_2783_cast_fp16 = mul(x = out_27_cast_fp16, y = tm_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2783_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_21_cast_fp16 = add(x = x_29_cast_fp16, y = var_2783_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16357312))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16619520))))[name = string("tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_17_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            string x_33_mode_0 = const()[name = string("x_33_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_33_cast_fp16 = gelu(mode = x_33_mode_0, x = linear_17_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> var_2803_begin_0 = const()[name = string("op_2803_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2803_end_0 = const()[name = string("op_2803_end_0"), val = tensor<int32, [4]>([1, 128, 2, 512])];
            tensor<bool, [4]> var_2803_end_mask_0 = const()[name = string("op_2803_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2803_squeeze_mask_0 = const()[name = string("op_2803_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_2803_cast_fp16 = slice_by_index(begin = var_2803_begin_0, end = var_2803_end_0, end_mask = var_2803_end_mask_0, squeeze_mask = var_2803_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_23_cast_fp16 = mul(x = x_33_cast_fp16, y = var_2803_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16620608))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16882816))))[name = string("tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2813_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2813_cast_fp16")];
            bool var_2816 = const()[name = string("op_2816"), val = bool(true)];
            tensor<int32, [1]> var_2817_axes_0 = const()[name = string("op_2817_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2817_cast_fp16 = reduce_max(axes = var_2817_axes_0, keep_dims = var_2816, x = var_2813_cast_fp16)[name = string("op_2817_cast_fp16")];
            fp16 var_2818_to_fp16 = const()[name = string("op_2818_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_33_cast_fp16 = maximum(x = var_2817_cast_fp16, y = var_2818_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_33_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_33_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2821_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<int32, [1]> var_2826_axes_0 = const()[name = string("op_2826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2826_keep_dims_0 = const()[name = string("op_2826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2826_cast_fp16 = reduce_mean(axes = var_2826_axes_0, keep_dims = var_2826_keep_dims_0, x = var_2821_cast_fp16)[name = string("op_2826_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2827_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp32 var_2828_epsilon_0 = const()[name = string("op_2828_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2828_cast_fp16 = inverse(epsilon = var_2828_epsilon_0, x = var_2827_cast_fp16)[name = string("op_2828_cast_fp16")];
            fp16 var_2829_to_fp16 = const()[name = string("op_2829_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2830_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_2829_to_fp16)[name = string("op_2830_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2832_cast_fp16 = add(x = var_2826_cast_fp16, y = var_2830_cast_fp16)[name = string("op_2832_cast_fp16")];
            fp32 var_2833_epsilon_0 = const()[name = string("op_2833_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2833_cast_fp16 = rsqrt(epsilon = var_2833_epsilon_0, x = var_2832_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2833_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16883904)))];
            tensor<fp16, [1, 128, 512]> var_2835_cast_fp16 = mul(x = out_29_cast_fp16, y = tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_3_cast_fp16 = add(x = input_21_cast_fp16, y = var_2835_cast_fp16)[name = string("h_3_cast_fp16")];
            tensor<fp16, [1]> tm_layers_1_layer_scalar_to_fp16 = const()[name = string("tm_layers_1_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.5ap-3])];
            tensor<fp16, [1, 128, 512]> x_37_cast_fp16 = mul(x = h_3_cast_fp16, y = tm_layers_1_layer_scalar_to_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2839_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2839_cast_fp16")];
            bool var_2842 = const()[name = string("op_2842"), val = bool(true)];
            tensor<int32, [1]> var_2843_axes_0 = const()[name = string("op_2843_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2843_cast_fp16 = reduce_max(axes = var_2843_axes_0, keep_dims = var_2842, x = var_2839_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp16 var_2844_to_fp16 = const()[name = string("op_2844_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_35_cast_fp16 = maximum(x = var_2843_cast_fp16, y = var_2844_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_35_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_35_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_2847_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2847_cast_fp16")];
            tensor<int32, [1]> var_2852_axes_0 = const()[name = string("op_2852_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2852_keep_dims_0 = const()[name = string("op_2852_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2852_cast_fp16 = reduce_mean(axes = var_2852_axes_0, keep_dims = var_2852_keep_dims_0, x = var_2847_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2853_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp32 var_2854_epsilon_0 = const()[name = string("op_2854_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2854_cast_fp16 = inverse(epsilon = var_2854_epsilon_0, x = var_2853_cast_fp16)[name = string("op_2854_cast_fp16")];
            fp16 var_2855_to_fp16 = const()[name = string("op_2855_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2856_cast_fp16 = mul(x = var_2854_cast_fp16, y = var_2855_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2858_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2856_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp32 var_2859_epsilon_0 = const()[name = string("op_2859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2859_cast_fp16 = rsqrt(epsilon = var_2859_epsilon_0, x = var_2858_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_31_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2859_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16884992)))];
            tensor<fp16, [1, 128, 512]> input_25_cast_fp16 = mul(x = out_31_cast_fp16, y = tm_layers_2_input_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16886080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17410432))))[name = string("tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [4]> var_2869 = const()[name = string("op_2869"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_2870_cast_fp16 = reshape(shape = var_2869, x = linear_19_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17412544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17674752))))[name = string("tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [4]> var_2881 = const()[name = string("op_2881"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_2882_cast_fp16 = reshape(shape = var_2881, x = linear_20_cast_fp16)[name = string("op_2882_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17675840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17938048))))[name = string("tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_2894_cast_fp16 = reshape(shape = var_2893, x = linear_21_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [4]> x_43_perm_0 = const()[name = string("x_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2870_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 4, 128, 256]> var_2898_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2898_cast_fp16")];
            bool var_2901 = const()[name = string("op_2901"), val = bool(true)];
            tensor<int32, [1]> var_2902_axes_0 = const()[name = string("op_2902_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_2902_cast_fp16 = reduce_max(axes = var_2902_axes_0, keep_dims = var_2901, x = var_2898_cast_fp16)[name = string("op_2902_cast_fp16")];
            fp16 var_2903_to_fp16 = const()[name = string("op_2903_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_37_cast_fp16 = maximum(x = var_2902_cast_fp16, y = var_2903_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_37_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_37_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_2906_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<int32, [1]> var_2911_axes_0 = const()[name = string("op_2911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2911_keep_dims_0 = const()[name = string("op_2911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_2911_cast_fp16 = reduce_mean(axes = var_2911_axes_0, keep_dims = var_2911_keep_dims_0, x = var_2906_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_2912_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2912_cast_fp16")];
            fp32 var_2913_epsilon_0 = const()[name = string("op_2913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_2913_cast_fp16 = inverse(epsilon = var_2913_epsilon_0, x = var_2912_cast_fp16)[name = string("op_2913_cast_fp16")];
            fp16 var_2914_to_fp16 = const()[name = string("op_2914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_2915_cast_fp16 = mul(x = var_2913_cast_fp16, y = var_2914_to_fp16)[name = string("op_2915_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_2917_cast_fp16 = add(x = var_2911_cast_fp16, y = var_2915_cast_fp16)[name = string("op_2917_cast_fp16")];
            fp32 var_2918_epsilon_0 = const()[name = string("op_2918_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_2918_cast_fp16 = rsqrt(epsilon = var_2918_epsilon_0, x = var_2917_cast_fp16)[name = string("op_2918_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_33_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2918_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_9_cast_fp16 = mul(x = out_33_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2882_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 2, 128, 256]> var_2921_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2921_cast_fp16")];
            bool var_2924 = const()[name = string("op_2924"), val = bool(true)];
            tensor<int32, [1]> var_2925_axes_0 = const()[name = string("op_2925_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_2925_cast_fp16 = reduce_max(axes = var_2925_axes_0, keep_dims = var_2924, x = var_2921_cast_fp16)[name = string("op_2925_cast_fp16")];
            fp16 var_2926_to_fp16 = const()[name = string("op_2926_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_39_cast_fp16 = maximum(x = var_2925_cast_fp16, y = var_2926_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_39_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_39_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2929_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2929_cast_fp16")];
            tensor<int32, [1]> var_2934_axes_0 = const()[name = string("op_2934_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2934_keep_dims_0 = const()[name = string("op_2934_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_2934_cast_fp16 = reduce_mean(axes = var_2934_axes_0, keep_dims = var_2934_keep_dims_0, x = var_2929_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2935_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2935_cast_fp16")];
            fp32 var_2936_epsilon_0 = const()[name = string("op_2936_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_2936_cast_fp16 = inverse(epsilon = var_2936_epsilon_0, x = var_2935_cast_fp16)[name = string("op_2936_cast_fp16")];
            fp16 var_2937_to_fp16 = const()[name = string("op_2937_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_2938_cast_fp16 = mul(x = var_2936_cast_fp16, y = var_2937_to_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2940_cast_fp16 = add(x = var_2934_cast_fp16, y = var_2938_cast_fp16)[name = string("op_2940_cast_fp16")];
            fp32 var_2941_epsilon_0 = const()[name = string("op_2941_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_2941_cast_fp16 = rsqrt(epsilon = var_2941_epsilon_0, x = var_2940_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_35_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2941_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_13_cast_fp16 = mul(x = out_35_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_43_cast_fp16 = transpose(perm = x_43_perm_0, x = var_2894_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 2, 128, 256]> var_2944_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2944_cast_fp16")];
            bool var_2947 = const()[name = string("op_2947"), val = bool(true)];
            tensor<int32, [1]> var_2948_axes_0 = const()[name = string("op_2948_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_2948_cast_fp16 = reduce_max(axes = var_2948_axes_0, keep_dims = var_2947, x = var_2944_cast_fp16)[name = string("op_2948_cast_fp16")];
            fp16 var_2949_to_fp16 = const()[name = string("op_2949_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_41_cast_fp16 = maximum(x = var_2948_cast_fp16, y = var_2949_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_41_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_41_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2952_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2952_cast_fp16")];
            tensor<int32, [1]> var_2957_axes_0 = const()[name = string("op_2957_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2957_keep_dims_0 = const()[name = string("op_2957_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_2957_cast_fp16 = reduce_mean(axes = var_2957_axes_0, keep_dims = var_2957_keep_dims_0, x = var_2952_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2958_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2958_cast_fp16")];
            fp32 var_2959_epsilon_0 = const()[name = string("op_2959_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_2959_cast_fp16 = inverse(epsilon = var_2959_epsilon_0, x = var_2958_cast_fp16)[name = string("op_2959_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_2961_cast_fp16 = mul(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_2963_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2961_cast_fp16)[name = string("op_2963_cast_fp16")];
            fp32 var_2964_epsilon_0 = const()[name = string("op_2964_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_2964_cast_fp16 = rsqrt(epsilon = var_2964_epsilon_0, x = var_2963_cast_fp16)[name = string("op_2964_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_9_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2964_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_2966_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2966_cast_fp16")];
            tensor<int32, [4]> var_2971_begin_0 = const()[name = string("op_2971_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2971_end_0 = const()[name = string("op_2971_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_2971_end_mask_0 = const()[name = string("op_2971_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_2971_cast_fp16 = slice_by_index(begin = var_2971_begin_0, end = var_2971_end_0, end_mask = var_2971_end_mask_0, x = q_9_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_2972_cast_fp16 = mul(x = var_2971_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2972_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            int32 var_2979 = const()[name = string("op_2979"), val = int32(-1)];
            bool var_2980_interleave_0 = const()[name = string("op_2980_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_2980_cast_fp16 = concat(axis = var_2979, interleave = var_2980_interleave_0, values = (var_2972_cast_fp16, var_2977_cast_fp16))[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_2981_cast_fp16 = mul(x = var_2980_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2981_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_11_cast_fp16 = add(x = var_2966_cast_fp16, y = var_2981_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2984_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2989_begin_0 = const()[name = string("op_2989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2989_end_0 = const()[name = string("op_2989_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_2989_end_mask_0 = const()[name = string("op_2989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_2989_cast_fp16 = slice_by_index(begin = var_2989_begin_0, end = var_2989_end_0, end_mask = var_2989_end_mask_0, x = k_13_cast_fp16)[name = string("op_2989_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_2990_cast_fp16 = mul(x = var_2989_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2990_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = k_13_cast_fp16)[name = string("op_2995_cast_fp16")];
            int32 var_2997 = const()[name = string("op_2997"), val = int32(-1)];
            bool var_2998_interleave_0 = const()[name = string("op_2998_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_2998_cast_fp16 = concat(axis = var_2997, interleave = var_2998_interleave_0, values = (var_2990_cast_fp16, var_2995_cast_fp16))[name = string("op_2998_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_2999_cast_fp16 = mul(x = var_2998_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2999_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_15_cast_fp16 = add(x = var_2984_cast_fp16, y = var_2999_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = k_15_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [4, 1, 128, 256]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
            tensor<int32, [5]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_8_cast_fp16 = reshape(shape = concat_8, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [4, 1, 128, 256]> reshape_9_cast_fp16 = reshape(shape = concat_9, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = v_9_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [4, 1, 128, 256]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
            tensor<int32, [5]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_10_cast_fp16 = reshape(shape = concat_10, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [4, 1, 128, 256]> reshape_11_cast_fp16 = reshape(shape = concat_11, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3013_transpose_x_0 = const()[name = string("op_3013_transpose_x_0"), val = bool(false)];
            bool var_3013_transpose_y_0 = const()[name = string("op_3013_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 4, 128, 128]> var_3013_cast_fp16 = matmul(transpose_x = var_3013_transpose_x_0, transpose_y = var_3013_transpose_y_0, x = q_11_cast_fp16, y = transpose_98_cast_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_3015_cast_fp16 = add(x = var_3013_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3015_cast_fp16")];
            int32 var_3016 = const()[name = string("op_3016"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_41_cast_fp16 = softmax(axis = var_3016, x = var_3015_cast_fp16)[name = string("w_41_cast_fp16")];
            bool var_3019_transpose_x_0 = const()[name = string("op_3019_transpose_x_0"), val = bool(false)];
            bool var_3019_transpose_y_0 = const()[name = string("op_3019_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 4, 128, 256]> var_3019_cast_fp16 = matmul(transpose_x = var_3019_transpose_x_0, transpose_y = var_3019_transpose_y_0, x = w_41_cast_fp16, y = v_11_cast_fp16)[name = string("op_3019_cast_fp16")];
            tensor<int32, [4]> var_3022_perm_0 = const()[name = string("op_3022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3026 = const()[name = string("op_3026"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_3022_cast_fp16 = transpose(perm = var_3022_perm_0, x = var_3019_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 128, 1024]> input_27_cast_fp16 = reshape(shape = var_3026, x = var_3022_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17939136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18463488))))[name = string("tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3031_cast_fp16 = abs(x = linear_22_cast_fp16)[name = string("op_3031_cast_fp16")];
            bool var_3034 = const()[name = string("op_3034"), val = bool(true)];
            tensor<int32, [1]> var_3035_axes_0 = const()[name = string("op_3035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3035_cast_fp16 = reduce_max(axes = var_3035_axes_0, keep_dims = var_3034, x = var_3031_cast_fp16)[name = string("op_3035_cast_fp16")];
            fp16 var_3036_to_fp16 = const()[name = string("op_3036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_43_cast_fp16 = maximum(x = var_3035_cast_fp16, y = var_3036_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_43_cast_fp16 = real_div(x = linear_22_cast_fp16, y = s_43_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3039_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3039_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3045_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_3045_cast_fp16")];
            fp32 var_3046_epsilon_0 = const()[name = string("op_3046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3046_cast_fp16 = inverse(epsilon = var_3046_epsilon_0, x = var_3045_cast_fp16)[name = string("op_3046_cast_fp16")];
            fp16 var_3047_to_fp16 = const()[name = string("op_3047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3048_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3047_to_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3050_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3048_cast_fp16)[name = string("op_3050_cast_fp16")];
            fp32 var_3051_epsilon_0 = const()[name = string("op_3051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3051_cast_fp16 = rsqrt(epsilon = var_3051_epsilon_0, x = var_3050_cast_fp16)[name = string("op_3051_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_3051_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18464576)))];
            tensor<fp16, [1, 128, 512]> var_3053_cast_fp16 = mul(x = out_37_cast_fp16, y = tm_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = var_3053_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3056_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_3056_cast_fp16")];
            bool var_3059 = const()[name = string("op_3059"), val = bool(true)];
            tensor<int32, [1]> var_3060_axes_0 = const()[name = string("op_3060_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3060_cast_fp16 = reduce_max(axes = var_3060_axes_0, keep_dims = var_3059, x = var_3056_cast_fp16)[name = string("op_3060_cast_fp16")];
            fp16 var_3061_to_fp16 = const()[name = string("op_3061_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_45_cast_fp16 = maximum(x = var_3060_cast_fp16, y = var_3061_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_45_cast_fp16 = real_div(x = x_47_cast_fp16, y = s_45_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3064_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<int32, [1]> var_3069_axes_0 = const()[name = string("op_3069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3069_keep_dims_0 = const()[name = string("op_3069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3069_cast_fp16 = reduce_mean(axes = var_3069_axes_0, keep_dims = var_3069_keep_dims_0, x = var_3064_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3070_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp32 var_3071_epsilon_0 = const()[name = string("op_3071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3071_cast_fp16 = inverse(epsilon = var_3071_epsilon_0, x = var_3070_cast_fp16)[name = string("op_3071_cast_fp16")];
            fp16 var_3072_to_fp16 = const()[name = string("op_3072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3073_cast_fp16 = mul(x = var_3071_cast_fp16, y = var_3072_to_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3075_cast_fp16 = add(x = var_3069_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3075_cast_fp16")];
            fp32 var_3076_epsilon_0 = const()[name = string("op_3076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3076_cast_fp16 = rsqrt(epsilon = var_3076_epsilon_0, x = var_3075_cast_fp16)[name = string("op_3076_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_3076_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18465664)))];
            tensor<fp16, [1, 128, 512]> input_29_cast_fp16 = mul(x = out_39_cast_fp16, y = tm_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18466752))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19515392))))[name = string("tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_23_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_23_cast_fp16")];
            string var_3083_mode_0 = const()[name = string("op_3083_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_3083_cast_fp16 = gelu(mode = var_3083_mode_0, x = linear_23_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19519552))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20568192))))[name = string("tm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_24_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_31_cast_fp16 = mul(x = var_3083_cast_fp16, y = linear_24_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20572352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21620992))))[name = string("tm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3091_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_3091_cast_fp16")];
            bool var_3094 = const()[name = string("op_3094"), val = bool(true)];
            tensor<int32, [1]> var_3095_axes_0 = const()[name = string("op_3095_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3095_cast_fp16 = reduce_max(axes = var_3095_axes_0, keep_dims = var_3094, x = var_3091_cast_fp16)[name = string("op_3095_cast_fp16")];
            fp16 var_3096_to_fp16 = const()[name = string("op_3096_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_47_cast_fp16 = maximum(x = var_3095_cast_fp16, y = var_3096_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_47_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_47_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3099_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3104_keep_dims_0 = const()[name = string("op_3104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3104_cast_fp16 = reduce_mean(axes = var_3104_axes_0, keep_dims = var_3104_keep_dims_0, x = var_3099_cast_fp16)[name = string("op_3104_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3105_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_3105_cast_fp16")];
            fp32 var_3106_epsilon_0 = const()[name = string("op_3106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3106_cast_fp16 = inverse(epsilon = var_3106_epsilon_0, x = var_3105_cast_fp16)[name = string("op_3106_cast_fp16")];
            fp16 var_3107_to_fp16 = const()[name = string("op_3107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3108_cast_fp16 = mul(x = var_3106_cast_fp16, y = var_3107_to_fp16)[name = string("op_3108_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3110_cast_fp16 = add(x = var_3104_cast_fp16, y = var_3108_cast_fp16)[name = string("op_3110_cast_fp16")];
            fp32 var_3111_epsilon_0 = const()[name = string("op_3111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3111_cast_fp16 = rsqrt(epsilon = var_3111_epsilon_0, x = var_3110_cast_fp16)[name = string("op_3111_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_3111_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21622080)))];
            tensor<fp16, [1, 128, 512]> var_3113_cast_fp16 = mul(x = out_41_cast_fp16, y = tm_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_33_cast_fp16 = add(x = x_47_cast_fp16, y = var_3113_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21623168))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21885376))))[name = string("tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_26_cast_fp16")];
            string x_51_mode_0 = const()[name = string("x_51_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_51_cast_fp16 = gelu(mode = x_51_mode_0, x = linear_26_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([1, 128, 3, 512])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3133_squeeze_mask_0 = const()[name = string("op_3133_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, squeeze_mask = var_3133_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_35_cast_fp16 = mul(x = x_51_cast_fp16, y = var_3133_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21886464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22148672))))[name = string("tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3143_cast_fp16 = abs(x = linear_27_cast_fp16)[name = string("op_3143_cast_fp16")];
            bool var_3146 = const()[name = string("op_3146"), val = bool(true)];
            tensor<int32, [1]> var_3147_axes_0 = const()[name = string("op_3147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3147_cast_fp16 = reduce_max(axes = var_3147_axes_0, keep_dims = var_3146, x = var_3143_cast_fp16)[name = string("op_3147_cast_fp16")];
            fp16 var_3148_to_fp16 = const()[name = string("op_3148_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_49_cast_fp16 = maximum(x = var_3147_cast_fp16, y = var_3148_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_49_cast_fp16 = real_div(x = linear_27_cast_fp16, y = s_49_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3151_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_3151_cast_fp16")];
            tensor<int32, [1]> var_3156_axes_0 = const()[name = string("op_3156_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3156_keep_dims_0 = const()[name = string("op_3156_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3156_cast_fp16 = reduce_mean(axes = var_3156_axes_0, keep_dims = var_3156_keep_dims_0, x = var_3151_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3157_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_3157_cast_fp16")];
            fp32 var_3158_epsilon_0 = const()[name = string("op_3158_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3158_cast_fp16 = inverse(epsilon = var_3158_epsilon_0, x = var_3157_cast_fp16)[name = string("op_3158_cast_fp16")];
            fp16 var_3159_to_fp16 = const()[name = string("op_3159_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3160_cast_fp16 = mul(x = var_3158_cast_fp16, y = var_3159_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3162_cast_fp16 = add(x = var_3156_cast_fp16, y = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp32 var_3163_epsilon_0 = const()[name = string("op_3163_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3163_cast_fp16 = rsqrt(epsilon = var_3163_epsilon_0, x = var_3162_cast_fp16)[name = string("op_3163_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_43_cast_fp16 = mul(x = y_49_cast_fp16, y = var_3163_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22149760)))];
            tensor<fp16, [1, 128, 512]> var_3165_cast_fp16 = mul(x = out_43_cast_fp16, y = tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_5_cast_fp16 = add(x = input_33_cast_fp16, y = var_3165_cast_fp16)[name = string("h_5_cast_fp16")];
            tensor<fp16, [1]> tm_layers_2_layer_scalar_to_fp16 = const()[name = string("tm_layers_2_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ep-1])];
            tensor<fp16, [1, 128, 512]> x_55_cast_fp16 = mul(x = h_5_cast_fp16, y = tm_layers_2_layer_scalar_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3169_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_3169_cast_fp16")];
            bool var_3172 = const()[name = string("op_3172"), val = bool(true)];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3173_cast_fp16 = reduce_max(axes = var_3173_axes_0, keep_dims = var_3172, x = var_3169_cast_fp16)[name = string("op_3173_cast_fp16")];
            fp16 var_3174_to_fp16 = const()[name = string("op_3174_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_51_cast_fp16 = maximum(x = var_3173_cast_fp16, y = var_3174_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_51_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_51_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3177_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_3177_cast_fp16")];
            tensor<int32, [1]> var_3182_axes_0 = const()[name = string("op_3182_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3182_keep_dims_0 = const()[name = string("op_3182_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3182_cast_fp16 = reduce_mean(axes = var_3182_axes_0, keep_dims = var_3182_keep_dims_0, x = var_3177_cast_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3183_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_3183_cast_fp16")];
            fp32 var_3184_epsilon_0 = const()[name = string("op_3184_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3184_cast_fp16 = inverse(epsilon = var_3184_epsilon_0, x = var_3183_cast_fp16)[name = string("op_3184_cast_fp16")];
            fp16 var_3185_to_fp16 = const()[name = string("op_3185_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3186_cast_fp16 = mul(x = var_3184_cast_fp16, y = var_3185_to_fp16)[name = string("op_3186_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3188_cast_fp16 = add(x = var_3182_cast_fp16, y = var_3186_cast_fp16)[name = string("op_3188_cast_fp16")];
            fp32 var_3189_epsilon_0 = const()[name = string("op_3189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3189_cast_fp16 = rsqrt(epsilon = var_3189_epsilon_0, x = var_3188_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_45_cast_fp16 = mul(x = y_51_cast_fp16, y = var_3189_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22150848)))];
            tensor<fp16, [1, 128, 512]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = tm_layers_3_input_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22151936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22676288))))[name = string("tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_3200_cast_fp16 = reshape(shape = var_3199, x = linear_28_cast_fp16)[name = string("op_3200_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22678400))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22940608))))[name = string("tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_29_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_59_perm_0 = const()[name = string("x_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22941696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23203904))))[name = string("tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_3223 = const()[name = string("op_3223"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_3224_cast_fp16 = reshape(shape = var_3223, x = linear_30_cast_fp16)[name = string("op_3224_cast_fp16")];
            tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_3200_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 4, 128, 256]> var_3228_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3231 = const()[name = string("op_3231"), val = bool(true)];
            tensor<int32, [1]> var_3232_axes_0 = const()[name = string("op_3232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_3232_cast_fp16 = reduce_max(axes = var_3232_axes_0, keep_dims = var_3231, x = var_3228_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_53_cast_fp16 = maximum(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_53_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_53_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_3236_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_3236_cast_fp16")];
            tensor<int32, [1]> var_3241_axes_0 = const()[name = string("op_3241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3241_keep_dims_0 = const()[name = string("op_3241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_3241_cast_fp16 = reduce_mean(axes = var_3241_axes_0, keep_dims = var_3241_keep_dims_0, x = var_3236_cast_fp16)[name = string("op_3241_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_3242_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_3242_cast_fp16")];
            fp32 var_3243_epsilon_0 = const()[name = string("op_3243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_3243_cast_fp16 = inverse(epsilon = var_3243_epsilon_0, x = var_3242_cast_fp16)[name = string("op_3243_cast_fp16")];
            fp16 var_3244_to_fp16 = const()[name = string("op_3244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_3245_cast_fp16 = mul(x = var_3243_cast_fp16, y = var_3244_to_fp16)[name = string("op_3245_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_3247_cast_fp16 = add(x = var_3241_cast_fp16, y = var_3245_cast_fp16)[name = string("op_3247_cast_fp16")];
            fp32 var_3248_epsilon_0 = const()[name = string("op_3248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_3248_cast_fp16 = rsqrt(epsilon = var_3248_epsilon_0, x = var_3247_cast_fp16)[name = string("op_3248_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_47_cast_fp16 = mul(x = y_53_cast_fp16, y = var_3248_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23204992)))];
            tensor<fp16, [1, 4, 128, 256]> q_13_cast_fp16 = mul(x = out_47_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_59_cast_fp16 = transpose(perm = x_59_perm_0, x = var_3212_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 2, 128, 256]> var_3251_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_3251_cast_fp16")];
            bool var_3254 = const()[name = string("op_3254"), val = bool(true)];
            tensor<int32, [1]> var_3255_axes_0 = const()[name = string("op_3255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_3255_cast_fp16 = reduce_max(axes = var_3255_axes_0, keep_dims = var_3254, x = var_3251_cast_fp16)[name = string("op_3255_cast_fp16")];
            fp16 var_3256_to_fp16 = const()[name = string("op_3256_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_55_cast_fp16 = maximum(x = var_3255_cast_fp16, y = var_3256_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_55_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_55_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_3259_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<int32, [1]> var_3264_axes_0 = const()[name = string("op_3264_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3264_keep_dims_0 = const()[name = string("op_3264_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_3264_cast_fp16 = reduce_mean(axes = var_3264_axes_0, keep_dims = var_3264_keep_dims_0, x = var_3259_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3265_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_3265_cast_fp16")];
            fp32 var_3266_epsilon_0 = const()[name = string("op_3266_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_3266_cast_fp16 = inverse(epsilon = var_3266_epsilon_0, x = var_3265_cast_fp16)[name = string("op_3266_cast_fp16")];
            fp16 var_3267_to_fp16 = const()[name = string("op_3267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_3268_cast_fp16 = mul(x = var_3266_cast_fp16, y = var_3267_to_fp16)[name = string("op_3268_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3270_cast_fp16 = add(x = var_3264_cast_fp16, y = var_3268_cast_fp16)[name = string("op_3270_cast_fp16")];
            fp32 var_3271_epsilon_0 = const()[name = string("op_3271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_3271_cast_fp16 = rsqrt(epsilon = var_3271_epsilon_0, x = var_3270_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_49_cast_fp16 = mul(x = y_55_cast_fp16, y = var_3271_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23205568)))];
            tensor<fp16, [1, 2, 128, 256]> k_19_cast_fp16 = mul(x = out_49_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_3224_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2, 128, 256]> var_3274_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_3274_cast_fp16")];
            bool var_3277 = const()[name = string("op_3277"), val = bool(true)];
            tensor<int32, [1]> var_3278_axes_0 = const()[name = string("op_3278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_3278_cast_fp16 = reduce_max(axes = var_3278_axes_0, keep_dims = var_3277, x = var_3274_cast_fp16)[name = string("op_3278_cast_fp16")];
            fp16 var_3279_to_fp16 = const()[name = string("op_3279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_57_cast_fp16 = maximum(x = var_3278_cast_fp16, y = var_3279_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_57_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_57_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_3282_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_3282_cast_fp16")];
            tensor<int32, [1]> var_3287_axes_0 = const()[name = string("op_3287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3287_keep_dims_0 = const()[name = string("op_3287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_3287_cast_fp16 = reduce_mean(axes = var_3287_axes_0, keep_dims = var_3287_keep_dims_0, x = var_3282_cast_fp16)[name = string("op_3287_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3288_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_3289_cast_fp16 = inverse(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            fp16 var_3290_to_fp16 = const()[name = string("op_3290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_3291_cast_fp16 = mul(x = var_3289_cast_fp16, y = var_3290_to_fp16)[name = string("op_3291_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3293_cast_fp16 = add(x = var_3287_cast_fp16, y = var_3291_cast_fp16)[name = string("op_3293_cast_fp16")];
            fp32 var_3294_epsilon_0 = const()[name = string("op_3294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_3294_cast_fp16 = rsqrt(epsilon = var_3294_epsilon_0, x = var_3293_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_13_cast_fp16 = mul(x = y_57_cast_fp16, y = var_3294_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_3296_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3296_cast_fp16")];
            tensor<int32, [4]> var_3301_begin_0 = const()[name = string("op_3301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3301_end_0 = const()[name = string("op_3301_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_3301_end_mask_0 = const()[name = string("op_3301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_3301_cast_fp16 = slice_by_index(begin = var_3301_begin_0, end = var_3301_end_0, end_mask = var_3301_end_mask_0, x = q_13_cast_fp16)[name = string("op_3301_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_3302_cast_fp16 = mul(x = var_3301_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3307_begin_0 = const()[name = string("op_3307_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3307_end_0 = const()[name = string("op_3307_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_3307_end_mask_0 = const()[name = string("op_3307_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_3307_cast_fp16 = slice_by_index(begin = var_3307_begin_0, end = var_3307_end_0, end_mask = var_3307_end_mask_0, x = q_13_cast_fp16)[name = string("op_3307_cast_fp16")];
            int32 var_3309 = const()[name = string("op_3309"), val = int32(-1)];
            bool var_3310_interleave_0 = const()[name = string("op_3310_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_3310_cast_fp16 = concat(axis = var_3309, interleave = var_3310_interleave_0, values = (var_3302_cast_fp16, var_3307_cast_fp16))[name = string("op_3310_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_3311_cast_fp16 = mul(x = var_3310_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3311_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_15_cast_fp16 = add(x = var_3296_cast_fp16, y = var_3311_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_3314_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3314_cast_fp16")];
            tensor<int32, [4]> var_3319_begin_0 = const()[name = string("op_3319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3319_end_0 = const()[name = string("op_3319_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_3319_end_mask_0 = const()[name = string("op_3319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_3319_cast_fp16 = slice_by_index(begin = var_3319_begin_0, end = var_3319_end_0, end_mask = var_3319_end_mask_0, x = k_19_cast_fp16)[name = string("op_3319_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_3320_cast_fp16 = mul(x = var_3319_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_3320_cast_fp16")];
            tensor<int32, [4]> var_3325_begin_0 = const()[name = string("op_3325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3325_end_0 = const()[name = string("op_3325_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_3325_end_mask_0 = const()[name = string("op_3325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_3325_cast_fp16 = slice_by_index(begin = var_3325_begin_0, end = var_3325_end_0, end_mask = var_3325_end_mask_0, x = k_19_cast_fp16)[name = string("op_3325_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), val = int32(-1)];
            bool var_3328_interleave_0 = const()[name = string("op_3328_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_3328_cast_fp16 = concat(axis = var_3327, interleave = var_3328_interleave_0, values = (var_3320_cast_fp16, var_3325_cast_fp16))[name = string("op_3328_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_3329_cast_fp16 = mul(x = var_3328_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3329_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_21_cast_fp16 = add(x = var_3314_cast_fp16, y = var_3329_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = k_21_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [4, 1, 128, 256]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_12_cast_fp16 = reshape(shape = concat_12, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [4, 1, 128, 256]> reshape_13_cast_fp16 = reshape(shape = concat_13, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = v_13_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [4, 1, 128, 256]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_14_cast_fp16 = reshape(shape = concat_14, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [4, 1, 128, 256]> reshape_15_cast_fp16 = reshape(shape = concat_15, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3343_transpose_x_0 = const()[name = string("op_3343_transpose_x_0"), val = bool(false)];
            bool var_3343_transpose_y_0 = const()[name = string("op_3343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 4, 128, 128]> var_3343_cast_fp16 = matmul(transpose_x = var_3343_transpose_x_0, transpose_y = var_3343_transpose_y_0, x = q_15_cast_fp16, y = transpose_99_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_3345_cast_fp16 = add(x = var_3343_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3345_cast_fp16")];
            int32 var_3346 = const()[name = string("op_3346"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_57_cast_fp16 = softmax(axis = var_3346, x = var_3345_cast_fp16)[name = string("w_57_cast_fp16")];
            bool var_3349_transpose_x_0 = const()[name = string("op_3349_transpose_x_0"), val = bool(false)];
            bool var_3349_transpose_y_0 = const()[name = string("op_3349_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 4, 128, 256]> var_3349_cast_fp16 = matmul(transpose_x = var_3349_transpose_x_0, transpose_y = var_3349_transpose_y_0, x = w_57_cast_fp16, y = v_15_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<int32, [4]> var_3352_perm_0 = const()[name = string("op_3352_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3356 = const()[name = string("op_3356"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_3352_cast_fp16 = transpose(perm = var_3352_perm_0, x = var_3349_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 128, 1024]> input_39_cast_fp16 = reshape(shape = var_3356, x = var_3352_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23206144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23730496))))[name = string("tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3361_cast_fp16 = abs(x = linear_31_cast_fp16)[name = string("op_3361_cast_fp16")];
            bool var_3364 = const()[name = string("op_3364"), val = bool(true)];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3365_cast_fp16 = reduce_max(axes = var_3365_axes_0, keep_dims = var_3364, x = var_3361_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_59_cast_fp16 = maximum(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_59_cast_fp16 = real_div(x = linear_31_cast_fp16, y = s_59_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3369_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<int32, [1]> var_3374_axes_0 = const()[name = string("op_3374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3374_keep_dims_0 = const()[name = string("op_3374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3374_cast_fp16 = reduce_mean(axes = var_3374_axes_0, keep_dims = var_3374_keep_dims_0, x = var_3369_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3375_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_3375_cast_fp16")];
            fp32 var_3376_epsilon_0 = const()[name = string("op_3376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3376_cast_fp16 = inverse(epsilon = var_3376_epsilon_0, x = var_3375_cast_fp16)[name = string("op_3376_cast_fp16")];
            fp16 var_3377_to_fp16 = const()[name = string("op_3377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3378_cast_fp16 = mul(x = var_3376_cast_fp16, y = var_3377_to_fp16)[name = string("op_3378_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3380_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3380_cast_fp16")];
            fp32 var_3381_epsilon_0 = const()[name = string("op_3381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3381_cast_fp16 = rsqrt(epsilon = var_3381_epsilon_0, x = var_3380_cast_fp16)[name = string("op_3381_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_3381_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23731584)))];
            tensor<fp16, [1, 128, 512]> var_3383_cast_fp16 = mul(x = out_51_cast_fp16, y = tm_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_65_cast_fp16 = add(x = x_55_cast_fp16, y = var_3383_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3386_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_3386_cast_fp16")];
            bool var_3389 = const()[name = string("op_3389"), val = bool(true)];
            tensor<int32, [1]> var_3390_axes_0 = const()[name = string("op_3390_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3390_cast_fp16 = reduce_max(axes = var_3390_axes_0, keep_dims = var_3389, x = var_3386_cast_fp16)[name = string("op_3390_cast_fp16")];
            fp16 var_3391_to_fp16 = const()[name = string("op_3391_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_61_cast_fp16 = maximum(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_61_cast_fp16 = real_div(x = x_65_cast_fp16, y = s_61_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3394_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<int32, [1]> var_3399_axes_0 = const()[name = string("op_3399_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3399_keep_dims_0 = const()[name = string("op_3399_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3399_cast_fp16 = reduce_mean(axes = var_3399_axes_0, keep_dims = var_3399_keep_dims_0, x = var_3394_cast_fp16)[name = string("op_3399_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3400_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_3400_cast_fp16")];
            fp32 var_3401_epsilon_0 = const()[name = string("op_3401_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3401_cast_fp16 = inverse(epsilon = var_3401_epsilon_0, x = var_3400_cast_fp16)[name = string("op_3401_cast_fp16")];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3403_cast_fp16 = mul(x = var_3401_cast_fp16, y = var_3402_to_fp16)[name = string("op_3403_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3405_cast_fp16 = add(x = var_3399_cast_fp16, y = var_3403_cast_fp16)[name = string("op_3405_cast_fp16")];
            fp32 var_3406_epsilon_0 = const()[name = string("op_3406_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3406_cast_fp16 = rsqrt(epsilon = var_3406_epsilon_0, x = var_3405_cast_fp16)[name = string("op_3406_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_3406_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23732672)))];
            tensor<fp16, [1, 128, 512]> input_41_cast_fp16 = mul(x = out_53_cast_fp16, y = tm_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23733760))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24782400))))[name = string("tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_32_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_32_cast_fp16")];
            string var_3413_mode_0 = const()[name = string("op_3413_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_3413_cast_fp16 = gelu(mode = var_3413_mode_0, x = linear_32_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24786560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25835200))))[name = string("tm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_43_cast_fp16 = mul(x = var_3413_cast_fp16, y = linear_33_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25839360))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26888000))))[name = string("tm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3421_cast_fp16 = abs(x = linear_34_cast_fp16)[name = string("op_3421_cast_fp16")];
            bool var_3424 = const()[name = string("op_3424"), val = bool(true)];
            tensor<int32, [1]> var_3425_axes_0 = const()[name = string("op_3425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3425_cast_fp16 = reduce_max(axes = var_3425_axes_0, keep_dims = var_3424, x = var_3421_cast_fp16)[name = string("op_3425_cast_fp16")];
            fp16 var_3426_to_fp16 = const()[name = string("op_3426_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_63_cast_fp16 = maximum(x = var_3425_cast_fp16, y = var_3426_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_63_cast_fp16 = real_div(x = linear_34_cast_fp16, y = s_63_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3429_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_3429_cast_fp16")];
            tensor<int32, [1]> var_3434_axes_0 = const()[name = string("op_3434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3434_keep_dims_0 = const()[name = string("op_3434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3434_cast_fp16 = reduce_mean(axes = var_3434_axes_0, keep_dims = var_3434_keep_dims_0, x = var_3429_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3435_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp32 var_3436_epsilon_0 = const()[name = string("op_3436_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3436_cast_fp16 = inverse(epsilon = var_3436_epsilon_0, x = var_3435_cast_fp16)[name = string("op_3436_cast_fp16")];
            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3438_cast_fp16 = mul(x = var_3436_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3440_cast_fp16 = add(x = var_3434_cast_fp16, y = var_3438_cast_fp16)[name = string("op_3440_cast_fp16")];
            fp32 var_3441_epsilon_0 = const()[name = string("op_3441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3441_cast_fp16 = rsqrt(epsilon = var_3441_epsilon_0, x = var_3440_cast_fp16)[name = string("op_3441_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_55_cast_fp16 = mul(x = y_63_cast_fp16, y = var_3441_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26889088)))];
            tensor<fp16, [1, 128, 512]> var_3443_cast_fp16 = mul(x = out_55_cast_fp16, y = tm_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_45_cast_fp16 = add(x = x_65_cast_fp16, y = var_3443_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26890176))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27152384))))[name = string("tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_35_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_35_cast_fp16")];
            string x_69_mode_0 = const()[name = string("x_69_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_69_cast_fp16 = gelu(mode = x_69_mode_0, x = linear_35_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [4]> var_3463_begin_0 = const()[name = string("op_3463_begin_0"), val = tensor<int32, [4]>([0, 0, 3, 0])];
            tensor<int32, [4]> var_3463_end_0 = const()[name = string("op_3463_end_0"), val = tensor<int32, [4]>([1, 128, 4, 512])];
            tensor<bool, [4]> var_3463_end_mask_0 = const()[name = string("op_3463_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3463_squeeze_mask_0 = const()[name = string("op_3463_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_3463_cast_fp16 = slice_by_index(begin = var_3463_begin_0, end = var_3463_end_0, end_mask = var_3463_end_mask_0, squeeze_mask = var_3463_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3463_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_47_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3463_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27153472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27415680))))[name = string("tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3473_cast_fp16 = abs(x = linear_36_cast_fp16)[name = string("op_3473_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<int32, [1]> var_3477_axes_0 = const()[name = string("op_3477_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3477_cast_fp16 = reduce_max(axes = var_3477_axes_0, keep_dims = var_3476, x = var_3473_cast_fp16)[name = string("op_3477_cast_fp16")];
            fp16 var_3478_to_fp16 = const()[name = string("op_3478_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_65_cast_fp16 = maximum(x = var_3477_cast_fp16, y = var_3478_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_65_cast_fp16 = real_div(x = linear_36_cast_fp16, y = s_65_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3481_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [1]> var_3486_axes_0 = const()[name = string("op_3486_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3486_keep_dims_0 = const()[name = string("op_3486_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3486_cast_fp16 = reduce_mean(axes = var_3486_axes_0, keep_dims = var_3486_keep_dims_0, x = var_3481_cast_fp16)[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3487_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp32 var_3488_epsilon_0 = const()[name = string("op_3488_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3488_cast_fp16 = inverse(epsilon = var_3488_epsilon_0, x = var_3487_cast_fp16)[name = string("op_3488_cast_fp16")];
            fp16 var_3489_to_fp16 = const()[name = string("op_3489_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3490_cast_fp16 = mul(x = var_3488_cast_fp16, y = var_3489_to_fp16)[name = string("op_3490_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3492_cast_fp16 = add(x = var_3486_cast_fp16, y = var_3490_cast_fp16)[name = string("op_3492_cast_fp16")];
            fp32 var_3493_epsilon_0 = const()[name = string("op_3493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3493_cast_fp16 = rsqrt(epsilon = var_3493_epsilon_0, x = var_3492_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_57_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3493_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27416768)))];
            tensor<fp16, [1, 128, 512]> var_3495_cast_fp16 = mul(x = out_57_cast_fp16, y = tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3495_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_7_cast_fp16 = add(x = input_45_cast_fp16, y = var_3495_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<fp16, [1]> tm_layers_3_layer_scalar_to_fp16 = const()[name = string("tm_layers_3_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 128, 512]> x_73_cast_fp16 = mul(x = h_7_cast_fp16, y = tm_layers_3_layer_scalar_to_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3499_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_3499_cast_fp16")];
            bool var_3502 = const()[name = string("op_3502"), val = bool(true)];
            tensor<int32, [1]> var_3503_axes_0 = const()[name = string("op_3503_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3503_cast_fp16 = reduce_max(axes = var_3503_axes_0, keep_dims = var_3502, x = var_3499_cast_fp16)[name = string("op_3503_cast_fp16")];
            fp16 var_3504_to_fp16 = const()[name = string("op_3504_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_67_cast_fp16 = maximum(x = var_3503_cast_fp16, y = var_3504_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_67_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_67_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3507_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<int32, [1]> var_3512_axes_0 = const()[name = string("op_3512_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3512_keep_dims_0 = const()[name = string("op_3512_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3512_cast_fp16 = reduce_mean(axes = var_3512_axes_0, keep_dims = var_3512_keep_dims_0, x = var_3507_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3513_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_3513_cast_fp16")];
            fp32 var_3514_epsilon_0 = const()[name = string("op_3514_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3514_cast_fp16 = inverse(epsilon = var_3514_epsilon_0, x = var_3513_cast_fp16)[name = string("op_3514_cast_fp16")];
            fp16 var_3515_to_fp16 = const()[name = string("op_3515_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3516_cast_fp16 = mul(x = var_3514_cast_fp16, y = var_3515_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3518_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp32 var_3519_epsilon_0 = const()[name = string("op_3519_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3519_cast_fp16 = rsqrt(epsilon = var_3519_epsilon_0, x = var_3518_cast_fp16)[name = string("op_3519_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_59_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3519_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27417856)))];
            tensor<fp16, [1, 128, 512]> input_49_cast_fp16 = mul(x = out_59_cast_fp16, y = tm_layers_4_input_layernorm_weight_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27418944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27943296))))[name = string("tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3529 = const()[name = string("op_3529"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_3530_cast_fp16 = reshape(shape = var_3529, x = linear_37_cast_fp16)[name = string("op_3530_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27945408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28207616))))[name = string("tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3541 = const()[name = string("op_3541"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_3542_cast_fp16 = reshape(shape = var_3541, x = linear_38_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28208704))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28470912))))[name = string("tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_3553 = const()[name = string("op_3553"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_3554_cast_fp16 = reshape(shape = var_3553, x = linear_39_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<int32, [4]> x_79_perm_0 = const()[name = string("x_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3530_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 4, 128, 256]> var_3558_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3558_cast_fp16")];
            bool var_3561 = const()[name = string("op_3561"), val = bool(true)];
            tensor<int32, [1]> var_3562_axes_0 = const()[name = string("op_3562_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_3562_cast_fp16 = reduce_max(axes = var_3562_axes_0, keep_dims = var_3561, x = var_3558_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp16 var_3563_to_fp16 = const()[name = string("op_3563_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_69_cast_fp16 = maximum(x = var_3562_cast_fp16, y = var_3563_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_69_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_69_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_3566_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3566_cast_fp16")];
            tensor<int32, [1]> var_3571_axes_0 = const()[name = string("op_3571_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3571_keep_dims_0 = const()[name = string("op_3571_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_3571_cast_fp16 = reduce_mean(axes = var_3571_axes_0, keep_dims = var_3571_keep_dims_0, x = var_3566_cast_fp16)[name = string("op_3571_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_3572_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3572_cast_fp16")];
            fp32 var_3573_epsilon_0 = const()[name = string("op_3573_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_3573_cast_fp16 = inverse(epsilon = var_3573_epsilon_0, x = var_3572_cast_fp16)[name = string("op_3573_cast_fp16")];
            fp16 var_3574_to_fp16 = const()[name = string("op_3574_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_3575_cast_fp16 = mul(x = var_3573_cast_fp16, y = var_3574_to_fp16)[name = string("op_3575_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_3577_cast_fp16 = add(x = var_3571_cast_fp16, y = var_3575_cast_fp16)[name = string("op_3577_cast_fp16")];
            fp32 var_3578_epsilon_0 = const()[name = string("op_3578_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_3578_cast_fp16 = rsqrt(epsilon = var_3578_epsilon_0, x = var_3577_cast_fp16)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_61_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3578_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [256]> tm_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472000)))];
            tensor<fp16, [1, 4, 128, 256]> q_17_cast_fp16 = mul(x = out_61_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3542_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2, 128, 256]> var_3581_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3581_cast_fp16")];
            bool var_3584 = const()[name = string("op_3584"), val = bool(true)];
            tensor<int32, [1]> var_3585_axes_0 = const()[name = string("op_3585_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_3585_cast_fp16 = reduce_max(axes = var_3585_axes_0, keep_dims = var_3584, x = var_3581_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp16 var_3586_to_fp16 = const()[name = string("op_3586_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_71_cast_fp16 = maximum(x = var_3585_cast_fp16, y = var_3586_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_71_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_71_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_3589_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3589_cast_fp16")];
            tensor<int32, [1]> var_3594_axes_0 = const()[name = string("op_3594_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3594_keep_dims_0 = const()[name = string("op_3594_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_3594_cast_fp16 = reduce_mean(axes = var_3594_axes_0, keep_dims = var_3594_keep_dims_0, x = var_3589_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3595_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp32 var_3596_epsilon_0 = const()[name = string("op_3596_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_3596_cast_fp16 = inverse(epsilon = var_3596_epsilon_0, x = var_3595_cast_fp16)[name = string("op_3596_cast_fp16")];
            fp16 var_3597_to_fp16 = const()[name = string("op_3597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_3598_cast_fp16 = mul(x = var_3596_cast_fp16, y = var_3597_to_fp16)[name = string("op_3598_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3600_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3598_cast_fp16)[name = string("op_3600_cast_fp16")];
            fp32 var_3601_epsilon_0 = const()[name = string("op_3601_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_3601_cast_fp16 = rsqrt(epsilon = var_3601_epsilon_0, x = var_3600_cast_fp16)[name = string("op_3601_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_63_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3601_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472576)))];
            tensor<fp16, [1, 2, 128, 256]> k_25_cast_fp16 = mul(x = out_63_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_79_cast_fp16 = transpose(perm = x_79_perm_0, x = var_3554_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 2, 128, 256]> var_3604_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3604_cast_fp16")];
            bool var_3607 = const()[name = string("op_3607"), val = bool(true)];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_3607, x = var_3604_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_73_cast_fp16 = maximum(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_73_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_73_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_3612_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<int32, [1]> var_3617_axes_0 = const()[name = string("op_3617_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3617_keep_dims_0 = const()[name = string("op_3617_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_3617_cast_fp16 = reduce_mean(axes = var_3617_axes_0, keep_dims = var_3617_keep_dims_0, x = var_3612_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3618_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_3619_cast_fp16 = inverse(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            fp16 var_3620_to_fp16 = const()[name = string("op_3620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_3621_cast_fp16 = mul(x = var_3619_cast_fp16, y = var_3620_to_fp16)[name = string("op_3621_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_3623_cast_fp16 = add(x = var_3617_cast_fp16, y = var_3621_cast_fp16)[name = string("op_3623_cast_fp16")];
            fp32 var_3624_epsilon_0 = const()[name = string("op_3624_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_3624_cast_fp16 = rsqrt(epsilon = var_3624_epsilon_0, x = var_3623_cast_fp16)[name = string("op_3624_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_17_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3624_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_3626_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3626_cast_fp16")];
            tensor<int32, [4]> var_3631_begin_0 = const()[name = string("op_3631_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3631_end_0 = const()[name = string("op_3631_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_3631_end_mask_0 = const()[name = string("op_3631_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_3631_cast_fp16 = slice_by_index(begin = var_3631_begin_0, end = var_3631_end_0, end_mask = var_3631_end_mask_0, x = q_17_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_3632_cast_fp16 = mul(x = var_3631_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> var_3637_begin_0 = const()[name = string("op_3637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3637_end_0 = const()[name = string("op_3637_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_3637_end_mask_0 = const()[name = string("op_3637_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_3637_cast_fp16 = slice_by_index(begin = var_3637_begin_0, end = var_3637_end_0, end_mask = var_3637_end_mask_0, x = q_17_cast_fp16)[name = string("op_3637_cast_fp16")];
            int32 var_3639 = const()[name = string("op_3639"), val = int32(-1)];
            bool var_3640_interleave_0 = const()[name = string("op_3640_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_3640_cast_fp16 = concat(axis = var_3639, interleave = var_3640_interleave_0, values = (var_3632_cast_fp16, var_3637_cast_fp16))[name = string("op_3640_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_3641_cast_fp16 = mul(x = var_3640_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3641_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_19_cast_fp16 = add(x = var_3626_cast_fp16, y = var_3641_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_3644_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = k_25_cast_fp16)[name = string("op_3649_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_3650_cast_fp16 = mul(x = var_3649_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            tensor<int32, [4]> var_3655_begin_0 = const()[name = string("op_3655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3655_end_0 = const()[name = string("op_3655_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_3655_end_mask_0 = const()[name = string("op_3655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_3655_cast_fp16 = slice_by_index(begin = var_3655_begin_0, end = var_3655_end_0, end_mask = var_3655_end_mask_0, x = k_25_cast_fp16)[name = string("op_3655_cast_fp16")];
            int32 var_3657 = const()[name = string("op_3657"), val = int32(-1)];
            bool var_3658_interleave_0 = const()[name = string("op_3658_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_3658_cast_fp16 = concat(axis = var_3657, interleave = var_3658_interleave_0, values = (var_3650_cast_fp16, var_3655_cast_fp16))[name = string("op_3658_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_3659_cast_fp16 = mul(x = var_3658_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3659_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_27_cast_fp16 = add(x = var_3644_cast_fp16, y = var_3659_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = k_27_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [4, 1, 128, 256]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
            tensor<int32, [5]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_16_cast_fp16 = reshape(shape = concat_16, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [4, 1, 128, 256]> reshape_17_cast_fp16 = reshape(shape = concat_17, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = v_17_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [4, 1, 128, 256]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
            tensor<int32, [5]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_18_cast_fp16 = reshape(shape = concat_18, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [4, 1, 128, 256]> reshape_19_cast_fp16 = reshape(shape = concat_19, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3673_transpose_x_0 = const()[name = string("op_3673_transpose_x_0"), val = bool(false)];
            bool var_3673_transpose_y_0 = const()[name = string("op_3673_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 4, 128, 128]> var_3673_cast_fp16 = matmul(transpose_x = var_3673_transpose_x_0, transpose_y = var_3673_transpose_y_0, x = q_19_cast_fp16, y = transpose_100_cast_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_3675_cast_fp16 = add(x = var_3673_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3675_cast_fp16")];
            int32 var_3676 = const()[name = string("op_3676"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_73_cast_fp16 = softmax(axis = var_3676, x = var_3675_cast_fp16)[name = string("w_73_cast_fp16")];
            bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)];
            bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 4, 128, 256]> var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = w_73_cast_fp16, y = v_19_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<int32, [4]> var_3682_perm_0 = const()[name = string("op_3682_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3686 = const()[name = string("op_3686"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_3682_cast_fp16 = transpose(perm = var_3682_perm_0, x = var_3679_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 128, 1024]> input_51_cast_fp16 = reshape(shape = var_3686, x = var_3682_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28473152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28997504))))[name = string("tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3691_cast_fp16 = abs(x = linear_40_cast_fp16)[name = string("op_3691_cast_fp16")];
            bool var_3694 = const()[name = string("op_3694"), val = bool(true)];
            tensor<int32, [1]> var_3695_axes_0 = const()[name = string("op_3695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3695_cast_fp16 = reduce_max(axes = var_3695_axes_0, keep_dims = var_3694, x = var_3691_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 var_3696_to_fp16 = const()[name = string("op_3696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_75_cast_fp16 = maximum(x = var_3695_cast_fp16, y = var_3696_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_75_cast_fp16 = real_div(x = linear_40_cast_fp16, y = s_75_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3699_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3699_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3704_keep_dims_0 = const()[name = string("op_3704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3704_cast_fp16 = reduce_mean(axes = var_3704_axes_0, keep_dims = var_3704_keep_dims_0, x = var_3699_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3705_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3705_cast_fp16")];
            fp32 var_3706_epsilon_0 = const()[name = string("op_3706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3706_cast_fp16 = inverse(epsilon = var_3706_epsilon_0, x = var_3705_cast_fp16)[name = string("op_3706_cast_fp16")];
            fp16 var_3707_to_fp16 = const()[name = string("op_3707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3708_cast_fp16 = mul(x = var_3706_cast_fp16, y = var_3707_to_fp16)[name = string("op_3708_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3710_cast_fp16 = add(x = var_3704_cast_fp16, y = var_3708_cast_fp16)[name = string("op_3710_cast_fp16")];
            fp32 var_3711_epsilon_0 = const()[name = string("op_3711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3711_cast_fp16 = rsqrt(epsilon = var_3711_epsilon_0, x = var_3710_cast_fp16)[name = string("op_3711_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3711_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28998592)))];
            tensor<fp16, [1, 128, 512]> var_3713_cast_fp16 = mul(x = out_65_cast_fp16, y = tm_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = var_3713_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3716_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3716_cast_fp16")];
            bool var_3719 = const()[name = string("op_3719"), val = bool(true)];
            tensor<int32, [1]> var_3720_axes_0 = const()[name = string("op_3720_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3720_cast_fp16 = reduce_max(axes = var_3720_axes_0, keep_dims = var_3719, x = var_3716_cast_fp16)[name = string("op_3720_cast_fp16")];
            fp16 var_3721_to_fp16 = const()[name = string("op_3721_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_77_cast_fp16 = maximum(x = var_3720_cast_fp16, y = var_3721_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_77_cast_fp16 = real_div(x = x_83_cast_fp16, y = s_77_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3724_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3729_axes_0 = const()[name = string("op_3729_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3729_keep_dims_0 = const()[name = string("op_3729_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3729_cast_fp16 = reduce_mean(axes = var_3729_axes_0, keep_dims = var_3729_keep_dims_0, x = var_3724_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3730_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3730_cast_fp16")];
            fp32 var_3731_epsilon_0 = const()[name = string("op_3731_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3731_cast_fp16 = inverse(epsilon = var_3731_epsilon_0, x = var_3730_cast_fp16)[name = string("op_3731_cast_fp16")];
            fp16 var_3732_to_fp16 = const()[name = string("op_3732_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3733_cast_fp16 = mul(x = var_3731_cast_fp16, y = var_3732_to_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3735_cast_fp16 = add(x = var_3729_cast_fp16, y = var_3733_cast_fp16)[name = string("op_3735_cast_fp16")];
            fp32 var_3736_epsilon_0 = const()[name = string("op_3736_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3736_cast_fp16 = rsqrt(epsilon = var_3736_epsilon_0, x = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_67_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3736_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28999680)))];
            tensor<fp16, [1, 128, 512]> input_53_cast_fp16 = mul(x = out_67_cast_fp16, y = tm_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29000768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30049408))))[name = string("tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_41_cast_fp16")];
            string var_3743_mode_0 = const()[name = string("op_3743_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_3743_cast_fp16 = gelu(mode = var_3743_mode_0, x = linear_41_cast_fp16)[name = string("op_3743_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30053568))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31102208))))[name = string("tm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_42_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_55_cast_fp16 = mul(x = var_3743_cast_fp16, y = linear_42_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31106368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32155008))))[name = string("tm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3751_cast_fp16 = abs(x = linear_43_cast_fp16)[name = string("op_3751_cast_fp16")];
            bool var_3754 = const()[name = string("op_3754"), val = bool(true)];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_3754, x = var_3751_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 var_3756_to_fp16 = const()[name = string("op_3756_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_79_cast_fp16 = maximum(x = var_3755_cast_fp16, y = var_3756_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_79_cast_fp16 = real_div(x = linear_43_cast_fp16, y = s_79_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3759_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<int32, [1]> var_3764_axes_0 = const()[name = string("op_3764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3764_keep_dims_0 = const()[name = string("op_3764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3764_cast_fp16 = reduce_mean(axes = var_3764_axes_0, keep_dims = var_3764_keep_dims_0, x = var_3759_cast_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3765_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3766_cast_fp16 = inverse(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3768_cast_fp16 = mul(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("op_3768_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3770_cast_fp16 = add(x = var_3764_cast_fp16, y = var_3768_cast_fp16)[name = string("op_3770_cast_fp16")];
            fp32 var_3771_epsilon_0 = const()[name = string("op_3771_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3771_cast_fp16 = rsqrt(epsilon = var_3771_epsilon_0, x = var_3770_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_69_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3771_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32156096)))];
            tensor<fp16, [1, 128, 512]> var_3773_cast_fp16 = mul(x = out_69_cast_fp16, y = tm_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3773_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_57_cast_fp16 = add(x = x_83_cast_fp16, y = var_3773_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32157184))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32419392))))[name = string("tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_44_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_44_cast_fp16")];
            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_87_cast_fp16 = gelu(mode = x_87_mode_0, x = linear_44_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [4]> var_3793_begin_0 = const()[name = string("op_3793_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_3793_end_0 = const()[name = string("op_3793_end_0"), val = tensor<int32, [4]>([1, 128, 5, 512])];
            tensor<bool, [4]> var_3793_end_mask_0 = const()[name = string("op_3793_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3793_squeeze_mask_0 = const()[name = string("op_3793_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_3793_cast_fp16 = slice_by_index(begin = var_3793_begin_0, end = var_3793_end_0, end_mask = var_3793_end_mask_0, squeeze_mask = var_3793_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_59_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3793_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32420480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32682688))))[name = string("tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3803_cast_fp16 = abs(x = linear_45_cast_fp16)[name = string("op_3803_cast_fp16")];
            bool var_3806 = const()[name = string("op_3806"), val = bool(true)];
            tensor<int32, [1]> var_3807_axes_0 = const()[name = string("op_3807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3807_cast_fp16 = reduce_max(axes = var_3807_axes_0, keep_dims = var_3806, x = var_3803_cast_fp16)[name = string("op_3807_cast_fp16")];
            fp16 var_3808_to_fp16 = const()[name = string("op_3808_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_81_cast_fp16 = maximum(x = var_3807_cast_fp16, y = var_3808_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_81_cast_fp16 = real_div(x = linear_45_cast_fp16, y = s_81_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3811_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [1]> var_3816_axes_0 = const()[name = string("op_3816_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3816_keep_dims_0 = const()[name = string("op_3816_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3816_cast_fp16 = reduce_mean(axes = var_3816_axes_0, keep_dims = var_3816_keep_dims_0, x = var_3811_cast_fp16)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3817_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3817_cast_fp16")];
            fp32 var_3818_epsilon_0 = const()[name = string("op_3818_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3818_cast_fp16 = inverse(epsilon = var_3818_epsilon_0, x = var_3817_cast_fp16)[name = string("op_3818_cast_fp16")];
            fp16 var_3819_to_fp16 = const()[name = string("op_3819_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3820_cast_fp16 = mul(x = var_3818_cast_fp16, y = var_3819_to_fp16)[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3822_cast_fp16 = add(x = var_3816_cast_fp16, y = var_3820_cast_fp16)[name = string("op_3822_cast_fp16")];
            fp32 var_3823_epsilon_0 = const()[name = string("op_3823_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3823_cast_fp16 = rsqrt(epsilon = var_3823_epsilon_0, x = var_3822_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_71_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3823_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32683776)))];
            tensor<fp16, [1, 128, 512]> var_3825_cast_fp16 = mul(x = out_71_cast_fp16, y = tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3825_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_9_cast_fp16 = add(x = input_57_cast_fp16, y = var_3825_cast_fp16)[name = string("h_9_cast_fp16")];
            tensor<fp16, [1]> tm_layers_4_layer_scalar_to_fp16 = const()[name = string("tm_layers_4_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.62p-1])];
            tensor<fp16, [1, 128, 512]> x_91_cast_fp16 = mul(x = h_9_cast_fp16, y = tm_layers_4_layer_scalar_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3829_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3829_cast_fp16")];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<int32, [1]> var_3833_axes_0 = const()[name = string("op_3833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3833_cast_fp16 = reduce_max(axes = var_3833_axes_0, keep_dims = var_3832, x = var_3829_cast_fp16)[name = string("op_3833_cast_fp16")];
            fp16 var_3834_to_fp16 = const()[name = string("op_3834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_83_cast_fp16 = maximum(x = var_3833_cast_fp16, y = var_3834_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_83_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_83_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_3837_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3837_cast_fp16")];
            tensor<int32, [1]> var_3842_axes_0 = const()[name = string("op_3842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3842_keep_dims_0 = const()[name = string("op_3842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3842_cast_fp16 = reduce_mean(axes = var_3842_axes_0, keep_dims = var_3842_keep_dims_0, x = var_3837_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3843_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp32 var_3844_epsilon_0 = const()[name = string("op_3844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3844_cast_fp16 = inverse(epsilon = var_3844_epsilon_0, x = var_3843_cast_fp16)[name = string("op_3844_cast_fp16")];
            fp16 var_3845_to_fp16 = const()[name = string("op_3845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3846_cast_fp16 = mul(x = var_3844_cast_fp16, y = var_3845_to_fp16)[name = string("op_3846_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3848_cast_fp16 = add(x = var_3842_cast_fp16, y = var_3846_cast_fp16)[name = string("op_3848_cast_fp16")];
            fp32 var_3849_epsilon_0 = const()[name = string("op_3849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3849_cast_fp16 = rsqrt(epsilon = var_3849_epsilon_0, x = var_3848_cast_fp16)[name = string("op_3849_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_73_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3849_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32684864)))];
            tensor<fp16, [1, 128, 512]> input_61_cast_fp16 = mul(x = out_73_cast_fp16, y = tm_layers_5_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32685952))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33734592))))[name = string("tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_46_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [4]> var_3859 = const()[name = string("op_3859"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 128, 4, 512]> var_3860_cast_fp16 = reshape(shape = var_3859, x = linear_46_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<int32, [4]> x_93_perm_0 = const()[name = string("x_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33738752))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34000960))))[name = string("tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<int32, [4]> var_3871 = const()[name = string("op_3871"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 128, 1, 512]> var_3872_cast_fp16 = reshape(shape = var_3871, x = linear_47_cast_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [4]> x_95_perm_0 = const()[name = string("x_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34002048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34264256))))[name = string("tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [4]> var_3883 = const()[name = string("op_3883"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 128, 1, 512]> var_3884_cast_fp16 = reshape(shape = var_3883, x = linear_48_cast_fp16)[name = string("op_3884_cast_fp16")];
            tensor<int32, [4]> x_97_perm_0 = const()[name = string("x_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 512]> x_93_cast_fp16 = transpose(perm = x_93_perm_0, x = var_3860_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 4, 128, 512]> var_3888_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3888_cast_fp16")];
            bool var_3891 = const()[name = string("op_3891"), val = bool(true)];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_3892_cast_fp16 = reduce_max(axes = var_3892_axes_0, keep_dims = var_3891, x = var_3888_cast_fp16)[name = string("op_3892_cast_fp16")];
            fp16 var_3893_to_fp16 = const()[name = string("op_3893_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_85_cast_fp16 = maximum(x = var_3892_cast_fp16, y = var_3893_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> y_85_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_85_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_3896_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3896_cast_fp16")];
            tensor<int32, [1]> var_3901_axes_0 = const()[name = string("op_3901_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3901_keep_dims_0 = const()[name = string("op_3901_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_3901_cast_fp16 = reduce_mean(axes = var_3901_axes_0, keep_dims = var_3901_keep_dims_0, x = var_3896_cast_fp16)[name = string("op_3901_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_3902_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3902_cast_fp16")];
            fp32 var_3903_epsilon_0 = const()[name = string("op_3903_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_3903_cast_fp16 = inverse(epsilon = var_3903_epsilon_0, x = var_3902_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_3905_cast_fp16 = mul(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_3907_cast_fp16 = add(x = var_3901_cast_fp16, y = var_3905_cast_fp16)[name = string("op_3907_cast_fp16")];
            fp32 var_3908_epsilon_0 = const()[name = string("op_3908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_3908_cast_fp16 = rsqrt(epsilon = var_3908_epsilon_0, x = var_3907_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> out_75_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3908_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34265344)))];
            tensor<fp16, [1, 4, 128, 512]> q_21_cast_fp16 = mul(x = out_75_cast_fp16, y = tm_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> x_95_cast_fp16 = transpose(perm = x_95_perm_0, x = var_3872_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1, 128, 512]> var_3911_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 128, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 128, 1]> s_87_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> y_87_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_87_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_3919_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_3925_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 128, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 128, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 128, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> out_77_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3931_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34266432)))];
            tensor<fp16, [1, 1, 128, 512]> k_31_cast_fp16 = mul(x = out_77_cast_fp16, y = tm_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = var_3884_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 1, 128, 512]> var_3934_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3934_cast_fp16")];
            bool var_3937 = const()[name = string("op_3937"), val = bool(true)];
            tensor<int32, [1]> var_3938_axes_0 = const()[name = string("op_3938_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 128, 1]> var_3938_cast_fp16 = reduce_max(axes = var_3938_axes_0, keep_dims = var_3937, x = var_3934_cast_fp16)[name = string("op_3938_cast_fp16")];
            fp16 var_3939_to_fp16 = const()[name = string("op_3939_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 128, 1]> s_89_cast_fp16 = maximum(x = var_3938_cast_fp16, y = var_3939_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> y_89_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_89_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_3942_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3942_cast_fp16")];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3947_keep_dims_0 = const()[name = string("op_3947_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_3947_cast_fp16 = reduce_mean(axes = var_3947_axes_0, keep_dims = var_3947_keep_dims_0, x = var_3942_cast_fp16)[name = string("op_3947_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_3948_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3948_cast_fp16")];
            fp32 var_3949_epsilon_0 = const()[name = string("op_3949_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 128, 1]> var_3949_cast_fp16 = inverse(epsilon = var_3949_epsilon_0, x = var_3948_cast_fp16)[name = string("op_3949_cast_fp16")];
            fp16 var_3950_to_fp16 = const()[name = string("op_3950_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 128, 1]> var_3951_cast_fp16 = mul(x = var_3949_cast_fp16, y = var_3950_to_fp16)[name = string("op_3951_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_3953_cast_fp16 = add(x = var_3947_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3953_cast_fp16")];
            fp32 var_3954_epsilon_0 = const()[name = string("op_3954_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 128, 1]> var_3954_cast_fp16 = rsqrt(epsilon = var_3954_epsilon_0, x = var_3953_cast_fp16)[name = string("op_3954_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> v_21_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3954_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<fp16, [128, 512]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138828416))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138794944))))[name = string("cos_to_fp16_quantized")];
            tensor<fp16, [1, 4, 128, 512]> var_3956_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3956_cast_fp16")];
            tensor<int32, [4]> var_3961_begin_0 = const()[name = string("op_3961_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3961_end_0 = const()[name = string("op_3961_end_0"), val = tensor<int32, [4]>([1, 4, 128, 512])];
            tensor<bool, [4]> var_3961_end_mask_0 = const()[name = string("op_3961_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 256]> var_3961_cast_fp16 = slice_by_index(begin = var_3961_begin_0, end = var_3961_end_0, end_mask = var_3961_end_mask_0, x = q_21_cast_fp16)[name = string("op_3961_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 256]> var_3962_cast_fp16 = mul(x = var_3961_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_3962_cast_fp16")];
            tensor<int32, [4]> var_3967_begin_0 = const()[name = string("op_3967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3967_end_0 = const()[name = string("op_3967_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_3967_end_mask_0 = const()[name = string("op_3967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 256]> var_3967_cast_fp16 = slice_by_index(begin = var_3967_begin_0, end = var_3967_end_0, end_mask = var_3967_end_mask_0, x = q_21_cast_fp16)[name = string("op_3967_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(-1)];
            bool var_3970_interleave_0 = const()[name = string("op_3970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 512]> var_3970_cast_fp16 = concat(axis = var_3969, interleave = var_3970_interleave_0, values = (var_3962_cast_fp16, var_3967_cast_fp16))[name = string("op_3970_cast_fp16")];
            tensor<fp16, [128, 512]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138894016))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138959616))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [1, 4, 128, 512]> var_3971_cast_fp16 = mul(x = var_3970_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3971_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> q_23_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3971_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_3974_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3974_cast_fp16")];
            tensor<int32, [4]> var_3979_begin_0 = const()[name = string("op_3979_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3979_end_0 = const()[name = string("op_3979_end_0"), val = tensor<int32, [4]>([1, 1, 128, 512])];
            tensor<bool, [4]> var_3979_end_mask_0 = const()[name = string("op_3979_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 128, 256]> var_3979_cast_fp16 = slice_by_index(begin = var_3979_begin_0, end = var_3979_end_0, end_mask = var_3979_end_mask_0, x = k_31_cast_fp16)[name = string("op_3979_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 128, 256]> var_3980_cast_fp16 = mul(x = var_3979_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_3980_cast_fp16")];
            tensor<int32, [4]> var_3985_begin_0 = const()[name = string("op_3985_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3985_end_0 = const()[name = string("op_3985_end_0"), val = tensor<int32, [4]>([1, 1, 128, 256])];
            tensor<bool, [4]> var_3985_end_mask_0 = const()[name = string("op_3985_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 128, 256]> var_3985_cast_fp16 = slice_by_index(begin = var_3985_begin_0, end = var_3985_end_0, end_mask = var_3985_end_mask_0, x = k_31_cast_fp16)[name = string("op_3985_cast_fp16")];
            int32 var_3987 = const()[name = string("op_3987"), val = int32(-1)];
            bool var_3988_interleave_0 = const()[name = string("op_3988_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 512]> var_3988_cast_fp16 = concat(axis = var_3987, interleave = var_3988_interleave_0, values = (var_3980_cast_fp16, var_3985_cast_fp16))[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_3989_cast_fp16 = mul(x = var_3988_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> k_33_cast_fp16 = add(x = var_3974_cast_fp16, y = var_3989_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 128, 512]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = k_33_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [4, 1, 128, 512]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([4, 1, 1, 128, 512])];
            tensor<fp16, [4, 1, 1, 128, 512]> reshape_20_cast_fp16 = reshape(shape = concat_20, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 128, 512])];
            tensor<fp16, [1, 4, 1, 128, 512]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [4, 1, 128, 512]> reshape_21_cast_fp16 = reshape(shape = concat_21, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 128, 512]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = v_21_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [4, 1, 128, 512]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([4, 1, 1, 128, 512])];
            tensor<fp16, [4, 1, 1, 128, 512]> reshape_22_cast_fp16 = reshape(shape = concat_22, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 128, 512])];
            tensor<fp16, [1, 4, 1, 128, 512]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [4, 1, 128, 512]> reshape_23_cast_fp16 = reshape(shape = concat_23, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4003_transpose_x_0 = const()[name = string("op_4003_transpose_x_0"), val = bool(false)];
            bool var_4003_transpose_y_0 = const()[name = string("op_4003_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 128]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 4, 128, 128]> var_4003_cast_fp16 = matmul(transpose_x = var_4003_transpose_x_0, transpose_y = var_4003_transpose_y_0, x = q_23_cast_fp16, y = transpose_101_cast_fp16)[name = string("op_4003_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_4005_cast_fp16 = add(x = var_4003_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4005_cast_fp16")];
            int32 var_4006 = const()[name = string("op_4006"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_89_cast_fp16 = softmax(axis = var_4006, x = var_4005_cast_fp16)[name = string("w_89_cast_fp16")];
            bool var_4009_transpose_x_0 = const()[name = string("op_4009_transpose_x_0"), val = bool(false)];
            bool var_4009_transpose_y_0 = const()[name = string("op_4009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 512]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 4, 128, 512]> var_4009_cast_fp16 = matmul(transpose_x = var_4009_transpose_x_0, transpose_y = var_4009_transpose_y_0, x = w_89_cast_fp16, y = v_23_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4012_perm_0 = const()[name = string("op_4012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4016 = const()[name = string("op_4016"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 128, 4, 512]> var_4012_cast_fp16 = transpose(perm = var_4012_perm_0, x = var_4009_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 128, 2048]> input_63_cast_fp16 = reshape(shape = var_4016, x = var_4012_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34267520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35316160))))[name = string("tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4021_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_4021_cast_fp16")];
            bool var_4024 = const()[name = string("op_4024"), val = bool(true)];
            tensor<int32, [1]> var_4025_axes_0 = const()[name = string("op_4025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4025_cast_fp16 = reduce_max(axes = var_4025_axes_0, keep_dims = var_4024, x = var_4021_cast_fp16)[name = string("op_4025_cast_fp16")];
            fp16 var_4026_to_fp16 = const()[name = string("op_4026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_91_cast_fp16 = maximum(x = var_4025_cast_fp16, y = var_4026_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_91_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_91_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4029_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<int32, [1]> var_4034_axes_0 = const()[name = string("op_4034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4034_keep_dims_0 = const()[name = string("op_4034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4034_cast_fp16 = reduce_mean(axes = var_4034_axes_0, keep_dims = var_4034_keep_dims_0, x = var_4029_cast_fp16)[name = string("op_4034_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4035_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_4035_cast_fp16")];
            fp32 var_4036_epsilon_0 = const()[name = string("op_4036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4036_cast_fp16 = inverse(epsilon = var_4036_epsilon_0, x = var_4035_cast_fp16)[name = string("op_4036_cast_fp16")];
            fp16 var_4037_to_fp16 = const()[name = string("op_4037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4038_cast_fp16 = mul(x = var_4036_cast_fp16, y = var_4037_to_fp16)[name = string("op_4038_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4040_cast_fp16 = add(x = var_4034_cast_fp16, y = var_4038_cast_fp16)[name = string("op_4040_cast_fp16")];
            fp32 var_4041_epsilon_0 = const()[name = string("op_4041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4041_cast_fp16 = rsqrt(epsilon = var_4041_epsilon_0, x = var_4040_cast_fp16)[name = string("op_4041_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_79_cast_fp16 = mul(x = y_91_cast_fp16, y = var_4041_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35317248)))];
            tensor<fp16, [1, 128, 512]> var_4043_cast_fp16 = mul(x = out_79_cast_fp16, y = tm_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_4043_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_101_cast_fp16 = add(x = x_91_cast_fp16, y = var_4043_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4046_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_4046_cast_fp16")];
            bool var_4049 = const()[name = string("op_4049"), val = bool(true)];
            tensor<int32, [1]> var_4050_axes_0 = const()[name = string("op_4050_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4050_cast_fp16 = reduce_max(axes = var_4050_axes_0, keep_dims = var_4049, x = var_4046_cast_fp16)[name = string("op_4050_cast_fp16")];
            fp16 var_4051_to_fp16 = const()[name = string("op_4051_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_93_cast_fp16 = maximum(x = var_4050_cast_fp16, y = var_4051_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_93_cast_fp16 = real_div(x = x_101_cast_fp16, y = s_93_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4054_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [1]> var_4059_axes_0 = const()[name = string("op_4059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4059_keep_dims_0 = const()[name = string("op_4059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4059_cast_fp16 = reduce_mean(axes = var_4059_axes_0, keep_dims = var_4059_keep_dims_0, x = var_4054_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4060_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_4060_cast_fp16")];
            fp32 var_4061_epsilon_0 = const()[name = string("op_4061_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4061_cast_fp16 = inverse(epsilon = var_4061_epsilon_0, x = var_4060_cast_fp16)[name = string("op_4061_cast_fp16")];
            fp16 var_4062_to_fp16 = const()[name = string("op_4062_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4063_cast_fp16 = mul(x = var_4061_cast_fp16, y = var_4062_to_fp16)[name = string("op_4063_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4065_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4063_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp32 var_4066_epsilon_0 = const()[name = string("op_4066_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4066_cast_fp16 = rsqrt(epsilon = var_4066_epsilon_0, x = var_4065_cast_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_81_cast_fp16 = mul(x = y_93_cast_fp16, y = var_4066_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35318336)))];
            tensor<fp16, [1, 128, 512]> input_65_cast_fp16 = mul(x = out_81_cast_fp16, y = tm_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35319424))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36368064))))[name = string("tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_50_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_50_cast_fp16")];
            string var_4073_mode_0 = const()[name = string("op_4073_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_4073_cast_fp16 = gelu(mode = var_4073_mode_0, x = linear_50_cast_fp16)[name = string("op_4073_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36372224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37420864))))[name = string("tm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_51_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_67_cast_fp16 = mul(x = var_4073_cast_fp16, y = linear_51_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37425024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38473664))))[name = string("tm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4081_cast_fp16 = abs(x = linear_52_cast_fp16)[name = string("op_4081_cast_fp16")];
            bool var_4084 = const()[name = string("op_4084"), val = bool(true)];
            tensor<int32, [1]> var_4085_axes_0 = const()[name = string("op_4085_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4085_cast_fp16 = reduce_max(axes = var_4085_axes_0, keep_dims = var_4084, x = var_4081_cast_fp16)[name = string("op_4085_cast_fp16")];
            fp16 var_4086_to_fp16 = const()[name = string("op_4086_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_95_cast_fp16 = maximum(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_95_cast_fp16 = real_div(x = linear_52_cast_fp16, y = s_95_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4089_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_4089_cast_fp16")];
            tensor<int32, [1]> var_4094_axes_0 = const()[name = string("op_4094_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4094_keep_dims_0 = const()[name = string("op_4094_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4094_cast_fp16 = reduce_mean(axes = var_4094_axes_0, keep_dims = var_4094_keep_dims_0, x = var_4089_cast_fp16)[name = string("op_4094_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4095_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_4095_cast_fp16")];
            fp32 var_4096_epsilon_0 = const()[name = string("op_4096_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4096_cast_fp16 = inverse(epsilon = var_4096_epsilon_0, x = var_4095_cast_fp16)[name = string("op_4096_cast_fp16")];
            fp16 var_4097_to_fp16 = const()[name = string("op_4097_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4098_cast_fp16 = mul(x = var_4096_cast_fp16, y = var_4097_to_fp16)[name = string("op_4098_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4100_cast_fp16 = add(x = var_4094_cast_fp16, y = var_4098_cast_fp16)[name = string("op_4100_cast_fp16")];
            fp32 var_4101_epsilon_0 = const()[name = string("op_4101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4101_cast_fp16 = rsqrt(epsilon = var_4101_epsilon_0, x = var_4100_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_83_cast_fp16 = mul(x = y_95_cast_fp16, y = var_4101_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38474752)))];
            tensor<fp16, [1, 128, 512]> var_4103_cast_fp16 = mul(x = out_83_cast_fp16, y = tm_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_69_cast_fp16 = add(x = x_101_cast_fp16, y = var_4103_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38475840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38738048))))[name = string("tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_53_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_53_cast_fp16")];
            string x_105_mode_0 = const()[name = string("x_105_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_105_cast_fp16 = gelu(mode = x_105_mode_0, x = linear_53_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [4]> var_4123_begin_0 = const()[name = string("op_4123_begin_0"), val = tensor<int32, [4]>([0, 0, 5, 0])];
            tensor<int32, [4]> var_4123_end_0 = const()[name = string("op_4123_end_0"), val = tensor<int32, [4]>([1, 128, 6, 512])];
            tensor<bool, [4]> var_4123_end_mask_0 = const()[name = string("op_4123_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4123_squeeze_mask_0 = const()[name = string("op_4123_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_4123_cast_fp16 = slice_by_index(begin = var_4123_begin_0, end = var_4123_end_0, end_mask = var_4123_end_mask_0, squeeze_mask = var_4123_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_71_cast_fp16 = mul(x = x_105_cast_fp16, y = var_4123_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38739136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39001344))))[name = string("tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4133_cast_fp16 = abs(x = linear_54_cast_fp16)[name = string("op_4133_cast_fp16")];
            bool var_4136 = const()[name = string("op_4136"), val = bool(true)];
            tensor<int32, [1]> var_4137_axes_0 = const()[name = string("op_4137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4137_cast_fp16 = reduce_max(axes = var_4137_axes_0, keep_dims = var_4136, x = var_4133_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 var_4138_to_fp16 = const()[name = string("op_4138_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_97_cast_fp16 = maximum(x = var_4137_cast_fp16, y = var_4138_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_97_cast_fp16 = real_div(x = linear_54_cast_fp16, y = s_97_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4141_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_4141_cast_fp16")];
            tensor<int32, [1]> var_4146_axes_0 = const()[name = string("op_4146_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4146_keep_dims_0 = const()[name = string("op_4146_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4146_cast_fp16 = reduce_mean(axes = var_4146_axes_0, keep_dims = var_4146_keep_dims_0, x = var_4141_cast_fp16)[name = string("op_4146_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4147_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_4147_cast_fp16")];
            fp32 var_4148_epsilon_0 = const()[name = string("op_4148_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4148_cast_fp16 = inverse(epsilon = var_4148_epsilon_0, x = var_4147_cast_fp16)[name = string("op_4148_cast_fp16")];
            fp16 var_4149_to_fp16 = const()[name = string("op_4149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4150_cast_fp16 = mul(x = var_4148_cast_fp16, y = var_4149_to_fp16)[name = string("op_4150_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4152_cast_fp16 = add(x = var_4146_cast_fp16, y = var_4150_cast_fp16)[name = string("op_4152_cast_fp16")];
            fp32 var_4153_epsilon_0 = const()[name = string("op_4153_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4153_cast_fp16 = rsqrt(epsilon = var_4153_epsilon_0, x = var_4152_cast_fp16)[name = string("op_4153_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_85_cast_fp16 = mul(x = y_97_cast_fp16, y = var_4153_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39002432)))];
            tensor<fp16, [1, 128, 512]> var_4155_cast_fp16 = mul(x = out_85_cast_fp16, y = tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4155_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_11_cast_fp16 = add(x = input_69_cast_fp16, y = var_4155_cast_fp16)[name = string("h_11_cast_fp16")];
            tensor<fp16, [1]> tm_layers_5_layer_scalar_to_fp16 = const()[name = string("tm_layers_5_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-2])];
            tensor<fp16, [1, 128, 512]> x_109_cast_fp16 = mul(x = h_11_cast_fp16, y = tm_layers_5_layer_scalar_to_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4159_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_4159_cast_fp16")];
            bool var_4162 = const()[name = string("op_4162"), val = bool(true)];
            tensor<int32, [1]> var_4163_axes_0 = const()[name = string("op_4163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4163_cast_fp16 = reduce_max(axes = var_4163_axes_0, keep_dims = var_4162, x = var_4159_cast_fp16)[name = string("op_4163_cast_fp16")];
            fp16 var_4164_to_fp16 = const()[name = string("op_4164_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_99_cast_fp16 = maximum(x = var_4163_cast_fp16, y = var_4164_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_99_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_99_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4167_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_4167_cast_fp16")];
            tensor<int32, [1]> var_4172_axes_0 = const()[name = string("op_4172_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4172_keep_dims_0 = const()[name = string("op_4172_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4172_cast_fp16 = reduce_mean(axes = var_4172_axes_0, keep_dims = var_4172_keep_dims_0, x = var_4167_cast_fp16)[name = string("op_4172_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4173_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4174_cast_fp16 = inverse(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            fp16 var_4175_to_fp16 = const()[name = string("op_4175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4176_cast_fp16 = mul(x = var_4174_cast_fp16, y = var_4175_to_fp16)[name = string("op_4176_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4178_cast_fp16 = add(x = var_4172_cast_fp16, y = var_4176_cast_fp16)[name = string("op_4178_cast_fp16")];
            fp32 var_4179_epsilon_0 = const()[name = string("op_4179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4179_cast_fp16 = rsqrt(epsilon = var_4179_epsilon_0, x = var_4178_cast_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_87_cast_fp16 = mul(x = y_99_cast_fp16, y = var_4179_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39003520)))];
            tensor<fp16, [1, 128, 512]> input_73_cast_fp16 = mul(x = out_87_cast_fp16, y = tm_layers_6_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39004608))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39528960))))[name = string("tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [4]> var_4189 = const()[name = string("op_4189"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_4190_cast_fp16 = reshape(shape = var_4189, x = linear_55_cast_fp16)[name = string("op_4190_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39531072))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39793280))))[name = string("tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_4201 = const()[name = string("op_4201"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_4202_cast_fp16 = reshape(shape = var_4201, x = linear_56_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39794368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40056576))))[name = string("tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_57_cast_fp16)[name = string("op_4214_cast_fp16")];
            tensor<int32, [4]> x_115_perm_0 = const()[name = string("x_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4190_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 4, 128, 256]> var_4218_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_4218_cast_fp16")];
            bool var_4221 = const()[name = string("op_4221"), val = bool(true)];
            tensor<int32, [1]> var_4222_axes_0 = const()[name = string("op_4222_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_4222_cast_fp16 = reduce_max(axes = var_4222_axes_0, keep_dims = var_4221, x = var_4218_cast_fp16)[name = string("op_4222_cast_fp16")];
            fp16 var_4223_to_fp16 = const()[name = string("op_4223_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_101_cast_fp16 = maximum(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_101_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_101_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4226_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [1]> var_4231_axes_0 = const()[name = string("op_4231_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4231_keep_dims_0 = const()[name = string("op_4231_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_4231_cast_fp16 = reduce_mean(axes = var_4231_axes_0, keep_dims = var_4231_keep_dims_0, x = var_4226_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_4232_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_4232_cast_fp16")];
            fp32 var_4233_epsilon_0 = const()[name = string("op_4233_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_4233_cast_fp16 = inverse(epsilon = var_4233_epsilon_0, x = var_4232_cast_fp16)[name = string("op_4233_cast_fp16")];
            fp16 var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_4235_cast_fp16 = mul(x = var_4233_cast_fp16, y = var_4234_to_fp16)[name = string("op_4235_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_4237_cast_fp16 = add(x = var_4231_cast_fp16, y = var_4235_cast_fp16)[name = string("op_4237_cast_fp16")];
            fp32 var_4238_epsilon_0 = const()[name = string("op_4238_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_4238_cast_fp16 = rsqrt(epsilon = var_4238_epsilon_0, x = var_4237_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_89_cast_fp16 = mul(x = y_101_cast_fp16, y = var_4238_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40057664)))];
            tensor<fp16, [1, 4, 128, 256]> q_25_cast_fp16 = mul(x = out_89_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4202_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 2, 128, 256]> var_4241_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_4241_cast_fp16")];
            bool var_4244 = const()[name = string("op_4244"), val = bool(true)];
            tensor<int32, [1]> var_4245_axes_0 = const()[name = string("op_4245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_4245_cast_fp16 = reduce_max(axes = var_4245_axes_0, keep_dims = var_4244, x = var_4241_cast_fp16)[name = string("op_4245_cast_fp16")];
            fp16 var_4246_to_fp16 = const()[name = string("op_4246_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_103_cast_fp16 = maximum(x = var_4245_cast_fp16, y = var_4246_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_103_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_103_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4249_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<int32, [1]> var_4254_axes_0 = const()[name = string("op_4254_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4254_keep_dims_0 = const()[name = string("op_4254_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_4254_cast_fp16 = reduce_mean(axes = var_4254_axes_0, keep_dims = var_4254_keep_dims_0, x = var_4249_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4255_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_4255_cast_fp16")];
            fp32 var_4256_epsilon_0 = const()[name = string("op_4256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_4256_cast_fp16 = inverse(epsilon = var_4256_epsilon_0, x = var_4255_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp16 var_4257_to_fp16 = const()[name = string("op_4257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_4258_cast_fp16 = mul(x = var_4256_cast_fp16, y = var_4257_to_fp16)[name = string("op_4258_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4260_cast_fp16 = add(x = var_4254_cast_fp16, y = var_4258_cast_fp16)[name = string("op_4260_cast_fp16")];
            fp32 var_4261_epsilon_0 = const()[name = string("op_4261_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_4261_cast_fp16 = rsqrt(epsilon = var_4261_epsilon_0, x = var_4260_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_91_cast_fp16 = mul(x = y_103_cast_fp16, y = var_4261_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058240)))];
            tensor<fp16, [1, 2, 128, 256]> k_37_cast_fp16 = mul(x = out_91_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_37_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_115_cast_fp16 = transpose(perm = x_115_perm_0, x = var_4214_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 2, 128, 256]> var_4264_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_4264_cast_fp16")];
            bool var_4267 = const()[name = string("op_4267"), val = bool(true)];
            tensor<int32, [1]> var_4268_axes_0 = const()[name = string("op_4268_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_4268_cast_fp16 = reduce_max(axes = var_4268_axes_0, keep_dims = var_4267, x = var_4264_cast_fp16)[name = string("op_4268_cast_fp16")];
            fp16 var_4269_to_fp16 = const()[name = string("op_4269_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_105_cast_fp16 = maximum(x = var_4268_cast_fp16, y = var_4269_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_105_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_105_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4272_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_4272_cast_fp16")];
            tensor<int32, [1]> var_4277_axes_0 = const()[name = string("op_4277_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4277_keep_dims_0 = const()[name = string("op_4277_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_4277_cast_fp16 = reduce_mean(axes = var_4277_axes_0, keep_dims = var_4277_keep_dims_0, x = var_4272_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4278_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp32 var_4279_epsilon_0 = const()[name = string("op_4279_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_4279_cast_fp16 = inverse(epsilon = var_4279_epsilon_0, x = var_4278_cast_fp16)[name = string("op_4279_cast_fp16")];
            fp16 var_4280_to_fp16 = const()[name = string("op_4280_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_4281_cast_fp16 = mul(x = var_4279_cast_fp16, y = var_4280_to_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4283_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4281_cast_fp16)[name = string("op_4283_cast_fp16")];
            fp32 var_4284_epsilon_0 = const()[name = string("op_4284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_4284_cast_fp16 = rsqrt(epsilon = var_4284_epsilon_0, x = var_4283_cast_fp16)[name = string("op_4284_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_25_cast_fp16 = mul(x = y_105_cast_fp16, y = var_4284_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4286_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4286_cast_fp16")];
            tensor<int32, [4]> var_4291_begin_0 = const()[name = string("op_4291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4291_end_0 = const()[name = string("op_4291_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_4291_end_mask_0 = const()[name = string("op_4291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_4291_cast_fp16 = slice_by_index(begin = var_4291_begin_0, end = var_4291_end_0, end_mask = var_4291_end_mask_0, x = q_25_cast_fp16)[name = string("op_4291_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_4292_cast_fp16 = mul(x = var_4291_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = q_25_cast_fp16)[name = string("op_4297_cast_fp16")];
            int32 var_4299 = const()[name = string("op_4299"), val = int32(-1)];
            bool var_4300_interleave_0 = const()[name = string("op_4300_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_4300_cast_fp16 = concat(axis = var_4299, interleave = var_4300_interleave_0, values = (var_4292_cast_fp16, var_4297_cast_fp16))[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4301_cast_fp16 = mul(x = var_4300_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4301_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_27_cast_fp16 = add(x = var_4286_cast_fp16, y = var_4301_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4304_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4304_cast_fp16")];
            tensor<int32, [4]> var_4309_begin_0 = const()[name = string("op_4309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4309_end_0 = const()[name = string("op_4309_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_4309_end_mask_0 = const()[name = string("op_4309_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_4309_cast_fp16 = slice_by_index(begin = var_4309_begin_0, end = var_4309_end_0, end_mask = var_4309_end_mask_0, x = k_37_cast_fp16)[name = string("op_4309_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_4310_cast_fp16 = mul(x = var_4309_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_4310_cast_fp16")];
            tensor<int32, [4]> var_4315_begin_0 = const()[name = string("op_4315_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4315_end_0 = const()[name = string("op_4315_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_4315_end_mask_0 = const()[name = string("op_4315_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_4315_cast_fp16 = slice_by_index(begin = var_4315_begin_0, end = var_4315_end_0, end_mask = var_4315_end_mask_0, x = k_37_cast_fp16)[name = string("op_4315_cast_fp16")];
            int32 var_4317 = const()[name = string("op_4317"), val = int32(-1)];
            bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_4318_cast_fp16 = concat(axis = var_4317, interleave = var_4318_interleave_0, values = (var_4310_cast_fp16, var_4315_cast_fp16))[name = string("op_4318_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4319_cast_fp16 = mul(x = var_4318_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_39_cast_fp16 = add(x = var_4304_cast_fp16, y = var_4319_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = k_39_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [4, 1, 128, 256]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
            tensor<int32, [5]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_24_cast_fp16 = reshape(shape = concat_24, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [4, 1, 128, 256]> reshape_25_cast_fp16 = reshape(shape = concat_25, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = v_25_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [4, 1, 128, 256]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
            tensor<int32, [5]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_26_cast_fp16 = reshape(shape = concat_26, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [4, 1, 128, 256]> reshape_27_cast_fp16 = reshape(shape = concat_27, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4333_transpose_x_0 = const()[name = string("op_4333_transpose_x_0"), val = bool(false)];
            bool var_4333_transpose_y_0 = const()[name = string("op_4333_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_102_cast_fp16 = transpose(perm = transpose_102_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 4, 128, 128]> var_4333_cast_fp16 = matmul(transpose_x = var_4333_transpose_x_0, transpose_y = var_4333_transpose_y_0, x = q_27_cast_fp16, y = transpose_102_cast_fp16)[name = string("op_4333_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_4335_cast_fp16 = add(x = var_4333_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4335_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_105_cast_fp16 = softmax(axis = var_4336, x = var_4335_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4339_transpose_x_0 = const()[name = string("op_4339_transpose_x_0"), val = bool(false)];
            bool var_4339_transpose_y_0 = const()[name = string("op_4339_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 4, 128, 256]> var_4339_cast_fp16 = matmul(transpose_x = var_4339_transpose_x_0, transpose_y = var_4339_transpose_y_0, x = w_105_cast_fp16, y = v_27_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<int32, [4]> var_4342_perm_0 = const()[name = string("op_4342_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4346 = const()[name = string("op_4346"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_4342_cast_fp16 = transpose(perm = var_4342_perm_0, x = var_4339_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 128, 1024]> input_75_cast_fp16 = reshape(shape = var_4346, x = var_4342_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40583168))))[name = string("tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4351_cast_fp16 = abs(x = linear_58_cast_fp16)[name = string("op_4351_cast_fp16")];
            bool var_4354 = const()[name = string("op_4354"), val = bool(true)];
            tensor<int32, [1]> var_4355_axes_0 = const()[name = string("op_4355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4355_cast_fp16 = reduce_max(axes = var_4355_axes_0, keep_dims = var_4354, x = var_4351_cast_fp16)[name = string("op_4355_cast_fp16")];
            fp16 var_4356_to_fp16 = const()[name = string("op_4356_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_107_cast_fp16 = maximum(x = var_4355_cast_fp16, y = var_4356_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_107_cast_fp16 = real_div(x = linear_58_cast_fp16, y = s_107_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4359_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<int32, [1]> var_4364_axes_0 = const()[name = string("op_4364_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4364_keep_dims_0 = const()[name = string("op_4364_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4364_cast_fp16 = reduce_mean(axes = var_4364_axes_0, keep_dims = var_4364_keep_dims_0, x = var_4359_cast_fp16)[name = string("op_4364_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4365_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_4365_cast_fp16")];
            fp32 var_4366_epsilon_0 = const()[name = string("op_4366_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4366_cast_fp16 = inverse(epsilon = var_4366_epsilon_0, x = var_4365_cast_fp16)[name = string("op_4366_cast_fp16")];
            fp16 var_4367_to_fp16 = const()[name = string("op_4367_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4368_cast_fp16 = mul(x = var_4366_cast_fp16, y = var_4367_to_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4370_cast_fp16 = add(x = var_4364_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            fp32 var_4371_epsilon_0 = const()[name = string("op_4371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4371_cast_fp16 = rsqrt(epsilon = var_4371_epsilon_0, x = var_4370_cast_fp16)[name = string("op_4371_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_93_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4371_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40584256)))];
            tensor<fp16, [1, 128, 512]> var_4373_cast_fp16 = mul(x = out_93_cast_fp16, y = tm_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_4373_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = var_4373_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4376_cast_fp16 = abs(x = x_119_cast_fp16)[name = string("op_4376_cast_fp16")];
            bool var_4379 = const()[name = string("op_4379"), val = bool(true)];
            tensor<int32, [1]> var_4380_axes_0 = const()[name = string("op_4380_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4380_cast_fp16 = reduce_max(axes = var_4380_axes_0, keep_dims = var_4379, x = var_4376_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 var_4381_to_fp16 = const()[name = string("op_4381_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_109_cast_fp16 = maximum(x = var_4380_cast_fp16, y = var_4381_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_109_cast_fp16 = real_div(x = x_119_cast_fp16, y = s_109_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4384_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4384_cast_fp16")];
            tensor<int32, [1]> var_4389_axes_0 = const()[name = string("op_4389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4389_keep_dims_0 = const()[name = string("op_4389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4389_cast_fp16 = reduce_mean(axes = var_4389_axes_0, keep_dims = var_4389_keep_dims_0, x = var_4384_cast_fp16)[name = string("op_4389_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4390_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_4390_cast_fp16")];
            fp32 var_4391_epsilon_0 = const()[name = string("op_4391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4391_cast_fp16 = inverse(epsilon = var_4391_epsilon_0, x = var_4390_cast_fp16)[name = string("op_4391_cast_fp16")];
            fp16 var_4392_to_fp16 = const()[name = string("op_4392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = string("op_4393_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4395_cast_fp16 = add(x = var_4389_cast_fp16, y = var_4393_cast_fp16)[name = string("op_4395_cast_fp16")];
            fp32 var_4396_epsilon_0 = const()[name = string("op_4396_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4396_cast_fp16 = rsqrt(epsilon = var_4396_epsilon_0, x = var_4395_cast_fp16)[name = string("op_4396_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_95_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4396_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40585344)))];
            tensor<fp16, [1, 128, 512]> input_77_cast_fp16 = mul(x = out_95_cast_fp16, y = tm_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40586432))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41635072))))[name = string("tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_59_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_59_cast_fp16")];
            string var_4403_mode_0 = const()[name = string("op_4403_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_4403_cast_fp16 = gelu(mode = var_4403_mode_0, x = linear_59_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41639232))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42687872))))[name = string("tm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_60_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_79_cast_fp16 = mul(x = var_4403_cast_fp16, y = linear_60_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42692032))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43740672))))[name = string("tm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4411_cast_fp16 = abs(x = linear_61_cast_fp16)[name = string("op_4411_cast_fp16")];
            bool var_4414 = const()[name = string("op_4414"), val = bool(true)];
            tensor<int32, [1]> var_4415_axes_0 = const()[name = string("op_4415_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4415_cast_fp16 = reduce_max(axes = var_4415_axes_0, keep_dims = var_4414, x = var_4411_cast_fp16)[name = string("op_4415_cast_fp16")];
            fp16 var_4416_to_fp16 = const()[name = string("op_4416_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_111_cast_fp16 = maximum(x = var_4415_cast_fp16, y = var_4416_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_111_cast_fp16 = real_div(x = linear_61_cast_fp16, y = s_111_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4419_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4419_cast_fp16")];
            tensor<int32, [1]> var_4424_axes_0 = const()[name = string("op_4424_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4424_keep_dims_0 = const()[name = string("op_4424_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4424_cast_fp16 = reduce_mean(axes = var_4424_axes_0, keep_dims = var_4424_keep_dims_0, x = var_4419_cast_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4425_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4426_cast_fp16 = inverse(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            fp16 var_4427_to_fp16 = const()[name = string("op_4427_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4428_cast_fp16 = mul(x = var_4426_cast_fp16, y = var_4427_to_fp16)[name = string("op_4428_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4430_cast_fp16 = add(x = var_4424_cast_fp16, y = var_4428_cast_fp16)[name = string("op_4430_cast_fp16")];
            fp32 var_4431_epsilon_0 = const()[name = string("op_4431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4431_cast_fp16 = rsqrt(epsilon = var_4431_epsilon_0, x = var_4430_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_97_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4431_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43741760)))];
            tensor<fp16, [1, 128, 512]> var_4433_cast_fp16 = mul(x = out_97_cast_fp16, y = tm_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4433_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_81_cast_fp16 = add(x = x_119_cast_fp16, y = var_4433_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43742848))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44005056))))[name = string("tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_62_cast_fp16")];
            string x_123_mode_0 = const()[name = string("x_123_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_123_cast_fp16 = gelu(mode = x_123_mode_0, x = linear_62_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [4]> var_4453_begin_0 = const()[name = string("op_4453_begin_0"), val = tensor<int32, [4]>([0, 0, 6, 0])];
            tensor<int32, [4]> var_4453_end_0 = const()[name = string("op_4453_end_0"), val = tensor<int32, [4]>([1, 128, 7, 512])];
            tensor<bool, [4]> var_4453_end_mask_0 = const()[name = string("op_4453_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4453_squeeze_mask_0 = const()[name = string("op_4453_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_4453_cast_fp16 = slice_by_index(begin = var_4453_begin_0, end = var_4453_end_0, end_mask = var_4453_end_mask_0, squeeze_mask = var_4453_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_83_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4453_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44006144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44268352))))[name = string("tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4463_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4463_cast_fp16")];
            bool var_4466 = const()[name = string("op_4466"), val = bool(true)];
            tensor<int32, [1]> var_4467_axes_0 = const()[name = string("op_4467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4467_cast_fp16 = reduce_max(axes = var_4467_axes_0, keep_dims = var_4466, x = var_4463_cast_fp16)[name = string("op_4467_cast_fp16")];
            fp16 var_4468_to_fp16 = const()[name = string("op_4468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_113_cast_fp16 = maximum(x = var_4467_cast_fp16, y = var_4468_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_113_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_113_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4471_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [1]> var_4476_axes_0 = const()[name = string("op_4476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4476_keep_dims_0 = const()[name = string("op_4476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4476_cast_fp16 = reduce_mean(axes = var_4476_axes_0, keep_dims = var_4476_keep_dims_0, x = var_4471_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4477_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_fp16)[name = string("op_4477_cast_fp16")];
            fp32 var_4478_epsilon_0 = const()[name = string("op_4478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4478_cast_fp16 = inverse(epsilon = var_4478_epsilon_0, x = var_4477_cast_fp16)[name = string("op_4478_cast_fp16")];
            fp16 var_4479_to_fp16 = const()[name = string("op_4479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4480_cast_fp16 = mul(x = var_4478_cast_fp16, y = var_4479_to_fp16)[name = string("op_4480_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4482_cast_fp16 = add(x = var_4476_cast_fp16, y = var_4480_cast_fp16)[name = string("op_4482_cast_fp16")];
            fp32 var_4483_epsilon_0 = const()[name = string("op_4483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4483_cast_fp16 = rsqrt(epsilon = var_4483_epsilon_0, x = var_4482_cast_fp16)[name = string("op_4483_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_99_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4483_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44269440)))];
            tensor<fp16, [1, 128, 512]> var_4485_cast_fp16 = mul(x = out_99_cast_fp16, y = tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_13_cast_fp16 = add(x = input_81_cast_fp16, y = var_4485_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<fp16, [1]> tm_layers_6_layer_scalar_to_fp16 = const()[name = string("tm_layers_6_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.cep-3])];
            tensor<fp16, [1, 128, 512]> x_127_cast_fp16 = mul(x = h_13_cast_fp16, y = tm_layers_6_layer_scalar_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4489_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_4489_cast_fp16")];
            bool var_4492 = const()[name = string("op_4492"), val = bool(true)];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4493_cast_fp16 = reduce_max(axes = var_4493_axes_0, keep_dims = var_4492, x = var_4489_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_115_cast_fp16 = maximum(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_115_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_115_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4497_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<int32, [1]> var_4502_axes_0 = const()[name = string("op_4502_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4502_keep_dims_0 = const()[name = string("op_4502_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4502_cast_fp16 = reduce_mean(axes = var_4502_axes_0, keep_dims = var_4502_keep_dims_0, x = var_4497_cast_fp16)[name = string("op_4502_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4503_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4504_cast_fp16 = inverse(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            fp16 var_4505_to_fp16 = const()[name = string("op_4505_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4506_cast_fp16 = mul(x = var_4504_cast_fp16, y = var_4505_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4508_cast_fp16 = add(x = var_4502_cast_fp16, y = var_4506_cast_fp16)[name = string("op_4508_cast_fp16")];
            fp32 var_4509_epsilon_0 = const()[name = string("op_4509_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4509_cast_fp16 = rsqrt(epsilon = var_4509_epsilon_0, x = var_4508_cast_fp16)[name = string("op_4509_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_101_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4509_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44270528)))];
            tensor<fp16, [1, 128, 512]> input_85_cast_fp16 = mul(x = out_101_cast_fp16, y = tm_layers_7_input_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44271616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44795968))))[name = string("tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4519 = const()[name = string("op_4519"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_4520_cast_fp16 = reshape(shape = var_4519, x = linear_64_cast_fp16)[name = string("op_4520_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44798080))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45060288))))[name = string("tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4531 = const()[name = string("op_4531"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_4532_cast_fp16 = reshape(shape = var_4531, x = linear_65_cast_fp16)[name = string("op_4532_cast_fp16")];
            tensor<int32, [4]> x_131_perm_0 = const()[name = string("x_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45061376))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45323584))))[name = string("tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4543 = const()[name = string("op_4543"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_4544_cast_fp16 = reshape(shape = var_4543, x = linear_66_cast_fp16)[name = string("op_4544_cast_fp16")];
            tensor<int32, [4]> x_133_perm_0 = const()[name = string("x_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_4520_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 4, 128, 256]> var_4548_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_4548_cast_fp16")];
            bool var_4551 = const()[name = string("op_4551"), val = bool(true)];
            tensor<int32, [1]> var_4552_axes_0 = const()[name = string("op_4552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_4552_cast_fp16 = reduce_max(axes = var_4552_axes_0, keep_dims = var_4551, x = var_4548_cast_fp16)[name = string("op_4552_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_117_cast_fp16 = maximum(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_117_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_117_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4556_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<int32, [1]> var_4561_axes_0 = const()[name = string("op_4561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4561_keep_dims_0 = const()[name = string("op_4561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_4561_cast_fp16 = reduce_mean(axes = var_4561_axes_0, keep_dims = var_4561_keep_dims_0, x = var_4556_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_4562_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4562_cast_fp16")];
            fp32 var_4563_epsilon_0 = const()[name = string("op_4563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_4563_cast_fp16 = inverse(epsilon = var_4563_epsilon_0, x = var_4562_cast_fp16)[name = string("op_4563_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_4565_cast_fp16 = mul(x = var_4563_cast_fp16, y = var_4564_to_fp16)[name = string("op_4565_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_4567_cast_fp16 = add(x = var_4561_cast_fp16, y = var_4565_cast_fp16)[name = string("op_4567_cast_fp16")];
            fp32 var_4568_epsilon_0 = const()[name = string("op_4568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_4568_cast_fp16 = rsqrt(epsilon = var_4568_epsilon_0, x = var_4567_cast_fp16)[name = string("op_4568_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_103_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4568_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_29_cast_fp16 = mul(x = out_103_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_131_cast_fp16 = transpose(perm = x_131_perm_0, x = var_4532_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2, 128, 256]> var_4571_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_4571_cast_fp16")];
            bool var_4574 = const()[name = string("op_4574"), val = bool(true)];
            tensor<int32, [1]> var_4575_axes_0 = const()[name = string("op_4575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_4575_cast_fp16 = reduce_max(axes = var_4575_axes_0, keep_dims = var_4574, x = var_4571_cast_fp16)[name = string("op_4575_cast_fp16")];
            fp16 var_4576_to_fp16 = const()[name = string("op_4576_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_119_cast_fp16 = maximum(x = var_4575_cast_fp16, y = var_4576_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_119_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_119_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4579_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4579_cast_fp16")];
            tensor<int32, [1]> var_4584_axes_0 = const()[name = string("op_4584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4584_keep_dims_0 = const()[name = string("op_4584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_4584_cast_fp16 = reduce_mean(axes = var_4584_axes_0, keep_dims = var_4584_keep_dims_0, x = var_4579_cast_fp16)[name = string("op_4584_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4585_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4585_cast_fp16")];
            fp32 var_4586_epsilon_0 = const()[name = string("op_4586_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_4586_cast_fp16 = inverse(epsilon = var_4586_epsilon_0, x = var_4585_cast_fp16)[name = string("op_4586_cast_fp16")];
            fp16 var_4587_to_fp16 = const()[name = string("op_4587_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_4588_cast_fp16 = mul(x = var_4586_cast_fp16, y = var_4587_to_fp16)[name = string("op_4588_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4590_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4588_cast_fp16)[name = string("op_4590_cast_fp16")];
            fp32 var_4591_epsilon_0 = const()[name = string("op_4591_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_4591_cast_fp16 = rsqrt(epsilon = var_4591_epsilon_0, x = var_4590_cast_fp16)[name = string("op_4591_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_105_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4591_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [256]> tm_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45324672)))];
            tensor<fp16, [1, 2, 128, 256]> k_43_cast_fp16 = mul(x = out_105_cast_fp16, y = tm_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_133_cast_fp16 = transpose(perm = x_133_perm_0, x = var_4544_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 2, 128, 256]> var_4594_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_4594_cast_fp16")];
            bool var_4597 = const()[name = string("op_4597"), val = bool(true)];
            tensor<int32, [1]> var_4598_axes_0 = const()[name = string("op_4598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_4598_cast_fp16 = reduce_max(axes = var_4598_axes_0, keep_dims = var_4597, x = var_4594_cast_fp16)[name = string("op_4598_cast_fp16")];
            fp16 var_4599_to_fp16 = const()[name = string("op_4599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_121_cast_fp16 = maximum(x = var_4598_cast_fp16, y = var_4599_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_121_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_121_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4602_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4602_cast_fp16")];
            tensor<int32, [1]> var_4607_axes_0 = const()[name = string("op_4607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4607_keep_dims_0 = const()[name = string("op_4607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_4607_cast_fp16 = reduce_mean(axes = var_4607_axes_0, keep_dims = var_4607_keep_dims_0, x = var_4602_cast_fp16)[name = string("op_4607_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4608_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4608_cast_fp16")];
            fp32 var_4609_epsilon_0 = const()[name = string("op_4609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_4609_cast_fp16 = inverse(epsilon = var_4609_epsilon_0, x = var_4608_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_4611_cast_fp16 = mul(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("op_4611_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4613_cast_fp16 = add(x = var_4607_cast_fp16, y = var_4611_cast_fp16)[name = string("op_4613_cast_fp16")];
            fp32 var_4614_epsilon_0 = const()[name = string("op_4614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_4614_cast_fp16 = rsqrt(epsilon = var_4614_epsilon_0, x = var_4613_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_29_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4614_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4616_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4616_cast_fp16")];
            tensor<int32, [4]> var_4621_begin_0 = const()[name = string("op_4621_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4621_end_0 = const()[name = string("op_4621_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_4621_end_mask_0 = const()[name = string("op_4621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_4621_cast_fp16 = slice_by_index(begin = var_4621_begin_0, end = var_4621_end_0, end_mask = var_4621_end_mask_0, x = q_29_cast_fp16)[name = string("op_4621_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_4622_cast_fp16 = mul(x = var_4621_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<int32, [4]> var_4627_begin_0 = const()[name = string("op_4627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4627_end_0 = const()[name = string("op_4627_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_4627_end_mask_0 = const()[name = string("op_4627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_4627_cast_fp16 = slice_by_index(begin = var_4627_begin_0, end = var_4627_end_0, end_mask = var_4627_end_mask_0, x = q_29_cast_fp16)[name = string("op_4627_cast_fp16")];
            int32 var_4629 = const()[name = string("op_4629"), val = int32(-1)];
            bool var_4630_interleave_0 = const()[name = string("op_4630_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_4630_cast_fp16 = concat(axis = var_4629, interleave = var_4630_interleave_0, values = (var_4622_cast_fp16, var_4627_cast_fp16))[name = string("op_4630_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4631_cast_fp16 = mul(x = var_4630_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4631_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_31_cast_fp16 = add(x = var_4616_cast_fp16, y = var_4631_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4634_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4634_cast_fp16")];
            tensor<int32, [4]> var_4639_begin_0 = const()[name = string("op_4639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4639_end_0 = const()[name = string("op_4639_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_4639_end_mask_0 = const()[name = string("op_4639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_4639_cast_fp16 = slice_by_index(begin = var_4639_begin_0, end = var_4639_end_0, end_mask = var_4639_end_mask_0, x = k_43_cast_fp16)[name = string("op_4639_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_4640_cast_fp16 = mul(x = var_4639_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4640_cast_fp16")];
            tensor<int32, [4]> var_4645_begin_0 = const()[name = string("op_4645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4645_end_0 = const()[name = string("op_4645_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_4645_end_mask_0 = const()[name = string("op_4645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_4645_cast_fp16 = slice_by_index(begin = var_4645_begin_0, end = var_4645_end_0, end_mask = var_4645_end_mask_0, x = k_43_cast_fp16)[name = string("op_4645_cast_fp16")];
            int32 var_4647 = const()[name = string("op_4647"), val = int32(-1)];
            bool var_4648_interleave_0 = const()[name = string("op_4648_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_4648_cast_fp16 = concat(axis = var_4647, interleave = var_4648_interleave_0, values = (var_4640_cast_fp16, var_4645_cast_fp16))[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4649_cast_fp16 = mul(x = var_4648_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4649_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_45_cast_fp16 = add(x = var_4634_cast_fp16, y = var_4649_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_14_reps_0 = const()[name = string("tile_14_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = k_45_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [4, 1, 128, 256]> tile_14_cast_fp16 = tile(reps = tile_14_reps_0, x = transpose_28_cast_fp16)[name = string("tile_14_cast_fp16")];
            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_28_cast_fp16 = reshape(shape = concat_28, x = tile_14_cast_fp16)[name = string("reshape_28_cast_fp16")];
            tensor<int32, [5]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_28_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [4, 1, 128, 256]> reshape_29_cast_fp16 = reshape(shape = concat_29, x = transpose_29_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_15_reps_0 = const()[name = string("tile_15_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = v_29_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [4, 1, 128, 256]> tile_15_cast_fp16 = tile(reps = tile_15_reps_0, x = transpose_30_cast_fp16)[name = string("tile_15_cast_fp16")];
            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_30_cast_fp16 = reshape(shape = concat_30, x = tile_15_cast_fp16)[name = string("reshape_30_cast_fp16")];
            tensor<int32, [5]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_30_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [4, 1, 128, 256]> reshape_31_cast_fp16 = reshape(shape = concat_31, x = transpose_31_cast_fp16)[name = string("reshape_31_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4663_transpose_x_0 = const()[name = string("op_4663_transpose_x_0"), val = bool(false)];
            bool var_4663_transpose_y_0 = const()[name = string("op_4663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 4, 128, 128]> var_4663_cast_fp16 = matmul(transpose_x = var_4663_transpose_x_0, transpose_y = var_4663_transpose_y_0, x = q_31_cast_fp16, y = transpose_103_cast_fp16)[name = string("op_4663_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_4665_cast_fp16 = add(x = var_4663_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4666 = const()[name = string("op_4666"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_121_cast_fp16 = softmax(axis = var_4666, x = var_4665_cast_fp16)[name = string("w_121_cast_fp16")];
            bool var_4669_transpose_x_0 = const()[name = string("op_4669_transpose_x_0"), val = bool(false)];
            bool var_4669_transpose_y_0 = const()[name = string("op_4669_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = reshape_31_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 4, 128, 256]> var_4669_cast_fp16 = matmul(transpose_x = var_4669_transpose_x_0, transpose_y = var_4669_transpose_y_0, x = w_121_cast_fp16, y = v_31_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<int32, [4]> var_4672_perm_0 = const()[name = string("op_4672_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4676 = const()[name = string("op_4676"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_4672_cast_fp16 = transpose(perm = var_4672_perm_0, x = var_4669_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 128, 1024]> input_87_cast_fp16 = reshape(shape = var_4676, x = var_4672_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45325248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45849600))))[name = string("tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4681_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4684 = const()[name = string("op_4684"), val = bool(true)];
            tensor<int32, [1]> var_4685_axes_0 = const()[name = string("op_4685_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4685_cast_fp16 = reduce_max(axes = var_4685_axes_0, keep_dims = var_4684, x = var_4681_cast_fp16)[name = string("op_4685_cast_fp16")];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_123_cast_fp16 = maximum(x = var_4685_cast_fp16, y = var_4686_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_123_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_123_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4689_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [1]> var_4694_axes_0 = const()[name = string("op_4694_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4694_keep_dims_0 = const()[name = string("op_4694_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4694_cast_fp16 = reduce_mean(axes = var_4694_axes_0, keep_dims = var_4694_keep_dims_0, x = var_4689_cast_fp16)[name = string("op_4694_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4695_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4695_cast_fp16")];
            fp32 var_4696_epsilon_0 = const()[name = string("op_4696_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4696_cast_fp16 = inverse(epsilon = var_4696_epsilon_0, x = var_4695_cast_fp16)[name = string("op_4696_cast_fp16")];
            fp16 var_4697_to_fp16 = const()[name = string("op_4697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4698_cast_fp16 = mul(x = var_4696_cast_fp16, y = var_4697_to_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4700_cast_fp16 = add(x = var_4694_cast_fp16, y = var_4698_cast_fp16)[name = string("op_4700_cast_fp16")];
            fp32 var_4701_epsilon_0 = const()[name = string("op_4701_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4701_cast_fp16 = rsqrt(epsilon = var_4701_epsilon_0, x = var_4700_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_107_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4701_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45850688)))];
            tensor<fp16, [1, 128, 512]> var_4703_cast_fp16 = mul(x = out_107_cast_fp16, y = tm_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4703_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_137_cast_fp16 = add(x = x_127_cast_fp16, y = var_4703_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4706_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_4706_cast_fp16")];
            bool var_4709 = const()[name = string("op_4709"), val = bool(true)];
            tensor<int32, [1]> var_4710_axes_0 = const()[name = string("op_4710_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4710_cast_fp16 = reduce_max(axes = var_4710_axes_0, keep_dims = var_4709, x = var_4706_cast_fp16)[name = string("op_4710_cast_fp16")];
            fp16 var_4711_to_fp16 = const()[name = string("op_4711_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_125_cast_fp16 = maximum(x = var_4710_cast_fp16, y = var_4711_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_125_cast_fp16 = real_div(x = x_137_cast_fp16, y = s_125_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4714_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4714_cast_fp16")];
            tensor<int32, [1]> var_4719_axes_0 = const()[name = string("op_4719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4719_keep_dims_0 = const()[name = string("op_4719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4719_cast_fp16 = reduce_mean(axes = var_4719_axes_0, keep_dims = var_4719_keep_dims_0, x = var_4714_cast_fp16)[name = string("op_4719_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4720_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4720_cast_fp16")];
            fp32 var_4721_epsilon_0 = const()[name = string("op_4721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4721_cast_fp16 = inverse(epsilon = var_4721_epsilon_0, x = var_4720_cast_fp16)[name = string("op_4721_cast_fp16")];
            fp16 var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4723_cast_fp16 = mul(x = var_4721_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4725_cast_fp16 = add(x = var_4719_cast_fp16, y = var_4723_cast_fp16)[name = string("op_4725_cast_fp16")];
            fp32 var_4726_epsilon_0 = const()[name = string("op_4726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4726_cast_fp16 = rsqrt(epsilon = var_4726_epsilon_0, x = var_4725_cast_fp16)[name = string("op_4726_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_109_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4726_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45851776)))];
            tensor<fp16, [1, 128, 512]> input_89_cast_fp16 = mul(x = out_109_cast_fp16, y = tm_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45852864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46901504))))[name = string("tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4733_mode_0 = const()[name = string("op_4733_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_4733_cast_fp16 = gelu(mode = var_4733_mode_0, x = linear_68_cast_fp16)[name = string("op_4733_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46905664))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47954304))))[name = string("tm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_91_cast_fp16 = mul(x = var_4733_cast_fp16, y = linear_69_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47958464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49007104))))[name = string("tm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4741_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4741_cast_fp16")];
            bool var_4744 = const()[name = string("op_4744"), val = bool(true)];
            tensor<int32, [1]> var_4745_axes_0 = const()[name = string("op_4745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4745_cast_fp16 = reduce_max(axes = var_4745_axes_0, keep_dims = var_4744, x = var_4741_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_127_cast_fp16 = maximum(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_127_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_127_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4749_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<int32, [1]> var_4754_axes_0 = const()[name = string("op_4754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4754_keep_dims_0 = const()[name = string("op_4754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4754_cast_fp16 = reduce_mean(axes = var_4754_axes_0, keep_dims = var_4754_keep_dims_0, x = var_4749_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4755_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4755_cast_fp16")];
            fp32 var_4756_epsilon_0 = const()[name = string("op_4756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4756_cast_fp16 = inverse(epsilon = var_4756_epsilon_0, x = var_4755_cast_fp16)[name = string("op_4756_cast_fp16")];
            fp16 var_4757_to_fp16 = const()[name = string("op_4757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4758_cast_fp16 = mul(x = var_4756_cast_fp16, y = var_4757_to_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4760_cast_fp16 = add(x = var_4754_cast_fp16, y = var_4758_cast_fp16)[name = string("op_4760_cast_fp16")];
            fp32 var_4761_epsilon_0 = const()[name = string("op_4761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4761_cast_fp16 = rsqrt(epsilon = var_4761_epsilon_0, x = var_4760_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_111_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4761_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49008192)))];
            tensor<fp16, [1, 128, 512]> var_4763_cast_fp16 = mul(x = out_111_cast_fp16, y = tm_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4763_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_93_cast_fp16 = add(x = x_137_cast_fp16, y = var_4763_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49009280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49271488))))[name = string("tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_71_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_71_cast_fp16")];
            string x_141_mode_0 = const()[name = string("x_141_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_141_cast_fp16 = gelu(mode = x_141_mode_0, x = linear_71_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_4783_begin_0 = const()[name = string("op_4783_begin_0"), val = tensor<int32, [4]>([0, 0, 7, 0])];
            tensor<int32, [4]> var_4783_end_0 = const()[name = string("op_4783_end_0"), val = tensor<int32, [4]>([1, 128, 8, 512])];
            tensor<bool, [4]> var_4783_end_mask_0 = const()[name = string("op_4783_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4783_squeeze_mask_0 = const()[name = string("op_4783_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_4783_cast_fp16 = slice_by_index(begin = var_4783_begin_0, end = var_4783_end_0, end_mask = var_4783_end_mask_0, squeeze_mask = var_4783_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_95_cast_fp16 = mul(x = x_141_cast_fp16, y = var_4783_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49272576))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49534784))))[name = string("tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4793_cast_fp16 = abs(x = linear_72_cast_fp16)[name = string("op_4793_cast_fp16")];
            bool var_4796 = const()[name = string("op_4796"), val = bool(true)];
            tensor<int32, [1]> var_4797_axes_0 = const()[name = string("op_4797_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4797_cast_fp16 = reduce_max(axes = var_4797_axes_0, keep_dims = var_4796, x = var_4793_cast_fp16)[name = string("op_4797_cast_fp16")];
            fp16 var_4798_to_fp16 = const()[name = string("op_4798_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_129_cast_fp16 = maximum(x = var_4797_cast_fp16, y = var_4798_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_129_cast_fp16 = real_div(x = linear_72_cast_fp16, y = s_129_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4801_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4801_cast_fp16")];
            tensor<int32, [1]> var_4806_axes_0 = const()[name = string("op_4806_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4806_keep_dims_0 = const()[name = string("op_4806_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4806_cast_fp16 = reduce_mean(axes = var_4806_axes_0, keep_dims = var_4806_keep_dims_0, x = var_4801_cast_fp16)[name = string("op_4806_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4807_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4807_cast_fp16")];
            fp32 var_4808_epsilon_0 = const()[name = string("op_4808_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4808_cast_fp16 = inverse(epsilon = var_4808_epsilon_0, x = var_4807_cast_fp16)[name = string("op_4808_cast_fp16")];
            fp16 var_4809_to_fp16 = const()[name = string("op_4809_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4810_cast_fp16 = mul(x = var_4808_cast_fp16, y = var_4809_to_fp16)[name = string("op_4810_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4812_cast_fp16 = add(x = var_4806_cast_fp16, y = var_4810_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp32 var_4813_epsilon_0 = const()[name = string("op_4813_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4813_cast_fp16 = rsqrt(epsilon = var_4813_epsilon_0, x = var_4812_cast_fp16)[name = string("op_4813_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_113_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4813_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49535872)))];
            tensor<fp16, [1, 128, 512]> var_4815_cast_fp16 = mul(x = out_113_cast_fp16, y = tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4815_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_15_cast_fp16 = add(x = input_93_cast_fp16, y = var_4815_cast_fp16)[name = string("h_15_cast_fp16")];
            tensor<fp16, [1]> tm_layers_7_layer_scalar_to_fp16 = const()[name = string("tm_layers_7_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-1])];
            tensor<fp16, [1, 128, 512]> x_145_cast_fp16 = mul(x = h_15_cast_fp16, y = tm_layers_7_layer_scalar_to_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4819_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4819_cast_fp16")];
            bool var_4822 = const()[name = string("op_4822"), val = bool(true)];
            tensor<int32, [1]> var_4823_axes_0 = const()[name = string("op_4823_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_4823_cast_fp16 = reduce_max(axes = var_4823_axes_0, keep_dims = var_4822, x = var_4819_cast_fp16)[name = string("op_4823_cast_fp16")];
            fp16 var_4824_to_fp16 = const()[name = string("op_4824_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_131_cast_fp16 = maximum(x = var_4823_cast_fp16, y = var_4824_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_131_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_131_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_4827_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [1]> var_4832_axes_0 = const()[name = string("op_4832_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4832_keep_dims_0 = const()[name = string("op_4832_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_4832_cast_fp16 = reduce_mean(axes = var_4832_axes_0, keep_dims = var_4832_keep_dims_0, x = var_4827_cast_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4833_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4833_cast_fp16")];
            fp32 var_4834_epsilon_0 = const()[name = string("op_4834_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_4834_cast_fp16 = inverse(epsilon = var_4834_epsilon_0, x = var_4833_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp16 var_4835_to_fp16 = const()[name = string("op_4835_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_4836_cast_fp16 = mul(x = var_4834_cast_fp16, y = var_4835_to_fp16)[name = string("op_4836_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_4838_cast_fp16 = add(x = var_4832_cast_fp16, y = var_4836_cast_fp16)[name = string("op_4838_cast_fp16")];
            fp32 var_4839_epsilon_0 = const()[name = string("op_4839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_4839_cast_fp16 = rsqrt(epsilon = var_4839_epsilon_0, x = var_4838_cast_fp16)[name = string("op_4839_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_115_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4839_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49536960)))];
            tensor<fp16, [1, 128, 512]> input_97_cast_fp16 = mul(x = out_115_cast_fp16, y = tm_layers_8_input_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49538048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50062400))))[name = string("tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4849 = const()[name = string("op_4849"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_4850_cast_fp16 = reshape(shape = var_4849, x = linear_73_cast_fp16)[name = string("op_4850_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50064512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50326720))))[name = string("tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [4]> var_4861 = const()[name = string("op_4861"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_4862_cast_fp16 = reshape(shape = var_4861, x = linear_74_cast_fp16)[name = string("op_4862_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50327808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50590016))))[name = string("tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [4]> var_4873 = const()[name = string("op_4873"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_4874_cast_fp16 = reshape(shape = var_4873, x = linear_75_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<int32, [4]> x_151_perm_0 = const()[name = string("x_151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_4850_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 4, 128, 256]> var_4878_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4878_cast_fp16")];
            bool var_4881 = const()[name = string("op_4881"), val = bool(true)];
            tensor<int32, [1]> var_4882_axes_0 = const()[name = string("op_4882_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_4882_cast_fp16 = reduce_max(axes = var_4882_axes_0, keep_dims = var_4881, x = var_4878_cast_fp16)[name = string("op_4882_cast_fp16")];
            fp16 var_4883_to_fp16 = const()[name = string("op_4883_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_133_cast_fp16 = maximum(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_133_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_133_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4886_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4886_cast_fp16")];
            tensor<int32, [1]> var_4891_axes_0 = const()[name = string("op_4891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4891_keep_dims_0 = const()[name = string("op_4891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_4891_cast_fp16 = reduce_mean(axes = var_4891_axes_0, keep_dims = var_4891_keep_dims_0, x = var_4886_cast_fp16)[name = string("op_4891_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_4892_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4892_cast_fp16")];
            fp32 var_4893_epsilon_0 = const()[name = string("op_4893_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_4893_cast_fp16 = inverse(epsilon = var_4893_epsilon_0, x = var_4892_cast_fp16)[name = string("op_4893_cast_fp16")];
            fp16 var_4894_to_fp16 = const()[name = string("op_4894_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_4895_cast_fp16 = mul(x = var_4893_cast_fp16, y = var_4894_to_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_4897_cast_fp16 = add(x = var_4891_cast_fp16, y = var_4895_cast_fp16)[name = string("op_4897_cast_fp16")];
            fp32 var_4898_epsilon_0 = const()[name = string("op_4898_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_4898_cast_fp16 = rsqrt(epsilon = var_4898_epsilon_0, x = var_4897_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_117_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4898_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591104)))];
            tensor<fp16, [1, 4, 128, 256]> q_33_cast_fp16 = mul(x = out_117_cast_fp16, y = tm_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_4862_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 2, 128, 256]> var_4901_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4901_cast_fp16")];
            bool var_4904 = const()[name = string("op_4904"), val = bool(true)];
            tensor<int32, [1]> var_4905_axes_0 = const()[name = string("op_4905_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_4905_cast_fp16 = reduce_max(axes = var_4905_axes_0, keep_dims = var_4904, x = var_4901_cast_fp16)[name = string("op_4905_cast_fp16")];
            fp16 var_4906_to_fp16 = const()[name = string("op_4906_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_135_cast_fp16 = maximum(x = var_4905_cast_fp16, y = var_4906_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_135_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_135_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4909_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [1]> var_4914_axes_0 = const()[name = string("op_4914_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4914_keep_dims_0 = const()[name = string("op_4914_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_4914_cast_fp16 = reduce_mean(axes = var_4914_axes_0, keep_dims = var_4914_keep_dims_0, x = var_4909_cast_fp16)[name = string("op_4914_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4915_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4915_cast_fp16")];
            fp32 var_4916_epsilon_0 = const()[name = string("op_4916_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_4916_cast_fp16 = inverse(epsilon = var_4916_epsilon_0, x = var_4915_cast_fp16)[name = string("op_4916_cast_fp16")];
            fp16 var_4917_to_fp16 = const()[name = string("op_4917_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_4918_cast_fp16 = mul(x = var_4916_cast_fp16, y = var_4917_to_fp16)[name = string("op_4918_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4920_cast_fp16 = add(x = var_4914_cast_fp16, y = var_4918_cast_fp16)[name = string("op_4920_cast_fp16")];
            fp32 var_4921_epsilon_0 = const()[name = string("op_4921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_4921_cast_fp16 = rsqrt(epsilon = var_4921_epsilon_0, x = var_4920_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_119_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4921_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591680)))];
            tensor<fp16, [1, 2, 128, 256]> k_49_cast_fp16 = mul(x = out_119_cast_fp16, y = tm_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_151_cast_fp16 = transpose(perm = x_151_perm_0, x = var_4874_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 2, 128, 256]> var_4924_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4924_cast_fp16")];
            bool var_4927 = const()[name = string("op_4927"), val = bool(true)];
            tensor<int32, [1]> var_4928_axes_0 = const()[name = string("op_4928_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_4928_cast_fp16 = reduce_max(axes = var_4928_axes_0, keep_dims = var_4927, x = var_4924_cast_fp16)[name = string("op_4928_cast_fp16")];
            fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_137_cast_fp16 = maximum(x = var_4928_cast_fp16, y = var_4929_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_137_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_137_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4932_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4932_cast_fp16")];
            tensor<int32, [1]> var_4937_axes_0 = const()[name = string("op_4937_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4937_keep_dims_0 = const()[name = string("op_4937_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_4937_cast_fp16 = reduce_mean(axes = var_4937_axes_0, keep_dims = var_4937_keep_dims_0, x = var_4932_cast_fp16)[name = string("op_4937_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4938_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4938_cast_fp16")];
            fp32 var_4939_epsilon_0 = const()[name = string("op_4939_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_4939_cast_fp16 = inverse(epsilon = var_4939_epsilon_0, x = var_4938_cast_fp16)[name = string("op_4939_cast_fp16")];
            fp16 var_4940_to_fp16 = const()[name = string("op_4940_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_4941_cast_fp16 = mul(x = var_4939_cast_fp16, y = var_4940_to_fp16)[name = string("op_4941_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_4943_cast_fp16 = add(x = var_4937_cast_fp16, y = var_4941_cast_fp16)[name = string("op_4943_cast_fp16")];
            fp32 var_4944_epsilon_0 = const()[name = string("op_4944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_4944_cast_fp16 = rsqrt(epsilon = var_4944_epsilon_0, x = var_4943_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_33_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4944_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4946_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = q_33_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_4952_cast_fp16 = mul(x = var_4951_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4957_begin_0 = const()[name = string("op_4957_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4957_end_0 = const()[name = string("op_4957_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_4957_end_mask_0 = const()[name = string("op_4957_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_4957_cast_fp16 = slice_by_index(begin = var_4957_begin_0, end = var_4957_end_0, end_mask = var_4957_end_mask_0, x = q_33_cast_fp16)[name = string("op_4957_cast_fp16")];
            int32 var_4959 = const()[name = string("op_4959"), val = int32(-1)];
            bool var_4960_interleave_0 = const()[name = string("op_4960_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_4960_cast_fp16 = concat(axis = var_4959, interleave = var_4960_interleave_0, values = (var_4952_cast_fp16, var_4957_cast_fp16))[name = string("op_4960_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_4961_cast_fp16 = mul(x = var_4960_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4961_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_35_cast_fp16 = add(x = var_4946_cast_fp16, y = var_4961_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4964_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4964_cast_fp16")];
            tensor<int32, [4]> var_4969_begin_0 = const()[name = string("op_4969_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4969_end_0 = const()[name = string("op_4969_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_4969_end_mask_0 = const()[name = string("op_4969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_4969_cast_fp16 = slice_by_index(begin = var_4969_begin_0, end = var_4969_end_0, end_mask = var_4969_end_mask_0, x = k_49_cast_fp16)[name = string("op_4969_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_4970_cast_fp16 = mul(x = var_4969_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_4970_cast_fp16")];
            tensor<int32, [4]> var_4975_begin_0 = const()[name = string("op_4975_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4975_end_0 = const()[name = string("op_4975_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_4975_end_mask_0 = const()[name = string("op_4975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_4975_cast_fp16 = slice_by_index(begin = var_4975_begin_0, end = var_4975_end_0, end_mask = var_4975_end_mask_0, x = k_49_cast_fp16)[name = string("op_4975_cast_fp16")];
            int32 var_4977 = const()[name = string("op_4977"), val = int32(-1)];
            bool var_4978_interleave_0 = const()[name = string("op_4978_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_4978_cast_fp16 = concat(axis = var_4977, interleave = var_4978_interleave_0, values = (var_4970_cast_fp16, var_4975_cast_fp16))[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_4979_cast_fp16 = mul(x = var_4978_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_51_cast_fp16 = add(x = var_4964_cast_fp16, y = var_4979_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_16_reps_0 = const()[name = string("tile_16_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = k_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [4, 1, 128, 256]> tile_16_cast_fp16 = tile(reps = tile_16_reps_0, x = transpose_32_cast_fp16)[name = string("tile_16_cast_fp16")];
            tensor<int32, [5]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_32_cast_fp16 = reshape(shape = concat_32, x = tile_16_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [5]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [4, 1, 128, 256]> reshape_33_cast_fp16 = reshape(shape = concat_33, x = transpose_33_cast_fp16)[name = string("reshape_33_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_17_reps_0 = const()[name = string("tile_17_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = v_33_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [4, 1, 128, 256]> tile_17_cast_fp16 = tile(reps = tile_17_reps_0, x = transpose_34_cast_fp16)[name = string("tile_17_cast_fp16")];
            tensor<int32, [5]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_34_cast_fp16 = reshape(shape = concat_34, x = tile_17_cast_fp16)[name = string("reshape_34_cast_fp16")];
            tensor<int32, [5]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = reshape_34_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [4, 1, 128, 256]> reshape_35_cast_fp16 = reshape(shape = concat_35, x = transpose_35_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = reshape_33_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 4, 128, 128]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = q_35_cast_fp16, y = transpose_104_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_4995_cast_fp16 = add(x = var_4993_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4995_cast_fp16")];
            int32 var_4996 = const()[name = string("op_4996"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_137_cast_fp16 = softmax(axis = var_4996, x = var_4995_cast_fp16)[name = string("w_137_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 4, 128, 256]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = w_137_cast_fp16, y = v_35_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5002_perm_0 = const()[name = string("op_5002_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5006 = const()[name = string("op_5006"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_5002_cast_fp16 = transpose(perm = var_5002_perm_0, x = var_4999_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 128, 1024]> input_99_cast_fp16 = reshape(shape = var_5006, x = var_5002_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50592256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51116608))))[name = string("tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5011_cast_fp16 = abs(x = linear_76_cast_fp16)[name = string("op_5011_cast_fp16")];
            bool var_5014 = const()[name = string("op_5014"), val = bool(true)];
            tensor<int32, [1]> var_5015_axes_0 = const()[name = string("op_5015_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5015_cast_fp16 = reduce_max(axes = var_5015_axes_0, keep_dims = var_5014, x = var_5011_cast_fp16)[name = string("op_5015_cast_fp16")];
            fp16 var_5016_to_fp16 = const()[name = string("op_5016_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_139_cast_fp16 = maximum(x = var_5015_cast_fp16, y = var_5016_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_139_cast_fp16 = real_div(x = linear_76_cast_fp16, y = s_139_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5019_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_5019_cast_fp16")];
            tensor<int32, [1]> var_5024_axes_0 = const()[name = string("op_5024_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5024_keep_dims_0 = const()[name = string("op_5024_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5024_cast_fp16 = reduce_mean(axes = var_5024_axes_0, keep_dims = var_5024_keep_dims_0, x = var_5019_cast_fp16)[name = string("op_5024_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5025_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_5025_cast_fp16")];
            fp32 var_5026_epsilon_0 = const()[name = string("op_5026_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5026_cast_fp16 = inverse(epsilon = var_5026_epsilon_0, x = var_5025_cast_fp16)[name = string("op_5026_cast_fp16")];
            fp16 var_5027_to_fp16 = const()[name = string("op_5027_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5028_cast_fp16 = mul(x = var_5026_cast_fp16, y = var_5027_to_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5030_cast_fp16 = add(x = var_5024_cast_fp16, y = var_5028_cast_fp16)[name = string("op_5030_cast_fp16")];
            fp32 var_5031_epsilon_0 = const()[name = string("op_5031_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5031_cast_fp16 = rsqrt(epsilon = var_5031_epsilon_0, x = var_5030_cast_fp16)[name = string("op_5031_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_121_cast_fp16 = mul(x = y_139_cast_fp16, y = var_5031_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51117696)))];
            tensor<fp16, [1, 128, 512]> var_5033_cast_fp16 = mul(x = out_121_cast_fp16, y = tm_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_5033_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = var_5033_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5036_cast_fp16 = abs(x = x_155_cast_fp16)[name = string("op_5036_cast_fp16")];
            bool var_5039 = const()[name = string("op_5039"), val = bool(true)];
            tensor<int32, [1]> var_5040_axes_0 = const()[name = string("op_5040_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5040_cast_fp16 = reduce_max(axes = var_5040_axes_0, keep_dims = var_5039, x = var_5036_cast_fp16)[name = string("op_5040_cast_fp16")];
            fp16 var_5041_to_fp16 = const()[name = string("op_5041_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_141_cast_fp16 = maximum(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_141_cast_fp16 = real_div(x = x_155_cast_fp16, y = s_141_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5044_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<int32, [1]> var_5049_axes_0 = const()[name = string("op_5049_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5049_keep_dims_0 = const()[name = string("op_5049_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5049_cast_fp16 = reduce_mean(axes = var_5049_axes_0, keep_dims = var_5049_keep_dims_0, x = var_5044_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5050_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_5050_cast_fp16")];
            fp32 var_5051_epsilon_0 = const()[name = string("op_5051_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5051_cast_fp16 = inverse(epsilon = var_5051_epsilon_0, x = var_5050_cast_fp16)[name = string("op_5051_cast_fp16")];
            fp16 var_5052_to_fp16 = const()[name = string("op_5052_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5053_cast_fp16 = mul(x = var_5051_cast_fp16, y = var_5052_to_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5055_cast_fp16 = add(x = var_5049_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            fp32 var_5056_epsilon_0 = const()[name = string("op_5056_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5056_cast_fp16 = rsqrt(epsilon = var_5056_epsilon_0, x = var_5055_cast_fp16)[name = string("op_5056_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_123_cast_fp16 = mul(x = y_141_cast_fp16, y = var_5056_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51118784)))];
            tensor<fp16, [1, 128, 512]> input_101_cast_fp16 = mul(x = out_123_cast_fp16, y = tm_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51119872))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52168512))))[name = string("tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_77_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_77_cast_fp16")];
            string var_5063_mode_0 = const()[name = string("op_5063_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_5063_cast_fp16 = gelu(mode = var_5063_mode_0, x = linear_77_cast_fp16)[name = string("op_5063_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52172672))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53221312))))[name = string("tm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_78_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_103_cast_fp16 = mul(x = var_5063_cast_fp16, y = linear_78_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53225472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54274112))))[name = string("tm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5071_cast_fp16 = abs(x = linear_79_cast_fp16)[name = string("op_5071_cast_fp16")];
            bool var_5074 = const()[name = string("op_5074"), val = bool(true)];
            tensor<int32, [1]> var_5075_axes_0 = const()[name = string("op_5075_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5075_cast_fp16 = reduce_max(axes = var_5075_axes_0, keep_dims = var_5074, x = var_5071_cast_fp16)[name = string("op_5075_cast_fp16")];
            fp16 var_5076_to_fp16 = const()[name = string("op_5076_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_143_cast_fp16 = maximum(x = var_5075_cast_fp16, y = var_5076_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_143_cast_fp16 = real_div(x = linear_79_cast_fp16, y = s_143_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5079_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_5079_cast_fp16")];
            tensor<int32, [1]> var_5084_axes_0 = const()[name = string("op_5084_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5084_keep_dims_0 = const()[name = string("op_5084_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5084_cast_fp16 = reduce_mean(axes = var_5084_axes_0, keep_dims = var_5084_keep_dims_0, x = var_5079_cast_fp16)[name = string("op_5084_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5085_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5086_cast_fp16 = inverse(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            fp16 var_5087_to_fp16 = const()[name = string("op_5087_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5088_cast_fp16 = mul(x = var_5086_cast_fp16, y = var_5087_to_fp16)[name = string("op_5088_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5090_cast_fp16 = add(x = var_5084_cast_fp16, y = var_5088_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp32 var_5091_epsilon_0 = const()[name = string("op_5091_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5091_cast_fp16 = rsqrt(epsilon = var_5091_epsilon_0, x = var_5090_cast_fp16)[name = string("op_5091_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_125_cast_fp16 = mul(x = y_143_cast_fp16, y = var_5091_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54275200)))];
            tensor<fp16, [1, 128, 512]> var_5093_cast_fp16 = mul(x = out_125_cast_fp16, y = tm_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5093_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_105_cast_fp16 = add(x = x_155_cast_fp16, y = var_5093_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54276288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54538496))))[name = string("tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_80_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_80_cast_fp16")];
            string x_159_mode_0 = const()[name = string("x_159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_159_cast_fp16 = gelu(mode = x_159_mode_0, x = linear_80_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [4]> var_5113_begin_0 = const()[name = string("op_5113_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_5113_end_0 = const()[name = string("op_5113_end_0"), val = tensor<int32, [4]>([1, 128, 9, 512])];
            tensor<bool, [4]> var_5113_end_mask_0 = const()[name = string("op_5113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5113_squeeze_mask_0 = const()[name = string("op_5113_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_5113_cast_fp16 = slice_by_index(begin = var_5113_begin_0, end = var_5113_end_0, end_mask = var_5113_end_mask_0, squeeze_mask = var_5113_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_107_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5113_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54539584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54801792))))[name = string("tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5123_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5123_cast_fp16")];
            bool var_5126 = const()[name = string("op_5126"), val = bool(true)];
            tensor<int32, [1]> var_5127_axes_0 = const()[name = string("op_5127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5127_cast_fp16 = reduce_max(axes = var_5127_axes_0, keep_dims = var_5126, x = var_5123_cast_fp16)[name = string("op_5127_cast_fp16")];
            fp16 var_5128_to_fp16 = const()[name = string("op_5128_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_145_cast_fp16 = maximum(x = var_5127_cast_fp16, y = var_5128_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_145_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_145_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5131_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5136_axes_0 = const()[name = string("op_5136_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5136_keep_dims_0 = const()[name = string("op_5136_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5136_cast_fp16 = reduce_mean(axes = var_5136_axes_0, keep_dims = var_5136_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5136_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5137_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5138_cast_fp16 = inverse(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp16 var_5139_to_fp16 = const()[name = string("op_5139_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5140_cast_fp16 = mul(x = var_5138_cast_fp16, y = var_5139_to_fp16)[name = string("op_5140_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5142_cast_fp16 = add(x = var_5136_cast_fp16, y = var_5140_cast_fp16)[name = string("op_5142_cast_fp16")];
            fp32 var_5143_epsilon_0 = const()[name = string("op_5143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5143_cast_fp16 = rsqrt(epsilon = var_5143_epsilon_0, x = var_5142_cast_fp16)[name = string("op_5143_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_127_cast_fp16 = mul(x = y_145_cast_fp16, y = var_5143_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54802880)))];
            tensor<fp16, [1, 128, 512]> var_5145_cast_fp16 = mul(x = out_127_cast_fp16, y = tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5145_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_17_cast_fp16 = add(x = input_105_cast_fp16, y = var_5145_cast_fp16)[name = string("h_17_cast_fp16")];
            tensor<fp16, [1]> tm_layers_8_layer_scalar_to_fp16 = const()[name = string("tm_layers_8_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bcp-1])];
            tensor<fp16, [1, 128, 512]> x_163_cast_fp16 = mul(x = h_17_cast_fp16, y = tm_layers_8_layer_scalar_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5149_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_5149_cast_fp16")];
            bool var_5152 = const()[name = string("op_5152"), val = bool(true)];
            tensor<int32, [1]> var_5153_axes_0 = const()[name = string("op_5153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5153_cast_fp16 = reduce_max(axes = var_5153_axes_0, keep_dims = var_5152, x = var_5149_cast_fp16)[name = string("op_5153_cast_fp16")];
            fp16 var_5154_to_fp16 = const()[name = string("op_5154_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_147_cast_fp16 = maximum(x = var_5153_cast_fp16, y = var_5154_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_147_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_147_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5157_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_5157_cast_fp16")];
            tensor<int32, [1]> var_5162_axes_0 = const()[name = string("op_5162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5162_keep_dims_0 = const()[name = string("op_5162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5162_cast_fp16 = reduce_mean(axes = var_5162_axes_0, keep_dims = var_5162_keep_dims_0, x = var_5157_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5163_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_5163_cast_fp16")];
            fp32 var_5164_epsilon_0 = const()[name = string("op_5164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5164_cast_fp16 = inverse(epsilon = var_5164_epsilon_0, x = var_5163_cast_fp16)[name = string("op_5164_cast_fp16")];
            fp16 var_5165_to_fp16 = const()[name = string("op_5165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5166_cast_fp16 = mul(x = var_5164_cast_fp16, y = var_5165_to_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5168_cast_fp16 = add(x = var_5162_cast_fp16, y = var_5166_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp32 var_5169_epsilon_0 = const()[name = string("op_5169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5169_cast_fp16 = rsqrt(epsilon = var_5169_epsilon_0, x = var_5168_cast_fp16)[name = string("op_5169_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_129_cast_fp16 = mul(x = y_147_cast_fp16, y = var_5169_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54803968)))];
            tensor<fp16, [1, 128, 512]> input_109_cast_fp16 = mul(x = out_129_cast_fp16, y = tm_layers_9_input_layernorm_weight_to_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54805056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55329408))))[name = string("tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [4]> var_5179 = const()[name = string("op_5179"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_5180_cast_fp16 = reshape(shape = var_5179, x = linear_82_cast_fp16)[name = string("op_5180_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55331520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55593728))))[name = string("tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [4]> var_5191 = const()[name = string("op_5191"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_5192_cast_fp16 = reshape(shape = var_5191, x = linear_83_cast_fp16)[name = string("op_5192_cast_fp16")];
            tensor<int32, [4]> x_167_perm_0 = const()[name = string("x_167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55594816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55857024))))[name = string("tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5203 = const()[name = string("op_5203"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_5204_cast_fp16 = reshape(shape = var_5203, x = linear_84_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_5180_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 4, 128, 256]> var_5208_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_5208_cast_fp16")];
            bool var_5211 = const()[name = string("op_5211"), val = bool(true)];
            tensor<int32, [1]> var_5212_axes_0 = const()[name = string("op_5212_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_5212_cast_fp16 = reduce_max(axes = var_5212_axes_0, keep_dims = var_5211, x = var_5208_cast_fp16)[name = string("op_5212_cast_fp16")];
            fp16 var_5213_to_fp16 = const()[name = string("op_5213_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_149_cast_fp16 = maximum(x = var_5212_cast_fp16, y = var_5213_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_149_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_149_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_5216_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<int32, [1]> var_5221_axes_0 = const()[name = string("op_5221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5221_keep_dims_0 = const()[name = string("op_5221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_5221_cast_fp16 = reduce_mean(axes = var_5221_axes_0, keep_dims = var_5221_keep_dims_0, x = var_5216_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_5222_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_5223_cast_fp16 = inverse(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            fp16 var_5224_to_fp16 = const()[name = string("op_5224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_5225_cast_fp16 = mul(x = var_5223_cast_fp16, y = var_5224_to_fp16)[name = string("op_5225_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_5227_cast_fp16 = add(x = var_5221_cast_fp16, y = var_5225_cast_fp16)[name = string("op_5227_cast_fp16")];
            fp32 var_5228_epsilon_0 = const()[name = string("op_5228_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_5228_cast_fp16 = rsqrt(epsilon = var_5228_epsilon_0, x = var_5227_cast_fp16)[name = string("op_5228_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_131_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5228_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = var_5192_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 2, 128, 256]> var_5231_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_5231_cast_fp16")];
            bool var_5234 = const()[name = string("op_5234"), val = bool(true)];
            tensor<int32, [1]> var_5235_axes_0 = const()[name = string("op_5235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_5235_cast_fp16 = reduce_max(axes = var_5235_axes_0, keep_dims = var_5234, x = var_5231_cast_fp16)[name = string("op_5235_cast_fp16")];
            fp16 var_5236_to_fp16 = const()[name = string("op_5236_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_151_cast_fp16 = maximum(x = var_5235_cast_fp16, y = var_5236_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_151_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_151_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_5239_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5239_cast_fp16")];
            tensor<int32, [1]> var_5244_axes_0 = const()[name = string("op_5244_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5244_keep_dims_0 = const()[name = string("op_5244_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_5244_cast_fp16 = reduce_mean(axes = var_5244_axes_0, keep_dims = var_5244_keep_dims_0, x = var_5239_cast_fp16)[name = string("op_5244_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_5245_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_5245_cast_fp16")];
            fp32 var_5246_epsilon_0 = const()[name = string("op_5246_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_5246_cast_fp16 = inverse(epsilon = var_5246_epsilon_0, x = var_5245_cast_fp16)[name = string("op_5246_cast_fp16")];
            fp16 var_5247_to_fp16 = const()[name = string("op_5247_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_5248_cast_fp16 = mul(x = var_5246_cast_fp16, y = var_5247_to_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_5250_cast_fp16 = add(x = var_5244_cast_fp16, y = var_5248_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp32 var_5251_epsilon_0 = const()[name = string("op_5251_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_5251_cast_fp16 = rsqrt(epsilon = var_5251_epsilon_0, x = var_5250_cast_fp16)[name = string("op_5251_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_133_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5251_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [256]> tm_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858112)))];
            tensor<fp16, [1, 2, 128, 256]> k_55_cast_fp16 = mul(x = out_133_cast_fp16, y = tm_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_5204_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2, 128, 256]> var_5254_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_5254_cast_fp16")];
            bool var_5257 = const()[name = string("op_5257"), val = bool(true)];
            tensor<int32, [1]> var_5258_axes_0 = const()[name = string("op_5258_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_5258_cast_fp16 = reduce_max(axes = var_5258_axes_0, keep_dims = var_5257, x = var_5254_cast_fp16)[name = string("op_5258_cast_fp16")];
            fp16 var_5259_to_fp16 = const()[name = string("op_5259_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_153_cast_fp16 = maximum(x = var_5258_cast_fp16, y = var_5259_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_153_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_153_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_5262_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5262_cast_fp16")];
            tensor<int32, [1]> var_5267_axes_0 = const()[name = string("op_5267_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5267_keep_dims_0 = const()[name = string("op_5267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_5267_cast_fp16 = reduce_mean(axes = var_5267_axes_0, keep_dims = var_5267_keep_dims_0, x = var_5262_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_5268_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5268_cast_fp16")];
            fp32 var_5269_epsilon_0 = const()[name = string("op_5269_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_5269_cast_fp16 = inverse(epsilon = var_5269_epsilon_0, x = var_5268_cast_fp16)[name = string("op_5269_cast_fp16")];
            fp16 var_5270_to_fp16 = const()[name = string("op_5270_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_5271_cast_fp16 = mul(x = var_5269_cast_fp16, y = var_5270_to_fp16)[name = string("op_5271_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_5273_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5271_cast_fp16)[name = string("op_5273_cast_fp16")];
            fp32 var_5274_epsilon_0 = const()[name = string("op_5274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_5274_cast_fp16 = rsqrt(epsilon = var_5274_epsilon_0, x = var_5273_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_37_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5274_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_5276_cast_fp16 = mul(x = out_131_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5276_cast_fp16")];
            tensor<int32, [4]> var_5281_begin_0 = const()[name = string("op_5281_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5281_end_0 = const()[name = string("op_5281_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_5281_end_mask_0 = const()[name = string("op_5281_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_5281_cast_fp16 = slice_by_index(begin = var_5281_begin_0, end = var_5281_end_0, end_mask = var_5281_end_mask_0, x = out_131_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_5282_cast_fp16 = mul(x = var_5281_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_5282_cast_fp16")];
            tensor<int32, [4]> var_5287_begin_0 = const()[name = string("op_5287_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5287_end_0 = const()[name = string("op_5287_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_5287_end_mask_0 = const()[name = string("op_5287_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_5287_cast_fp16 = slice_by_index(begin = var_5287_begin_0, end = var_5287_end_0, end_mask = var_5287_end_mask_0, x = out_131_cast_fp16)[name = string("op_5287_cast_fp16")];
            int32 var_5289 = const()[name = string("op_5289"), val = int32(-1)];
            bool var_5290_interleave_0 = const()[name = string("op_5290_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_5290_cast_fp16 = concat(axis = var_5289, interleave = var_5290_interleave_0, values = (var_5282_cast_fp16, var_5287_cast_fp16))[name = string("op_5290_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_5291_cast_fp16 = mul(x = var_5290_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5291_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_39_cast_fp16 = add(x = var_5276_cast_fp16, y = var_5291_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_5294_cast_fp16 = mul(x = k_55_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5294_cast_fp16")];
            tensor<int32, [4]> var_5299_begin_0 = const()[name = string("op_5299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5299_end_0 = const()[name = string("op_5299_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_5299_end_mask_0 = const()[name = string("op_5299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_5299_cast_fp16 = slice_by_index(begin = var_5299_begin_0, end = var_5299_end_0, end_mask = var_5299_end_mask_0, x = k_55_cast_fp16)[name = string("op_5299_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_5300_cast_fp16 = mul(x = var_5299_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_5300_cast_fp16")];
            tensor<int32, [4]> var_5305_begin_0 = const()[name = string("op_5305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5305_end_0 = const()[name = string("op_5305_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_5305_end_mask_0 = const()[name = string("op_5305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_5305_cast_fp16 = slice_by_index(begin = var_5305_begin_0, end = var_5305_end_0, end_mask = var_5305_end_mask_0, x = k_55_cast_fp16)[name = string("op_5305_cast_fp16")];
            int32 var_5307 = const()[name = string("op_5307"), val = int32(-1)];
            bool var_5308_interleave_0 = const()[name = string("op_5308_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_5308_cast_fp16 = concat(axis = var_5307, interleave = var_5308_interleave_0, values = (var_5300_cast_fp16, var_5305_cast_fp16))[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_5309_cast_fp16 = mul(x = var_5308_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_57_cast_fp16 = add(x = var_5294_cast_fp16, y = var_5309_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_18_reps_0 = const()[name = string("tile_18_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = k_57_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [4, 1, 128, 256]> tile_18_cast_fp16 = tile(reps = tile_18_reps_0, x = transpose_36_cast_fp16)[name = string("tile_18_cast_fp16")];
            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_36_cast_fp16 = reshape(shape = concat_36, x = tile_18_cast_fp16)[name = string("reshape_36_cast_fp16")];
            tensor<int32, [5]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = reshape_36_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [4, 1, 128, 256]> reshape_37_cast_fp16 = reshape(shape = concat_37, x = transpose_37_cast_fp16)[name = string("reshape_37_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_19_reps_0 = const()[name = string("tile_19_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = v_37_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [4, 1, 128, 256]> tile_19_cast_fp16 = tile(reps = tile_19_reps_0, x = transpose_38_cast_fp16)[name = string("tile_19_cast_fp16")];
            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_38_cast_fp16 = reshape(shape = concat_38, x = tile_19_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [5]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [4, 1, 128, 256]> reshape_39_cast_fp16 = reshape(shape = concat_39, x = transpose_39_cast_fp16)[name = string("reshape_39_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5323_transpose_x_0 = const()[name = string("op_5323_transpose_x_0"), val = bool(false)];
            bool var_5323_transpose_y_0 = const()[name = string("op_5323_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = reshape_37_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 4, 128, 128]> var_5323_cast_fp16 = matmul(transpose_x = var_5323_transpose_x_0, transpose_y = var_5323_transpose_y_0, x = q_39_cast_fp16, y = transpose_105_cast_fp16)[name = string("op_5323_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_5325_cast_fp16 = add(x = var_5323_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5325_cast_fp16")];
            int32 var_5326 = const()[name = string("op_5326"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_153_cast_fp16 = softmax(axis = var_5326, x = var_5325_cast_fp16)[name = string("w_153_cast_fp16")];
            bool var_5329_transpose_x_0 = const()[name = string("op_5329_transpose_x_0"), val = bool(false)];
            bool var_5329_transpose_y_0 = const()[name = string("op_5329_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = reshape_39_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 4, 128, 256]> var_5329_cast_fp16 = matmul(transpose_x = var_5329_transpose_x_0, transpose_y = var_5329_transpose_y_0, x = w_153_cast_fp16, y = v_39_cast_fp16)[name = string("op_5329_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = var_5329_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 128, 1024]> input_111_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56383040))))[name = string("tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5341_cast_fp16 = abs(x = linear_85_cast_fp16)[name = string("op_5341_cast_fp16")];
            bool var_5344 = const()[name = string("op_5344"), val = bool(true)];
            tensor<int32, [1]> var_5345_axes_0 = const()[name = string("op_5345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5345_cast_fp16 = reduce_max(axes = var_5345_axes_0, keep_dims = var_5344, x = var_5341_cast_fp16)[name = string("op_5345_cast_fp16")];
            fp16 var_5346_to_fp16 = const()[name = string("op_5346_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_155_cast_fp16 = maximum(x = var_5345_cast_fp16, y = var_5346_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_155_cast_fp16 = real_div(x = linear_85_cast_fp16, y = s_155_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5349_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5354_axes_0 = const()[name = string("op_5354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5354_keep_dims_0 = const()[name = string("op_5354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5354_cast_fp16 = reduce_mean(axes = var_5354_axes_0, keep_dims = var_5354_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5355_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5355_cast_fp16")];
            fp32 var_5356_epsilon_0 = const()[name = string("op_5356_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5356_cast_fp16 = inverse(epsilon = var_5356_epsilon_0, x = var_5355_cast_fp16)[name = string("op_5356_cast_fp16")];
            fp16 var_5357_to_fp16 = const()[name = string("op_5357_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5358_cast_fp16 = mul(x = var_5356_cast_fp16, y = var_5357_to_fp16)[name = string("op_5358_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5360_cast_fp16 = add(x = var_5354_cast_fp16, y = var_5358_cast_fp16)[name = string("op_5360_cast_fp16")];
            fp32 var_5361_epsilon_0 = const()[name = string("op_5361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5361_cast_fp16 = rsqrt(epsilon = var_5361_epsilon_0, x = var_5360_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_135_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5361_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56384128)))];
            tensor<fp16, [1, 128, 512]> var_5363_cast_fp16 = mul(x = out_135_cast_fp16, y = tm_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_5363_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_173_cast_fp16 = add(x = x_163_cast_fp16, y = var_5363_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5366_cast_fp16 = abs(x = x_173_cast_fp16)[name = string("op_5366_cast_fp16")];
            bool var_5369 = const()[name = string("op_5369"), val = bool(true)];
            tensor<int32, [1]> var_5370_axes_0 = const()[name = string("op_5370_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5370_cast_fp16 = reduce_max(axes = var_5370_axes_0, keep_dims = var_5369, x = var_5366_cast_fp16)[name = string("op_5370_cast_fp16")];
            fp16 var_5371_to_fp16 = const()[name = string("op_5371_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_157_cast_fp16 = maximum(x = var_5370_cast_fp16, y = var_5371_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_157_cast_fp16 = real_div(x = x_173_cast_fp16, y = s_157_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5374_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [1]> var_5379_axes_0 = const()[name = string("op_5379_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5379_keep_dims_0 = const()[name = string("op_5379_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5379_cast_fp16 = reduce_mean(axes = var_5379_axes_0, keep_dims = var_5379_keep_dims_0, x = var_5374_cast_fp16)[name = string("op_5379_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5380_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5381_cast_fp16 = inverse(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            fp16 var_5382_to_fp16 = const()[name = string("op_5382_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5383_cast_fp16 = mul(x = var_5381_cast_fp16, y = var_5382_to_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5385_cast_fp16 = add(x = var_5379_cast_fp16, y = var_5383_cast_fp16)[name = string("op_5385_cast_fp16")];
            fp32 var_5386_epsilon_0 = const()[name = string("op_5386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5386_cast_fp16 = rsqrt(epsilon = var_5386_epsilon_0, x = var_5385_cast_fp16)[name = string("op_5386_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_137_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5386_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56385216)))];
            tensor<fp16, [1, 128, 512]> input_113_cast_fp16 = mul(x = out_137_cast_fp16, y = tm_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56386304))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57434944))))[name = string("tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_86_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_86_cast_fp16")];
            string var_5393_mode_0 = const()[name = string("op_5393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_5393_cast_fp16 = gelu(mode = var_5393_mode_0, x = linear_86_cast_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57439104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58487744))))[name = string("tm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_87_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_115_cast_fp16 = mul(x = var_5393_cast_fp16, y = linear_87_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58491904))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59540544))))[name = string("tm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5401_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5401_cast_fp16")];
            bool var_5404 = const()[name = string("op_5404"), val = bool(true)];
            tensor<int32, [1]> var_5405_axes_0 = const()[name = string("op_5405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5405_cast_fp16 = reduce_max(axes = var_5405_axes_0, keep_dims = var_5404, x = var_5401_cast_fp16)[name = string("op_5405_cast_fp16")];
            fp16 var_5406_to_fp16 = const()[name = string("op_5406_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_159_cast_fp16 = maximum(x = var_5405_cast_fp16, y = var_5406_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_159_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_159_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5409_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5409_cast_fp16")];
            tensor<int32, [1]> var_5414_axes_0 = const()[name = string("op_5414_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5414_keep_dims_0 = const()[name = string("op_5414_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5414_cast_fp16 = reduce_mean(axes = var_5414_axes_0, keep_dims = var_5414_keep_dims_0, x = var_5409_cast_fp16)[name = string("op_5414_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5415_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5415_cast_fp16")];
            fp32 var_5416_epsilon_0 = const()[name = string("op_5416_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5416_cast_fp16 = inverse(epsilon = var_5416_epsilon_0, x = var_5415_cast_fp16)[name = string("op_5416_cast_fp16")];
            fp16 var_5417_to_fp16 = const()[name = string("op_5417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5418_cast_fp16 = mul(x = var_5416_cast_fp16, y = var_5417_to_fp16)[name = string("op_5418_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5420_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5418_cast_fp16)[name = string("op_5420_cast_fp16")];
            fp32 var_5421_epsilon_0 = const()[name = string("op_5421_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5421_cast_fp16 = rsqrt(epsilon = var_5421_epsilon_0, x = var_5420_cast_fp16)[name = string("op_5421_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_139_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5421_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59541632)))];
            tensor<fp16, [1, 128, 512]> var_5423_cast_fp16 = mul(x = out_139_cast_fp16, y = tm_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5423_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_117_cast_fp16 = add(x = x_173_cast_fp16, y = var_5423_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59542720))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59804928))))[name = string("tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_89_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_89_cast_fp16")];
            string x_177_mode_0 = const()[name = string("x_177_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_177_cast_fp16 = gelu(mode = x_177_mode_0, x = linear_89_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [4]> var_5443_begin_0 = const()[name = string("op_5443_begin_0"), val = tensor<int32, [4]>([0, 0, 9, 0])];
            tensor<int32, [4]> var_5443_end_0 = const()[name = string("op_5443_end_0"), val = tensor<int32, [4]>([1, 128, 10, 512])];
            tensor<bool, [4]> var_5443_end_mask_0 = const()[name = string("op_5443_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5443_squeeze_mask_0 = const()[name = string("op_5443_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_5443_cast_fp16 = slice_by_index(begin = var_5443_begin_0, end = var_5443_end_0, end_mask = var_5443_end_mask_0, squeeze_mask = var_5443_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5443_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_119_cast_fp16 = mul(x = x_177_cast_fp16, y = var_5443_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59806016))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60068224))))[name = string("tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5453_cast_fp16 = abs(x = linear_90_cast_fp16)[name = string("op_5453_cast_fp16")];
            bool var_5456 = const()[name = string("op_5456"), val = bool(true)];
            tensor<int32, [1]> var_5457_axes_0 = const()[name = string("op_5457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5457_cast_fp16 = reduce_max(axes = var_5457_axes_0, keep_dims = var_5456, x = var_5453_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_161_cast_fp16 = maximum(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_161_cast_fp16 = real_div(x = linear_90_cast_fp16, y = s_161_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5461_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<int32, [1]> var_5466_axes_0 = const()[name = string("op_5466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5466_keep_dims_0 = const()[name = string("op_5466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5466_cast_fp16 = reduce_mean(axes = var_5466_axes_0, keep_dims = var_5466_keep_dims_0, x = var_5461_cast_fp16)[name = string("op_5466_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5467_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5467_cast_fp16")];
            fp32 var_5468_epsilon_0 = const()[name = string("op_5468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5468_cast_fp16 = inverse(epsilon = var_5468_epsilon_0, x = var_5467_cast_fp16)[name = string("op_5468_cast_fp16")];
            fp16 var_5469_to_fp16 = const()[name = string("op_5469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5470_cast_fp16 = mul(x = var_5468_cast_fp16, y = var_5469_to_fp16)[name = string("op_5470_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5472_cast_fp16 = add(x = var_5466_cast_fp16, y = var_5470_cast_fp16)[name = string("op_5472_cast_fp16")];
            fp32 var_5473_epsilon_0 = const()[name = string("op_5473_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5473_cast_fp16 = rsqrt(epsilon = var_5473_epsilon_0, x = var_5472_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_141_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5473_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60069312)))];
            tensor<fp16, [1, 128, 512]> var_5475_cast_fp16 = mul(x = out_141_cast_fp16, y = tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5475_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_19_cast_fp16 = add(x = input_117_cast_fp16, y = var_5475_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<fp16, [1]> tm_layers_9_layer_scalar_to_fp16 = const()[name = string("tm_layers_9_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9ep-1])];
            tensor<fp16, [1, 128, 512]> x_181_cast_fp16 = mul(x = h_19_cast_fp16, y = tm_layers_9_layer_scalar_to_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5479_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_5479_cast_fp16")];
            bool var_5482 = const()[name = string("op_5482"), val = bool(true)];
            tensor<int32, [1]> var_5483_axes_0 = const()[name = string("op_5483_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5483_cast_fp16 = reduce_max(axes = var_5483_axes_0, keep_dims = var_5482, x = var_5479_cast_fp16)[name = string("op_5483_cast_fp16")];
            fp16 var_5484_to_fp16 = const()[name = string("op_5484_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_163_cast_fp16 = maximum(x = var_5483_cast_fp16, y = var_5484_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_163_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_163_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5487_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5492_axes_0 = const()[name = string("op_5492_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5492_keep_dims_0 = const()[name = string("op_5492_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5492_cast_fp16 = reduce_mean(axes = var_5492_axes_0, keep_dims = var_5492_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5493_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5493_cast_fp16")];
            fp32 var_5494_epsilon_0 = const()[name = string("op_5494_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5494_cast_fp16 = inverse(epsilon = var_5494_epsilon_0, x = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp16 var_5495_to_fp16 = const()[name = string("op_5495_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5496_cast_fp16 = mul(x = var_5494_cast_fp16, y = var_5495_to_fp16)[name = string("op_5496_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5498_cast_fp16 = add(x = var_5492_cast_fp16, y = var_5496_cast_fp16)[name = string("op_5498_cast_fp16")];
            fp32 var_5499_epsilon_0 = const()[name = string("op_5499_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5499_cast_fp16 = rsqrt(epsilon = var_5499_epsilon_0, x = var_5498_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_143_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5499_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60070400)))];
            tensor<fp16, [1, 128, 512]> input_121_cast_fp16 = mul(x = out_143_cast_fp16, y = tm_layers_10_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60071488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60595840))))[name = string("tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5509 = const()[name = string("op_5509"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_5510_cast_fp16 = reshape(shape = var_5509, x = linear_91_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60597952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60860160))))[name = string("tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5521 = const()[name = string("op_5521"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_5522_cast_fp16 = reshape(shape = var_5521, x = linear_92_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60861248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61123456))))[name = string("tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5533 = const()[name = string("op_5533"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_5534_cast_fp16 = reshape(shape = var_5533, x = linear_93_cast_fp16)[name = string("op_5534_cast_fp16")];
            tensor<int32, [4]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5510_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 4, 128, 256]> var_5538_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_5538_cast_fp16")];
            bool var_5541 = const()[name = string("op_5541"), val = bool(true)];
            tensor<int32, [1]> var_5542_axes_0 = const()[name = string("op_5542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_5542_cast_fp16 = reduce_max(axes = var_5542_axes_0, keep_dims = var_5541, x = var_5538_cast_fp16)[name = string("op_5542_cast_fp16")];
            fp16 var_5543_to_fp16 = const()[name = string("op_5543_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_165_cast_fp16 = maximum(x = var_5542_cast_fp16, y = var_5543_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_165_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_165_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_5546_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<int32, [1]> var_5551_axes_0 = const()[name = string("op_5551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5551_keep_dims_0 = const()[name = string("op_5551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_5551_cast_fp16 = reduce_mean(axes = var_5551_axes_0, keep_dims = var_5551_keep_dims_0, x = var_5546_cast_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_5552_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5552_cast_fp16")];
            fp32 var_5553_epsilon_0 = const()[name = string("op_5553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_5553_cast_fp16 = inverse(epsilon = var_5553_epsilon_0, x = var_5552_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp16 var_5554_to_fp16 = const()[name = string("op_5554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_5555_cast_fp16 = mul(x = var_5553_cast_fp16, y = var_5554_to_fp16)[name = string("op_5555_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_5557_cast_fp16 = add(x = var_5551_cast_fp16, y = var_5555_cast_fp16)[name = string("op_5557_cast_fp16")];
            fp32 var_5558_epsilon_0 = const()[name = string("op_5558_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_5558_cast_fp16 = rsqrt(epsilon = var_5558_epsilon_0, x = var_5557_cast_fp16)[name = string("op_5558_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_145_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5558_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61124544)))];
            tensor<fp16, [1, 4, 128, 256]> q_41_cast_fp16 = mul(x = out_145_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_41_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_5522_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2, 128, 256]> var_5561_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_5561_cast_fp16")];
            bool var_5564 = const()[name = string("op_5564"), val = bool(true)];
            tensor<int32, [1]> var_5565_axes_0 = const()[name = string("op_5565_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_5565_cast_fp16 = reduce_max(axes = var_5565_axes_0, keep_dims = var_5564, x = var_5561_cast_fp16)[name = string("op_5565_cast_fp16")];
            fp16 var_5566_to_fp16 = const()[name = string("op_5566_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_167_cast_fp16 = maximum(x = var_5565_cast_fp16, y = var_5566_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_167_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_167_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_5569_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5569_cast_fp16")];
            tensor<int32, [1]> var_5574_axes_0 = const()[name = string("op_5574_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5574_keep_dims_0 = const()[name = string("op_5574_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_5574_cast_fp16 = reduce_mean(axes = var_5574_axes_0, keep_dims = var_5574_keep_dims_0, x = var_5569_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_5575_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5575_cast_fp16")];
            fp32 var_5576_epsilon_0 = const()[name = string("op_5576_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_5576_cast_fp16 = inverse(epsilon = var_5576_epsilon_0, x = var_5575_cast_fp16)[name = string("op_5576_cast_fp16")];
            fp16 var_5577_to_fp16 = const()[name = string("op_5577_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_5578_cast_fp16 = mul(x = var_5576_cast_fp16, y = var_5577_to_fp16)[name = string("op_5578_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_5580_cast_fp16 = add(x = var_5574_cast_fp16, y = var_5578_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_5581_cast_fp16 = rsqrt(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_147_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5581_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [256]> tm_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125120)))];
            tensor<fp16, [1, 2, 128, 256]> k_61_cast_fp16 = mul(x = out_147_cast_fp16, y = tm_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = var_5534_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 2, 128, 256]> var_5584_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_5584_cast_fp16")];
            bool var_5587 = const()[name = string("op_5587"), val = bool(true)];
            tensor<int32, [1]> var_5588_axes_0 = const()[name = string("op_5588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_5588_cast_fp16 = reduce_max(axes = var_5588_axes_0, keep_dims = var_5587, x = var_5584_cast_fp16)[name = string("op_5588_cast_fp16")];
            fp16 var_5589_to_fp16 = const()[name = string("op_5589_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_169_cast_fp16 = maximum(x = var_5588_cast_fp16, y = var_5589_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_169_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_169_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_5592_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5592_cast_fp16")];
            tensor<int32, [1]> var_5597_axes_0 = const()[name = string("op_5597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5597_keep_dims_0 = const()[name = string("op_5597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_5597_cast_fp16 = reduce_mean(axes = var_5597_axes_0, keep_dims = var_5597_keep_dims_0, x = var_5592_cast_fp16)[name = string("op_5597_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_5598_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5598_cast_fp16")];
            fp32 var_5599_epsilon_0 = const()[name = string("op_5599_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_5599_cast_fp16 = inverse(epsilon = var_5599_epsilon_0, x = var_5598_cast_fp16)[name = string("op_5599_cast_fp16")];
            fp16 var_5600_to_fp16 = const()[name = string("op_5600_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_5601_cast_fp16 = mul(x = var_5599_cast_fp16, y = var_5600_to_fp16)[name = string("op_5601_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_5603_cast_fp16 = add(x = var_5597_cast_fp16, y = var_5601_cast_fp16)[name = string("op_5603_cast_fp16")];
            fp32 var_5604_epsilon_0 = const()[name = string("op_5604_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_5604_cast_fp16 = rsqrt(epsilon = var_5604_epsilon_0, x = var_5603_cast_fp16)[name = string("op_5604_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_41_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5604_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_5606_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5611_begin_0 = const()[name = string("op_5611_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5611_end_0 = const()[name = string("op_5611_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_5611_end_mask_0 = const()[name = string("op_5611_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_5611_cast_fp16 = slice_by_index(begin = var_5611_begin_0, end = var_5611_end_0, end_mask = var_5611_end_mask_0, x = q_41_cast_fp16)[name = string("op_5611_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_5612_cast_fp16 = mul(x = var_5611_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_5612_cast_fp16")];
            tensor<int32, [4]> var_5617_begin_0 = const()[name = string("op_5617_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5617_end_0 = const()[name = string("op_5617_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_5617_end_mask_0 = const()[name = string("op_5617_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_5617_cast_fp16 = slice_by_index(begin = var_5617_begin_0, end = var_5617_end_0, end_mask = var_5617_end_mask_0, x = q_41_cast_fp16)[name = string("op_5617_cast_fp16")];
            int32 var_5619 = const()[name = string("op_5619"), val = int32(-1)];
            bool var_5620_interleave_0 = const()[name = string("op_5620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_5620_cast_fp16 = concat(axis = var_5619, interleave = var_5620_interleave_0, values = (var_5612_cast_fp16, var_5617_cast_fp16))[name = string("op_5620_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_5621_cast_fp16 = mul(x = var_5620_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5621_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_43_cast_fp16 = add(x = var_5606_cast_fp16, y = var_5621_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_5624_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = k_61_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = k_61_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool var_5638_interleave_0 = const()[name = string("op_5638_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_5638_cast_fp16 = concat(axis = var_5637, interleave = var_5638_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("op_5638_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_5639_cast_fp16 = mul(x = var_5638_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_63_cast_fp16 = add(x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_20_reps_0 = const()[name = string("tile_20_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = k_63_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [4, 1, 128, 256]> tile_20_cast_fp16 = tile(reps = tile_20_reps_0, x = transpose_40_cast_fp16)[name = string("tile_20_cast_fp16")];
            tensor<int32, [5]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_40_cast_fp16 = reshape(shape = concat_40, x = tile_20_cast_fp16)[name = string("reshape_40_cast_fp16")];
            tensor<int32, [5]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = reshape_40_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [4, 1, 128, 256]> reshape_41_cast_fp16 = reshape(shape = concat_41, x = transpose_41_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_21_reps_0 = const()[name = string("tile_21_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = v_41_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [4, 1, 128, 256]> tile_21_cast_fp16 = tile(reps = tile_21_reps_0, x = transpose_42_cast_fp16)[name = string("tile_21_cast_fp16")];
            tensor<int32, [5]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_42_cast_fp16 = reshape(shape = concat_42, x = tile_21_cast_fp16)[name = string("reshape_42_cast_fp16")];
            tensor<int32, [5]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = reshape_42_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [4, 1, 128, 256]> reshape_43_cast_fp16 = reshape(shape = concat_43, x = transpose_43_cast_fp16)[name = string("reshape_43_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5653_transpose_x_0 = const()[name = string("op_5653_transpose_x_0"), val = bool(false)];
            bool var_5653_transpose_y_0 = const()[name = string("op_5653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 4, 128, 128]> var_5653_cast_fp16 = matmul(transpose_x = var_5653_transpose_x_0, transpose_y = var_5653_transpose_y_0, x = q_43_cast_fp16, y = transpose_106_cast_fp16)[name = string("op_5653_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_5655_cast_fp16 = add(x = var_5653_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5655_cast_fp16")];
            int32 var_5656 = const()[name = string("op_5656"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_169_cast_fp16 = softmax(axis = var_5656, x = var_5655_cast_fp16)[name = string("w_169_cast_fp16")];
            bool var_5659_transpose_x_0 = const()[name = string("op_5659_transpose_x_0"), val = bool(false)];
            bool var_5659_transpose_y_0 = const()[name = string("op_5659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = reshape_43_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 4, 128, 256]> var_5659_cast_fp16 = matmul(transpose_x = var_5659_transpose_x_0, transpose_y = var_5659_transpose_y_0, x = w_169_cast_fp16, y = v_43_cast_fp16)[name = string("op_5659_cast_fp16")];
            tensor<int32, [4]> var_5662_perm_0 = const()[name = string("op_5662_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5666 = const()[name = string("op_5666"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_5662_cast_fp16 = transpose(perm = var_5662_perm_0, x = var_5659_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 128, 1024]> input_123_cast_fp16 = reshape(shape = var_5666, x = var_5662_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61650048))))[name = string("tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5671_cast_fp16 = abs(x = linear_94_cast_fp16)[name = string("op_5671_cast_fp16")];
            bool var_5674 = const()[name = string("op_5674"), val = bool(true)];
            tensor<int32, [1]> var_5675_axes_0 = const()[name = string("op_5675_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5675_cast_fp16 = reduce_max(axes = var_5675_axes_0, keep_dims = var_5674, x = var_5671_cast_fp16)[name = string("op_5675_cast_fp16")];
            fp16 var_5676_to_fp16 = const()[name = string("op_5676_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_171_cast_fp16 = maximum(x = var_5675_cast_fp16, y = var_5676_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_171_cast_fp16 = real_div(x = linear_94_cast_fp16, y = s_171_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5679_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5679_cast_fp16")];
            tensor<int32, [1]> var_5684_axes_0 = const()[name = string("op_5684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5684_keep_dims_0 = const()[name = string("op_5684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5684_cast_fp16 = reduce_mean(axes = var_5684_axes_0, keep_dims = var_5684_keep_dims_0, x = var_5679_cast_fp16)[name = string("op_5684_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5685_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5685_cast_fp16")];
            fp32 var_5686_epsilon_0 = const()[name = string("op_5686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5686_cast_fp16 = inverse(epsilon = var_5686_epsilon_0, x = var_5685_cast_fp16)[name = string("op_5686_cast_fp16")];
            fp16 var_5687_to_fp16 = const()[name = string("op_5687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5688_cast_fp16 = mul(x = var_5686_cast_fp16, y = var_5687_to_fp16)[name = string("op_5688_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5690_cast_fp16 = add(x = var_5684_cast_fp16, y = var_5688_cast_fp16)[name = string("op_5690_cast_fp16")];
            fp32 var_5691_epsilon_0 = const()[name = string("op_5691_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5691_cast_fp16 = rsqrt(epsilon = var_5691_epsilon_0, x = var_5690_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_149_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5691_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61651136)))];
            tensor<fp16, [1, 128, 512]> var_5693_cast_fp16 = mul(x = out_149_cast_fp16, y = tm_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5693_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = var_5693_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5696_cast_fp16 = abs(x = x_191_cast_fp16)[name = string("op_5696_cast_fp16")];
            bool var_5699 = const()[name = string("op_5699"), val = bool(true)];
            tensor<int32, [1]> var_5700_axes_0 = const()[name = string("op_5700_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5700_cast_fp16 = reduce_max(axes = var_5700_axes_0, keep_dims = var_5699, x = var_5696_cast_fp16)[name = string("op_5700_cast_fp16")];
            fp16 var_5701_to_fp16 = const()[name = string("op_5701_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_173_cast_fp16 = maximum(x = var_5700_cast_fp16, y = var_5701_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_173_cast_fp16 = real_div(x = x_191_cast_fp16, y = s_173_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5704_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5704_cast_fp16")];
            tensor<int32, [1]> var_5709_axes_0 = const()[name = string("op_5709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5709_keep_dims_0 = const()[name = string("op_5709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5709_cast_fp16 = reduce_mean(axes = var_5709_axes_0, keep_dims = var_5709_keep_dims_0, x = var_5704_cast_fp16)[name = string("op_5709_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5710_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5710_cast_fp16")];
            fp32 var_5711_epsilon_0 = const()[name = string("op_5711_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5711_cast_fp16 = inverse(epsilon = var_5711_epsilon_0, x = var_5710_cast_fp16)[name = string("op_5711_cast_fp16")];
            fp16 var_5712_to_fp16 = const()[name = string("op_5712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = string("op_5713_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5715_cast_fp16 = add(x = var_5709_cast_fp16, y = var_5713_cast_fp16)[name = string("op_5715_cast_fp16")];
            fp32 var_5716_epsilon_0 = const()[name = string("op_5716_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5716_cast_fp16 = rsqrt(epsilon = var_5716_epsilon_0, x = var_5715_cast_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_151_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5716_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61652224)))];
            tensor<fp16, [1, 128, 512]> input_125_cast_fp16 = mul(x = out_151_cast_fp16, y = tm_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61653312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62701952))))[name = string("tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_95_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_95_cast_fp16")];
            string var_5723_mode_0 = const()[name = string("op_5723_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_5723_cast_fp16 = gelu(mode = var_5723_mode_0, x = linear_95_cast_fp16)[name = string("op_5723_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62706112))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63754752))))[name = string("tm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_127_cast_fp16 = mul(x = var_5723_cast_fp16, y = linear_96_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63758912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64807552))))[name = string("tm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5731_cast_fp16 = abs(x = linear_97_cast_fp16)[name = string("op_5731_cast_fp16")];
            bool var_5734 = const()[name = string("op_5734"), val = bool(true)];
            tensor<int32, [1]> var_5735_axes_0 = const()[name = string("op_5735_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5735_cast_fp16 = reduce_max(axes = var_5735_axes_0, keep_dims = var_5734, x = var_5731_cast_fp16)[name = string("op_5735_cast_fp16")];
            fp16 var_5736_to_fp16 = const()[name = string("op_5736_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_175_cast_fp16 = maximum(x = var_5735_cast_fp16, y = var_5736_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_175_cast_fp16 = real_div(x = linear_97_cast_fp16, y = s_175_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5739_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [1]> var_5744_axes_0 = const()[name = string("op_5744_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5744_keep_dims_0 = const()[name = string("op_5744_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5744_cast_fp16 = reduce_mean(axes = var_5744_axes_0, keep_dims = var_5744_keep_dims_0, x = var_5739_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5745_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5745_cast_fp16")];
            fp32 var_5746_epsilon_0 = const()[name = string("op_5746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5746_cast_fp16 = inverse(epsilon = var_5746_epsilon_0, x = var_5745_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5747_to_fp16 = const()[name = string("op_5747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5748_cast_fp16 = mul(x = var_5746_cast_fp16, y = var_5747_to_fp16)[name = string("op_5748_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5750_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5748_cast_fp16)[name = string("op_5750_cast_fp16")];
            fp32 var_5751_epsilon_0 = const()[name = string("op_5751_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5751_cast_fp16 = rsqrt(epsilon = var_5751_epsilon_0, x = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_153_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5751_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64808640)))];
            tensor<fp16, [1, 128, 512]> var_5753_cast_fp16 = mul(x = out_153_cast_fp16, y = tm_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_129_cast_fp16 = add(x = x_191_cast_fp16, y = var_5753_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64809728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65071936))))[name = string("tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_98_cast_fp16")];
            string x_195_mode_0 = const()[name = string("x_195_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_195_cast_fp16 = gelu(mode = x_195_mode_0, x = linear_98_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [4]> var_5773_begin_0 = const()[name = string("op_5773_begin_0"), val = tensor<int32, [4]>([0, 0, 10, 0])];
            tensor<int32, [4]> var_5773_end_0 = const()[name = string("op_5773_end_0"), val = tensor<int32, [4]>([1, 128, 11, 512])];
            tensor<bool, [4]> var_5773_end_mask_0 = const()[name = string("op_5773_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5773_squeeze_mask_0 = const()[name = string("op_5773_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_5773_cast_fp16 = slice_by_index(begin = var_5773_begin_0, end = var_5773_end_0, end_mask = var_5773_end_mask_0, squeeze_mask = var_5773_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_131_cast_fp16 = mul(x = x_195_cast_fp16, y = var_5773_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65073024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65335232))))[name = string("tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5783_cast_fp16 = abs(x = linear_99_cast_fp16)[name = string("op_5783_cast_fp16")];
            bool var_5786 = const()[name = string("op_5786"), val = bool(true)];
            tensor<int32, [1]> var_5787_axes_0 = const()[name = string("op_5787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5787_cast_fp16 = reduce_max(axes = var_5787_axes_0, keep_dims = var_5786, x = var_5783_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp16 var_5788_to_fp16 = const()[name = string("op_5788_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_177_cast_fp16 = maximum(x = var_5787_cast_fp16, y = var_5788_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_177_cast_fp16 = real_div(x = linear_99_cast_fp16, y = s_177_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5791_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5791_cast_fp16")];
            tensor<int32, [1]> var_5796_axes_0 = const()[name = string("op_5796_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5796_keep_dims_0 = const()[name = string("op_5796_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5796_cast_fp16 = reduce_mean(axes = var_5796_axes_0, keep_dims = var_5796_keep_dims_0, x = var_5791_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5797_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5797_cast_fp16")];
            fp32 var_5798_epsilon_0 = const()[name = string("op_5798_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5798_cast_fp16 = inverse(epsilon = var_5798_epsilon_0, x = var_5797_cast_fp16)[name = string("op_5798_cast_fp16")];
            fp16 var_5799_to_fp16 = const()[name = string("op_5799_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5800_cast_fp16 = mul(x = var_5798_cast_fp16, y = var_5799_to_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5802_cast_fp16 = add(x = var_5796_cast_fp16, y = var_5800_cast_fp16)[name = string("op_5802_cast_fp16")];
            fp32 var_5803_epsilon_0 = const()[name = string("op_5803_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5803_cast_fp16 = rsqrt(epsilon = var_5803_epsilon_0, x = var_5802_cast_fp16)[name = string("op_5803_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_155_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5803_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65336320)))];
            tensor<fp16, [1, 128, 512]> var_5805_cast_fp16 = mul(x = out_155_cast_fp16, y = tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5805_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_21_cast_fp16 = add(x = input_129_cast_fp16, y = var_5805_cast_fp16)[name = string("h_21_cast_fp16")];
            tensor<fp16, [1]> tm_layers_10_layer_scalar_to_fp16 = const()[name = string("tm_layers_10_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.1cp-1])];
            tensor<fp16, [1, 128, 512]> x_199_cast_fp16 = mul(x = h_21_cast_fp16, y = tm_layers_10_layer_scalar_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5809_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5809_cast_fp16")];
            bool var_5812 = const()[name = string("op_5812"), val = bool(true)];
            tensor<int32, [1]> var_5813_axes_0 = const()[name = string("op_5813_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_5813_cast_fp16 = reduce_max(axes = var_5813_axes_0, keep_dims = var_5812, x = var_5809_cast_fp16)[name = string("op_5813_cast_fp16")];
            fp16 var_5814_to_fp16 = const()[name = string("op_5814_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_179_cast_fp16 = maximum(x = var_5813_cast_fp16, y = var_5814_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_179_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_179_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_5817_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5817_cast_fp16")];
            tensor<int32, [1]> var_5822_axes_0 = const()[name = string("op_5822_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5822_keep_dims_0 = const()[name = string("op_5822_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5822_cast_fp16 = reduce_mean(axes = var_5822_axes_0, keep_dims = var_5822_keep_dims_0, x = var_5817_cast_fp16)[name = string("op_5822_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5823_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp32 var_5824_epsilon_0 = const()[name = string("op_5824_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_5824_cast_fp16 = inverse(epsilon = var_5824_epsilon_0, x = var_5823_cast_fp16)[name = string("op_5824_cast_fp16")];
            fp16 var_5825_to_fp16 = const()[name = string("op_5825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_5826_cast_fp16 = mul(x = var_5824_cast_fp16, y = var_5825_to_fp16)[name = string("op_5826_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_5828_cast_fp16 = add(x = var_5822_cast_fp16, y = var_5826_cast_fp16)[name = string("op_5828_cast_fp16")];
            fp32 var_5829_epsilon_0 = const()[name = string("op_5829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_5829_cast_fp16 = rsqrt(epsilon = var_5829_epsilon_0, x = var_5828_cast_fp16)[name = string("op_5829_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_157_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5829_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65337408)))];
            tensor<fp16, [1, 128, 512]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = tm_layers_11_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65338496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66387136))))[name = string("tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_100_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5839 = const()[name = string("op_5839"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 128, 4, 512]> var_5840_cast_fp16 = reshape(shape = var_5839, x = linear_100_cast_fp16)[name = string("op_5840_cast_fp16")];
            tensor<int32, [4]> x_201_perm_0 = const()[name = string("x_201_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66391296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66653504))))[name = string("tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_5851 = const()[name = string("op_5851"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 128, 1, 512]> var_5852_cast_fp16 = reshape(shape = var_5851, x = linear_101_cast_fp16)[name = string("op_5852_cast_fp16")];
            tensor<int32, [4]> x_203_perm_0 = const()[name = string("x_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66654592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66916800))))[name = string("tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 128, 1, 512]> var_5864_cast_fp16 = reshape(shape = var_5863, x = linear_102_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 512]> x_201_cast_fp16 = transpose(perm = x_201_perm_0, x = var_5840_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 4, 128, 512]> var_5868_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5868_cast_fp16")];
            bool var_5871 = const()[name = string("op_5871"), val = bool(true)];
            tensor<int32, [1]> var_5872_axes_0 = const()[name = string("op_5872_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_5872_cast_fp16 = reduce_max(axes = var_5872_axes_0, keep_dims = var_5871, x = var_5868_cast_fp16)[name = string("op_5872_cast_fp16")];
            fp16 var_5873_to_fp16 = const()[name = string("op_5873_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_181_cast_fp16 = maximum(x = var_5872_cast_fp16, y = var_5873_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> y_181_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_181_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_5876_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5876_cast_fp16")];
            tensor<int32, [1]> var_5881_axes_0 = const()[name = string("op_5881_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5881_keep_dims_0 = const()[name = string("op_5881_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_5881_cast_fp16 = reduce_mean(axes = var_5881_axes_0, keep_dims = var_5881_keep_dims_0, x = var_5876_cast_fp16)[name = string("op_5881_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_5882_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5882_cast_fp16")];
            fp32 var_5883_epsilon_0 = const()[name = string("op_5883_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_5883_cast_fp16 = inverse(epsilon = var_5883_epsilon_0, x = var_5882_cast_fp16)[name = string("op_5883_cast_fp16")];
            fp16 var_5884_to_fp16 = const()[name = string("op_5884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_5885_cast_fp16 = mul(x = var_5883_cast_fp16, y = var_5884_to_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_5887_cast_fp16 = add(x = var_5881_cast_fp16, y = var_5885_cast_fp16)[name = string("op_5887_cast_fp16")];
            fp32 var_5888_epsilon_0 = const()[name = string("op_5888_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_5888_cast_fp16 = rsqrt(epsilon = var_5888_epsilon_0, x = var_5887_cast_fp16)[name = string("op_5888_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> out_159_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5888_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66917888)))];
            tensor<fp16, [1, 4, 128, 512]> q_45_cast_fp16 = mul(x = out_159_cast_fp16, y = tm_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> x_203_cast_fp16 = transpose(perm = x_203_perm_0, x = var_5852_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1, 128, 512]> var_5891_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5891_cast_fp16")];
            bool var_5894 = const()[name = string("op_5894"), val = bool(true)];
            tensor<int32, [1]> var_5895_axes_0 = const()[name = string("op_5895_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 128, 1]> var_5895_cast_fp16 = reduce_max(axes = var_5895_axes_0, keep_dims = var_5894, x = var_5891_cast_fp16)[name = string("op_5895_cast_fp16")];
            fp16 var_5896_to_fp16 = const()[name = string("op_5896_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 128, 1]> s_183_cast_fp16 = maximum(x = var_5895_cast_fp16, y = var_5896_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> y_183_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_183_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_5899_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5899_cast_fp16")];
            tensor<int32, [1]> var_5904_axes_0 = const()[name = string("op_5904_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5904_keep_dims_0 = const()[name = string("op_5904_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_5904_cast_fp16 = reduce_mean(axes = var_5904_axes_0, keep_dims = var_5904_keep_dims_0, x = var_5899_cast_fp16)[name = string("op_5904_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_5905_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5905_cast_fp16")];
            fp32 var_5906_epsilon_0 = const()[name = string("op_5906_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 128, 1]> var_5906_cast_fp16 = inverse(epsilon = var_5906_epsilon_0, x = var_5905_cast_fp16)[name = string("op_5906_cast_fp16")];
            fp16 var_5907_to_fp16 = const()[name = string("op_5907_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 128, 1]> var_5908_cast_fp16 = mul(x = var_5906_cast_fp16, y = var_5907_to_fp16)[name = string("op_5908_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_5910_cast_fp16 = add(x = var_5904_cast_fp16, y = var_5908_cast_fp16)[name = string("op_5910_cast_fp16")];
            fp32 var_5911_epsilon_0 = const()[name = string("op_5911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 128, 1]> var_5911_cast_fp16 = rsqrt(epsilon = var_5911_epsilon_0, x = var_5910_cast_fp16)[name = string("op_5911_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> out_161_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5911_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66918976)))];
            tensor<fp16, [1, 1, 128, 512]> k_67_cast_fp16 = mul(x = out_161_cast_fp16, y = tm_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_5864_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1, 128, 512]> var_5914_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5914_cast_fp16")];
            bool var_5917 = const()[name = string("op_5917"), val = bool(true)];
            tensor<int32, [1]> var_5918_axes_0 = const()[name = string("op_5918_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 128, 1]> var_5918_cast_fp16 = reduce_max(axes = var_5918_axes_0, keep_dims = var_5917, x = var_5914_cast_fp16)[name = string("op_5918_cast_fp16")];
            fp16 var_5919_to_fp16 = const()[name = string("op_5919_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 128, 1]> s_185_cast_fp16 = maximum(x = var_5918_cast_fp16, y = var_5919_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> y_185_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_185_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_5922_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5922_cast_fp16")];
            tensor<int32, [1]> var_5927_axes_0 = const()[name = string("op_5927_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5927_keep_dims_0 = const()[name = string("op_5927_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_5927_cast_fp16 = reduce_mean(axes = var_5927_axes_0, keep_dims = var_5927_keep_dims_0, x = var_5922_cast_fp16)[name = string("op_5927_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_5928_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5928_cast_fp16")];
            fp32 var_5929_epsilon_0 = const()[name = string("op_5929_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 128, 1]> var_5929_cast_fp16 = inverse(epsilon = var_5929_epsilon_0, x = var_5928_cast_fp16)[name = string("op_5929_cast_fp16")];
            fp16 var_5930_to_fp16 = const()[name = string("op_5930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 128, 1]> var_5931_cast_fp16 = mul(x = var_5929_cast_fp16, y = var_5930_to_fp16)[name = string("op_5931_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_5933_cast_fp16 = add(x = var_5927_cast_fp16, y = var_5931_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp32 var_5934_epsilon_0 = const()[name = string("op_5934_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 128, 1]> var_5934_cast_fp16 = rsqrt(epsilon = var_5934_epsilon_0, x = var_5933_cast_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> v_45_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5934_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_5936_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5936_cast_fp16")];
            tensor<int32, [4]> var_5941_begin_0 = const()[name = string("op_5941_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5941_end_0 = const()[name = string("op_5941_end_0"), val = tensor<int32, [4]>([1, 4, 128, 512])];
            tensor<bool, [4]> var_5941_end_mask_0 = const()[name = string("op_5941_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 256]> var_5941_cast_fp16 = slice_by_index(begin = var_5941_begin_0, end = var_5941_end_0, end_mask = var_5941_end_mask_0, x = q_45_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 256]> var_5942_cast_fp16 = mul(x = var_5941_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_5942_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 256]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = q_45_cast_fp16)[name = string("op_5947_cast_fp16")];
            int32 var_5949 = const()[name = string("op_5949"), val = int32(-1)];
            bool var_5950_interleave_0 = const()[name = string("op_5950_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 512]> var_5950_cast_fp16 = concat(axis = var_5949, interleave = var_5950_interleave_0, values = (var_5942_cast_fp16, var_5947_cast_fp16))[name = string("op_5950_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_5951_cast_fp16 = mul(x = var_5950_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5951_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> q_47_cast_fp16 = add(x = var_5936_cast_fp16, y = var_5951_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_5954_cast_fp16 = mul(x = k_67_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5954_cast_fp16")];
            tensor<int32, [4]> var_5959_begin_0 = const()[name = string("op_5959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5959_end_0 = const()[name = string("op_5959_end_0"), val = tensor<int32, [4]>([1, 1, 128, 512])];
            tensor<bool, [4]> var_5959_end_mask_0 = const()[name = string("op_5959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 128, 256]> var_5959_cast_fp16 = slice_by_index(begin = var_5959_begin_0, end = var_5959_end_0, end_mask = var_5959_end_mask_0, x = k_67_cast_fp16)[name = string("op_5959_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 128, 256]> var_5960_cast_fp16 = mul(x = var_5959_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_5960_cast_fp16")];
            tensor<int32, [4]> var_5965_begin_0 = const()[name = string("op_5965_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5965_end_0 = const()[name = string("op_5965_end_0"), val = tensor<int32, [4]>([1, 1, 128, 256])];
            tensor<bool, [4]> var_5965_end_mask_0 = const()[name = string("op_5965_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 128, 256]> var_5965_cast_fp16 = slice_by_index(begin = var_5965_begin_0, end = var_5965_end_0, end_mask = var_5965_end_mask_0, x = k_67_cast_fp16)[name = string("op_5965_cast_fp16")];
            int32 var_5967 = const()[name = string("op_5967"), val = int32(-1)];
            bool var_5968_interleave_0 = const()[name = string("op_5968_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 512]> var_5968_cast_fp16 = concat(axis = var_5967, interleave = var_5968_interleave_0, values = (var_5960_cast_fp16, var_5965_cast_fp16))[name = string("op_5968_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_5969_cast_fp16 = mul(x = var_5968_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5969_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> k_69_cast_fp16 = add(x = var_5954_cast_fp16, y = var_5969_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_22_reps_0 = const()[name = string("tile_22_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 128, 512]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = k_69_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [4, 1, 128, 512]> tile_22_cast_fp16 = tile(reps = tile_22_reps_0, x = transpose_44_cast_fp16)[name = string("tile_22_cast_fp16")];
            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([4, 1, 1, 128, 512])];
            tensor<fp16, [4, 1, 1, 128, 512]> reshape_44_cast_fp16 = reshape(shape = concat_44, x = tile_22_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [5]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 128, 512])];
            tensor<fp16, [1, 4, 1, 128, 512]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [4, 1, 128, 512]> reshape_45_cast_fp16 = reshape(shape = concat_45, x = transpose_45_cast_fp16)[name = string("reshape_45_cast_fp16")];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_23_reps_0 = const()[name = string("tile_23_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 128, 512]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = v_45_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [4, 1, 128, 512]> tile_23_cast_fp16 = tile(reps = tile_23_reps_0, x = transpose_46_cast_fp16)[name = string("tile_23_cast_fp16")];
            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([4, 1, 1, 128, 512])];
            tensor<fp16, [4, 1, 1, 128, 512]> reshape_46_cast_fp16 = reshape(shape = concat_46, x = tile_23_cast_fp16)[name = string("reshape_46_cast_fp16")];
            tensor<int32, [5]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 128, 512])];
            tensor<fp16, [1, 4, 1, 128, 512]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = reshape_46_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [4, 1, 128, 512]> reshape_47_cast_fp16 = reshape(shape = concat_47, x = transpose_47_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5983_transpose_x_0 = const()[name = string("op_5983_transpose_x_0"), val = bool(false)];
            bool var_5983_transpose_y_0 = const()[name = string("op_5983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 128]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = reshape_45_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 4, 128, 128]> var_5983_cast_fp16 = matmul(transpose_x = var_5983_transpose_x_0, transpose_y = var_5983_transpose_y_0, x = q_47_cast_fp16, y = transpose_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_5985_cast_fp16 = add(x = var_5983_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5985_cast_fp16")];
            int32 var_5986 = const()[name = string("op_5986"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_185_cast_fp16 = softmax(axis = var_5986, x = var_5985_cast_fp16)[name = string("w_185_cast_fp16")];
            bool var_5989_transpose_x_0 = const()[name = string("op_5989_transpose_x_0"), val = bool(false)];
            bool var_5989_transpose_y_0 = const()[name = string("op_5989_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 512]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 4, 128, 512]> var_5989_cast_fp16 = matmul(transpose_x = var_5989_transpose_x_0, transpose_y = var_5989_transpose_y_0, x = w_185_cast_fp16, y = v_47_cast_fp16)[name = string("op_5989_cast_fp16")];
            tensor<int32, [4]> var_5992_perm_0 = const()[name = string("op_5992_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5996 = const()[name = string("op_5996"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 128, 4, 512]> var_5992_cast_fp16 = transpose(perm = var_5992_perm_0, x = var_5989_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 128, 2048]> input_135_cast_fp16 = reshape(shape = var_5996, x = var_5992_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66920064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67968704))))[name = string("tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6001_cast_fp16 = abs(x = linear_103_cast_fp16)[name = string("op_6001_cast_fp16")];
            bool var_6004 = const()[name = string("op_6004"), val = bool(true)];
            tensor<int32, [1]> var_6005_axes_0 = const()[name = string("op_6005_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6005_cast_fp16 = reduce_max(axes = var_6005_axes_0, keep_dims = var_6004, x = var_6001_cast_fp16)[name = string("op_6005_cast_fp16")];
            fp16 var_6006_to_fp16 = const()[name = string("op_6006_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_187_cast_fp16 = maximum(x = var_6005_cast_fp16, y = var_6006_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_187_cast_fp16 = real_div(x = linear_103_cast_fp16, y = s_187_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6009_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_6009_cast_fp16")];
            tensor<int32, [1]> var_6014_axes_0 = const()[name = string("op_6014_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6014_keep_dims_0 = const()[name = string("op_6014_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6014_cast_fp16 = reduce_mean(axes = var_6014_axes_0, keep_dims = var_6014_keep_dims_0, x = var_6009_cast_fp16)[name = string("op_6014_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6015_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_6015_cast_fp16")];
            fp32 var_6016_epsilon_0 = const()[name = string("op_6016_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6016_cast_fp16 = inverse(epsilon = var_6016_epsilon_0, x = var_6015_cast_fp16)[name = string("op_6016_cast_fp16")];
            fp16 var_6017_to_fp16 = const()[name = string("op_6017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6018_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6020_cast_fp16 = add(x = var_6014_cast_fp16, y = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp32 var_6021_epsilon_0 = const()[name = string("op_6021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6021_cast_fp16 = rsqrt(epsilon = var_6021_epsilon_0, x = var_6020_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_163_cast_fp16 = mul(x = y_187_cast_fp16, y = var_6021_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969792)))];
            tensor<fp16, [1, 128, 512]> var_6023_cast_fp16 = mul(x = out_163_cast_fp16, y = tm_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_209_cast_fp16 = add(x = x_199_cast_fp16, y = var_6023_cast_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6026_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_6026_cast_fp16")];
            bool var_6029 = const()[name = string("op_6029"), val = bool(true)];
            tensor<int32, [1]> var_6030_axes_0 = const()[name = string("op_6030_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6030_cast_fp16 = reduce_max(axes = var_6030_axes_0, keep_dims = var_6029, x = var_6026_cast_fp16)[name = string("op_6030_cast_fp16")];
            fp16 var_6031_to_fp16 = const()[name = string("op_6031_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_189_cast_fp16 = maximum(x = var_6030_cast_fp16, y = var_6031_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_189_cast_fp16 = real_div(x = x_209_cast_fp16, y = s_189_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6034_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [1]> var_6039_axes_0 = const()[name = string("op_6039_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6039_keep_dims_0 = const()[name = string("op_6039_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6039_cast_fp16 = reduce_mean(axes = var_6039_axes_0, keep_dims = var_6039_keep_dims_0, x = var_6034_cast_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6040_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6041_cast_fp16 = inverse(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            fp16 var_6042_to_fp16 = const()[name = string("op_6042_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6043_cast_fp16 = mul(x = var_6041_cast_fp16, y = var_6042_to_fp16)[name = string("op_6043_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6045_cast_fp16 = add(x = var_6039_cast_fp16, y = var_6043_cast_fp16)[name = string("op_6045_cast_fp16")];
            fp32 var_6046_epsilon_0 = const()[name = string("op_6046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6046_cast_fp16 = rsqrt(epsilon = var_6046_epsilon_0, x = var_6045_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_165_cast_fp16 = mul(x = y_189_cast_fp16, y = var_6046_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67970880)))];
            tensor<fp16, [1, 128, 512]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = tm_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971968))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69020608))))[name = string("tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_104_cast_fp16")];
            string var_6053_mode_0 = const()[name = string("op_6053_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_6053_cast_fp16 = gelu(mode = var_6053_mode_0, x = linear_104_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70073408))))[name = string("tm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_105_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_139_cast_fp16 = mul(x = var_6053_cast_fp16, y = linear_105_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70077568))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71126208))))[name = string("tm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6061_cast_fp16 = abs(x = linear_106_cast_fp16)[name = string("op_6061_cast_fp16")];
            bool var_6064 = const()[name = string("op_6064"), val = bool(true)];
            tensor<int32, [1]> var_6065_axes_0 = const()[name = string("op_6065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6065_cast_fp16 = reduce_max(axes = var_6065_axes_0, keep_dims = var_6064, x = var_6061_cast_fp16)[name = string("op_6065_cast_fp16")];
            fp16 var_6066_to_fp16 = const()[name = string("op_6066_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_191_cast_fp16 = maximum(x = var_6065_cast_fp16, y = var_6066_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_191_cast_fp16 = real_div(x = linear_106_cast_fp16, y = s_191_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6069_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<int32, [1]> var_6074_axes_0 = const()[name = string("op_6074_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6074_keep_dims_0 = const()[name = string("op_6074_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6074_cast_fp16 = reduce_mean(axes = var_6074_axes_0, keep_dims = var_6074_keep_dims_0, x = var_6069_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6075_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_6075_cast_fp16")];
            fp32 var_6076_epsilon_0 = const()[name = string("op_6076_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6076_cast_fp16 = inverse(epsilon = var_6076_epsilon_0, x = var_6075_cast_fp16)[name = string("op_6076_cast_fp16")];
            fp16 var_6077_to_fp16 = const()[name = string("op_6077_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6078_cast_fp16 = mul(x = var_6076_cast_fp16, y = var_6077_to_fp16)[name = string("op_6078_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6080_cast_fp16 = add(x = var_6074_cast_fp16, y = var_6078_cast_fp16)[name = string("op_6080_cast_fp16")];
            fp32 var_6081_epsilon_0 = const()[name = string("op_6081_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6081_cast_fp16 = rsqrt(epsilon = var_6081_epsilon_0, x = var_6080_cast_fp16)[name = string("op_6081_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_167_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6081_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71127296)))];
            tensor<fp16, [1, 128, 512]> var_6083_cast_fp16 = mul(x = out_167_cast_fp16, y = tm_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6083_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_141_cast_fp16 = add(x = x_209_cast_fp16, y = var_6083_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71128384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71390592))))[name = string("tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_107_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_107_cast_fp16")];
            string x_213_mode_0 = const()[name = string("x_213_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_213_cast_fp16 = gelu(mode = x_213_mode_0, x = linear_107_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [4]> var_6103_begin_0 = const()[name = string("op_6103_begin_0"), val = tensor<int32, [4]>([0, 0, 11, 0])];
            tensor<int32, [4]> var_6103_end_0 = const()[name = string("op_6103_end_0"), val = tensor<int32, [4]>([1, 128, 12, 512])];
            tensor<bool, [4]> var_6103_end_mask_0 = const()[name = string("op_6103_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6103_squeeze_mask_0 = const()[name = string("op_6103_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_6103_cast_fp16 = slice_by_index(begin = var_6103_begin_0, end = var_6103_end_0, end_mask = var_6103_end_mask_0, squeeze_mask = var_6103_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6103_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_143_cast_fp16 = mul(x = x_213_cast_fp16, y = var_6103_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71391680))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71653888))))[name = string("tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6113_cast_fp16 = abs(x = linear_108_cast_fp16)[name = string("op_6113_cast_fp16")];
            bool var_6116 = const()[name = string("op_6116"), val = bool(true)];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6117_cast_fp16 = reduce_max(axes = var_6117_axes_0, keep_dims = var_6116, x = var_6113_cast_fp16)[name = string("op_6117_cast_fp16")];
            fp16 var_6118_to_fp16 = const()[name = string("op_6118_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_193_cast_fp16 = maximum(x = var_6117_cast_fp16, y = var_6118_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_193_cast_fp16 = real_div(x = linear_108_cast_fp16, y = s_193_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6121_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6121_cast_fp16")];
            tensor<int32, [1]> var_6126_axes_0 = const()[name = string("op_6126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6126_keep_dims_0 = const()[name = string("op_6126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6126_cast_fp16 = reduce_mean(axes = var_6126_axes_0, keep_dims = var_6126_keep_dims_0, x = var_6121_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6127_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_6127_cast_fp16")];
            fp32 var_6128_epsilon_0 = const()[name = string("op_6128_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6128_cast_fp16 = inverse(epsilon = var_6128_epsilon_0, x = var_6127_cast_fp16)[name = string("op_6128_cast_fp16")];
            fp16 var_6129_to_fp16 = const()[name = string("op_6129_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6130_cast_fp16 = mul(x = var_6128_cast_fp16, y = var_6129_to_fp16)[name = string("op_6130_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6132_cast_fp16 = add(x = var_6126_cast_fp16, y = var_6130_cast_fp16)[name = string("op_6132_cast_fp16")];
            fp32 var_6133_epsilon_0 = const()[name = string("op_6133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6133_cast_fp16 = rsqrt(epsilon = var_6133_epsilon_0, x = var_6132_cast_fp16)[name = string("op_6133_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_169_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6133_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71654976)))];
            tensor<fp16, [1, 128, 512]> var_6135_cast_fp16 = mul(x = out_169_cast_fp16, y = tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6135_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_23_cast_fp16 = add(x = input_141_cast_fp16, y = var_6135_cast_fp16)[name = string("h_23_cast_fp16")];
            tensor<fp16, [1]> tm_layers_11_layer_scalar_to_fp16 = const()[name = string("tm_layers_11_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.86p-2])];
            tensor<fp16, [1, 128, 512]> x_217_cast_fp16 = mul(x = h_23_cast_fp16, y = tm_layers_11_layer_scalar_to_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6139_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_195_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_195_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_195_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6153_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_171_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71656064)))];
            tensor<fp16, [1, 128, 512]> input_145_cast_fp16 = mul(x = out_171_cast_fp16, y = tm_layers_12_input_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71657152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72181504))))[name = string("tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<int32, [4]> var_6169 = const()[name = string("op_6169"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_6170_cast_fp16 = reshape(shape = var_6169, x = linear_109_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72183616))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72445824))))[name = string("tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [4]> var_6181 = const()[name = string("op_6181"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_6182_cast_fp16 = reshape(shape = var_6181, x = linear_110_cast_fp16)[name = string("op_6182_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72446912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72709120))))[name = string("tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<int32, [4]> var_6193 = const()[name = string("op_6193"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_6194_cast_fp16 = reshape(shape = var_6193, x = linear_111_cast_fp16)[name = string("op_6194_cast_fp16")];
            tensor<int32, [4]> x_223_perm_0 = const()[name = string("x_223_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6170_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 4, 128, 256]> var_6198_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_6198_cast_fp16")];
            bool var_6201 = const()[name = string("op_6201"), val = bool(true)];
            tensor<int32, [1]> var_6202_axes_0 = const()[name = string("op_6202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_6202_cast_fp16 = reduce_max(axes = var_6202_axes_0, keep_dims = var_6201, x = var_6198_cast_fp16)[name = string("op_6202_cast_fp16")];
            fp16 var_6203_to_fp16 = const()[name = string("op_6203_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_197_cast_fp16 = maximum(x = var_6202_cast_fp16, y = var_6203_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_197_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_197_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6206_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6206_cast_fp16")];
            tensor<int32, [1]> var_6211_axes_0 = const()[name = string("op_6211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6211_keep_dims_0 = const()[name = string("op_6211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_6211_cast_fp16 = reduce_mean(axes = var_6211_axes_0, keep_dims = var_6211_keep_dims_0, x = var_6206_cast_fp16)[name = string("op_6211_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_6212_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6212_cast_fp16")];
            fp32 var_6213_epsilon_0 = const()[name = string("op_6213_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_6213_cast_fp16 = inverse(epsilon = var_6213_epsilon_0, x = var_6212_cast_fp16)[name = string("op_6213_cast_fp16")];
            fp16 var_6214_to_fp16 = const()[name = string("op_6214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_6215_cast_fp16 = mul(x = var_6213_cast_fp16, y = var_6214_to_fp16)[name = string("op_6215_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_6217_cast_fp16 = add(x = var_6211_cast_fp16, y = var_6215_cast_fp16)[name = string("op_6217_cast_fp16")];
            fp32 var_6218_epsilon_0 = const()[name = string("op_6218_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_6218_cast_fp16 = rsqrt(epsilon = var_6218_epsilon_0, x = var_6217_cast_fp16)[name = string("op_6218_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_173_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6218_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_49_cast_fp16 = mul(x = out_173_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_49_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6182_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 2, 128, 256]> var_6221_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_6221_cast_fp16")];
            bool var_6224 = const()[name = string("op_6224"), val = bool(true)];
            tensor<int32, [1]> var_6225_axes_0 = const()[name = string("op_6225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_6225_cast_fp16 = reduce_max(axes = var_6225_axes_0, keep_dims = var_6224, x = var_6221_cast_fp16)[name = string("op_6225_cast_fp16")];
            fp16 var_6226_to_fp16 = const()[name = string("op_6226_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_199_cast_fp16 = maximum(x = var_6225_cast_fp16, y = var_6226_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_199_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_199_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6229_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6229_cast_fp16")];
            tensor<int32, [1]> var_6234_axes_0 = const()[name = string("op_6234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6234_keep_dims_0 = const()[name = string("op_6234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_6234_cast_fp16 = reduce_mean(axes = var_6234_axes_0, keep_dims = var_6234_keep_dims_0, x = var_6229_cast_fp16)[name = string("op_6234_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6235_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6235_cast_fp16")];
            fp32 var_6236_epsilon_0 = const()[name = string("op_6236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_6236_cast_fp16 = inverse(epsilon = var_6236_epsilon_0, x = var_6235_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_6238_cast_fp16 = mul(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6240_cast_fp16 = add(x = var_6234_cast_fp16, y = var_6238_cast_fp16)[name = string("op_6240_cast_fp16")];
            fp32 var_6241_epsilon_0 = const()[name = string("op_6241_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_6241_cast_fp16 = rsqrt(epsilon = var_6241_epsilon_0, x = var_6240_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_175_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6241_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_73_cast_fp16 = mul(x = out_175_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_223_cast_fp16 = transpose(perm = x_223_perm_0, x = var_6194_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2, 128, 256]> var_6244_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_6244_cast_fp16")];
            bool var_6247 = const()[name = string("op_6247"), val = bool(true)];
            tensor<int32, [1]> var_6248_axes_0 = const()[name = string("op_6248_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_6248_cast_fp16 = reduce_max(axes = var_6248_axes_0, keep_dims = var_6247, x = var_6244_cast_fp16)[name = string("op_6248_cast_fp16")];
            fp16 var_6249_to_fp16 = const()[name = string("op_6249_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_201_cast_fp16 = maximum(x = var_6248_cast_fp16, y = var_6249_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_201_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_201_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6252_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6252_cast_fp16")];
            tensor<int32, [1]> var_6257_axes_0 = const()[name = string("op_6257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6257_keep_dims_0 = const()[name = string("op_6257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_6257_cast_fp16 = reduce_mean(axes = var_6257_axes_0, keep_dims = var_6257_keep_dims_0, x = var_6252_cast_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6258_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6258_cast_fp16")];
            fp32 var_6259_epsilon_0 = const()[name = string("op_6259_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_6259_cast_fp16 = inverse(epsilon = var_6259_epsilon_0, x = var_6258_cast_fp16)[name = string("op_6259_cast_fp16")];
            fp16 var_6260_to_fp16 = const()[name = string("op_6260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_6261_cast_fp16 = mul(x = var_6259_cast_fp16, y = var_6260_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6263_cast_fp16 = add(x = var_6257_cast_fp16, y = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp32 var_6264_epsilon_0 = const()[name = string("op_6264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_6264_cast_fp16 = rsqrt(epsilon = var_6264_epsilon_0, x = var_6263_cast_fp16)[name = string("op_6264_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_49_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6264_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6266_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6266_cast_fp16")];
            tensor<int32, [4]> var_6271_begin_0 = const()[name = string("op_6271_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6271_end_0 = const()[name = string("op_6271_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_6271_end_mask_0 = const()[name = string("op_6271_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_6271_cast_fp16 = slice_by_index(begin = var_6271_begin_0, end = var_6271_end_0, end_mask = var_6271_end_mask_0, x = q_49_cast_fp16)[name = string("op_6271_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_6272_cast_fp16 = mul(x = var_6271_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_6272_cast_fp16")];
            tensor<int32, [4]> var_6277_begin_0 = const()[name = string("op_6277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6277_end_0 = const()[name = string("op_6277_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_6277_end_mask_0 = const()[name = string("op_6277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_6277_cast_fp16 = slice_by_index(begin = var_6277_begin_0, end = var_6277_end_0, end_mask = var_6277_end_mask_0, x = q_49_cast_fp16)[name = string("op_6277_cast_fp16")];
            int32 var_6279 = const()[name = string("op_6279"), val = int32(-1)];
            bool var_6280_interleave_0 = const()[name = string("op_6280_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_6280_cast_fp16 = concat(axis = var_6279, interleave = var_6280_interleave_0, values = (var_6272_cast_fp16, var_6277_cast_fp16))[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6281_cast_fp16 = mul(x = var_6280_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6281_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_51_cast_fp16 = add(x = var_6266_cast_fp16, y = var_6281_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6284_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6284_cast_fp16")];
            tensor<int32, [4]> var_6289_begin_0 = const()[name = string("op_6289_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6289_end_0 = const()[name = string("op_6289_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_6289_end_mask_0 = const()[name = string("op_6289_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_6289_cast_fp16 = slice_by_index(begin = var_6289_begin_0, end = var_6289_end_0, end_mask = var_6289_end_mask_0, x = k_73_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_6290_cast_fp16 = mul(x = var_6289_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_6290_cast_fp16")];
            tensor<int32, [4]> var_6295_begin_0 = const()[name = string("op_6295_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6295_end_0 = const()[name = string("op_6295_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_6295_end_mask_0 = const()[name = string("op_6295_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_6295_cast_fp16 = slice_by_index(begin = var_6295_begin_0, end = var_6295_end_0, end_mask = var_6295_end_mask_0, x = k_73_cast_fp16)[name = string("op_6295_cast_fp16")];
            int32 var_6297 = const()[name = string("op_6297"), val = int32(-1)];
            bool var_6298_interleave_0 = const()[name = string("op_6298_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_6298_cast_fp16 = concat(axis = var_6297, interleave = var_6298_interleave_0, values = (var_6290_cast_fp16, var_6295_cast_fp16))[name = string("op_6298_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6299_cast_fp16 = mul(x = var_6298_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6299_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_75_cast_fp16 = add(x = var_6284_cast_fp16, y = var_6299_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_24_reps_0 = const()[name = string("tile_24_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = k_75_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [4, 1, 128, 256]> tile_24_cast_fp16 = tile(reps = tile_24_reps_0, x = transpose_48_cast_fp16)[name = string("tile_24_cast_fp16")];
            tensor<int32, [5]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_48_cast_fp16 = reshape(shape = concat_48, x = tile_24_cast_fp16)[name = string("reshape_48_cast_fp16")];
            tensor<int32, [5]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = reshape_48_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [4, 1, 128, 256]> reshape_49_cast_fp16 = reshape(shape = concat_49, x = transpose_49_cast_fp16)[name = string("reshape_49_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_25_reps_0 = const()[name = string("tile_25_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = v_49_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [4, 1, 128, 256]> tile_25_cast_fp16 = tile(reps = tile_25_reps_0, x = transpose_50_cast_fp16)[name = string("tile_25_cast_fp16")];
            tensor<int32, [5]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_50_cast_fp16 = reshape(shape = concat_50, x = tile_25_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [5]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [4, 1, 128, 256]> reshape_51_cast_fp16 = reshape(shape = concat_51, x = transpose_51_cast_fp16)[name = string("reshape_51_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6313_transpose_x_0 = const()[name = string("op_6313_transpose_x_0"), val = bool(false)];
            bool var_6313_transpose_y_0 = const()[name = string("op_6313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = reshape_49_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 4, 128, 128]> var_6313_cast_fp16 = matmul(transpose_x = var_6313_transpose_x_0, transpose_y = var_6313_transpose_y_0, x = q_51_cast_fp16, y = transpose_108_cast_fp16)[name = string("op_6313_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_6315_cast_fp16 = add(x = var_6313_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6315_cast_fp16")];
            int32 var_6316 = const()[name = string("op_6316"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_201_cast_fp16 = softmax(axis = var_6316, x = var_6315_cast_fp16)[name = string("w_201_cast_fp16")];
            bool var_6319_transpose_x_0 = const()[name = string("op_6319_transpose_x_0"), val = bool(false)];
            bool var_6319_transpose_y_0 = const()[name = string("op_6319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = reshape_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 4, 128, 256]> var_6319_cast_fp16 = matmul(transpose_x = var_6319_transpose_x_0, transpose_y = var_6319_transpose_y_0, x = w_201_cast_fp16, y = v_51_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<int32, [4]> var_6322_perm_0 = const()[name = string("op_6322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6326 = const()[name = string("op_6326"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_6322_cast_fp16 = transpose(perm = var_6322_perm_0, x = var_6319_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 128, 1024]> input_147_cast_fp16 = reshape(shape = var_6326, x = var_6322_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72710208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73234560))))[name = string("tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6331_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6331_cast_fp16")];
            bool var_6334 = const()[name = string("op_6334"), val = bool(true)];
            tensor<int32, [1]> var_6335_axes_0 = const()[name = string("op_6335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6335_cast_fp16 = reduce_max(axes = var_6335_axes_0, keep_dims = var_6334, x = var_6331_cast_fp16)[name = string("op_6335_cast_fp16")];
            fp16 var_6336_to_fp16 = const()[name = string("op_6336_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_203_cast_fp16 = maximum(x = var_6335_cast_fp16, y = var_6336_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_203_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_203_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6339_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6339_cast_fp16")];
            tensor<int32, [1]> var_6344_axes_0 = const()[name = string("op_6344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6344_keep_dims_0 = const()[name = string("op_6344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6344_cast_fp16 = reduce_mean(axes = var_6344_axes_0, keep_dims = var_6344_keep_dims_0, x = var_6339_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6345_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6345_cast_fp16")];
            fp32 var_6346_epsilon_0 = const()[name = string("op_6346_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6346_cast_fp16 = inverse(epsilon = var_6346_epsilon_0, x = var_6345_cast_fp16)[name = string("op_6346_cast_fp16")];
            fp16 var_6347_to_fp16 = const()[name = string("op_6347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6348_cast_fp16 = mul(x = var_6346_cast_fp16, y = var_6347_to_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6350_cast_fp16 = add(x = var_6344_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            fp32 var_6351_epsilon_0 = const()[name = string("op_6351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6351_cast_fp16 = rsqrt(epsilon = var_6351_epsilon_0, x = var_6350_cast_fp16)[name = string("op_6351_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_177_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6351_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73235648)))];
            tensor<fp16, [1, 128, 512]> var_6353_cast_fp16 = mul(x = out_177_cast_fp16, y = tm_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = var_6353_cast_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6356_cast_fp16 = abs(x = x_227_cast_fp16)[name = string("op_6356_cast_fp16")];
            bool var_6359 = const()[name = string("op_6359"), val = bool(true)];
            tensor<int32, [1]> var_6360_axes_0 = const()[name = string("op_6360_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6360_cast_fp16 = reduce_max(axes = var_6360_axes_0, keep_dims = var_6359, x = var_6356_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_205_cast_fp16 = maximum(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_205_cast_fp16 = real_div(x = x_227_cast_fp16, y = s_205_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6364_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<int32, [1]> var_6369_axes_0 = const()[name = string("op_6369_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6369_keep_dims_0 = const()[name = string("op_6369_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6369_cast_fp16 = reduce_mean(axes = var_6369_axes_0, keep_dims = var_6369_keep_dims_0, x = var_6364_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6370_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6370_cast_fp16")];
            fp32 var_6371_epsilon_0 = const()[name = string("op_6371_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6371_cast_fp16 = inverse(epsilon = var_6371_epsilon_0, x = var_6370_cast_fp16)[name = string("op_6371_cast_fp16")];
            fp16 var_6372_to_fp16 = const()[name = string("op_6372_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6373_cast_fp16 = mul(x = var_6371_cast_fp16, y = var_6372_to_fp16)[name = string("op_6373_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6375_cast_fp16 = add(x = var_6369_cast_fp16, y = var_6373_cast_fp16)[name = string("op_6375_cast_fp16")];
            fp32 var_6376_epsilon_0 = const()[name = string("op_6376_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6376_cast_fp16 = rsqrt(epsilon = var_6376_epsilon_0, x = var_6375_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_179_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6376_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73236736)))];
            tensor<fp16, [1, 128, 512]> input_149_cast_fp16 = mul(x = out_179_cast_fp16, y = tm_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73237824))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74286464))))[name = string("tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_113_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_113_cast_fp16")];
            string var_6383_mode_0 = const()[name = string("op_6383_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_6383_cast_fp16 = gelu(mode = var_6383_mode_0, x = linear_113_cast_fp16)[name = string("op_6383_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74290624))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339264))))[name = string("tm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_114_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_151_cast_fp16 = mul(x = var_6383_cast_fp16, y = linear_114_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75343424))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76392064))))[name = string("tm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6391_cast_fp16 = abs(x = linear_115_cast_fp16)[name = string("op_6391_cast_fp16")];
            bool var_6394 = const()[name = string("op_6394"), val = bool(true)];
            tensor<int32, [1]> var_6395_axes_0 = const()[name = string("op_6395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6395_cast_fp16 = reduce_max(axes = var_6395_axes_0, keep_dims = var_6394, x = var_6391_cast_fp16)[name = string("op_6395_cast_fp16")];
            fp16 var_6396_to_fp16 = const()[name = string("op_6396_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_207_cast_fp16 = maximum(x = var_6395_cast_fp16, y = var_6396_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_207_cast_fp16 = real_div(x = linear_115_cast_fp16, y = s_207_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6399_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6399_cast_fp16")];
            tensor<int32, [1]> var_6404_axes_0 = const()[name = string("op_6404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6404_keep_dims_0 = const()[name = string("op_6404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6404_cast_fp16 = reduce_mean(axes = var_6404_axes_0, keep_dims = var_6404_keep_dims_0, x = var_6399_cast_fp16)[name = string("op_6404_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6405_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6405_cast_fp16")];
            fp32 var_6406_epsilon_0 = const()[name = string("op_6406_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6406_cast_fp16 = inverse(epsilon = var_6406_epsilon_0, x = var_6405_cast_fp16)[name = string("op_6406_cast_fp16")];
            fp16 var_6407_to_fp16 = const()[name = string("op_6407_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6408_cast_fp16 = mul(x = var_6406_cast_fp16, y = var_6407_to_fp16)[name = string("op_6408_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6410_cast_fp16 = add(x = var_6404_cast_fp16, y = var_6408_cast_fp16)[name = string("op_6410_cast_fp16")];
            fp32 var_6411_epsilon_0 = const()[name = string("op_6411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6411_cast_fp16 = rsqrt(epsilon = var_6411_epsilon_0, x = var_6410_cast_fp16)[name = string("op_6411_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_181_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6411_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76393152)))];
            tensor<fp16, [1, 128, 512]> var_6413_cast_fp16 = mul(x = out_181_cast_fp16, y = tm_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6413_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_153_cast_fp16 = add(x = x_227_cast_fp16, y = var_6413_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76394240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76656448))))[name = string("tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_116_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_116_cast_fp16")];
            string x_231_mode_0 = const()[name = string("x_231_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_231_cast_fp16 = gelu(mode = x_231_mode_0, x = linear_116_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_6433_begin_0 = const()[name = string("op_6433_begin_0"), val = tensor<int32, [4]>([0, 0, 12, 0])];
            tensor<int32, [4]> var_6433_end_0 = const()[name = string("op_6433_end_0"), val = tensor<int32, [4]>([1, 128, 13, 512])];
            tensor<bool, [4]> var_6433_end_mask_0 = const()[name = string("op_6433_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6433_squeeze_mask_0 = const()[name = string("op_6433_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_6433_cast_fp16 = slice_by_index(begin = var_6433_begin_0, end = var_6433_end_0, end_mask = var_6433_end_mask_0, squeeze_mask = var_6433_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6433_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_155_cast_fp16 = mul(x = x_231_cast_fp16, y = var_6433_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76657536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76919744))))[name = string("tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6443_cast_fp16 = abs(x = linear_117_cast_fp16)[name = string("op_6443_cast_fp16")];
            bool var_6446 = const()[name = string("op_6446"), val = bool(true)];
            tensor<int32, [1]> var_6447_axes_0 = const()[name = string("op_6447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6447_cast_fp16 = reduce_max(axes = var_6447_axes_0, keep_dims = var_6446, x = var_6443_cast_fp16)[name = string("op_6447_cast_fp16")];
            fp16 var_6448_to_fp16 = const()[name = string("op_6448_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_209_cast_fp16 = maximum(x = var_6447_cast_fp16, y = var_6448_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_209_cast_fp16 = real_div(x = linear_117_cast_fp16, y = s_209_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6451_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6451_cast_fp16")];
            tensor<int32, [1]> var_6456_axes_0 = const()[name = string("op_6456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6456_keep_dims_0 = const()[name = string("op_6456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6456_cast_fp16 = reduce_mean(axes = var_6456_axes_0, keep_dims = var_6456_keep_dims_0, x = var_6451_cast_fp16)[name = string("op_6456_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6457_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6457_cast_fp16")];
            fp32 var_6458_epsilon_0 = const()[name = string("op_6458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6458_cast_fp16 = inverse(epsilon = var_6458_epsilon_0, x = var_6457_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6459_to_fp16 = const()[name = string("op_6459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6460_cast_fp16 = mul(x = var_6458_cast_fp16, y = var_6459_to_fp16)[name = string("op_6460_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6462_cast_fp16 = add(x = var_6456_cast_fp16, y = var_6460_cast_fp16)[name = string("op_6462_cast_fp16")];
            fp32 var_6463_epsilon_0 = const()[name = string("op_6463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6463_cast_fp16 = rsqrt(epsilon = var_6463_epsilon_0, x = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_183_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6463_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76920832)))];
            tensor<fp16, [1, 128, 512]> var_6465_cast_fp16 = mul(x = out_183_cast_fp16, y = tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6465_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_25_cast_fp16 = add(x = input_153_cast_fp16, y = var_6465_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<fp16, [1]> tm_layers_12_layer_scalar_to_fp16 = const()[name = string("tm_layers_12_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.8cp-1])];
            tensor<fp16, [1, 128, 512]> x_235_cast_fp16 = mul(x = h_25_cast_fp16, y = tm_layers_12_layer_scalar_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6469_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_6469_cast_fp16")];
            bool var_6472 = const()[name = string("op_6472"), val = bool(true)];
            tensor<int32, [1]> var_6473_axes_0 = const()[name = string("op_6473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6473_cast_fp16 = reduce_max(axes = var_6473_axes_0, keep_dims = var_6472, x = var_6469_cast_fp16)[name = string("op_6473_cast_fp16")];
            fp16 var_6474_to_fp16 = const()[name = string("op_6474_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_211_cast_fp16 = maximum(x = var_6473_cast_fp16, y = var_6474_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_211_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_211_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6477_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6477_cast_fp16")];
            tensor<int32, [1]> var_6482_axes_0 = const()[name = string("op_6482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6482_keep_dims_0 = const()[name = string("op_6482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6482_cast_fp16 = reduce_mean(axes = var_6482_axes_0, keep_dims = var_6482_keep_dims_0, x = var_6477_cast_fp16)[name = string("op_6482_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6483_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6483_cast_fp16")];
            fp32 var_6484_epsilon_0 = const()[name = string("op_6484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6484_cast_fp16 = inverse(epsilon = var_6484_epsilon_0, x = var_6483_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp16 var_6485_to_fp16 = const()[name = string("op_6485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6486_cast_fp16 = mul(x = var_6484_cast_fp16, y = var_6485_to_fp16)[name = string("op_6486_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6488_cast_fp16 = add(x = var_6482_cast_fp16, y = var_6486_cast_fp16)[name = string("op_6488_cast_fp16")];
            fp32 var_6489_epsilon_0 = const()[name = string("op_6489_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6489_cast_fp16 = rsqrt(epsilon = var_6489_epsilon_0, x = var_6488_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_185_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6489_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76921920)))];
            tensor<fp16, [1, 128, 512]> input_157_cast_fp16 = mul(x = out_185_cast_fp16, y = tm_layers_13_input_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76923008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77447360))))[name = string("tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<int32, [4]> var_6499 = const()[name = string("op_6499"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_6500_cast_fp16 = reshape(shape = var_6499, x = linear_118_cast_fp16)[name = string("op_6500_cast_fp16")];
            tensor<int32, [4]> x_237_perm_0 = const()[name = string("x_237_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77449472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77711680))))[name = string("tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6511 = const()[name = string("op_6511"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_6512_cast_fp16 = reshape(shape = var_6511, x = linear_119_cast_fp16)[name = string("op_6512_cast_fp16")];
            tensor<int32, [4]> x_239_perm_0 = const()[name = string("x_239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77712768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77974976))))[name = string("tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6523 = const()[name = string("op_6523"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_6524_cast_fp16 = reshape(shape = var_6523, x = linear_120_cast_fp16)[name = string("op_6524_cast_fp16")];
            tensor<int32, [4]> x_241_perm_0 = const()[name = string("x_241_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_237_cast_fp16 = transpose(perm = x_237_perm_0, x = var_6500_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 4, 128, 256]> var_6528_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_6528_cast_fp16")];
            bool var_6531 = const()[name = string("op_6531"), val = bool(true)];
            tensor<int32, [1]> var_6532_axes_0 = const()[name = string("op_6532_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_6532_cast_fp16 = reduce_max(axes = var_6532_axes_0, keep_dims = var_6531, x = var_6528_cast_fp16)[name = string("op_6532_cast_fp16")];
            fp16 var_6533_to_fp16 = const()[name = string("op_6533_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_213_cast_fp16 = maximum(x = var_6532_cast_fp16, y = var_6533_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_213_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_213_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6536_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6536_cast_fp16")];
            tensor<int32, [1]> var_6541_axes_0 = const()[name = string("op_6541_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6541_keep_dims_0 = const()[name = string("op_6541_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_6541_cast_fp16 = reduce_mean(axes = var_6541_axes_0, keep_dims = var_6541_keep_dims_0, x = var_6536_cast_fp16)[name = string("op_6541_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_6542_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp32 var_6543_epsilon_0 = const()[name = string("op_6543_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_6543_cast_fp16 = inverse(epsilon = var_6543_epsilon_0, x = var_6542_cast_fp16)[name = string("op_6543_cast_fp16")];
            fp16 var_6544_to_fp16 = const()[name = string("op_6544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_6545_cast_fp16 = mul(x = var_6543_cast_fp16, y = var_6544_to_fp16)[name = string("op_6545_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_6547_cast_fp16 = add(x = var_6541_cast_fp16, y = var_6545_cast_fp16)[name = string("op_6547_cast_fp16")];
            fp32 var_6548_epsilon_0 = const()[name = string("op_6548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_6548_cast_fp16 = rsqrt(epsilon = var_6548_epsilon_0, x = var_6547_cast_fp16)[name = string("op_6548_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_187_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6548_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976064)))];
            tensor<fp16, [1, 4, 128, 256]> q_53_cast_fp16 = mul(x = out_187_cast_fp16, y = tm_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("q_53_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_239_cast_fp16 = transpose(perm = x_239_perm_0, x = var_6512_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2, 128, 256]> var_6551_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_6551_cast_fp16")];
            bool var_6554 = const()[name = string("op_6554"), val = bool(true)];
            tensor<int32, [1]> var_6555_axes_0 = const()[name = string("op_6555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_6555_cast_fp16 = reduce_max(axes = var_6555_axes_0, keep_dims = var_6554, x = var_6551_cast_fp16)[name = string("op_6555_cast_fp16")];
            fp16 var_6556_to_fp16 = const()[name = string("op_6556_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_215_cast_fp16 = maximum(x = var_6555_cast_fp16, y = var_6556_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_215_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_215_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6559_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6559_cast_fp16")];
            tensor<int32, [1]> var_6564_axes_0 = const()[name = string("op_6564_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6564_keep_dims_0 = const()[name = string("op_6564_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_6564_cast_fp16 = reduce_mean(axes = var_6564_axes_0, keep_dims = var_6564_keep_dims_0, x = var_6559_cast_fp16)[name = string("op_6564_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6565_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6565_cast_fp16")];
            fp32 var_6566_epsilon_0 = const()[name = string("op_6566_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_6566_cast_fp16 = inverse(epsilon = var_6566_epsilon_0, x = var_6565_cast_fp16)[name = string("op_6566_cast_fp16")];
            fp16 var_6567_to_fp16 = const()[name = string("op_6567_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_6568_cast_fp16 = mul(x = var_6566_cast_fp16, y = var_6567_to_fp16)[name = string("op_6568_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6570_cast_fp16 = add(x = var_6564_cast_fp16, y = var_6568_cast_fp16)[name = string("op_6570_cast_fp16")];
            fp32 var_6571_epsilon_0 = const()[name = string("op_6571_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_6571_cast_fp16 = rsqrt(epsilon = var_6571_epsilon_0, x = var_6570_cast_fp16)[name = string("op_6571_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_189_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6571_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976640)))];
            tensor<fp16, [1, 2, 128, 256]> k_79_cast_fp16 = mul(x = out_189_cast_fp16, y = tm_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("k_79_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_241_cast_fp16 = transpose(perm = x_241_perm_0, x = var_6524_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 2, 128, 256]> var_6574_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_6574_cast_fp16")];
            bool var_6577 = const()[name = string("op_6577"), val = bool(true)];
            tensor<int32, [1]> var_6578_axes_0 = const()[name = string("op_6578_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_6578_cast_fp16 = reduce_max(axes = var_6578_axes_0, keep_dims = var_6577, x = var_6574_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_217_cast_fp16 = maximum(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_217_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_217_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6582_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6587_axes_0 = const()[name = string("op_6587_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6587_keep_dims_0 = const()[name = string("op_6587_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_6587_cast_fp16 = reduce_mean(axes = var_6587_axes_0, keep_dims = var_6587_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6588_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6588_cast_fp16")];
            fp32 var_6589_epsilon_0 = const()[name = string("op_6589_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_6589_cast_fp16 = inverse(epsilon = var_6589_epsilon_0, x = var_6588_cast_fp16)[name = string("op_6589_cast_fp16")];
            fp16 var_6590_to_fp16 = const()[name = string("op_6590_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_6591_cast_fp16 = mul(x = var_6589_cast_fp16, y = var_6590_to_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6593_cast_fp16 = add(x = var_6587_cast_fp16, y = var_6591_cast_fp16)[name = string("op_6593_cast_fp16")];
            fp32 var_6594_epsilon_0 = const()[name = string("op_6594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_6594_cast_fp16 = rsqrt(epsilon = var_6594_epsilon_0, x = var_6593_cast_fp16)[name = string("op_6594_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_53_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6594_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6596_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6596_cast_fp16")];
            tensor<int32, [4]> var_6601_begin_0 = const()[name = string("op_6601_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6601_end_0 = const()[name = string("op_6601_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_6601_end_mask_0 = const()[name = string("op_6601_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_6601_cast_fp16 = slice_by_index(begin = var_6601_begin_0, end = var_6601_end_0, end_mask = var_6601_end_mask_0, x = q_53_cast_fp16)[name = string("op_6601_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_6602_cast_fp16 = mul(x = var_6601_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_6602_cast_fp16")];
            tensor<int32, [4]> var_6607_begin_0 = const()[name = string("op_6607_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6607_end_0 = const()[name = string("op_6607_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_6607_end_mask_0 = const()[name = string("op_6607_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_6607_cast_fp16 = slice_by_index(begin = var_6607_begin_0, end = var_6607_end_0, end_mask = var_6607_end_mask_0, x = q_53_cast_fp16)[name = string("op_6607_cast_fp16")];
            int32 var_6609 = const()[name = string("op_6609"), val = int32(-1)];
            bool var_6610_interleave_0 = const()[name = string("op_6610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_6610_cast_fp16 = concat(axis = var_6609, interleave = var_6610_interleave_0, values = (var_6602_cast_fp16, var_6607_cast_fp16))[name = string("op_6610_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6611_cast_fp16 = mul(x = var_6610_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_55_cast_fp16 = add(x = var_6596_cast_fp16, y = var_6611_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6614_cast_fp16 = mul(x = k_79_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = k_79_cast_fp16)[name = string("op_6619_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_6620_cast_fp16 = mul(x = var_6619_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_6620_cast_fp16")];
            tensor<int32, [4]> var_6625_begin_0 = const()[name = string("op_6625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6625_end_0 = const()[name = string("op_6625_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_6625_end_mask_0 = const()[name = string("op_6625_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_6625_cast_fp16 = slice_by_index(begin = var_6625_begin_0, end = var_6625_end_0, end_mask = var_6625_end_mask_0, x = k_79_cast_fp16)[name = string("op_6625_cast_fp16")];
            int32 var_6627 = const()[name = string("op_6627"), val = int32(-1)];
            bool var_6628_interleave_0 = const()[name = string("op_6628_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_6628_cast_fp16 = concat(axis = var_6627, interleave = var_6628_interleave_0, values = (var_6620_cast_fp16, var_6625_cast_fp16))[name = string("op_6628_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6629_cast_fp16 = mul(x = var_6628_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6629_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_81_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6629_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_26_reps_0 = const()[name = string("tile_26_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = k_81_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [4, 1, 128, 256]> tile_26_cast_fp16 = tile(reps = tile_26_reps_0, x = transpose_52_cast_fp16)[name = string("tile_26_cast_fp16")];
            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_52_cast_fp16 = reshape(shape = concat_52, x = tile_26_cast_fp16)[name = string("reshape_52_cast_fp16")];
            tensor<int32, [5]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = reshape_52_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [4, 1, 128, 256]> reshape_53_cast_fp16 = reshape(shape = concat_53, x = transpose_53_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_27_reps_0 = const()[name = string("tile_27_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = v_53_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [4, 1, 128, 256]> tile_27_cast_fp16 = tile(reps = tile_27_reps_0, x = transpose_54_cast_fp16)[name = string("tile_27_cast_fp16")];
            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_54_cast_fp16 = reshape(shape = concat_54, x = tile_27_cast_fp16)[name = string("reshape_54_cast_fp16")];
            tensor<int32, [5]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = reshape_54_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [4, 1, 128, 256]> reshape_55_cast_fp16 = reshape(shape = concat_55, x = transpose_55_cast_fp16)[name = string("reshape_55_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6643_transpose_x_0 = const()[name = string("op_6643_transpose_x_0"), val = bool(false)];
            bool var_6643_transpose_y_0 = const()[name = string("op_6643_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 4, 128, 128]> var_6643_cast_fp16 = matmul(transpose_x = var_6643_transpose_x_0, transpose_y = var_6643_transpose_y_0, x = q_55_cast_fp16, y = transpose_109_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_6645_cast_fp16 = add(x = var_6643_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6646 = const()[name = string("op_6646"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_217_cast_fp16 = softmax(axis = var_6646, x = var_6645_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6649_transpose_x_0 = const()[name = string("op_6649_transpose_x_0"), val = bool(false)];
            bool var_6649_transpose_y_0 = const()[name = string("op_6649_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = reshape_55_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 4, 128, 256]> var_6649_cast_fp16 = matmul(transpose_x = var_6649_transpose_x_0, transpose_y = var_6649_transpose_y_0, x = w_217_cast_fp16, y = v_55_cast_fp16)[name = string("op_6649_cast_fp16")];
            tensor<int32, [4]> var_6652_perm_0 = const()[name = string("op_6652_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6656 = const()[name = string("op_6656"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_6652_cast_fp16 = transpose(perm = var_6652_perm_0, x = var_6649_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 128, 1024]> input_159_cast_fp16 = reshape(shape = var_6656, x = var_6652_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77977216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78501568))))[name = string("tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6661_cast_fp16 = abs(x = linear_121_cast_fp16)[name = string("op_6661_cast_fp16")];
            bool var_6664 = const()[name = string("op_6664"), val = bool(true)];
            tensor<int32, [1]> var_6665_axes_0 = const()[name = string("op_6665_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6665_cast_fp16 = reduce_max(axes = var_6665_axes_0, keep_dims = var_6664, x = var_6661_cast_fp16)[name = string("op_6665_cast_fp16")];
            fp16 var_6666_to_fp16 = const()[name = string("op_6666_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_219_cast_fp16 = maximum(x = var_6665_cast_fp16, y = var_6666_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_219_cast_fp16 = real_div(x = linear_121_cast_fp16, y = s_219_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6669_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6669_cast_fp16")];
            tensor<int32, [1]> var_6674_axes_0 = const()[name = string("op_6674_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6674_keep_dims_0 = const()[name = string("op_6674_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6674_cast_fp16 = reduce_mean(axes = var_6674_axes_0, keep_dims = var_6674_keep_dims_0, x = var_6669_cast_fp16)[name = string("op_6674_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6675_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6675_cast_fp16")];
            fp32 var_6676_epsilon_0 = const()[name = string("op_6676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6676_cast_fp16 = inverse(epsilon = var_6676_epsilon_0, x = var_6675_cast_fp16)[name = string("op_6676_cast_fp16")];
            fp16 var_6677_to_fp16 = const()[name = string("op_6677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6678_cast_fp16 = mul(x = var_6676_cast_fp16, y = var_6677_to_fp16)[name = string("op_6678_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6680_cast_fp16 = add(x = var_6674_cast_fp16, y = var_6678_cast_fp16)[name = string("op_6680_cast_fp16")];
            fp32 var_6681_epsilon_0 = const()[name = string("op_6681_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6681_cast_fp16 = rsqrt(epsilon = var_6681_epsilon_0, x = var_6680_cast_fp16)[name = string("op_6681_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_191_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6681_cast_fp16)[name = string("out_191_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78502656)))];
            tensor<fp16, [1, 128, 512]> var_6683_cast_fp16 = mul(x = out_191_cast_fp16, y = tm_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6683_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_245_cast_fp16 = add(x = x_235_cast_fp16, y = var_6683_cast_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6686_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_6686_cast_fp16")];
            bool var_6689 = const()[name = string("op_6689"), val = bool(true)];
            tensor<int32, [1]> var_6690_axes_0 = const()[name = string("op_6690_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6690_cast_fp16 = reduce_max(axes = var_6690_axes_0, keep_dims = var_6689, x = var_6686_cast_fp16)[name = string("op_6690_cast_fp16")];
            fp16 var_6691_to_fp16 = const()[name = string("op_6691_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_221_cast_fp16 = maximum(x = var_6690_cast_fp16, y = var_6691_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_221_cast_fp16 = real_div(x = x_245_cast_fp16, y = s_221_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6694_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6694_cast_fp16")];
            tensor<int32, [1]> var_6699_axes_0 = const()[name = string("op_6699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6699_keep_dims_0 = const()[name = string("op_6699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6699_cast_fp16 = reduce_mean(axes = var_6699_axes_0, keep_dims = var_6699_keep_dims_0, x = var_6694_cast_fp16)[name = string("op_6699_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6700_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6700_cast_fp16")];
            fp32 var_6701_epsilon_0 = const()[name = string("op_6701_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6701_cast_fp16 = inverse(epsilon = var_6701_epsilon_0, x = var_6700_cast_fp16)[name = string("op_6701_cast_fp16")];
            fp16 var_6702_to_fp16 = const()[name = string("op_6702_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6703_cast_fp16 = mul(x = var_6701_cast_fp16, y = var_6702_to_fp16)[name = string("op_6703_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6705_cast_fp16 = add(x = var_6699_cast_fp16, y = var_6703_cast_fp16)[name = string("op_6705_cast_fp16")];
            fp32 var_6706_epsilon_0 = const()[name = string("op_6706_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6706_cast_fp16 = rsqrt(epsilon = var_6706_epsilon_0, x = var_6705_cast_fp16)[name = string("op_6706_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_193_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6706_cast_fp16)[name = string("out_193_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78503744)))];
            tensor<fp16, [1, 128, 512]> input_161_cast_fp16 = mul(x = out_193_cast_fp16, y = tm_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78504832))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79553472))))[name = string("tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_122_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_122_cast_fp16")];
            string var_6713_mode_0 = const()[name = string("op_6713_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_6713_cast_fp16 = gelu(mode = var_6713_mode_0, x = linear_122_cast_fp16)[name = string("op_6713_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79557632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80606272))))[name = string("tm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_123_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_163_cast_fp16 = mul(x = var_6713_cast_fp16, y = linear_123_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80610432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81659072))))[name = string("tm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6721_cast_fp16 = abs(x = linear_124_cast_fp16)[name = string("op_6721_cast_fp16")];
            bool var_6724 = const()[name = string("op_6724"), val = bool(true)];
            tensor<int32, [1]> var_6725_axes_0 = const()[name = string("op_6725_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6725_cast_fp16 = reduce_max(axes = var_6725_axes_0, keep_dims = var_6724, x = var_6721_cast_fp16)[name = string("op_6725_cast_fp16")];
            fp16 var_6726_to_fp16 = const()[name = string("op_6726_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_223_cast_fp16 = maximum(x = var_6725_cast_fp16, y = var_6726_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_223_cast_fp16 = real_div(x = linear_124_cast_fp16, y = s_223_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6729_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [1]> var_6734_axes_0 = const()[name = string("op_6734_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6734_keep_dims_0 = const()[name = string("op_6734_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6734_cast_fp16 = reduce_mean(axes = var_6734_axes_0, keep_dims = var_6734_keep_dims_0, x = var_6729_cast_fp16)[name = string("op_6734_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6735_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6735_cast_fp16")];
            fp32 var_6736_epsilon_0 = const()[name = string("op_6736_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6736_cast_fp16 = inverse(epsilon = var_6736_epsilon_0, x = var_6735_cast_fp16)[name = string("op_6736_cast_fp16")];
            fp16 var_6737_to_fp16 = const()[name = string("op_6737_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6738_cast_fp16 = mul(x = var_6736_cast_fp16, y = var_6737_to_fp16)[name = string("op_6738_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6740_cast_fp16 = add(x = var_6734_cast_fp16, y = var_6738_cast_fp16)[name = string("op_6740_cast_fp16")];
            fp32 var_6741_epsilon_0 = const()[name = string("op_6741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6741_cast_fp16 = rsqrt(epsilon = var_6741_epsilon_0, x = var_6740_cast_fp16)[name = string("op_6741_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_195_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6741_cast_fp16)[name = string("out_195_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81660160)))];
            tensor<fp16, [1, 128, 512]> var_6743_cast_fp16 = mul(x = out_195_cast_fp16, y = tm_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6743_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_165_cast_fp16 = add(x = x_245_cast_fp16, y = var_6743_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81661248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81923456))))[name = string("tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_125_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_125_cast_fp16")];
            string x_249_mode_0 = const()[name = string("x_249_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_249_cast_fp16 = gelu(mode = x_249_mode_0, x = linear_125_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<int32, [4]> var_6763_begin_0 = const()[name = string("op_6763_begin_0"), val = tensor<int32, [4]>([0, 0, 13, 0])];
            tensor<int32, [4]> var_6763_end_0 = const()[name = string("op_6763_end_0"), val = tensor<int32, [4]>([1, 128, 14, 512])];
            tensor<bool, [4]> var_6763_end_mask_0 = const()[name = string("op_6763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6763_squeeze_mask_0 = const()[name = string("op_6763_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_6763_cast_fp16 = slice_by_index(begin = var_6763_begin_0, end = var_6763_end_0, end_mask = var_6763_end_mask_0, squeeze_mask = var_6763_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_167_cast_fp16 = mul(x = x_249_cast_fp16, y = var_6763_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81924544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82186752))))[name = string("tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6773_cast_fp16 = abs(x = linear_126_cast_fp16)[name = string("op_6773_cast_fp16")];
            bool var_6776 = const()[name = string("op_6776"), val = bool(true)];
            tensor<int32, [1]> var_6777_axes_0 = const()[name = string("op_6777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6777_cast_fp16 = reduce_max(axes = var_6777_axes_0, keep_dims = var_6776, x = var_6773_cast_fp16)[name = string("op_6777_cast_fp16")];
            fp16 var_6778_to_fp16 = const()[name = string("op_6778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_225_cast_fp16 = maximum(x = var_6777_cast_fp16, y = var_6778_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_225_cast_fp16 = real_div(x = linear_126_cast_fp16, y = s_225_cast_fp16)[name = string("y_225_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6781_cast_fp16 = mul(x = y_225_cast_fp16, y = y_225_cast_fp16)[name = string("op_6781_cast_fp16")];
            tensor<int32, [1]> var_6786_axes_0 = const()[name = string("op_6786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6786_keep_dims_0 = const()[name = string("op_6786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6786_cast_fp16 = reduce_mean(axes = var_6786_axes_0, keep_dims = var_6786_keep_dims_0, x = var_6781_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6787_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6787_cast_fp16")];
            fp32 var_6788_epsilon_0 = const()[name = string("op_6788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6788_cast_fp16 = inverse(epsilon = var_6788_epsilon_0, x = var_6787_cast_fp16)[name = string("op_6788_cast_fp16")];
            fp16 var_6789_to_fp16 = const()[name = string("op_6789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6790_cast_fp16 = mul(x = var_6788_cast_fp16, y = var_6789_to_fp16)[name = string("op_6790_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6792_cast_fp16 = add(x = var_6786_cast_fp16, y = var_6790_cast_fp16)[name = string("op_6792_cast_fp16")];
            fp32 var_6793_epsilon_0 = const()[name = string("op_6793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6793_cast_fp16 = rsqrt(epsilon = var_6793_epsilon_0, x = var_6792_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_197_cast_fp16 = mul(x = y_225_cast_fp16, y = var_6793_cast_fp16)[name = string("out_197_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82187840)))];
            tensor<fp16, [1, 128, 512]> var_6795_cast_fp16 = mul(x = out_197_cast_fp16, y = tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6795_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_27_cast_fp16 = add(x = input_165_cast_fp16, y = var_6795_cast_fp16)[name = string("h_27_cast_fp16")];
            tensor<fp16, [1]> tm_layers_13_layer_scalar_to_fp16 = const()[name = string("tm_layers_13_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9cp-1])];
            tensor<fp16, [1, 128, 512]> x_253_cast_fp16 = mul(x = h_27_cast_fp16, y = tm_layers_13_layer_scalar_to_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6799_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6799_cast_fp16")];
            bool var_6802 = const()[name = string("op_6802"), val = bool(true)];
            tensor<int32, [1]> var_6803_axes_0 = const()[name = string("op_6803_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6803_cast_fp16 = reduce_max(axes = var_6803_axes_0, keep_dims = var_6802, x = var_6799_cast_fp16)[name = string("op_6803_cast_fp16")];
            fp16 var_6804_to_fp16 = const()[name = string("op_6804_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_227_cast_fp16 = maximum(x = var_6803_cast_fp16, y = var_6804_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_227_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_227_cast_fp16)[name = string("y_227_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6807_cast_fp16 = mul(x = y_227_cast_fp16, y = y_227_cast_fp16)[name = string("op_6807_cast_fp16")];
            tensor<int32, [1]> var_6812_axes_0 = const()[name = string("op_6812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6812_keep_dims_0 = const()[name = string("op_6812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_6812_cast_fp16 = reduce_mean(axes = var_6812_axes_0, keep_dims = var_6812_keep_dims_0, x = var_6807_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6813_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6813_cast_fp16")];
            fp32 var_6814_epsilon_0 = const()[name = string("op_6814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_6814_cast_fp16 = inverse(epsilon = var_6814_epsilon_0, x = var_6813_cast_fp16)[name = string("op_6814_cast_fp16")];
            fp16 var_6815_to_fp16 = const()[name = string("op_6815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_6816_cast_fp16 = mul(x = var_6814_cast_fp16, y = var_6815_to_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_6818_cast_fp16 = add(x = var_6812_cast_fp16, y = var_6816_cast_fp16)[name = string("op_6818_cast_fp16")];
            fp32 var_6819_epsilon_0 = const()[name = string("op_6819_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_6819_cast_fp16 = rsqrt(epsilon = var_6819_epsilon_0, x = var_6818_cast_fp16)[name = string("op_6819_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_199_cast_fp16 = mul(x = y_227_cast_fp16, y = var_6819_cast_fp16)[name = string("out_199_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82188928)))];
            tensor<fp16, [1, 128, 512]> input_169_cast_fp16 = mul(x = out_199_cast_fp16, y = tm_layers_14_input_layernorm_weight_to_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82190016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82714368))))[name = string("tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6829 = const()[name = string("op_6829"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_6830_cast_fp16 = reshape(shape = var_6829, x = linear_127_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82716480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82978688))))[name = string("tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6841 = const()[name = string("op_6841"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_6842_cast_fp16 = reshape(shape = var_6841, x = linear_128_cast_fp16)[name = string("op_6842_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82979776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83241984))))[name = string("tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_6853 = const()[name = string("op_6853"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_6854_cast_fp16 = reshape(shape = var_6853, x = linear_129_cast_fp16)[name = string("op_6854_cast_fp16")];
            tensor<int32, [4]> x_259_perm_0 = const()[name = string("x_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_6830_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 4, 128, 256]> var_6858_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6858_cast_fp16")];
            bool var_6861 = const()[name = string("op_6861"), val = bool(true)];
            tensor<int32, [1]> var_6862_axes_0 = const()[name = string("op_6862_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_6862_cast_fp16 = reduce_max(axes = var_6862_axes_0, keep_dims = var_6861, x = var_6858_cast_fp16)[name = string("op_6862_cast_fp16")];
            fp16 var_6863_to_fp16 = const()[name = string("op_6863_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_229_cast_fp16 = maximum(x = var_6862_cast_fp16, y = var_6863_to_fp16)[name = string("s_229_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_229_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_229_cast_fp16)[name = string("y_229_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6866_cast_fp16 = mul(x = y_229_cast_fp16, y = y_229_cast_fp16)[name = string("op_6866_cast_fp16")];
            tensor<int32, [1]> var_6871_axes_0 = const()[name = string("op_6871_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6871_keep_dims_0 = const()[name = string("op_6871_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_6871_cast_fp16 = reduce_mean(axes = var_6871_axes_0, keep_dims = var_6871_keep_dims_0, x = var_6866_cast_fp16)[name = string("op_6871_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_6872_cast_fp16 = mul(x = s_229_cast_fp16, y = s_229_cast_fp16)[name = string("op_6872_cast_fp16")];
            fp32 var_6873_epsilon_0 = const()[name = string("op_6873_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_6873_cast_fp16 = inverse(epsilon = var_6873_epsilon_0, x = var_6872_cast_fp16)[name = string("op_6873_cast_fp16")];
            fp16 var_6874_to_fp16 = const()[name = string("op_6874_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_6875_cast_fp16 = mul(x = var_6873_cast_fp16, y = var_6874_to_fp16)[name = string("op_6875_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_6877_cast_fp16 = add(x = var_6871_cast_fp16, y = var_6875_cast_fp16)[name = string("op_6877_cast_fp16")];
            fp32 var_6878_epsilon_0 = const()[name = string("op_6878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_6878_cast_fp16 = rsqrt(epsilon = var_6878_epsilon_0, x = var_6877_cast_fp16)[name = string("op_6878_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_201_cast_fp16 = mul(x = y_229_cast_fp16, y = var_6878_cast_fp16)[name = string("out_201_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243072)))];
            tensor<fp16, [1, 4, 128, 256]> q_57_cast_fp16 = mul(x = out_201_cast_fp16, y = tm_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("q_57_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_6842_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, 128, 256]> var_6881_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6884 = const()[name = string("op_6884"), val = bool(true)];
            tensor<int32, [1]> var_6885_axes_0 = const()[name = string("op_6885_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_6885_cast_fp16 = reduce_max(axes = var_6885_axes_0, keep_dims = var_6884, x = var_6881_cast_fp16)[name = string("op_6885_cast_fp16")];
            fp16 var_6886_to_fp16 = const()[name = string("op_6886_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_231_cast_fp16 = maximum(x = var_6885_cast_fp16, y = var_6886_to_fp16)[name = string("s_231_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_231_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_231_cast_fp16)[name = string("y_231_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6889_cast_fp16 = mul(x = y_231_cast_fp16, y = y_231_cast_fp16)[name = string("op_6889_cast_fp16")];
            tensor<int32, [1]> var_6894_axes_0 = const()[name = string("op_6894_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6894_keep_dims_0 = const()[name = string("op_6894_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_6894_cast_fp16 = reduce_mean(axes = var_6894_axes_0, keep_dims = var_6894_keep_dims_0, x = var_6889_cast_fp16)[name = string("op_6894_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6895_cast_fp16 = mul(x = s_231_cast_fp16, y = s_231_cast_fp16)[name = string("op_6895_cast_fp16")];
            fp32 var_6896_epsilon_0 = const()[name = string("op_6896_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_6896_cast_fp16 = inverse(epsilon = var_6896_epsilon_0, x = var_6895_cast_fp16)[name = string("op_6896_cast_fp16")];
            fp16 var_6897_to_fp16 = const()[name = string("op_6897_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_6898_cast_fp16 = mul(x = var_6896_cast_fp16, y = var_6897_to_fp16)[name = string("op_6898_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6900_cast_fp16 = add(x = var_6894_cast_fp16, y = var_6898_cast_fp16)[name = string("op_6900_cast_fp16")];
            fp32 var_6901_epsilon_0 = const()[name = string("op_6901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_6901_cast_fp16 = rsqrt(epsilon = var_6901_epsilon_0, x = var_6900_cast_fp16)[name = string("op_6901_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_203_cast_fp16 = mul(x = y_231_cast_fp16, y = var_6901_cast_fp16)[name = string("out_203_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243648)))];
            tensor<fp16, [1, 2, 128, 256]> k_85_cast_fp16 = mul(x = out_203_cast_fp16, y = tm_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("k_85_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_259_cast_fp16 = transpose(perm = x_259_perm_0, x = var_6854_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 2, 128, 256]> var_6904_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6904_cast_fp16")];
            bool var_6907 = const()[name = string("op_6907"), val = bool(true)];
            tensor<int32, [1]> var_6908_axes_0 = const()[name = string("op_6908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_6908_cast_fp16 = reduce_max(axes = var_6908_axes_0, keep_dims = var_6907, x = var_6904_cast_fp16)[name = string("op_6908_cast_fp16")];
            fp16 var_6909_to_fp16 = const()[name = string("op_6909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_233_cast_fp16 = maximum(x = var_6908_cast_fp16, y = var_6909_to_fp16)[name = string("s_233_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_233_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_233_cast_fp16)[name = string("y_233_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6912_cast_fp16 = mul(x = y_233_cast_fp16, y = y_233_cast_fp16)[name = string("op_6912_cast_fp16")];
            tensor<int32, [1]> var_6917_axes_0 = const()[name = string("op_6917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6917_keep_dims_0 = const()[name = string("op_6917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_6917_cast_fp16 = reduce_mean(axes = var_6917_axes_0, keep_dims = var_6917_keep_dims_0, x = var_6912_cast_fp16)[name = string("op_6917_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6918_cast_fp16 = mul(x = s_233_cast_fp16, y = s_233_cast_fp16)[name = string("op_6918_cast_fp16")];
            fp32 var_6919_epsilon_0 = const()[name = string("op_6919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_6919_cast_fp16 = inverse(epsilon = var_6919_epsilon_0, x = var_6918_cast_fp16)[name = string("op_6919_cast_fp16")];
            fp16 var_6920_to_fp16 = const()[name = string("op_6920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_6921_cast_fp16 = mul(x = var_6919_cast_fp16, y = var_6920_to_fp16)[name = string("op_6921_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_6923_cast_fp16 = add(x = var_6917_cast_fp16, y = var_6921_cast_fp16)[name = string("op_6923_cast_fp16")];
            fp32 var_6924_epsilon_0 = const()[name = string("op_6924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_6924_cast_fp16 = rsqrt(epsilon = var_6924_epsilon_0, x = var_6923_cast_fp16)[name = string("op_6924_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_57_cast_fp16 = mul(x = y_233_cast_fp16, y = var_6924_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6926_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6926_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = q_57_cast_fp16)[name = string("op_6931_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_6932_cast_fp16 = mul(x = var_6931_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_6932_cast_fp16")];
            tensor<int32, [4]> var_6937_begin_0 = const()[name = string("op_6937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6937_end_0 = const()[name = string("op_6937_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_6937_end_mask_0 = const()[name = string("op_6937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_6937_cast_fp16 = slice_by_index(begin = var_6937_begin_0, end = var_6937_end_0, end_mask = var_6937_end_mask_0, x = q_57_cast_fp16)[name = string("op_6937_cast_fp16")];
            int32 var_6939 = const()[name = string("op_6939"), val = int32(-1)];
            bool var_6940_interleave_0 = const()[name = string("op_6940_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_6940_cast_fp16 = concat(axis = var_6939, interleave = var_6940_interleave_0, values = (var_6932_cast_fp16, var_6937_cast_fp16))[name = string("op_6940_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_6941_cast_fp16 = mul(x = var_6940_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6941_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_59_cast_fp16 = add(x = var_6926_cast_fp16, y = var_6941_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6944_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6944_cast_fp16")];
            tensor<int32, [4]> var_6949_begin_0 = const()[name = string("op_6949_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6949_end_0 = const()[name = string("op_6949_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_6949_end_mask_0 = const()[name = string("op_6949_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_6949_cast_fp16 = slice_by_index(begin = var_6949_begin_0, end = var_6949_end_0, end_mask = var_6949_end_mask_0, x = k_85_cast_fp16)[name = string("op_6949_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_6950_cast_fp16 = mul(x = var_6949_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_6950_cast_fp16")];
            tensor<int32, [4]> var_6955_begin_0 = const()[name = string("op_6955_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6955_end_0 = const()[name = string("op_6955_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_6955_end_mask_0 = const()[name = string("op_6955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_6955_cast_fp16 = slice_by_index(begin = var_6955_begin_0, end = var_6955_end_0, end_mask = var_6955_end_mask_0, x = k_85_cast_fp16)[name = string("op_6955_cast_fp16")];
            int32 var_6957 = const()[name = string("op_6957"), val = int32(-1)];
            bool var_6958_interleave_0 = const()[name = string("op_6958_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_6958_cast_fp16 = concat(axis = var_6957, interleave = var_6958_interleave_0, values = (var_6950_cast_fp16, var_6955_cast_fp16))[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_6959_cast_fp16 = mul(x = var_6958_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_87_cast_fp16 = add(x = var_6944_cast_fp16, y = var_6959_cast_fp16)[name = string("k_87_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_28_reps_0 = const()[name = string("tile_28_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = k_87_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 1, 128, 256]> tile_28_cast_fp16 = tile(reps = tile_28_reps_0, x = transpose_56_cast_fp16)[name = string("tile_28_cast_fp16")];
            tensor<int32, [5]> concat_56 = const()[name = string("concat_56"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_56_cast_fp16 = reshape(shape = concat_56, x = tile_28_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [5]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_57 = const()[name = string("concat_57"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 1, 128, 256]> reshape_57_cast_fp16 = reshape(shape = concat_57, x = transpose_57_cast_fp16)[name = string("reshape_57_cast_fp16")];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_29_reps_0 = const()[name = string("tile_29_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = v_57_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 1, 128, 256]> tile_29_cast_fp16 = tile(reps = tile_29_reps_0, x = transpose_58_cast_fp16)[name = string("tile_29_cast_fp16")];
            tensor<int32, [5]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_58_cast_fp16 = reshape(shape = concat_58, x = tile_29_cast_fp16)[name = string("reshape_58_cast_fp16")];
            tensor<int32, [5]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = reshape_58_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 1, 128, 256]> reshape_59_cast_fp16 = reshape(shape = concat_59, x = transpose_59_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6973_transpose_x_0 = const()[name = string("op_6973_transpose_x_0"), val = bool(false)];
            bool var_6973_transpose_y_0 = const()[name = string("op_6973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_110_cast_fp16 = transpose(perm = transpose_110_perm_0, x = reshape_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 4, 128, 128]> var_6973_cast_fp16 = matmul(transpose_x = var_6973_transpose_x_0, transpose_y = var_6973_transpose_y_0, x = q_59_cast_fp16, y = transpose_110_cast_fp16)[name = string("op_6973_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_6975_cast_fp16 = add(x = var_6973_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6976 = const()[name = string("op_6976"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_233_cast_fp16 = softmax(axis = var_6976, x = var_6975_cast_fp16)[name = string("w_233_cast_fp16")];
            bool var_6979_transpose_x_0 = const()[name = string("op_6979_transpose_x_0"), val = bool(false)];
            bool var_6979_transpose_y_0 = const()[name = string("op_6979_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 4, 128, 256]> var_6979_cast_fp16 = matmul(transpose_x = var_6979_transpose_x_0, transpose_y = var_6979_transpose_y_0, x = w_233_cast_fp16, y = v_59_cast_fp16)[name = string("op_6979_cast_fp16")];
            tensor<int32, [4]> var_6982_perm_0 = const()[name = string("op_6982_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6986 = const()[name = string("op_6986"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_6982_cast_fp16 = transpose(perm = var_6982_perm_0, x = var_6979_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 128, 1024]> input_171_cast_fp16 = reshape(shape = var_6986, x = var_6982_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83244224))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83768576))))[name = string("tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6991_cast_fp16 = abs(x = linear_130_cast_fp16)[name = string("op_6991_cast_fp16")];
            bool var_6994 = const()[name = string("op_6994"), val = bool(true)];
            tensor<int32, [1]> var_6995_axes_0 = const()[name = string("op_6995_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_6995_cast_fp16 = reduce_max(axes = var_6995_axes_0, keep_dims = var_6994, x = var_6991_cast_fp16)[name = string("op_6995_cast_fp16")];
            fp16 var_6996_to_fp16 = const()[name = string("op_6996_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_235_cast_fp16 = maximum(x = var_6995_cast_fp16, y = var_6996_to_fp16)[name = string("s_235_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_235_cast_fp16 = real_div(x = linear_130_cast_fp16, y = s_235_cast_fp16)[name = string("y_235_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_6999_cast_fp16 = mul(x = y_235_cast_fp16, y = y_235_cast_fp16)[name = string("op_6999_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_6999_cast_fp16)[name = string("op_7004_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7005_cast_fp16 = mul(x = s_235_cast_fp16, y = s_235_cast_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7010_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7010_cast_fp16")];
            fp32 var_7011_epsilon_0 = const()[name = string("op_7011_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7011_cast_fp16 = rsqrt(epsilon = var_7011_epsilon_0, x = var_7010_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_205_cast_fp16 = mul(x = y_235_cast_fp16, y = var_7011_cast_fp16)[name = string("out_205_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83769664)))];
            tensor<fp16, [1, 128, 512]> var_7013_cast_fp16 = mul(x = out_205_cast_fp16, y = tm_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_7013_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = var_7013_cast_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7016_cast_fp16 = abs(x = x_263_cast_fp16)[name = string("op_7016_cast_fp16")];
            bool var_7019 = const()[name = string("op_7019"), val = bool(true)];
            tensor<int32, [1]> var_7020_axes_0 = const()[name = string("op_7020_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7020_cast_fp16 = reduce_max(axes = var_7020_axes_0, keep_dims = var_7019, x = var_7016_cast_fp16)[name = string("op_7020_cast_fp16")];
            fp16 var_7021_to_fp16 = const()[name = string("op_7021_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_237_cast_fp16 = maximum(x = var_7020_cast_fp16, y = var_7021_to_fp16)[name = string("s_237_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_237_cast_fp16 = real_div(x = x_263_cast_fp16, y = s_237_cast_fp16)[name = string("y_237_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7024_cast_fp16 = mul(x = y_237_cast_fp16, y = y_237_cast_fp16)[name = string("op_7024_cast_fp16")];
            tensor<int32, [1]> var_7029_axes_0 = const()[name = string("op_7029_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7029_keep_dims_0 = const()[name = string("op_7029_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7029_cast_fp16 = reduce_mean(axes = var_7029_axes_0, keep_dims = var_7029_keep_dims_0, x = var_7024_cast_fp16)[name = string("op_7029_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7030_cast_fp16 = mul(x = s_237_cast_fp16, y = s_237_cast_fp16)[name = string("op_7030_cast_fp16")];
            fp32 var_7031_epsilon_0 = const()[name = string("op_7031_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7031_cast_fp16 = inverse(epsilon = var_7031_epsilon_0, x = var_7030_cast_fp16)[name = string("op_7031_cast_fp16")];
            fp16 var_7032_to_fp16 = const()[name = string("op_7032_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7033_cast_fp16 = mul(x = var_7031_cast_fp16, y = var_7032_to_fp16)[name = string("op_7033_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7035_cast_fp16 = add(x = var_7029_cast_fp16, y = var_7033_cast_fp16)[name = string("op_7035_cast_fp16")];
            fp32 var_7036_epsilon_0 = const()[name = string("op_7036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7036_cast_fp16 = rsqrt(epsilon = var_7036_epsilon_0, x = var_7035_cast_fp16)[name = string("op_7036_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_207_cast_fp16 = mul(x = y_237_cast_fp16, y = var_7036_cast_fp16)[name = string("out_207_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83770752)))];
            tensor<fp16, [1, 128, 512]> input_173_cast_fp16 = mul(x = out_207_cast_fp16, y = tm_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83771840))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84820480))))[name = string("tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_131_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_131_cast_fp16")];
            string var_7043_mode_0 = const()[name = string("op_7043_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_7043_cast_fp16 = gelu(mode = var_7043_mode_0, x = linear_131_cast_fp16)[name = string("op_7043_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84824640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85873280))))[name = string("tm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_132_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_175_cast_fp16 = mul(x = var_7043_cast_fp16, y = linear_132_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85877440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86926080))))[name = string("tm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7051_cast_fp16 = abs(x = linear_133_cast_fp16)[name = string("op_7051_cast_fp16")];
            bool var_7054 = const()[name = string("op_7054"), val = bool(true)];
            tensor<int32, [1]> var_7055_axes_0 = const()[name = string("op_7055_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7055_cast_fp16 = reduce_max(axes = var_7055_axes_0, keep_dims = var_7054, x = var_7051_cast_fp16)[name = string("op_7055_cast_fp16")];
            fp16 var_7056_to_fp16 = const()[name = string("op_7056_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_239_cast_fp16 = maximum(x = var_7055_cast_fp16, y = var_7056_to_fp16)[name = string("s_239_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_239_cast_fp16 = real_div(x = linear_133_cast_fp16, y = s_239_cast_fp16)[name = string("y_239_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7059_cast_fp16 = mul(x = y_239_cast_fp16, y = y_239_cast_fp16)[name = string("op_7059_cast_fp16")];
            tensor<int32, [1]> var_7064_axes_0 = const()[name = string("op_7064_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7064_keep_dims_0 = const()[name = string("op_7064_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7064_cast_fp16 = reduce_mean(axes = var_7064_axes_0, keep_dims = var_7064_keep_dims_0, x = var_7059_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7065_cast_fp16 = mul(x = s_239_cast_fp16, y = s_239_cast_fp16)[name = string("op_7065_cast_fp16")];
            fp32 var_7066_epsilon_0 = const()[name = string("op_7066_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7066_cast_fp16 = inverse(epsilon = var_7066_epsilon_0, x = var_7065_cast_fp16)[name = string("op_7066_cast_fp16")];
            fp16 var_7067_to_fp16 = const()[name = string("op_7067_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7068_cast_fp16 = mul(x = var_7066_cast_fp16, y = var_7067_to_fp16)[name = string("op_7068_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7070_cast_fp16 = add(x = var_7064_cast_fp16, y = var_7068_cast_fp16)[name = string("op_7070_cast_fp16")];
            fp32 var_7071_epsilon_0 = const()[name = string("op_7071_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7071_cast_fp16 = rsqrt(epsilon = var_7071_epsilon_0, x = var_7070_cast_fp16)[name = string("op_7071_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_209_cast_fp16 = mul(x = y_239_cast_fp16, y = var_7071_cast_fp16)[name = string("out_209_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86927168)))];
            tensor<fp16, [1, 128, 512]> var_7073_cast_fp16 = mul(x = out_209_cast_fp16, y = tm_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7073_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_177_cast_fp16 = add(x = x_263_cast_fp16, y = var_7073_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86928256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87190464))))[name = string("tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_134_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_134_cast_fp16")];
            string x_267_mode_0 = const()[name = string("x_267_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_267_cast_fp16 = gelu(mode = x_267_mode_0, x = linear_134_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<int32, [4]> var_7093_begin_0 = const()[name = string("op_7093_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
            tensor<int32, [4]> var_7093_end_0 = const()[name = string("op_7093_end_0"), val = tensor<int32, [4]>([1, 128, 15, 512])];
            tensor<bool, [4]> var_7093_end_mask_0 = const()[name = string("op_7093_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7093_squeeze_mask_0 = const()[name = string("op_7093_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_7093_cast_fp16 = slice_by_index(begin = var_7093_begin_0, end = var_7093_end_0, end_mask = var_7093_end_mask_0, squeeze_mask = var_7093_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7093_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_179_cast_fp16 = mul(x = x_267_cast_fp16, y = var_7093_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87191552))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87453760))))[name = string("tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7103_cast_fp16 = abs(x = linear_135_cast_fp16)[name = string("op_7103_cast_fp16")];
            bool var_7106 = const()[name = string("op_7106"), val = bool(true)];
            tensor<int32, [1]> var_7107_axes_0 = const()[name = string("op_7107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7107_cast_fp16 = reduce_max(axes = var_7107_axes_0, keep_dims = var_7106, x = var_7103_cast_fp16)[name = string("op_7107_cast_fp16")];
            fp16 var_7108_to_fp16 = const()[name = string("op_7108_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_241_cast_fp16 = maximum(x = var_7107_cast_fp16, y = var_7108_to_fp16)[name = string("s_241_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_241_cast_fp16 = real_div(x = linear_135_cast_fp16, y = s_241_cast_fp16)[name = string("y_241_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7111_cast_fp16 = mul(x = y_241_cast_fp16, y = y_241_cast_fp16)[name = string("op_7111_cast_fp16")];
            tensor<int32, [1]> var_7116_axes_0 = const()[name = string("op_7116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7116_keep_dims_0 = const()[name = string("op_7116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7116_cast_fp16 = reduce_mean(axes = var_7116_axes_0, keep_dims = var_7116_keep_dims_0, x = var_7111_cast_fp16)[name = string("op_7116_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7117_cast_fp16 = mul(x = s_241_cast_fp16, y = s_241_cast_fp16)[name = string("op_7117_cast_fp16")];
            fp32 var_7118_epsilon_0 = const()[name = string("op_7118_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7118_cast_fp16 = inverse(epsilon = var_7118_epsilon_0, x = var_7117_cast_fp16)[name = string("op_7118_cast_fp16")];
            fp16 var_7119_to_fp16 = const()[name = string("op_7119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7120_cast_fp16 = mul(x = var_7118_cast_fp16, y = var_7119_to_fp16)[name = string("op_7120_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7122_cast_fp16 = add(x = var_7116_cast_fp16, y = var_7120_cast_fp16)[name = string("op_7122_cast_fp16")];
            fp32 var_7123_epsilon_0 = const()[name = string("op_7123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7123_cast_fp16 = rsqrt(epsilon = var_7123_epsilon_0, x = var_7122_cast_fp16)[name = string("op_7123_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_211_cast_fp16 = mul(x = y_241_cast_fp16, y = var_7123_cast_fp16)[name = string("out_211_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87454848)))];
            tensor<fp16, [1, 128, 512]> var_7125_cast_fp16 = mul(x = out_211_cast_fp16, y = tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_29_cast_fp16 = add(x = input_177_cast_fp16, y = var_7125_cast_fp16)[name = string("h_29_cast_fp16")];
            tensor<fp16, [1]> tm_layers_14_layer_scalar_to_fp16 = const()[name = string("tm_layers_14_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1p-1])];
            tensor<fp16, [1, 128, 512]> x_271_cast_fp16 = mul(x = h_29_cast_fp16, y = tm_layers_14_layer_scalar_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7129_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_7129_cast_fp16")];
            bool var_7132 = const()[name = string("op_7132"), val = bool(true)];
            tensor<int32, [1]> var_7133_axes_0 = const()[name = string("op_7133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7133_cast_fp16 = reduce_max(axes = var_7133_axes_0, keep_dims = var_7132, x = var_7129_cast_fp16)[name = string("op_7133_cast_fp16")];
            fp16 var_7134_to_fp16 = const()[name = string("op_7134_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_243_cast_fp16 = maximum(x = var_7133_cast_fp16, y = var_7134_to_fp16)[name = string("s_243_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_243_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_243_cast_fp16)[name = string("y_243_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7137_cast_fp16 = mul(x = y_243_cast_fp16, y = y_243_cast_fp16)[name = string("op_7137_cast_fp16")];
            tensor<int32, [1]> var_7142_axes_0 = const()[name = string("op_7142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7142_keep_dims_0 = const()[name = string("op_7142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7142_cast_fp16 = reduce_mean(axes = var_7142_axes_0, keep_dims = var_7142_keep_dims_0, x = var_7137_cast_fp16)[name = string("op_7142_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7143_cast_fp16 = mul(x = s_243_cast_fp16, y = s_243_cast_fp16)[name = string("op_7143_cast_fp16")];
            fp32 var_7144_epsilon_0 = const()[name = string("op_7144_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7144_cast_fp16 = inverse(epsilon = var_7144_epsilon_0, x = var_7143_cast_fp16)[name = string("op_7144_cast_fp16")];
            fp16 var_7145_to_fp16 = const()[name = string("op_7145_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7146_cast_fp16 = mul(x = var_7144_cast_fp16, y = var_7145_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7148_cast_fp16 = add(x = var_7142_cast_fp16, y = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp32 var_7149_epsilon_0 = const()[name = string("op_7149_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7149_cast_fp16 = rsqrt(epsilon = var_7149_epsilon_0, x = var_7148_cast_fp16)[name = string("op_7149_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_213_cast_fp16 = mul(x = y_243_cast_fp16, y = var_7149_cast_fp16)[name = string("out_213_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87455936)))];
            tensor<fp16, [1, 128, 512]> input_181_cast_fp16 = mul(x = out_213_cast_fp16, y = tm_layers_15_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87457024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87981376))))[name = string("tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<int32, [4]> var_7159 = const()[name = string("op_7159"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_7160_cast_fp16 = reshape(shape = var_7159, x = linear_136_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<int32, [4]> x_273_perm_0 = const()[name = string("x_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87983488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88245696))))[name = string("tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<int32, [4]> var_7171 = const()[name = string("op_7171"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_7172_cast_fp16 = reshape(shape = var_7171, x = linear_137_cast_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [4]> x_275_perm_0 = const()[name = string("x_275_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88246784))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88508992))))[name = string("tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_7183 = const()[name = string("op_7183"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_7184_cast_fp16 = reshape(shape = var_7183, x = linear_138_cast_fp16)[name = string("op_7184_cast_fp16")];
            tensor<int32, [4]> x_277_perm_0 = const()[name = string("x_277_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_273_cast_fp16 = transpose(perm = x_273_perm_0, x = var_7160_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 4, 128, 256]> var_7188_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_7188_cast_fp16")];
            bool var_7191 = const()[name = string("op_7191"), val = bool(true)];
            tensor<int32, [1]> var_7192_axes_0 = const()[name = string("op_7192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_7192_cast_fp16 = reduce_max(axes = var_7192_axes_0, keep_dims = var_7191, x = var_7188_cast_fp16)[name = string("op_7192_cast_fp16")];
            fp16 var_7193_to_fp16 = const()[name = string("op_7193_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_245_cast_fp16 = maximum(x = var_7192_cast_fp16, y = var_7193_to_fp16)[name = string("s_245_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_245_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_245_cast_fp16)[name = string("y_245_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_7196_cast_fp16 = mul(x = y_245_cast_fp16, y = y_245_cast_fp16)[name = string("op_7196_cast_fp16")];
            tensor<int32, [1]> var_7201_axes_0 = const()[name = string("op_7201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7201_keep_dims_0 = const()[name = string("op_7201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_7201_cast_fp16 = reduce_mean(axes = var_7201_axes_0, keep_dims = var_7201_keep_dims_0, x = var_7196_cast_fp16)[name = string("op_7201_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_7202_cast_fp16 = mul(x = s_245_cast_fp16, y = s_245_cast_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_7203_cast_fp16 = inverse(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            fp16 var_7204_to_fp16 = const()[name = string("op_7204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_7205_cast_fp16 = mul(x = var_7203_cast_fp16, y = var_7204_to_fp16)[name = string("op_7205_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_7207_cast_fp16 = add(x = var_7201_cast_fp16, y = var_7205_cast_fp16)[name = string("op_7207_cast_fp16")];
            fp32 var_7208_epsilon_0 = const()[name = string("op_7208_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_7208_cast_fp16 = rsqrt(epsilon = var_7208_epsilon_0, x = var_7207_cast_fp16)[name = string("op_7208_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_215_cast_fp16 = mul(x = y_245_cast_fp16, y = var_7208_cast_fp16)[name = string("out_215_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_61_cast_fp16 = mul(x = out_215_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_61_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_275_cast_fp16 = transpose(perm = x_275_perm_0, x = var_7172_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 2, 128, 256]> var_7211_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_7211_cast_fp16")];
            bool var_7214 = const()[name = string("op_7214"), val = bool(true)];
            tensor<int32, [1]> var_7215_axes_0 = const()[name = string("op_7215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_7215_cast_fp16 = reduce_max(axes = var_7215_axes_0, keep_dims = var_7214, x = var_7211_cast_fp16)[name = string("op_7215_cast_fp16")];
            fp16 var_7216_to_fp16 = const()[name = string("op_7216_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_247_cast_fp16 = maximum(x = var_7215_cast_fp16, y = var_7216_to_fp16)[name = string("s_247_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_247_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_247_cast_fp16)[name = string("y_247_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_7219_cast_fp16 = mul(x = y_247_cast_fp16, y = y_247_cast_fp16)[name = string("op_7219_cast_fp16")];
            tensor<int32, [1]> var_7224_axes_0 = const()[name = string("op_7224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7224_keep_dims_0 = const()[name = string("op_7224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_7224_cast_fp16 = reduce_mean(axes = var_7224_axes_0, keep_dims = var_7224_keep_dims_0, x = var_7219_cast_fp16)[name = string("op_7224_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_7225_cast_fp16 = mul(x = s_247_cast_fp16, y = s_247_cast_fp16)[name = string("op_7225_cast_fp16")];
            fp32 var_7226_epsilon_0 = const()[name = string("op_7226_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_7226_cast_fp16 = inverse(epsilon = var_7226_epsilon_0, x = var_7225_cast_fp16)[name = string("op_7226_cast_fp16")];
            fp16 var_7227_to_fp16 = const()[name = string("op_7227_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_7228_cast_fp16 = mul(x = var_7226_cast_fp16, y = var_7227_to_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_7230_cast_fp16 = add(x = var_7224_cast_fp16, y = var_7228_cast_fp16)[name = string("op_7230_cast_fp16")];
            fp32 var_7231_epsilon_0 = const()[name = string("op_7231_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_7231_cast_fp16 = rsqrt(epsilon = var_7231_epsilon_0, x = var_7230_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_217_cast_fp16 = mul(x = y_247_cast_fp16, y = var_7231_cast_fp16)[name = string("out_217_cast_fp16")];
            tensor<fp16, [256]> tm_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510080)))];
            tensor<fp16, [1, 2, 128, 256]> k_91_cast_fp16 = mul(x = out_217_cast_fp16, y = tm_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_277_cast_fp16 = transpose(perm = x_277_perm_0, x = var_7184_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2, 128, 256]> var_7234_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_7234_cast_fp16")];
            bool var_7237 = const()[name = string("op_7237"), val = bool(true)];
            tensor<int32, [1]> var_7238_axes_0 = const()[name = string("op_7238_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_7238_cast_fp16 = reduce_max(axes = var_7238_axes_0, keep_dims = var_7237, x = var_7234_cast_fp16)[name = string("op_7238_cast_fp16")];
            fp16 var_7239_to_fp16 = const()[name = string("op_7239_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_249_cast_fp16 = maximum(x = var_7238_cast_fp16, y = var_7239_to_fp16)[name = string("s_249_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_249_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_249_cast_fp16)[name = string("y_249_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_7242_cast_fp16 = mul(x = y_249_cast_fp16, y = y_249_cast_fp16)[name = string("op_7242_cast_fp16")];
            tensor<int32, [1]> var_7247_axes_0 = const()[name = string("op_7247_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7247_keep_dims_0 = const()[name = string("op_7247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_7247_cast_fp16 = reduce_mean(axes = var_7247_axes_0, keep_dims = var_7247_keep_dims_0, x = var_7242_cast_fp16)[name = string("op_7247_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_7248_cast_fp16 = mul(x = s_249_cast_fp16, y = s_249_cast_fp16)[name = string("op_7248_cast_fp16")];
            fp32 var_7249_epsilon_0 = const()[name = string("op_7249_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_7249_cast_fp16 = inverse(epsilon = var_7249_epsilon_0, x = var_7248_cast_fp16)[name = string("op_7249_cast_fp16")];
            fp16 var_7250_to_fp16 = const()[name = string("op_7250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_7251_cast_fp16 = mul(x = var_7249_cast_fp16, y = var_7250_to_fp16)[name = string("op_7251_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_7253_cast_fp16 = add(x = var_7247_cast_fp16, y = var_7251_cast_fp16)[name = string("op_7253_cast_fp16")];
            fp32 var_7254_epsilon_0 = const()[name = string("op_7254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_7254_cast_fp16 = rsqrt(epsilon = var_7254_epsilon_0, x = var_7253_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_61_cast_fp16 = mul(x = y_249_cast_fp16, y = var_7254_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_7256_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7256_cast_fp16")];
            tensor<int32, [4]> var_7261_begin_0 = const()[name = string("op_7261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7261_end_0 = const()[name = string("op_7261_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_7261_end_mask_0 = const()[name = string("op_7261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_7261_cast_fp16 = slice_by_index(begin = var_7261_begin_0, end = var_7261_end_0, end_mask = var_7261_end_mask_0, x = q_61_cast_fp16)[name = string("op_7261_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_7262_cast_fp16 = mul(x = var_7261_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_7262_cast_fp16")];
            tensor<int32, [4]> var_7267_begin_0 = const()[name = string("op_7267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7267_end_0 = const()[name = string("op_7267_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_7267_end_mask_0 = const()[name = string("op_7267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_7267_cast_fp16 = slice_by_index(begin = var_7267_begin_0, end = var_7267_end_0, end_mask = var_7267_end_mask_0, x = q_61_cast_fp16)[name = string("op_7267_cast_fp16")];
            int32 var_7269 = const()[name = string("op_7269"), val = int32(-1)];
            bool var_7270_interleave_0 = const()[name = string("op_7270_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_7270_cast_fp16 = concat(axis = var_7269, interleave = var_7270_interleave_0, values = (var_7262_cast_fp16, var_7267_cast_fp16))[name = string("op_7270_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_7271_cast_fp16 = mul(x = var_7270_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7271_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_63_cast_fp16 = add(x = var_7256_cast_fp16, y = var_7271_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_7274_cast_fp16 = mul(x = k_91_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7274_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, x = k_91_cast_fp16)[name = string("op_7279_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_7280_cast_fp16 = mul(x = var_7279_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_7280_cast_fp16")];
            tensor<int32, [4]> var_7285_begin_0 = const()[name = string("op_7285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7285_end_0 = const()[name = string("op_7285_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_7285_end_mask_0 = const()[name = string("op_7285_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, x = k_91_cast_fp16)[name = string("op_7285_cast_fp16")];
            int32 var_7287 = const()[name = string("op_7287"), val = int32(-1)];
            bool var_7288_interleave_0 = const()[name = string("op_7288_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_7288_cast_fp16 = concat(axis = var_7287, interleave = var_7288_interleave_0, values = (var_7280_cast_fp16, var_7285_cast_fp16))[name = string("op_7288_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_7289_cast_fp16 = mul(x = var_7288_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7289_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_93_cast_fp16 = add(x = var_7274_cast_fp16, y = var_7289_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_30_reps_0 = const()[name = string("tile_30_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = k_93_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 1, 128, 256]> tile_30_cast_fp16 = tile(reps = tile_30_reps_0, x = transpose_60_cast_fp16)[name = string("tile_30_cast_fp16")];
            tensor<int32, [5]> concat_60 = const()[name = string("concat_60"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_60_cast_fp16 = reshape(shape = concat_60, x = tile_30_cast_fp16)[name = string("reshape_60_cast_fp16")];
            tensor<int32, [5]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = reshape_60_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 1, 128, 256]> reshape_61_cast_fp16 = reshape(shape = concat_61, x = transpose_61_cast_fp16)[name = string("reshape_61_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_31_reps_0 = const()[name = string("tile_31_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = v_61_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 1, 128, 256]> tile_31_cast_fp16 = tile(reps = tile_31_reps_0, x = transpose_62_cast_fp16)[name = string("tile_31_cast_fp16")];
            tensor<int32, [5]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_62_cast_fp16 = reshape(shape = concat_62, x = tile_31_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [5]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 1, 128, 256]> reshape_63_cast_fp16 = reshape(shape = concat_63, x = transpose_63_cast_fp16)[name = string("reshape_63_cast_fp16")];
            tensor<int32, [4]> v_63_perm_0 = const()[name = string("v_63_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7303_transpose_x_0 = const()[name = string("op_7303_transpose_x_0"), val = bool(false)];
            bool var_7303_transpose_y_0 = const()[name = string("op_7303_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = reshape_61_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 4, 128, 128]> var_7303_cast_fp16 = matmul(transpose_x = var_7303_transpose_x_0, transpose_y = var_7303_transpose_y_0, x = q_63_cast_fp16, y = transpose_111_cast_fp16)[name = string("op_7303_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_7305_cast_fp16 = add(x = var_7303_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7305_cast_fp16")];
            int32 var_7306 = const()[name = string("op_7306"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_249_cast_fp16 = softmax(axis = var_7306, x = var_7305_cast_fp16)[name = string("w_249_cast_fp16")];
            bool var_7309_transpose_x_0 = const()[name = string("op_7309_transpose_x_0"), val = bool(false)];
            bool var_7309_transpose_y_0 = const()[name = string("op_7309_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = reshape_63_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 4, 128, 256]> var_7309_cast_fp16 = matmul(transpose_x = var_7309_transpose_x_0, transpose_y = var_7309_transpose_y_0, x = w_249_cast_fp16, y = v_63_cast_fp16)[name = string("op_7309_cast_fp16")];
            tensor<int32, [4]> var_7312_perm_0 = const()[name = string("op_7312_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7316 = const()[name = string("op_7316"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_7312_cast_fp16 = transpose(perm = var_7312_perm_0, x = var_7309_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 128, 1024]> input_183_cast_fp16 = reshape(shape = var_7316, x = var_7312_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510656))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89035008))))[name = string("tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7321_cast_fp16 = abs(x = linear_139_cast_fp16)[name = string("op_7321_cast_fp16")];
            bool var_7324 = const()[name = string("op_7324"), val = bool(true)];
            tensor<int32, [1]> var_7325_axes_0 = const()[name = string("op_7325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7325_cast_fp16 = reduce_max(axes = var_7325_axes_0, keep_dims = var_7324, x = var_7321_cast_fp16)[name = string("op_7325_cast_fp16")];
            fp16 var_7326_to_fp16 = const()[name = string("op_7326_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_251_cast_fp16 = maximum(x = var_7325_cast_fp16, y = var_7326_to_fp16)[name = string("s_251_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_251_cast_fp16 = real_div(x = linear_139_cast_fp16, y = s_251_cast_fp16)[name = string("y_251_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7329_cast_fp16 = mul(x = y_251_cast_fp16, y = y_251_cast_fp16)[name = string("op_7329_cast_fp16")];
            tensor<int32, [1]> var_7334_axes_0 = const()[name = string("op_7334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7334_keep_dims_0 = const()[name = string("op_7334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7334_cast_fp16 = reduce_mean(axes = var_7334_axes_0, keep_dims = var_7334_keep_dims_0, x = var_7329_cast_fp16)[name = string("op_7334_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7335_cast_fp16 = mul(x = s_251_cast_fp16, y = s_251_cast_fp16)[name = string("op_7335_cast_fp16")];
            fp32 var_7336_epsilon_0 = const()[name = string("op_7336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7336_cast_fp16 = inverse(epsilon = var_7336_epsilon_0, x = var_7335_cast_fp16)[name = string("op_7336_cast_fp16")];
            fp16 var_7337_to_fp16 = const()[name = string("op_7337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7338_cast_fp16 = mul(x = var_7336_cast_fp16, y = var_7337_to_fp16)[name = string("op_7338_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7340_cast_fp16 = add(x = var_7334_cast_fp16, y = var_7338_cast_fp16)[name = string("op_7340_cast_fp16")];
            fp32 var_7341_epsilon_0 = const()[name = string("op_7341_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7341_cast_fp16 = rsqrt(epsilon = var_7341_epsilon_0, x = var_7340_cast_fp16)[name = string("op_7341_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_219_cast_fp16 = mul(x = y_251_cast_fp16, y = var_7341_cast_fp16)[name = string("out_219_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89036096)))];
            tensor<fp16, [1, 128, 512]> var_7343_cast_fp16 = mul(x = out_219_cast_fp16, y = tm_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_7343_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_281_cast_fp16 = add(x = x_271_cast_fp16, y = var_7343_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7346_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_7346_cast_fp16")];
            bool var_7349 = const()[name = string("op_7349"), val = bool(true)];
            tensor<int32, [1]> var_7350_axes_0 = const()[name = string("op_7350_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7350_cast_fp16 = reduce_max(axes = var_7350_axes_0, keep_dims = var_7349, x = var_7346_cast_fp16)[name = string("op_7350_cast_fp16")];
            fp16 var_7351_to_fp16 = const()[name = string("op_7351_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_253_cast_fp16 = maximum(x = var_7350_cast_fp16, y = var_7351_to_fp16)[name = string("s_253_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_253_cast_fp16 = real_div(x = x_281_cast_fp16, y = s_253_cast_fp16)[name = string("y_253_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7354_cast_fp16 = mul(x = y_253_cast_fp16, y = y_253_cast_fp16)[name = string("op_7354_cast_fp16")];
            tensor<int32, [1]> var_7359_axes_0 = const()[name = string("op_7359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7359_keep_dims_0 = const()[name = string("op_7359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7359_cast_fp16 = reduce_mean(axes = var_7359_axes_0, keep_dims = var_7359_keep_dims_0, x = var_7354_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7360_cast_fp16 = mul(x = s_253_cast_fp16, y = s_253_cast_fp16)[name = string("op_7360_cast_fp16")];
            fp32 var_7361_epsilon_0 = const()[name = string("op_7361_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7361_cast_fp16 = inverse(epsilon = var_7361_epsilon_0, x = var_7360_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7362_to_fp16 = const()[name = string("op_7362_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7363_cast_fp16 = mul(x = var_7361_cast_fp16, y = var_7362_to_fp16)[name = string("op_7363_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7365_cast_fp16 = add(x = var_7359_cast_fp16, y = var_7363_cast_fp16)[name = string("op_7365_cast_fp16")];
            fp32 var_7366_epsilon_0 = const()[name = string("op_7366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7366_cast_fp16 = rsqrt(epsilon = var_7366_epsilon_0, x = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_221_cast_fp16 = mul(x = y_253_cast_fp16, y = var_7366_cast_fp16)[name = string("out_221_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89037184)))];
            tensor<fp16, [1, 128, 512]> input_185_cast_fp16 = mul(x = out_221_cast_fp16, y = tm_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89038272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90086912))))[name = string("tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_140_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_140_cast_fp16")];
            string var_7373_mode_0 = const()[name = string("op_7373_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_7373_cast_fp16 = gelu(mode = var_7373_mode_0, x = linear_140_cast_fp16)[name = string("op_7373_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90091072))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91139712))))[name = string("tm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_141_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_187_cast_fp16 = mul(x = var_7373_cast_fp16, y = linear_141_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91143872))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92192512))))[name = string("tm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7381_cast_fp16 = abs(x = linear_142_cast_fp16)[name = string("op_7381_cast_fp16")];
            bool var_7384 = const()[name = string("op_7384"), val = bool(true)];
            tensor<int32, [1]> var_7385_axes_0 = const()[name = string("op_7385_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7385_cast_fp16 = reduce_max(axes = var_7385_axes_0, keep_dims = var_7384, x = var_7381_cast_fp16)[name = string("op_7385_cast_fp16")];
            fp16 var_7386_to_fp16 = const()[name = string("op_7386_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_255_cast_fp16 = maximum(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = string("s_255_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_255_cast_fp16 = real_div(x = linear_142_cast_fp16, y = s_255_cast_fp16)[name = string("y_255_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7389_cast_fp16 = mul(x = y_255_cast_fp16, y = y_255_cast_fp16)[name = string("op_7389_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7389_cast_fp16)[name = string("op_7394_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7395_cast_fp16 = mul(x = s_255_cast_fp16, y = s_255_cast_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7400_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7400_cast_fp16")];
            fp32 var_7401_epsilon_0 = const()[name = string("op_7401_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7401_cast_fp16 = rsqrt(epsilon = var_7401_epsilon_0, x = var_7400_cast_fp16)[name = string("op_7401_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_223_cast_fp16 = mul(x = y_255_cast_fp16, y = var_7401_cast_fp16)[name = string("out_223_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92193600)))];
            tensor<fp16, [1, 128, 512]> var_7403_cast_fp16 = mul(x = out_223_cast_fp16, y = tm_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7403_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_189_cast_fp16 = add(x = x_281_cast_fp16, y = var_7403_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92194688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92456896))))[name = string("tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_143_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_143_cast_fp16")];
            string x_285_mode_0 = const()[name = string("x_285_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_285_cast_fp16 = gelu(mode = x_285_mode_0, x = linear_143_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [4]> var_7423_begin_0 = const()[name = string("op_7423_begin_0"), val = tensor<int32, [4]>([0, 0, 15, 0])];
            tensor<int32, [4]> var_7423_end_0 = const()[name = string("op_7423_end_0"), val = tensor<int32, [4]>([1, 128, 16, 512])];
            tensor<bool, [4]> var_7423_end_mask_0 = const()[name = string("op_7423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7423_squeeze_mask_0 = const()[name = string("op_7423_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_7423_cast_fp16 = slice_by_index(begin = var_7423_begin_0, end = var_7423_end_0, end_mask = var_7423_end_mask_0, squeeze_mask = var_7423_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7423_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_191_cast_fp16 = mul(x = x_285_cast_fp16, y = var_7423_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92457984))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92720192))))[name = string("tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7433_cast_fp16 = abs(x = linear_144_cast_fp16)[name = string("op_7433_cast_fp16")];
            bool var_7436 = const()[name = string("op_7436"), val = bool(true)];
            tensor<int32, [1]> var_7437_axes_0 = const()[name = string("op_7437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7437_cast_fp16 = reduce_max(axes = var_7437_axes_0, keep_dims = var_7436, x = var_7433_cast_fp16)[name = string("op_7437_cast_fp16")];
            fp16 var_7438_to_fp16 = const()[name = string("op_7438_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_257_cast_fp16 = maximum(x = var_7437_cast_fp16, y = var_7438_to_fp16)[name = string("s_257_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_257_cast_fp16 = real_div(x = linear_144_cast_fp16, y = s_257_cast_fp16)[name = string("y_257_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7441_cast_fp16 = mul(x = y_257_cast_fp16, y = y_257_cast_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7446_axes_0 = const()[name = string("op_7446_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7446_keep_dims_0 = const()[name = string("op_7446_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7446_cast_fp16 = reduce_mean(axes = var_7446_axes_0, keep_dims = var_7446_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7447_cast_fp16 = mul(x = s_257_cast_fp16, y = s_257_cast_fp16)[name = string("op_7447_cast_fp16")];
            fp32 var_7448_epsilon_0 = const()[name = string("op_7448_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7448_cast_fp16 = inverse(epsilon = var_7448_epsilon_0, x = var_7447_cast_fp16)[name = string("op_7448_cast_fp16")];
            fp16 var_7449_to_fp16 = const()[name = string("op_7449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7450_cast_fp16 = mul(x = var_7448_cast_fp16, y = var_7449_to_fp16)[name = string("op_7450_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7452_cast_fp16 = add(x = var_7446_cast_fp16, y = var_7450_cast_fp16)[name = string("op_7452_cast_fp16")];
            fp32 var_7453_epsilon_0 = const()[name = string("op_7453_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7453_cast_fp16 = rsqrt(epsilon = var_7453_epsilon_0, x = var_7452_cast_fp16)[name = string("op_7453_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_225_cast_fp16 = mul(x = y_257_cast_fp16, y = var_7453_cast_fp16)[name = string("out_225_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92721280)))];
            tensor<fp16, [1, 128, 512]> var_7455_cast_fp16 = mul(x = out_225_cast_fp16, y = tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7455_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_31_cast_fp16 = add(x = input_189_cast_fp16, y = var_7455_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<fp16, [1]> tm_layers_15_layer_scalar_to_fp16 = const()[name = string("tm_layers_15_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7ap-1])];
            tensor<fp16, [1, 128, 512]> x_289_cast_fp16 = mul(x = h_31_cast_fp16, y = tm_layers_15_layer_scalar_to_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7459_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_7459_cast_fp16")];
            bool var_7462 = const()[name = string("op_7462"), val = bool(true)];
            tensor<int32, [1]> var_7463_axes_0 = const()[name = string("op_7463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7463_cast_fp16 = reduce_max(axes = var_7463_axes_0, keep_dims = var_7462, x = var_7459_cast_fp16)[name = string("op_7463_cast_fp16")];
            fp16 var_7464_to_fp16 = const()[name = string("op_7464_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_259_cast_fp16 = maximum(x = var_7463_cast_fp16, y = var_7464_to_fp16)[name = string("s_259_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_259_cast_fp16 = real_div(x = x_289_cast_fp16, y = s_259_cast_fp16)[name = string("y_259_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7467_cast_fp16 = mul(x = y_259_cast_fp16, y = y_259_cast_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7472_axes_0 = const()[name = string("op_7472_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7472_keep_dims_0 = const()[name = string("op_7472_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7472_cast_fp16 = reduce_mean(axes = var_7472_axes_0, keep_dims = var_7472_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7473_cast_fp16 = mul(x = s_259_cast_fp16, y = s_259_cast_fp16)[name = string("op_7473_cast_fp16")];
            fp32 var_7474_epsilon_0 = const()[name = string("op_7474_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7474_cast_fp16 = inverse(epsilon = var_7474_epsilon_0, x = var_7473_cast_fp16)[name = string("op_7474_cast_fp16")];
            fp16 var_7475_to_fp16 = const()[name = string("op_7475_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7476_cast_fp16 = mul(x = var_7474_cast_fp16, y = var_7475_to_fp16)[name = string("op_7476_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7478_cast_fp16 = add(x = var_7472_cast_fp16, y = var_7476_cast_fp16)[name = string("op_7478_cast_fp16")];
            fp32 var_7479_epsilon_0 = const()[name = string("op_7479_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7479_cast_fp16 = rsqrt(epsilon = var_7479_epsilon_0, x = var_7478_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_227_cast_fp16 = mul(x = y_259_cast_fp16, y = var_7479_cast_fp16)[name = string("out_227_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92722368)))];
            tensor<fp16, [1, 128, 512]> input_193_cast_fp16 = mul(x = out_227_cast_fp16, y = tm_layers_16_input_layernorm_weight_to_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92723456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93247808))))[name = string("tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<int32, [4]> var_7489 = const()[name = string("op_7489"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_7490_cast_fp16 = reshape(shape = var_7489, x = linear_145_cast_fp16)[name = string("op_7490_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93249920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93512128))))[name = string("tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<int32, [4]> var_7501 = const()[name = string("op_7501"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_7502_cast_fp16 = reshape(shape = var_7501, x = linear_146_cast_fp16)[name = string("op_7502_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93513216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93775424))))[name = string("tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7513 = const()[name = string("op_7513"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_7514_cast_fp16 = reshape(shape = var_7513, x = linear_147_cast_fp16)[name = string("op_7514_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_7490_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 4, 128, 256]> var_7518_cast_fp16 = abs(x = x_291_cast_fp16)[name = string("op_7518_cast_fp16")];
            bool var_7521 = const()[name = string("op_7521"), val = bool(true)];
            tensor<int32, [1]> var_7522_axes_0 = const()[name = string("op_7522_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_7522_cast_fp16 = reduce_max(axes = var_7522_axes_0, keep_dims = var_7521, x = var_7518_cast_fp16)[name = string("op_7522_cast_fp16")];
            fp16 var_7523_to_fp16 = const()[name = string("op_7523_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_261_cast_fp16 = maximum(x = var_7522_cast_fp16, y = var_7523_to_fp16)[name = string("s_261_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_261_cast_fp16 = real_div(x = x_291_cast_fp16, y = s_261_cast_fp16)[name = string("y_261_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_7526_cast_fp16 = mul(x = y_261_cast_fp16, y = y_261_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<int32, [1]> var_7531_axes_0 = const()[name = string("op_7531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7531_keep_dims_0 = const()[name = string("op_7531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_7531_cast_fp16 = reduce_mean(axes = var_7531_axes_0, keep_dims = var_7531_keep_dims_0, x = var_7526_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_7532_cast_fp16 = mul(x = s_261_cast_fp16, y = s_261_cast_fp16)[name = string("op_7532_cast_fp16")];
            fp32 var_7533_epsilon_0 = const()[name = string("op_7533_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_7533_cast_fp16 = inverse(epsilon = var_7533_epsilon_0, x = var_7532_cast_fp16)[name = string("op_7533_cast_fp16")];
            fp16 var_7534_to_fp16 = const()[name = string("op_7534_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_7535_cast_fp16 = mul(x = var_7533_cast_fp16, y = var_7534_to_fp16)[name = string("op_7535_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_7537_cast_fp16 = add(x = var_7531_cast_fp16, y = var_7535_cast_fp16)[name = string("op_7537_cast_fp16")];
            fp32 var_7538_epsilon_0 = const()[name = string("op_7538_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_7538_cast_fp16 = rsqrt(epsilon = var_7538_epsilon_0, x = var_7537_cast_fp16)[name = string("op_7538_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_229_cast_fp16 = mul(x = y_261_cast_fp16, y = var_7538_cast_fp16)[name = string("out_229_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_65_cast_fp16 = mul(x = out_229_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_65_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_7502_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 128, 256]> var_7541_cast_fp16 = abs(x = x_293_cast_fp16)[name = string("op_7541_cast_fp16")];
            bool var_7544 = const()[name = string("op_7544"), val = bool(true)];
            tensor<int32, [1]> var_7545_axes_0 = const()[name = string("op_7545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_7545_cast_fp16 = reduce_max(axes = var_7545_axes_0, keep_dims = var_7544, x = var_7541_cast_fp16)[name = string("op_7545_cast_fp16")];
            fp16 var_7546_to_fp16 = const()[name = string("op_7546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_263_cast_fp16 = maximum(x = var_7545_cast_fp16, y = var_7546_to_fp16)[name = string("s_263_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_263_cast_fp16 = real_div(x = x_293_cast_fp16, y = s_263_cast_fp16)[name = string("y_263_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_7549_cast_fp16 = mul(x = y_263_cast_fp16, y = y_263_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<int32, [1]> var_7554_axes_0 = const()[name = string("op_7554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7554_keep_dims_0 = const()[name = string("op_7554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_7554_cast_fp16 = reduce_mean(axes = var_7554_axes_0, keep_dims = var_7554_keep_dims_0, x = var_7549_cast_fp16)[name = string("op_7554_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_7555_cast_fp16 = mul(x = s_263_cast_fp16, y = s_263_cast_fp16)[name = string("op_7555_cast_fp16")];
            fp32 var_7556_epsilon_0 = const()[name = string("op_7556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_7556_cast_fp16 = inverse(epsilon = var_7556_epsilon_0, x = var_7555_cast_fp16)[name = string("op_7556_cast_fp16")];
            fp16 var_7557_to_fp16 = const()[name = string("op_7557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_7558_cast_fp16 = mul(x = var_7556_cast_fp16, y = var_7557_to_fp16)[name = string("op_7558_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_7560_cast_fp16 = add(x = var_7554_cast_fp16, y = var_7558_cast_fp16)[name = string("op_7560_cast_fp16")];
            fp32 var_7561_epsilon_0 = const()[name = string("op_7561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_7561_cast_fp16 = rsqrt(epsilon = var_7561_epsilon_0, x = var_7560_cast_fp16)[name = string("op_7561_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_231_cast_fp16 = mul(x = y_263_cast_fp16, y = var_7561_cast_fp16)[name = string("out_231_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_97_cast_fp16 = mul(x = out_231_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_97_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_7514_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 128, 256]> var_7564_cast_fp16 = abs(x = x_295_cast_fp16)[name = string("op_7564_cast_fp16")];
            bool var_7567 = const()[name = string("op_7567"), val = bool(true)];
            tensor<int32, [1]> var_7568_axes_0 = const()[name = string("op_7568_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_7568_cast_fp16 = reduce_max(axes = var_7568_axes_0, keep_dims = var_7567, x = var_7564_cast_fp16)[name = string("op_7568_cast_fp16")];
            fp16 var_7569_to_fp16 = const()[name = string("op_7569_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_265_cast_fp16 = maximum(x = var_7568_cast_fp16, y = var_7569_to_fp16)[name = string("s_265_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_265_cast_fp16 = real_div(x = x_295_cast_fp16, y = s_265_cast_fp16)[name = string("y_265_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_7572_cast_fp16 = mul(x = y_265_cast_fp16, y = y_265_cast_fp16)[name = string("op_7572_cast_fp16")];
            tensor<int32, [1]> var_7577_axes_0 = const()[name = string("op_7577_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7577_keep_dims_0 = const()[name = string("op_7577_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_7577_cast_fp16 = reduce_mean(axes = var_7577_axes_0, keep_dims = var_7577_keep_dims_0, x = var_7572_cast_fp16)[name = string("op_7577_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_7578_cast_fp16 = mul(x = s_265_cast_fp16, y = s_265_cast_fp16)[name = string("op_7578_cast_fp16")];
            fp32 var_7579_epsilon_0 = const()[name = string("op_7579_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_7579_cast_fp16 = inverse(epsilon = var_7579_epsilon_0, x = var_7578_cast_fp16)[name = string("op_7579_cast_fp16")];
            fp16 var_7580_to_fp16 = const()[name = string("op_7580_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_7581_cast_fp16 = mul(x = var_7579_cast_fp16, y = var_7580_to_fp16)[name = string("op_7581_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_7583_cast_fp16 = add(x = var_7577_cast_fp16, y = var_7581_cast_fp16)[name = string("op_7583_cast_fp16")];
            fp32 var_7584_epsilon_0 = const()[name = string("op_7584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_7584_cast_fp16 = rsqrt(epsilon = var_7584_epsilon_0, x = var_7583_cast_fp16)[name = string("op_7584_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_65_cast_fp16 = mul(x = y_265_cast_fp16, y = var_7584_cast_fp16)[name = string("v_65_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_7586_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7586_cast_fp16")];
            tensor<int32, [4]> var_7591_begin_0 = const()[name = string("op_7591_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7591_end_0 = const()[name = string("op_7591_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_7591_end_mask_0 = const()[name = string("op_7591_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_7591_cast_fp16 = slice_by_index(begin = var_7591_begin_0, end = var_7591_end_0, end_mask = var_7591_end_mask_0, x = q_65_cast_fp16)[name = string("op_7591_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_7592_cast_fp16 = mul(x = var_7591_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_7592_cast_fp16")];
            tensor<int32, [4]> var_7597_begin_0 = const()[name = string("op_7597_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7597_end_0 = const()[name = string("op_7597_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_7597_end_mask_0 = const()[name = string("op_7597_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, x = q_65_cast_fp16)[name = string("op_7597_cast_fp16")];
            int32 var_7599 = const()[name = string("op_7599"), val = int32(-1)];
            bool var_7600_interleave_0 = const()[name = string("op_7600_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_7600_cast_fp16 = concat(axis = var_7599, interleave = var_7600_interleave_0, values = (var_7592_cast_fp16, var_7597_cast_fp16))[name = string("op_7600_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_7601_cast_fp16 = mul(x = var_7600_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7601_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_67_cast_fp16 = add(x = var_7586_cast_fp16, y = var_7601_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_7604_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7604_cast_fp16")];
            tensor<int32, [4]> var_7609_begin_0 = const()[name = string("op_7609_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7609_end_0 = const()[name = string("op_7609_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_7609_end_mask_0 = const()[name = string("op_7609_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_7609_cast_fp16 = slice_by_index(begin = var_7609_begin_0, end = var_7609_end_0, end_mask = var_7609_end_mask_0, x = k_97_cast_fp16)[name = string("op_7609_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_7610_cast_fp16 = mul(x = var_7609_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_7610_cast_fp16")];
            tensor<int32, [4]> var_7615_begin_0 = const()[name = string("op_7615_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7615_end_0 = const()[name = string("op_7615_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_7615_end_mask_0 = const()[name = string("op_7615_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_7615_cast_fp16 = slice_by_index(begin = var_7615_begin_0, end = var_7615_end_0, end_mask = var_7615_end_mask_0, x = k_97_cast_fp16)[name = string("op_7615_cast_fp16")];
            int32 var_7617 = const()[name = string("op_7617"), val = int32(-1)];
            bool var_7618_interleave_0 = const()[name = string("op_7618_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_7618_cast_fp16 = concat(axis = var_7617, interleave = var_7618_interleave_0, values = (var_7610_cast_fp16, var_7615_cast_fp16))[name = string("op_7618_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_7619_cast_fp16 = mul(x = var_7618_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7619_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_99_cast_fp16 = add(x = var_7604_cast_fp16, y = var_7619_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_32_reps_0 = const()[name = string("tile_32_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = k_99_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 1, 128, 256]> tile_32_cast_fp16 = tile(reps = tile_32_reps_0, x = transpose_64_cast_fp16)[name = string("tile_32_cast_fp16")];
            tensor<int32, [5]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_64_cast_fp16 = reshape(shape = concat_64, x = tile_32_cast_fp16)[name = string("reshape_64_cast_fp16")];
            tensor<int32, [5]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = reshape_64_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 1, 128, 256]> reshape_65_cast_fp16 = reshape(shape = concat_65, x = transpose_65_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_33_reps_0 = const()[name = string("tile_33_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = v_65_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 1, 128, 256]> tile_33_cast_fp16 = tile(reps = tile_33_reps_0, x = transpose_66_cast_fp16)[name = string("tile_33_cast_fp16")];
            tensor<int32, [5]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_66_cast_fp16 = reshape(shape = concat_66, x = tile_33_cast_fp16)[name = string("reshape_66_cast_fp16")];
            tensor<int32, [5]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = reshape_66_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 1, 128, 256]> reshape_67_cast_fp16 = reshape(shape = concat_67, x = transpose_67_cast_fp16)[name = string("reshape_67_cast_fp16")];
            tensor<int32, [4]> v_67_perm_0 = const()[name = string("v_67_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7633_transpose_x_0 = const()[name = string("op_7633_transpose_x_0"), val = bool(false)];
            bool var_7633_transpose_y_0 = const()[name = string("op_7633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_112_cast_fp16 = transpose(perm = transpose_112_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 4, 128, 128]> var_7633_cast_fp16 = matmul(transpose_x = var_7633_transpose_x_0, transpose_y = var_7633_transpose_y_0, x = q_67_cast_fp16, y = transpose_112_cast_fp16)[name = string("op_7633_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_7635_cast_fp16 = add(x = var_7633_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7635_cast_fp16")];
            int32 var_7636 = const()[name = string("op_7636"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_265_cast_fp16 = softmax(axis = var_7636, x = var_7635_cast_fp16)[name = string("w_265_cast_fp16")];
            bool var_7639_transpose_x_0 = const()[name = string("op_7639_transpose_x_0"), val = bool(false)];
            bool var_7639_transpose_y_0 = const()[name = string("op_7639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = reshape_67_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 4, 128, 256]> var_7639_cast_fp16 = matmul(transpose_x = var_7639_transpose_x_0, transpose_y = var_7639_transpose_y_0, x = w_265_cast_fp16, y = v_67_cast_fp16)[name = string("op_7639_cast_fp16")];
            tensor<int32, [4]> var_7642_perm_0 = const()[name = string("op_7642_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7646 = const()[name = string("op_7646"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_7642_cast_fp16 = transpose(perm = var_7642_perm_0, x = var_7639_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 128, 1024]> input_195_cast_fp16 = reshape(shape = var_7646, x = var_7642_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93776512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94300864))))[name = string("tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7651_cast_fp16 = abs(x = linear_148_cast_fp16)[name = string("op_7651_cast_fp16")];
            bool var_7654 = const()[name = string("op_7654"), val = bool(true)];
            tensor<int32, [1]> var_7655_axes_0 = const()[name = string("op_7655_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7655_cast_fp16 = reduce_max(axes = var_7655_axes_0, keep_dims = var_7654, x = var_7651_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp16 var_7656_to_fp16 = const()[name = string("op_7656_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_267_cast_fp16 = maximum(x = var_7655_cast_fp16, y = var_7656_to_fp16)[name = string("s_267_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_267_cast_fp16 = real_div(x = linear_148_cast_fp16, y = s_267_cast_fp16)[name = string("y_267_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7659_cast_fp16 = mul(x = y_267_cast_fp16, y = y_267_cast_fp16)[name = string("op_7659_cast_fp16")];
            tensor<int32, [1]> var_7664_axes_0 = const()[name = string("op_7664_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7664_keep_dims_0 = const()[name = string("op_7664_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7664_cast_fp16 = reduce_mean(axes = var_7664_axes_0, keep_dims = var_7664_keep_dims_0, x = var_7659_cast_fp16)[name = string("op_7664_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7665_cast_fp16 = mul(x = s_267_cast_fp16, y = s_267_cast_fp16)[name = string("op_7665_cast_fp16")];
            fp32 var_7666_epsilon_0 = const()[name = string("op_7666_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7666_cast_fp16 = inverse(epsilon = var_7666_epsilon_0, x = var_7665_cast_fp16)[name = string("op_7666_cast_fp16")];
            fp16 var_7667_to_fp16 = const()[name = string("op_7667_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7668_cast_fp16 = mul(x = var_7666_cast_fp16, y = var_7667_to_fp16)[name = string("op_7668_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7670_cast_fp16 = add(x = var_7664_cast_fp16, y = var_7668_cast_fp16)[name = string("op_7670_cast_fp16")];
            fp32 var_7671_epsilon_0 = const()[name = string("op_7671_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7671_cast_fp16 = rsqrt(epsilon = var_7671_epsilon_0, x = var_7670_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_233_cast_fp16 = mul(x = y_267_cast_fp16, y = var_7671_cast_fp16)[name = string("out_233_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94301952)))];
            tensor<fp16, [1, 128, 512]> var_7673_cast_fp16 = mul(x = out_233_cast_fp16, y = tm_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("op_7673_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = var_7673_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7676_cast_fp16 = abs(x = x_299_cast_fp16)[name = string("op_7676_cast_fp16")];
            bool var_7679 = const()[name = string("op_7679"), val = bool(true)];
            tensor<int32, [1]> var_7680_axes_0 = const()[name = string("op_7680_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7680_cast_fp16 = reduce_max(axes = var_7680_axes_0, keep_dims = var_7679, x = var_7676_cast_fp16)[name = string("op_7680_cast_fp16")];
            fp16 var_7681_to_fp16 = const()[name = string("op_7681_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_269_cast_fp16 = maximum(x = var_7680_cast_fp16, y = var_7681_to_fp16)[name = string("s_269_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_269_cast_fp16 = real_div(x = x_299_cast_fp16, y = s_269_cast_fp16)[name = string("y_269_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7684_cast_fp16 = mul(x = y_269_cast_fp16, y = y_269_cast_fp16)[name = string("op_7684_cast_fp16")];
            tensor<int32, [1]> var_7689_axes_0 = const()[name = string("op_7689_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7689_keep_dims_0 = const()[name = string("op_7689_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7689_cast_fp16 = reduce_mean(axes = var_7689_axes_0, keep_dims = var_7689_keep_dims_0, x = var_7684_cast_fp16)[name = string("op_7689_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7690_cast_fp16 = mul(x = s_269_cast_fp16, y = s_269_cast_fp16)[name = string("op_7690_cast_fp16")];
            fp32 var_7691_epsilon_0 = const()[name = string("op_7691_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7691_cast_fp16 = inverse(epsilon = var_7691_epsilon_0, x = var_7690_cast_fp16)[name = string("op_7691_cast_fp16")];
            fp16 var_7692_to_fp16 = const()[name = string("op_7692_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7693_cast_fp16 = mul(x = var_7691_cast_fp16, y = var_7692_to_fp16)[name = string("op_7693_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7695_cast_fp16 = add(x = var_7689_cast_fp16, y = var_7693_cast_fp16)[name = string("op_7695_cast_fp16")];
            fp32 var_7696_epsilon_0 = const()[name = string("op_7696_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7696_cast_fp16 = rsqrt(epsilon = var_7696_epsilon_0, x = var_7695_cast_fp16)[name = string("op_7696_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_235_cast_fp16 = mul(x = y_269_cast_fp16, y = var_7696_cast_fp16)[name = string("out_235_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94303040)))];
            tensor<fp16, [1, 128, 512]> input_197_cast_fp16 = mul(x = out_235_cast_fp16, y = tm_layers_16_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94304128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95352768))))[name = string("tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_149_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_149_cast_fp16")];
            string var_7703_mode_0 = const()[name = string("op_7703_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_7703_cast_fp16 = gelu(mode = var_7703_mode_0, x = linear_149_cast_fp16)[name = string("op_7703_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95356928))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96405568))))[name = string("tm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_150_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_199_cast_fp16 = mul(x = var_7703_cast_fp16, y = linear_150_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96409728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97458368))))[name = string("tm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7711_cast_fp16 = abs(x = linear_151_cast_fp16)[name = string("op_7711_cast_fp16")];
            bool var_7714 = const()[name = string("op_7714"), val = bool(true)];
            tensor<int32, [1]> var_7715_axes_0 = const()[name = string("op_7715_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7715_cast_fp16 = reduce_max(axes = var_7715_axes_0, keep_dims = var_7714, x = var_7711_cast_fp16)[name = string("op_7715_cast_fp16")];
            fp16 var_7716_to_fp16 = const()[name = string("op_7716_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_271_cast_fp16 = maximum(x = var_7715_cast_fp16, y = var_7716_to_fp16)[name = string("s_271_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_271_cast_fp16 = real_div(x = linear_151_cast_fp16, y = s_271_cast_fp16)[name = string("y_271_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7719_cast_fp16 = mul(x = y_271_cast_fp16, y = y_271_cast_fp16)[name = string("op_7719_cast_fp16")];
            tensor<int32, [1]> var_7724_axes_0 = const()[name = string("op_7724_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7724_keep_dims_0 = const()[name = string("op_7724_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7724_cast_fp16 = reduce_mean(axes = var_7724_axes_0, keep_dims = var_7724_keep_dims_0, x = var_7719_cast_fp16)[name = string("op_7724_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7725_cast_fp16 = mul(x = s_271_cast_fp16, y = s_271_cast_fp16)[name = string("op_7725_cast_fp16")];
            fp32 var_7726_epsilon_0 = const()[name = string("op_7726_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7726_cast_fp16 = inverse(epsilon = var_7726_epsilon_0, x = var_7725_cast_fp16)[name = string("op_7726_cast_fp16")];
            fp16 var_7727_to_fp16 = const()[name = string("op_7727_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7728_cast_fp16 = mul(x = var_7726_cast_fp16, y = var_7727_to_fp16)[name = string("op_7728_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7730_cast_fp16 = add(x = var_7724_cast_fp16, y = var_7728_cast_fp16)[name = string("op_7730_cast_fp16")];
            fp32 var_7731_epsilon_0 = const()[name = string("op_7731_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7731_cast_fp16 = rsqrt(epsilon = var_7731_epsilon_0, x = var_7730_cast_fp16)[name = string("op_7731_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_237_cast_fp16 = mul(x = y_271_cast_fp16, y = var_7731_cast_fp16)[name = string("out_237_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97459456)))];
            tensor<fp16, [1, 128, 512]> var_7733_cast_fp16 = mul(x = out_237_cast_fp16, y = tm_layers_16_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7733_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_201_cast_fp16 = add(x = x_299_cast_fp16, y = var_7733_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97460544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97722752))))[name = string("tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_152_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_152_cast_fp16")];
            string x_303_mode_0 = const()[name = string("x_303_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_303_cast_fp16 = gelu(mode = x_303_mode_0, x = linear_152_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [4]> var_7753_begin_0 = const()[name = string("op_7753_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7753_end_0 = const()[name = string("op_7753_end_0"), val = tensor<int32, [4]>([1, 128, 17, 512])];
            tensor<bool, [4]> var_7753_end_mask_0 = const()[name = string("op_7753_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7753_squeeze_mask_0 = const()[name = string("op_7753_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_7753_cast_fp16 = slice_by_index(begin = var_7753_begin_0, end = var_7753_end_0, end_mask = var_7753_end_mask_0, squeeze_mask = var_7753_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7753_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_203_cast_fp16 = mul(x = x_303_cast_fp16, y = var_7753_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97723840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97986048))))[name = string("tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7763_cast_fp16 = abs(x = linear_153_cast_fp16)[name = string("op_7763_cast_fp16")];
            bool var_7766 = const()[name = string("op_7766"), val = bool(true)];
            tensor<int32, [1]> var_7767_axes_0 = const()[name = string("op_7767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7767_cast_fp16 = reduce_max(axes = var_7767_axes_0, keep_dims = var_7766, x = var_7763_cast_fp16)[name = string("op_7767_cast_fp16")];
            fp16 var_7768_to_fp16 = const()[name = string("op_7768_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_273_cast_fp16 = maximum(x = var_7767_cast_fp16, y = var_7768_to_fp16)[name = string("s_273_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_273_cast_fp16 = real_div(x = linear_153_cast_fp16, y = s_273_cast_fp16)[name = string("y_273_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7771_cast_fp16 = mul(x = y_273_cast_fp16, y = y_273_cast_fp16)[name = string("op_7771_cast_fp16")];
            tensor<int32, [1]> var_7776_axes_0 = const()[name = string("op_7776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7776_keep_dims_0 = const()[name = string("op_7776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7776_cast_fp16 = reduce_mean(axes = var_7776_axes_0, keep_dims = var_7776_keep_dims_0, x = var_7771_cast_fp16)[name = string("op_7776_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7777_cast_fp16 = mul(x = s_273_cast_fp16, y = s_273_cast_fp16)[name = string("op_7777_cast_fp16")];
            fp32 var_7778_epsilon_0 = const()[name = string("op_7778_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7778_cast_fp16 = inverse(epsilon = var_7778_epsilon_0, x = var_7777_cast_fp16)[name = string("op_7778_cast_fp16")];
            fp16 var_7779_to_fp16 = const()[name = string("op_7779_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7780_cast_fp16 = mul(x = var_7778_cast_fp16, y = var_7779_to_fp16)[name = string("op_7780_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7782_cast_fp16 = add(x = var_7776_cast_fp16, y = var_7780_cast_fp16)[name = string("op_7782_cast_fp16")];
            fp32 var_7783_epsilon_0 = const()[name = string("op_7783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7783_cast_fp16 = rsqrt(epsilon = var_7783_epsilon_0, x = var_7782_cast_fp16)[name = string("op_7783_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_239_cast_fp16 = mul(x = y_273_cast_fp16, y = var_7783_cast_fp16)[name = string("out_239_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97987136)))];
            tensor<fp16, [1, 128, 512]> var_7785_cast_fp16 = mul(x = out_239_cast_fp16, y = tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_33_cast_fp16 = add(x = input_201_cast_fp16, y = var_7785_cast_fp16)[name = string("h_33_cast_fp16")];
            tensor<fp16, [1]> tm_layers_16_layer_scalar_to_fp16 = const()[name = string("tm_layers_16_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7p-2])];
            tensor<fp16, [1, 128, 512]> x_307_cast_fp16 = mul(x = h_33_cast_fp16, y = tm_layers_16_layer_scalar_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7789_cast_fp16 = abs(x = x_307_cast_fp16)[name = string("op_7789_cast_fp16")];
            bool var_7792 = const()[name = string("op_7792"), val = bool(true)];
            tensor<int32, [1]> var_7793_axes_0 = const()[name = string("op_7793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7793_cast_fp16 = reduce_max(axes = var_7793_axes_0, keep_dims = var_7792, x = var_7789_cast_fp16)[name = string("op_7793_cast_fp16")];
            fp16 var_7794_to_fp16 = const()[name = string("op_7794_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_275_cast_fp16 = maximum(x = var_7793_cast_fp16, y = var_7794_to_fp16)[name = string("s_275_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_275_cast_fp16 = real_div(x = x_307_cast_fp16, y = s_275_cast_fp16)[name = string("y_275_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7797_cast_fp16 = mul(x = y_275_cast_fp16, y = y_275_cast_fp16)[name = string("op_7797_cast_fp16")];
            tensor<int32, [1]> var_7802_axes_0 = const()[name = string("op_7802_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7802_keep_dims_0 = const()[name = string("op_7802_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7802_cast_fp16 = reduce_mean(axes = var_7802_axes_0, keep_dims = var_7802_keep_dims_0, x = var_7797_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7803_cast_fp16 = mul(x = s_275_cast_fp16, y = s_275_cast_fp16)[name = string("op_7803_cast_fp16")];
            fp32 var_7804_epsilon_0 = const()[name = string("op_7804_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7804_cast_fp16 = inverse(epsilon = var_7804_epsilon_0, x = var_7803_cast_fp16)[name = string("op_7804_cast_fp16")];
            fp16 var_7805_to_fp16 = const()[name = string("op_7805_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7806_cast_fp16 = mul(x = var_7804_cast_fp16, y = var_7805_to_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7808_cast_fp16 = add(x = var_7802_cast_fp16, y = var_7806_cast_fp16)[name = string("op_7808_cast_fp16")];
            fp32 var_7809_epsilon_0 = const()[name = string("op_7809_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_7809_cast_fp16 = rsqrt(epsilon = var_7809_epsilon_0, x = var_7808_cast_fp16)[name = string("op_7809_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_241_cast_fp16 = mul(x = y_275_cast_fp16, y = var_7809_cast_fp16)[name = string("out_241_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97988224)))];
            tensor<fp16, [1, 128, 512]> input_205_cast_fp16 = mul(x = out_241_cast_fp16, y = tm_layers_17_input_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97989312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99037952))))[name = string("tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_154_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_7819 = const()[name = string("op_7819"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 128, 4, 512]> var_7820_cast_fp16 = reshape(shape = var_7819, x = linear_154_cast_fp16)[name = string("op_7820_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99042112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99304320))))[name = string("tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_7831 = const()[name = string("op_7831"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 128, 1, 512]> var_7832_cast_fp16 = reshape(shape = var_7831, x = linear_155_cast_fp16)[name = string("op_7832_cast_fp16")];
            tensor<int32, [4]> x_311_perm_0 = const()[name = string("x_311_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99305408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99567616))))[name = string("tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_7843 = const()[name = string("op_7843"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 128, 1, 512]> var_7844_cast_fp16 = reshape(shape = var_7843, x = linear_156_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<int32, [4]> x_313_perm_0 = const()[name = string("x_313_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 512]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_7820_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 4, 128, 512]> var_7848_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_7848_cast_fp16")];
            bool var_7851 = const()[name = string("op_7851"), val = bool(true)];
            tensor<int32, [1]> var_7852_axes_0 = const()[name = string("op_7852_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_7852_cast_fp16 = reduce_max(axes = var_7852_axes_0, keep_dims = var_7851, x = var_7848_cast_fp16)[name = string("op_7852_cast_fp16")];
            fp16 var_7853_to_fp16 = const()[name = string("op_7853_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_277_cast_fp16 = maximum(x = var_7852_cast_fp16, y = var_7853_to_fp16)[name = string("s_277_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> y_277_cast_fp16 = real_div(x = x_309_cast_fp16, y = s_277_cast_fp16)[name = string("y_277_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_7856_cast_fp16 = mul(x = y_277_cast_fp16, y = y_277_cast_fp16)[name = string("op_7856_cast_fp16")];
            tensor<int32, [1]> var_7861_axes_0 = const()[name = string("op_7861_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7861_keep_dims_0 = const()[name = string("op_7861_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_7861_cast_fp16 = reduce_mean(axes = var_7861_axes_0, keep_dims = var_7861_keep_dims_0, x = var_7856_cast_fp16)[name = string("op_7861_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_7862_cast_fp16 = mul(x = s_277_cast_fp16, y = s_277_cast_fp16)[name = string("op_7862_cast_fp16")];
            fp32 var_7863_epsilon_0 = const()[name = string("op_7863_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_7863_cast_fp16 = inverse(epsilon = var_7863_epsilon_0, x = var_7862_cast_fp16)[name = string("op_7863_cast_fp16")];
            fp16 var_7864_to_fp16 = const()[name = string("op_7864_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_7865_cast_fp16 = mul(x = var_7863_cast_fp16, y = var_7864_to_fp16)[name = string("op_7865_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_7867_cast_fp16 = add(x = var_7861_cast_fp16, y = var_7865_cast_fp16)[name = string("op_7867_cast_fp16")];
            fp32 var_7868_epsilon_0 = const()[name = string("op_7868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_7868_cast_fp16 = rsqrt(epsilon = var_7868_epsilon_0, x = var_7867_cast_fp16)[name = string("op_7868_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> out_243_cast_fp16 = mul(x = y_277_cast_fp16, y = var_7868_cast_fp16)[name = string("out_243_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99568704)))];
            tensor<fp16, [1, 4, 128, 512]> q_69_cast_fp16 = mul(x = out_243_cast_fp16, y = tm_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("q_69_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> x_311_cast_fp16 = transpose(perm = x_311_perm_0, x = var_7832_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 1, 128, 512]> var_7871_cast_fp16 = abs(x = x_311_cast_fp16)[name = string("op_7871_cast_fp16")];
            bool var_7874 = const()[name = string("op_7874"), val = bool(true)];
            tensor<int32, [1]> var_7875_axes_0 = const()[name = string("op_7875_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 128, 1]> var_7875_cast_fp16 = reduce_max(axes = var_7875_axes_0, keep_dims = var_7874, x = var_7871_cast_fp16)[name = string("op_7875_cast_fp16")];
            fp16 var_7876_to_fp16 = const()[name = string("op_7876_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 128, 1]> s_279_cast_fp16 = maximum(x = var_7875_cast_fp16, y = var_7876_to_fp16)[name = string("s_279_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> y_279_cast_fp16 = real_div(x = x_311_cast_fp16, y = s_279_cast_fp16)[name = string("y_279_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_7879_cast_fp16 = mul(x = y_279_cast_fp16, y = y_279_cast_fp16)[name = string("op_7879_cast_fp16")];
            tensor<int32, [1]> var_7884_axes_0 = const()[name = string("op_7884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7884_keep_dims_0 = const()[name = string("op_7884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_7884_cast_fp16 = reduce_mean(axes = var_7884_axes_0, keep_dims = var_7884_keep_dims_0, x = var_7879_cast_fp16)[name = string("op_7884_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_7885_cast_fp16 = mul(x = s_279_cast_fp16, y = s_279_cast_fp16)[name = string("op_7885_cast_fp16")];
            fp32 var_7886_epsilon_0 = const()[name = string("op_7886_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 128, 1]> var_7886_cast_fp16 = inverse(epsilon = var_7886_epsilon_0, x = var_7885_cast_fp16)[name = string("op_7886_cast_fp16")];
            fp16 var_7887_to_fp16 = const()[name = string("op_7887_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 128, 1]> var_7888_cast_fp16 = mul(x = var_7886_cast_fp16, y = var_7887_to_fp16)[name = string("op_7888_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_7890_cast_fp16 = add(x = var_7884_cast_fp16, y = var_7888_cast_fp16)[name = string("op_7890_cast_fp16")];
            fp32 var_7891_epsilon_0 = const()[name = string("op_7891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 128, 1]> var_7891_cast_fp16 = rsqrt(epsilon = var_7891_epsilon_0, x = var_7890_cast_fp16)[name = string("op_7891_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> out_245_cast_fp16 = mul(x = y_279_cast_fp16, y = var_7891_cast_fp16)[name = string("out_245_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99569792)))];
            tensor<fp16, [1, 1, 128, 512]> k_103_cast_fp16 = mul(x = out_245_cast_fp16, y = tm_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("k_103_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> x_313_cast_fp16 = transpose(perm = x_313_perm_0, x = var_7844_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1, 128, 512]> var_7894_cast_fp16 = abs(x = x_313_cast_fp16)[name = string("op_7894_cast_fp16")];
            bool var_7897 = const()[name = string("op_7897"), val = bool(true)];
            tensor<int32, [1]> var_7898_axes_0 = const()[name = string("op_7898_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 128, 1]> var_7898_cast_fp16 = reduce_max(axes = var_7898_axes_0, keep_dims = var_7897, x = var_7894_cast_fp16)[name = string("op_7898_cast_fp16")];
            fp16 var_7899_to_fp16 = const()[name = string("op_7899_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 128, 1]> s_281_cast_fp16 = maximum(x = var_7898_cast_fp16, y = var_7899_to_fp16)[name = string("s_281_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> y_281_cast_fp16 = real_div(x = x_313_cast_fp16, y = s_281_cast_fp16)[name = string("y_281_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_7902_cast_fp16 = mul(x = y_281_cast_fp16, y = y_281_cast_fp16)[name = string("op_7902_cast_fp16")];
            tensor<int32, [1]> var_7907_axes_0 = const()[name = string("op_7907_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7907_keep_dims_0 = const()[name = string("op_7907_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_7907_cast_fp16 = reduce_mean(axes = var_7907_axes_0, keep_dims = var_7907_keep_dims_0, x = var_7902_cast_fp16)[name = string("op_7907_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_7908_cast_fp16 = mul(x = s_281_cast_fp16, y = s_281_cast_fp16)[name = string("op_7908_cast_fp16")];
            fp32 var_7909_epsilon_0 = const()[name = string("op_7909_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 128, 1]> var_7909_cast_fp16 = inverse(epsilon = var_7909_epsilon_0, x = var_7908_cast_fp16)[name = string("op_7909_cast_fp16")];
            fp16 var_7910_to_fp16 = const()[name = string("op_7910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 128, 1]> var_7911_cast_fp16 = mul(x = var_7909_cast_fp16, y = var_7910_to_fp16)[name = string("op_7911_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_7913_cast_fp16 = add(x = var_7907_cast_fp16, y = var_7911_cast_fp16)[name = string("op_7913_cast_fp16")];
            fp32 var_7914_epsilon_0 = const()[name = string("op_7914_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 128, 1]> var_7914_cast_fp16 = rsqrt(epsilon = var_7914_epsilon_0, x = var_7913_cast_fp16)[name = string("op_7914_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> v_69_cast_fp16 = mul(x = y_281_cast_fp16, y = var_7914_cast_fp16)[name = string("v_69_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_7916_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7916_cast_fp16")];
            tensor<int32, [4]> var_7921_begin_0 = const()[name = string("op_7921_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7921_end_0 = const()[name = string("op_7921_end_0"), val = tensor<int32, [4]>([1, 4, 128, 512])];
            tensor<bool, [4]> var_7921_end_mask_0 = const()[name = string("op_7921_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 256]> var_7921_cast_fp16 = slice_by_index(begin = var_7921_begin_0, end = var_7921_end_0, end_mask = var_7921_end_mask_0, x = q_69_cast_fp16)[name = string("op_7921_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 256]> var_7922_cast_fp16 = mul(x = var_7921_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 256]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = q_69_cast_fp16)[name = string("op_7927_cast_fp16")];
            int32 var_7929 = const()[name = string("op_7929"), val = int32(-1)];
            bool var_7930_interleave_0 = const()[name = string("op_7930_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 512]> var_7930_cast_fp16 = concat(axis = var_7929, interleave = var_7930_interleave_0, values = (var_7922_cast_fp16, var_7927_cast_fp16))[name = string("op_7930_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_7931_cast_fp16 = mul(x = var_7930_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7931_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> q_71_cast_fp16 = add(x = var_7916_cast_fp16, y = var_7931_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_7934_cast_fp16 = mul(x = k_103_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7934_cast_fp16")];
            tensor<int32, [4]> var_7939_begin_0 = const()[name = string("op_7939_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7939_end_0 = const()[name = string("op_7939_end_0"), val = tensor<int32, [4]>([1, 1, 128, 512])];
            tensor<bool, [4]> var_7939_end_mask_0 = const()[name = string("op_7939_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 128, 256]> var_7939_cast_fp16 = slice_by_index(begin = var_7939_begin_0, end = var_7939_end_0, end_mask = var_7939_end_mask_0, x = k_103_cast_fp16)[name = string("op_7939_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 128, 256]> var_7940_cast_fp16 = mul(x = var_7939_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_7940_cast_fp16")];
            tensor<int32, [4]> var_7945_begin_0 = const()[name = string("op_7945_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7945_end_0 = const()[name = string("op_7945_end_0"), val = tensor<int32, [4]>([1, 1, 128, 256])];
            tensor<bool, [4]> var_7945_end_mask_0 = const()[name = string("op_7945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 128, 256]> var_7945_cast_fp16 = slice_by_index(begin = var_7945_begin_0, end = var_7945_end_0, end_mask = var_7945_end_mask_0, x = k_103_cast_fp16)[name = string("op_7945_cast_fp16")];
            int32 var_7947 = const()[name = string("op_7947"), val = int32(-1)];
            bool var_7948_interleave_0 = const()[name = string("op_7948_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 512]> var_7948_cast_fp16 = concat(axis = var_7947, interleave = var_7948_interleave_0, values = (var_7940_cast_fp16, var_7945_cast_fp16))[name = string("op_7948_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_7949_cast_fp16 = mul(x = var_7948_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> k_105_cast_fp16 = add(x = var_7934_cast_fp16, y = var_7949_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_34_reps_0 = const()[name = string("tile_34_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 128, 512]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = k_105_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 1, 128, 512]> tile_34_cast_fp16 = tile(reps = tile_34_reps_0, x = transpose_68_cast_fp16)[name = string("tile_34_cast_fp16")];
            tensor<int32, [5]> concat_68 = const()[name = string("concat_68"), val = tensor<int32, [5]>([4, 1, 1, 128, 512])];
            tensor<fp16, [4, 1, 1, 128, 512]> reshape_68_cast_fp16 = reshape(shape = concat_68, x = tile_34_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [5]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([-1, 1, 128, 512])];
            tensor<fp16, [1, 4, 1, 128, 512]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 1, 128, 512]> reshape_69_cast_fp16 = reshape(shape = concat_69, x = transpose_69_cast_fp16)[name = string("reshape_69_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_35_reps_0 = const()[name = string("tile_35_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 128, 512]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = v_69_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 1, 128, 512]> tile_35_cast_fp16 = tile(reps = tile_35_reps_0, x = transpose_70_cast_fp16)[name = string("tile_35_cast_fp16")];
            tensor<int32, [5]> concat_70 = const()[name = string("concat_70"), val = tensor<int32, [5]>([4, 1, 1, 128, 512])];
            tensor<fp16, [4, 1, 1, 128, 512]> reshape_70_cast_fp16 = reshape(shape = concat_70, x = tile_35_cast_fp16)[name = string("reshape_70_cast_fp16")];
            tensor<int32, [5]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([-1, 1, 128, 512])];
            tensor<fp16, [1, 4, 1, 128, 512]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = reshape_70_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 1, 128, 512]> reshape_71_cast_fp16 = reshape(shape = concat_71, x = transpose_71_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7963_transpose_x_0 = const()[name = string("op_7963_transpose_x_0"), val = bool(false)];
            bool var_7963_transpose_y_0 = const()[name = string("op_7963_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 128]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = reshape_69_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 4, 128, 128]> var_7963_cast_fp16 = matmul(transpose_x = var_7963_transpose_x_0, transpose_y = var_7963_transpose_y_0, x = q_71_cast_fp16, y = transpose_113_cast_fp16)[name = string("op_7963_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_7965_cast_fp16 = add(x = var_7963_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7965_cast_fp16")];
            int32 var_7966 = const()[name = string("op_7966"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_281_cast_fp16 = softmax(axis = var_7966, x = var_7965_cast_fp16)[name = string("w_281_cast_fp16")];
            bool var_7969_transpose_x_0 = const()[name = string("op_7969_transpose_x_0"), val = bool(false)];
            bool var_7969_transpose_y_0 = const()[name = string("op_7969_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 512]> v_71_cast_fp16 = transpose(perm = v_71_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 4, 128, 512]> var_7969_cast_fp16 = matmul(transpose_x = var_7969_transpose_x_0, transpose_y = var_7969_transpose_y_0, x = w_281_cast_fp16, y = v_71_cast_fp16)[name = string("op_7969_cast_fp16")];
            tensor<int32, [4]> var_7972_perm_0 = const()[name = string("op_7972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7976 = const()[name = string("op_7976"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 128, 4, 512]> var_7972_cast_fp16 = transpose(perm = var_7972_perm_0, x = var_7969_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 128, 2048]> input_207_cast_fp16 = reshape(shape = var_7976, x = var_7972_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99570880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619520))))[name = string("tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7981_cast_fp16 = abs(x = linear_157_cast_fp16)[name = string("op_7981_cast_fp16")];
            bool var_7984 = const()[name = string("op_7984"), val = bool(true)];
            tensor<int32, [1]> var_7985_axes_0 = const()[name = string("op_7985_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_7985_cast_fp16 = reduce_max(axes = var_7985_axes_0, keep_dims = var_7984, x = var_7981_cast_fp16)[name = string("op_7985_cast_fp16")];
            fp16 var_7986_to_fp16 = const()[name = string("op_7986_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_283_cast_fp16 = maximum(x = var_7985_cast_fp16, y = var_7986_to_fp16)[name = string("s_283_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_283_cast_fp16 = real_div(x = linear_157_cast_fp16, y = s_283_cast_fp16)[name = string("y_283_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_7989_cast_fp16 = mul(x = y_283_cast_fp16, y = y_283_cast_fp16)[name = string("op_7989_cast_fp16")];
            tensor<int32, [1]> var_7994_axes_0 = const()[name = string("op_7994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7994_keep_dims_0 = const()[name = string("op_7994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7994_cast_fp16 = reduce_mean(axes = var_7994_axes_0, keep_dims = var_7994_keep_dims_0, x = var_7989_cast_fp16)[name = string("op_7994_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_7995_cast_fp16 = mul(x = s_283_cast_fp16, y = s_283_cast_fp16)[name = string("op_7995_cast_fp16")];
            fp32 var_7996_epsilon_0 = const()[name = string("op_7996_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_7996_cast_fp16 = inverse(epsilon = var_7996_epsilon_0, x = var_7995_cast_fp16)[name = string("op_7996_cast_fp16")];
            fp16 var_7997_to_fp16 = const()[name = string("op_7997_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_7998_cast_fp16 = mul(x = var_7996_cast_fp16, y = var_7997_to_fp16)[name = string("op_7998_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8000_cast_fp16 = add(x = var_7994_cast_fp16, y = var_7998_cast_fp16)[name = string("op_8000_cast_fp16")];
            fp32 var_8001_epsilon_0 = const()[name = string("op_8001_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8001_cast_fp16 = rsqrt(epsilon = var_8001_epsilon_0, x = var_8000_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_247_cast_fp16 = mul(x = y_283_cast_fp16, y = var_8001_cast_fp16)[name = string("out_247_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100620608)))];
            tensor<fp16, [1, 128, 512]> var_8003_cast_fp16 = mul(x = out_247_cast_fp16, y = tm_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("op_8003_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_317_cast_fp16 = add(x = x_307_cast_fp16, y = var_8003_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8006_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_8006_cast_fp16")];
            bool var_8009 = const()[name = string("op_8009"), val = bool(true)];
            tensor<int32, [1]> var_8010_axes_0 = const()[name = string("op_8010_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8010_cast_fp16 = reduce_max(axes = var_8010_axes_0, keep_dims = var_8009, x = var_8006_cast_fp16)[name = string("op_8010_cast_fp16")];
            fp16 var_8011_to_fp16 = const()[name = string("op_8011_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_285_cast_fp16 = maximum(x = var_8010_cast_fp16, y = var_8011_to_fp16)[name = string("s_285_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_285_cast_fp16 = real_div(x = x_317_cast_fp16, y = s_285_cast_fp16)[name = string("y_285_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8014_cast_fp16 = mul(x = y_285_cast_fp16, y = y_285_cast_fp16)[name = string("op_8014_cast_fp16")];
            tensor<int32, [1]> var_8019_axes_0 = const()[name = string("op_8019_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8019_keep_dims_0 = const()[name = string("op_8019_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8019_cast_fp16 = reduce_mean(axes = var_8019_axes_0, keep_dims = var_8019_keep_dims_0, x = var_8014_cast_fp16)[name = string("op_8019_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8020_cast_fp16 = mul(x = s_285_cast_fp16, y = s_285_cast_fp16)[name = string("op_8020_cast_fp16")];
            fp32 var_8021_epsilon_0 = const()[name = string("op_8021_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8021_cast_fp16 = inverse(epsilon = var_8021_epsilon_0, x = var_8020_cast_fp16)[name = string("op_8021_cast_fp16")];
            fp16 var_8022_to_fp16 = const()[name = string("op_8022_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8023_cast_fp16 = mul(x = var_8021_cast_fp16, y = var_8022_to_fp16)[name = string("op_8023_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8025_cast_fp16 = add(x = var_8019_cast_fp16, y = var_8023_cast_fp16)[name = string("op_8025_cast_fp16")];
            fp32 var_8026_epsilon_0 = const()[name = string("op_8026_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8026_cast_fp16 = rsqrt(epsilon = var_8026_epsilon_0, x = var_8025_cast_fp16)[name = string("op_8026_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_249_cast_fp16 = mul(x = y_285_cast_fp16, y = var_8026_cast_fp16)[name = string("out_249_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621696)))];
            tensor<fp16, [1, 128, 512]> input_209_cast_fp16 = mul(x = out_249_cast_fp16, y = tm_layers_17_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100622784))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101671424))))[name = string("tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_158_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_158_cast_fp16")];
            string var_8033_mode_0 = const()[name = string("op_8033_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_8033_cast_fp16 = gelu(mode = var_8033_mode_0, x = linear_158_cast_fp16)[name = string("op_8033_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101675584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102724224))))[name = string("tm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_159_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_211_cast_fp16 = mul(x = var_8033_cast_fp16, y = linear_159_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102728384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103777024))))[name = string("tm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8041_cast_fp16 = abs(x = linear_160_cast_fp16)[name = string("op_8041_cast_fp16")];
            bool var_8044 = const()[name = string("op_8044"), val = bool(true)];
            tensor<int32, [1]> var_8045_axes_0 = const()[name = string("op_8045_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8045_cast_fp16 = reduce_max(axes = var_8045_axes_0, keep_dims = var_8044, x = var_8041_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp16 var_8046_to_fp16 = const()[name = string("op_8046_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_287_cast_fp16 = maximum(x = var_8045_cast_fp16, y = var_8046_to_fp16)[name = string("s_287_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_287_cast_fp16 = real_div(x = linear_160_cast_fp16, y = s_287_cast_fp16)[name = string("y_287_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8049_cast_fp16 = mul(x = y_287_cast_fp16, y = y_287_cast_fp16)[name = string("op_8049_cast_fp16")];
            tensor<int32, [1]> var_8054_axes_0 = const()[name = string("op_8054_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8054_keep_dims_0 = const()[name = string("op_8054_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8054_cast_fp16 = reduce_mean(axes = var_8054_axes_0, keep_dims = var_8054_keep_dims_0, x = var_8049_cast_fp16)[name = string("op_8054_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8055_cast_fp16 = mul(x = s_287_cast_fp16, y = s_287_cast_fp16)[name = string("op_8055_cast_fp16")];
            fp32 var_8056_epsilon_0 = const()[name = string("op_8056_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8056_cast_fp16 = inverse(epsilon = var_8056_epsilon_0, x = var_8055_cast_fp16)[name = string("op_8056_cast_fp16")];
            fp16 var_8057_to_fp16 = const()[name = string("op_8057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8058_cast_fp16 = mul(x = var_8056_cast_fp16, y = var_8057_to_fp16)[name = string("op_8058_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8060_cast_fp16 = add(x = var_8054_cast_fp16, y = var_8058_cast_fp16)[name = string("op_8060_cast_fp16")];
            fp32 var_8061_epsilon_0 = const()[name = string("op_8061_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8061_cast_fp16 = rsqrt(epsilon = var_8061_epsilon_0, x = var_8060_cast_fp16)[name = string("op_8061_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_251_cast_fp16 = mul(x = y_287_cast_fp16, y = var_8061_cast_fp16)[name = string("out_251_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103778112)))];
            tensor<fp16, [1, 128, 512]> var_8063_cast_fp16 = mul(x = out_251_cast_fp16, y = tm_layers_17_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_213_cast_fp16 = add(x = x_317_cast_fp16, y = var_8063_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103779200))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041408))))[name = string("tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_161_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_161_cast_fp16")];
            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_321_cast_fp16 = gelu(mode = x_321_mode_0, x = linear_161_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [4]> var_8083_begin_0 = const()[name = string("op_8083_begin_0"), val = tensor<int32, [4]>([0, 0, 17, 0])];
            tensor<int32, [4]> var_8083_end_0 = const()[name = string("op_8083_end_0"), val = tensor<int32, [4]>([1, 128, 18, 512])];
            tensor<bool, [4]> var_8083_end_mask_0 = const()[name = string("op_8083_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8083_squeeze_mask_0 = const()[name = string("op_8083_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_8083_cast_fp16 = slice_by_index(begin = var_8083_begin_0, end = var_8083_end_0, end_mask = var_8083_end_mask_0, squeeze_mask = var_8083_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8083_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_215_cast_fp16 = mul(x = x_321_cast_fp16, y = var_8083_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104042496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104304704))))[name = string("tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8093_cast_fp16 = abs(x = linear_162_cast_fp16)[name = string("op_8093_cast_fp16")];
            bool var_8096 = const()[name = string("op_8096"), val = bool(true)];
            tensor<int32, [1]> var_8097_axes_0 = const()[name = string("op_8097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8097_cast_fp16 = reduce_max(axes = var_8097_axes_0, keep_dims = var_8096, x = var_8093_cast_fp16)[name = string("op_8097_cast_fp16")];
            fp16 var_8098_to_fp16 = const()[name = string("op_8098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_289_cast_fp16 = maximum(x = var_8097_cast_fp16, y = var_8098_to_fp16)[name = string("s_289_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_289_cast_fp16 = real_div(x = linear_162_cast_fp16, y = s_289_cast_fp16)[name = string("y_289_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8101_cast_fp16 = mul(x = y_289_cast_fp16, y = y_289_cast_fp16)[name = string("op_8101_cast_fp16")];
            tensor<int32, [1]> var_8106_axes_0 = const()[name = string("op_8106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8106_keep_dims_0 = const()[name = string("op_8106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8106_cast_fp16 = reduce_mean(axes = var_8106_axes_0, keep_dims = var_8106_keep_dims_0, x = var_8101_cast_fp16)[name = string("op_8106_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8107_cast_fp16 = mul(x = s_289_cast_fp16, y = s_289_cast_fp16)[name = string("op_8107_cast_fp16")];
            fp32 var_8108_epsilon_0 = const()[name = string("op_8108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8108_cast_fp16 = inverse(epsilon = var_8108_epsilon_0, x = var_8107_cast_fp16)[name = string("op_8108_cast_fp16")];
            fp16 var_8109_to_fp16 = const()[name = string("op_8109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8110_cast_fp16 = mul(x = var_8108_cast_fp16, y = var_8109_to_fp16)[name = string("op_8110_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8112_cast_fp16 = add(x = var_8106_cast_fp16, y = var_8110_cast_fp16)[name = string("op_8112_cast_fp16")];
            fp32 var_8113_epsilon_0 = const()[name = string("op_8113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8113_cast_fp16 = rsqrt(epsilon = var_8113_epsilon_0, x = var_8112_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_253_cast_fp16 = mul(x = y_289_cast_fp16, y = var_8113_cast_fp16)[name = string("out_253_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104305792)))];
            tensor<fp16, [1, 128, 512]> var_8115_cast_fp16 = mul(x = out_253_cast_fp16, y = tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8115_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_35_cast_fp16 = add(x = input_213_cast_fp16, y = var_8115_cast_fp16)[name = string("h_35_cast_fp16")];
            tensor<fp16, [1]> tm_layers_17_layer_scalar_to_fp16 = const()[name = string("tm_layers_17_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e6p-1])];
            tensor<fp16, [1, 128, 512]> x_325_cast_fp16 = mul(x = h_35_cast_fp16, y = tm_layers_17_layer_scalar_to_fp16)[name = string("x_325_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8119_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_8119_cast_fp16")];
            bool var_8122 = const()[name = string("op_8122"), val = bool(true)];
            tensor<int32, [1]> var_8123_axes_0 = const()[name = string("op_8123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8123_cast_fp16 = reduce_max(axes = var_8123_axes_0, keep_dims = var_8122, x = var_8119_cast_fp16)[name = string("op_8123_cast_fp16")];
            fp16 var_8124_to_fp16 = const()[name = string("op_8124_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_291_cast_fp16 = maximum(x = var_8123_cast_fp16, y = var_8124_to_fp16)[name = string("s_291_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_291_cast_fp16 = real_div(x = x_325_cast_fp16, y = s_291_cast_fp16)[name = string("y_291_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8127_cast_fp16 = mul(x = y_291_cast_fp16, y = y_291_cast_fp16)[name = string("op_8127_cast_fp16")];
            tensor<int32, [1]> var_8132_axes_0 = const()[name = string("op_8132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8132_keep_dims_0 = const()[name = string("op_8132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8132_cast_fp16 = reduce_mean(axes = var_8132_axes_0, keep_dims = var_8132_keep_dims_0, x = var_8127_cast_fp16)[name = string("op_8132_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8133_cast_fp16 = mul(x = s_291_cast_fp16, y = s_291_cast_fp16)[name = string("op_8133_cast_fp16")];
            fp32 var_8134_epsilon_0 = const()[name = string("op_8134_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8134_cast_fp16 = inverse(epsilon = var_8134_epsilon_0, x = var_8133_cast_fp16)[name = string("op_8134_cast_fp16")];
            fp16 var_8135_to_fp16 = const()[name = string("op_8135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = string("op_8136_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8138_cast_fp16 = add(x = var_8132_cast_fp16, y = var_8136_cast_fp16)[name = string("op_8138_cast_fp16")];
            fp32 var_8139_epsilon_0 = const()[name = string("op_8139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8139_cast_fp16 = rsqrt(epsilon = var_8139_epsilon_0, x = var_8138_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_255_cast_fp16 = mul(x = y_291_cast_fp16, y = var_8139_cast_fp16)[name = string("out_255_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104306880)))];
            tensor<fp16, [1, 128, 512]> input_217_cast_fp16 = mul(x = out_255_cast_fp16, y = tm_layers_18_input_layernorm_weight_to_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104307968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104832320))))[name = string("tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8149 = const()[name = string("op_8149"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_8150_cast_fp16 = reshape(shape = var_8149, x = linear_163_cast_fp16)[name = string("op_8150_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104834432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105096640))))[name = string("tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [4]> var_8161 = const()[name = string("op_8161"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_8162_cast_fp16 = reshape(shape = var_8161, x = linear_164_cast_fp16)[name = string("op_8162_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105097728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105359936))))[name = string("tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<int32, [4]> var_8173 = const()[name = string("op_8173"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_8174_cast_fp16 = reshape(shape = var_8173, x = linear_165_cast_fp16)[name = string("op_8174_cast_fp16")];
            tensor<int32, [4]> x_331_perm_0 = const()[name = string("x_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_8150_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 4, 128, 256]> var_8178_cast_fp16 = abs(x = x_327_cast_fp16)[name = string("op_8178_cast_fp16")];
            bool var_8181 = const()[name = string("op_8181"), val = bool(true)];
            tensor<int32, [1]> var_8182_axes_0 = const()[name = string("op_8182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_8182_cast_fp16 = reduce_max(axes = var_8182_axes_0, keep_dims = var_8181, x = var_8178_cast_fp16)[name = string("op_8182_cast_fp16")];
            fp16 var_8183_to_fp16 = const()[name = string("op_8183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_293_cast_fp16 = maximum(x = var_8182_cast_fp16, y = var_8183_to_fp16)[name = string("s_293_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_293_cast_fp16 = real_div(x = x_327_cast_fp16, y = s_293_cast_fp16)[name = string("y_293_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8186_cast_fp16 = mul(x = y_293_cast_fp16, y = y_293_cast_fp16)[name = string("op_8186_cast_fp16")];
            tensor<int32, [1]> var_8191_axes_0 = const()[name = string("op_8191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8191_keep_dims_0 = const()[name = string("op_8191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_8191_cast_fp16 = reduce_mean(axes = var_8191_axes_0, keep_dims = var_8191_keep_dims_0, x = var_8186_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_8192_cast_fp16 = mul(x = s_293_cast_fp16, y = s_293_cast_fp16)[name = string("op_8192_cast_fp16")];
            fp32 var_8193_epsilon_0 = const()[name = string("op_8193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_8193_cast_fp16 = inverse(epsilon = var_8193_epsilon_0, x = var_8192_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_8197_cast_fp16 = add(x = var_8191_cast_fp16, y = var_8195_cast_fp16)[name = string("op_8197_cast_fp16")];
            fp32 var_8198_epsilon_0 = const()[name = string("op_8198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_8198_cast_fp16 = rsqrt(epsilon = var_8198_epsilon_0, x = var_8197_cast_fp16)[name = string("op_8198_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_257_cast_fp16 = mul(x = y_293_cast_fp16, y = var_8198_cast_fp16)[name = string("out_257_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_73_cast_fp16 = mul(x = out_257_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_73_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_8162_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 128, 256]> var_8201_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_8201_cast_fp16")];
            bool var_8204 = const()[name = string("op_8204"), val = bool(true)];
            tensor<int32, [1]> var_8205_axes_0 = const()[name = string("op_8205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_8205_cast_fp16 = reduce_max(axes = var_8205_axes_0, keep_dims = var_8204, x = var_8201_cast_fp16)[name = string("op_8205_cast_fp16")];
            fp16 var_8206_to_fp16 = const()[name = string("op_8206_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_295_cast_fp16 = maximum(x = var_8205_cast_fp16, y = var_8206_to_fp16)[name = string("s_295_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_295_cast_fp16 = real_div(x = x_329_cast_fp16, y = s_295_cast_fp16)[name = string("y_295_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8209_cast_fp16 = mul(x = y_295_cast_fp16, y = y_295_cast_fp16)[name = string("op_8209_cast_fp16")];
            tensor<int32, [1]> var_8214_axes_0 = const()[name = string("op_8214_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8214_keep_dims_0 = const()[name = string("op_8214_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_8214_cast_fp16 = reduce_mean(axes = var_8214_axes_0, keep_dims = var_8214_keep_dims_0, x = var_8209_cast_fp16)[name = string("op_8214_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8215_cast_fp16 = mul(x = s_295_cast_fp16, y = s_295_cast_fp16)[name = string("op_8215_cast_fp16")];
            fp32 var_8216_epsilon_0 = const()[name = string("op_8216_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_8216_cast_fp16 = inverse(epsilon = var_8216_epsilon_0, x = var_8215_cast_fp16)[name = string("op_8216_cast_fp16")];
            fp16 var_8217_to_fp16 = const()[name = string("op_8217_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_8218_cast_fp16 = mul(x = var_8216_cast_fp16, y = var_8217_to_fp16)[name = string("op_8218_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8220_cast_fp16 = add(x = var_8214_cast_fp16, y = var_8218_cast_fp16)[name = string("op_8220_cast_fp16")];
            fp32 var_8221_epsilon_0 = const()[name = string("op_8221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_8221_cast_fp16 = rsqrt(epsilon = var_8221_epsilon_0, x = var_8220_cast_fp16)[name = string("op_8221_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_259_cast_fp16 = mul(x = y_295_cast_fp16, y = var_8221_cast_fp16)[name = string("out_259_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_109_cast_fp16 = mul(x = out_259_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_109_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_331_cast_fp16 = transpose(perm = x_331_perm_0, x = var_8174_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 128, 256]> var_8224_cast_fp16 = abs(x = x_331_cast_fp16)[name = string("op_8224_cast_fp16")];
            bool var_8227 = const()[name = string("op_8227"), val = bool(true)];
            tensor<int32, [1]> var_8228_axes_0 = const()[name = string("op_8228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_8228_cast_fp16 = reduce_max(axes = var_8228_axes_0, keep_dims = var_8227, x = var_8224_cast_fp16)[name = string("op_8228_cast_fp16")];
            fp16 var_8229_to_fp16 = const()[name = string("op_8229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_297_cast_fp16 = maximum(x = var_8228_cast_fp16, y = var_8229_to_fp16)[name = string("s_297_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_297_cast_fp16 = real_div(x = x_331_cast_fp16, y = s_297_cast_fp16)[name = string("y_297_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8232_cast_fp16 = mul(x = y_297_cast_fp16, y = y_297_cast_fp16)[name = string("op_8232_cast_fp16")];
            tensor<int32, [1]> var_8237_axes_0 = const()[name = string("op_8237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8237_keep_dims_0 = const()[name = string("op_8237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_8237_cast_fp16 = reduce_mean(axes = var_8237_axes_0, keep_dims = var_8237_keep_dims_0, x = var_8232_cast_fp16)[name = string("op_8237_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8238_cast_fp16 = mul(x = s_297_cast_fp16, y = s_297_cast_fp16)[name = string("op_8238_cast_fp16")];
            fp32 var_8239_epsilon_0 = const()[name = string("op_8239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_8239_cast_fp16 = inverse(epsilon = var_8239_epsilon_0, x = var_8238_cast_fp16)[name = string("op_8239_cast_fp16")];
            fp16 var_8240_to_fp16 = const()[name = string("op_8240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_8241_cast_fp16 = mul(x = var_8239_cast_fp16, y = var_8240_to_fp16)[name = string("op_8241_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8243_cast_fp16 = add(x = var_8237_cast_fp16, y = var_8241_cast_fp16)[name = string("op_8243_cast_fp16")];
            fp32 var_8244_epsilon_0 = const()[name = string("op_8244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_8244_cast_fp16 = rsqrt(epsilon = var_8244_epsilon_0, x = var_8243_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_73_cast_fp16 = mul(x = y_297_cast_fp16, y = var_8244_cast_fp16)[name = string("v_73_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8246_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8246_cast_fp16")];
            tensor<int32, [4]> var_8251_begin_0 = const()[name = string("op_8251_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8251_end_0 = const()[name = string("op_8251_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_8251_end_mask_0 = const()[name = string("op_8251_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_8251_cast_fp16 = slice_by_index(begin = var_8251_begin_0, end = var_8251_end_0, end_mask = var_8251_end_mask_0, x = q_73_cast_fp16)[name = string("op_8251_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_8252_cast_fp16 = mul(x = var_8251_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8257_begin_0 = const()[name = string("op_8257_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8257_end_0 = const()[name = string("op_8257_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_8257_end_mask_0 = const()[name = string("op_8257_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_8257_cast_fp16 = slice_by_index(begin = var_8257_begin_0, end = var_8257_end_0, end_mask = var_8257_end_mask_0, x = q_73_cast_fp16)[name = string("op_8257_cast_fp16")];
            int32 var_8259 = const()[name = string("op_8259"), val = int32(-1)];
            bool var_8260_interleave_0 = const()[name = string("op_8260_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_8260_cast_fp16 = concat(axis = var_8259, interleave = var_8260_interleave_0, values = (var_8252_cast_fp16, var_8257_cast_fp16))[name = string("op_8260_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8261_cast_fp16 = mul(x = var_8260_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8261_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_75_cast_fp16 = add(x = var_8246_cast_fp16, y = var_8261_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8264_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8264_cast_fp16")];
            tensor<int32, [4]> var_8269_begin_0 = const()[name = string("op_8269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8269_end_0 = const()[name = string("op_8269_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_8269_end_mask_0 = const()[name = string("op_8269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_8269_cast_fp16 = slice_by_index(begin = var_8269_begin_0, end = var_8269_end_0, end_mask = var_8269_end_mask_0, x = k_109_cast_fp16)[name = string("op_8269_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_8270_cast_fp16 = mul(x = var_8269_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_8270_cast_fp16")];
            tensor<int32, [4]> var_8275_begin_0 = const()[name = string("op_8275_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8275_end_0 = const()[name = string("op_8275_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_8275_end_mask_0 = const()[name = string("op_8275_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_8275_cast_fp16 = slice_by_index(begin = var_8275_begin_0, end = var_8275_end_0, end_mask = var_8275_end_mask_0, x = k_109_cast_fp16)[name = string("op_8275_cast_fp16")];
            int32 var_8277 = const()[name = string("op_8277"), val = int32(-1)];
            bool var_8278_interleave_0 = const()[name = string("op_8278_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_8278_cast_fp16 = concat(axis = var_8277, interleave = var_8278_interleave_0, values = (var_8270_cast_fp16, var_8275_cast_fp16))[name = string("op_8278_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8279_cast_fp16 = mul(x = var_8278_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8279_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_111_cast_fp16 = add(x = var_8264_cast_fp16, y = var_8279_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_36_reps_0 = const()[name = string("tile_36_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_72_cast_fp16 = transpose(perm = transpose_72_perm_0, x = k_111_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 1, 128, 256]> tile_36_cast_fp16 = tile(reps = tile_36_reps_0, x = transpose_72_cast_fp16)[name = string("tile_36_cast_fp16")];
            tensor<int32, [5]> concat_72 = const()[name = string("concat_72"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_72_cast_fp16 = reshape(shape = concat_72, x = tile_36_cast_fp16)[name = string("reshape_72_cast_fp16")];
            tensor<int32, [5]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_73 = const()[name = string("concat_73"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = reshape_72_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 1, 128, 256]> reshape_73_cast_fp16 = reshape(shape = concat_73, x = transpose_73_cast_fp16)[name = string("reshape_73_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_37_reps_0 = const()[name = string("tile_37_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = v_73_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 1, 128, 256]> tile_37_cast_fp16 = tile(reps = tile_37_reps_0, x = transpose_74_cast_fp16)[name = string("tile_37_cast_fp16")];
            tensor<int32, [5]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_74_cast_fp16 = reshape(shape = concat_74, x = tile_37_cast_fp16)[name = string("reshape_74_cast_fp16")];
            tensor<int32, [5]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = reshape_74_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 1, 128, 256]> reshape_75_cast_fp16 = reshape(shape = concat_75, x = transpose_75_cast_fp16)[name = string("reshape_75_cast_fp16")];
            tensor<int32, [4]> v_75_perm_0 = const()[name = string("v_75_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8293_transpose_x_0 = const()[name = string("op_8293_transpose_x_0"), val = bool(false)];
            bool var_8293_transpose_y_0 = const()[name = string("op_8293_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = reshape_73_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 4, 128, 128]> var_8293_cast_fp16 = matmul(transpose_x = var_8293_transpose_x_0, transpose_y = var_8293_transpose_y_0, x = q_75_cast_fp16, y = transpose_114_cast_fp16)[name = string("op_8293_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_8295_cast_fp16 = add(x = var_8293_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8295_cast_fp16")];
            int32 var_8296 = const()[name = string("op_8296"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_297_cast_fp16 = softmax(axis = var_8296, x = var_8295_cast_fp16)[name = string("w_297_cast_fp16")];
            bool var_8299_transpose_x_0 = const()[name = string("op_8299_transpose_x_0"), val = bool(false)];
            bool var_8299_transpose_y_0 = const()[name = string("op_8299_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_75_cast_fp16 = transpose(perm = v_75_perm_0, x = reshape_75_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 4, 128, 256]> var_8299_cast_fp16 = matmul(transpose_x = var_8299_transpose_x_0, transpose_y = var_8299_transpose_y_0, x = w_297_cast_fp16, y = v_75_cast_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [4]> var_8302_perm_0 = const()[name = string("op_8302_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8306 = const()[name = string("op_8306"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_8302_cast_fp16 = transpose(perm = var_8302_perm_0, x = var_8299_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 128, 1024]> input_219_cast_fp16 = reshape(shape = var_8306, x = var_8302_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105361024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105885376))))[name = string("tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8311_cast_fp16 = abs(x = linear_166_cast_fp16)[name = string("op_8311_cast_fp16")];
            bool var_8314 = const()[name = string("op_8314"), val = bool(true)];
            tensor<int32, [1]> var_8315_axes_0 = const()[name = string("op_8315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8315_cast_fp16 = reduce_max(axes = var_8315_axes_0, keep_dims = var_8314, x = var_8311_cast_fp16)[name = string("op_8315_cast_fp16")];
            fp16 var_8316_to_fp16 = const()[name = string("op_8316_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_299_cast_fp16 = maximum(x = var_8315_cast_fp16, y = var_8316_to_fp16)[name = string("s_299_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_299_cast_fp16 = real_div(x = linear_166_cast_fp16, y = s_299_cast_fp16)[name = string("y_299_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8319_cast_fp16 = mul(x = y_299_cast_fp16, y = y_299_cast_fp16)[name = string("op_8319_cast_fp16")];
            tensor<int32, [1]> var_8324_axes_0 = const()[name = string("op_8324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8324_keep_dims_0 = const()[name = string("op_8324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8324_cast_fp16 = reduce_mean(axes = var_8324_axes_0, keep_dims = var_8324_keep_dims_0, x = var_8319_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8325_cast_fp16 = mul(x = s_299_cast_fp16, y = s_299_cast_fp16)[name = string("op_8325_cast_fp16")];
            fp32 var_8326_epsilon_0 = const()[name = string("op_8326_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8326_cast_fp16 = inverse(epsilon = var_8326_epsilon_0, x = var_8325_cast_fp16)[name = string("op_8326_cast_fp16")];
            fp16 var_8327_to_fp16 = const()[name = string("op_8327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8328_cast_fp16 = mul(x = var_8326_cast_fp16, y = var_8327_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8330_cast_fp16 = add(x = var_8324_cast_fp16, y = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_261_cast_fp16 = mul(x = y_299_cast_fp16, y = var_8331_cast_fp16)[name = string("out_261_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105886464)))];
            tensor<fp16, [1, 128, 512]> var_8333_cast_fp16 = mul(x = out_261_cast_fp16, y = tm_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("op_8333_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = var_8333_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8336_cast_fp16 = abs(x = x_335_cast_fp16)[name = string("op_8336_cast_fp16")];
            bool var_8339 = const()[name = string("op_8339"), val = bool(true)];
            tensor<int32, [1]> var_8340_axes_0 = const()[name = string("op_8340_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8340_cast_fp16 = reduce_max(axes = var_8340_axes_0, keep_dims = var_8339, x = var_8336_cast_fp16)[name = string("op_8340_cast_fp16")];
            fp16 var_8341_to_fp16 = const()[name = string("op_8341_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_301_cast_fp16 = maximum(x = var_8340_cast_fp16, y = var_8341_to_fp16)[name = string("s_301_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_301_cast_fp16 = real_div(x = x_335_cast_fp16, y = s_301_cast_fp16)[name = string("y_301_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8344_cast_fp16 = mul(x = y_301_cast_fp16, y = y_301_cast_fp16)[name = string("op_8344_cast_fp16")];
            tensor<int32, [1]> var_8349_axes_0 = const()[name = string("op_8349_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8349_keep_dims_0 = const()[name = string("op_8349_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8349_cast_fp16 = reduce_mean(axes = var_8349_axes_0, keep_dims = var_8349_keep_dims_0, x = var_8344_cast_fp16)[name = string("op_8349_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8350_cast_fp16 = mul(x = s_301_cast_fp16, y = s_301_cast_fp16)[name = string("op_8350_cast_fp16")];
            fp32 var_8351_epsilon_0 = const()[name = string("op_8351_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8351_cast_fp16 = inverse(epsilon = var_8351_epsilon_0, x = var_8350_cast_fp16)[name = string("op_8351_cast_fp16")];
            fp16 var_8352_to_fp16 = const()[name = string("op_8352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8353_cast_fp16 = mul(x = var_8351_cast_fp16, y = var_8352_to_fp16)[name = string("op_8353_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8355_cast_fp16 = add(x = var_8349_cast_fp16, y = var_8353_cast_fp16)[name = string("op_8355_cast_fp16")];
            fp32 var_8356_epsilon_0 = const()[name = string("op_8356_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8356_cast_fp16 = rsqrt(epsilon = var_8356_epsilon_0, x = var_8355_cast_fp16)[name = string("op_8356_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_263_cast_fp16 = mul(x = y_301_cast_fp16, y = var_8356_cast_fp16)[name = string("out_263_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105887552)))];
            tensor<fp16, [1, 128, 512]> input_221_cast_fp16 = mul(x = out_263_cast_fp16, y = tm_layers_18_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105888640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106937280))))[name = string("tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_167_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_167_cast_fp16")];
            string var_8363_mode_0 = const()[name = string("op_8363_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_8363_cast_fp16 = gelu(mode = var_8363_mode_0, x = linear_167_cast_fp16)[name = string("op_8363_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106941440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107990080))))[name = string("tm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_168_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_223_cast_fp16 = mul(x = var_8363_cast_fp16, y = linear_168_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107994240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109042880))))[name = string("tm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8371_cast_fp16 = abs(x = linear_169_cast_fp16)[name = string("op_8371_cast_fp16")];
            bool var_8374 = const()[name = string("op_8374"), val = bool(true)];
            tensor<int32, [1]> var_8375_axes_0 = const()[name = string("op_8375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8375_cast_fp16 = reduce_max(axes = var_8375_axes_0, keep_dims = var_8374, x = var_8371_cast_fp16)[name = string("op_8375_cast_fp16")];
            fp16 var_8376_to_fp16 = const()[name = string("op_8376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_303_cast_fp16 = maximum(x = var_8375_cast_fp16, y = var_8376_to_fp16)[name = string("s_303_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_303_cast_fp16 = real_div(x = linear_169_cast_fp16, y = s_303_cast_fp16)[name = string("y_303_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8379_cast_fp16 = mul(x = y_303_cast_fp16, y = y_303_cast_fp16)[name = string("op_8379_cast_fp16")];
            tensor<int32, [1]> var_8384_axes_0 = const()[name = string("op_8384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8384_keep_dims_0 = const()[name = string("op_8384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8384_cast_fp16 = reduce_mean(axes = var_8384_axes_0, keep_dims = var_8384_keep_dims_0, x = var_8379_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8385_cast_fp16 = mul(x = s_303_cast_fp16, y = s_303_cast_fp16)[name = string("op_8385_cast_fp16")];
            fp32 var_8386_epsilon_0 = const()[name = string("op_8386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8386_cast_fp16 = inverse(epsilon = var_8386_epsilon_0, x = var_8385_cast_fp16)[name = string("op_8386_cast_fp16")];
            fp16 var_8387_to_fp16 = const()[name = string("op_8387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8388_cast_fp16 = mul(x = var_8386_cast_fp16, y = var_8387_to_fp16)[name = string("op_8388_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8390_cast_fp16 = add(x = var_8384_cast_fp16, y = var_8388_cast_fp16)[name = string("op_8390_cast_fp16")];
            fp32 var_8391_epsilon_0 = const()[name = string("op_8391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8391_cast_fp16 = rsqrt(epsilon = var_8391_epsilon_0, x = var_8390_cast_fp16)[name = string("op_8391_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_265_cast_fp16 = mul(x = y_303_cast_fp16, y = var_8391_cast_fp16)[name = string("out_265_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109043968)))];
            tensor<fp16, [1, 128, 512]> var_8393_cast_fp16 = mul(x = out_265_cast_fp16, y = tm_layers_18_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8393_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_225_cast_fp16 = add(x = x_335_cast_fp16, y = var_8393_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109045056))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109307264))))[name = string("tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_170_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_170_cast_fp16")];
            string x_339_mode_0 = const()[name = string("x_339_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_339_cast_fp16 = gelu(mode = x_339_mode_0, x = linear_170_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [4]> var_8413_begin_0 = const()[name = string("op_8413_begin_0"), val = tensor<int32, [4]>([0, 0, 18, 0])];
            tensor<int32, [4]> var_8413_end_0 = const()[name = string("op_8413_end_0"), val = tensor<int32, [4]>([1, 128, 19, 512])];
            tensor<bool, [4]> var_8413_end_mask_0 = const()[name = string("op_8413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8413_squeeze_mask_0 = const()[name = string("op_8413_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_8413_cast_fp16 = slice_by_index(begin = var_8413_begin_0, end = var_8413_end_0, end_mask = var_8413_end_mask_0, squeeze_mask = var_8413_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_227_cast_fp16 = mul(x = x_339_cast_fp16, y = var_8413_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109308352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109570560))))[name = string("tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8423_cast_fp16 = abs(x = linear_171_cast_fp16)[name = string("op_8423_cast_fp16")];
            bool var_8426 = const()[name = string("op_8426"), val = bool(true)];
            tensor<int32, [1]> var_8427_axes_0 = const()[name = string("op_8427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8427_cast_fp16 = reduce_max(axes = var_8427_axes_0, keep_dims = var_8426, x = var_8423_cast_fp16)[name = string("op_8427_cast_fp16")];
            fp16 var_8428_to_fp16 = const()[name = string("op_8428_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_305_cast_fp16 = maximum(x = var_8427_cast_fp16, y = var_8428_to_fp16)[name = string("s_305_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_305_cast_fp16 = real_div(x = linear_171_cast_fp16, y = s_305_cast_fp16)[name = string("y_305_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8431_cast_fp16 = mul(x = y_305_cast_fp16, y = y_305_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<int32, [1]> var_8436_axes_0 = const()[name = string("op_8436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8436_keep_dims_0 = const()[name = string("op_8436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8436_cast_fp16 = reduce_mean(axes = var_8436_axes_0, keep_dims = var_8436_keep_dims_0, x = var_8431_cast_fp16)[name = string("op_8436_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8437_cast_fp16 = mul(x = s_305_cast_fp16, y = s_305_cast_fp16)[name = string("op_8437_cast_fp16")];
            fp32 var_8438_epsilon_0 = const()[name = string("op_8438_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8438_cast_fp16 = inverse(epsilon = var_8438_epsilon_0, x = var_8437_cast_fp16)[name = string("op_8438_cast_fp16")];
            fp16 var_8439_to_fp16 = const()[name = string("op_8439_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8440_cast_fp16 = mul(x = var_8438_cast_fp16, y = var_8439_to_fp16)[name = string("op_8440_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8442_cast_fp16 = add(x = var_8436_cast_fp16, y = var_8440_cast_fp16)[name = string("op_8442_cast_fp16")];
            fp32 var_8443_epsilon_0 = const()[name = string("op_8443_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8443_cast_fp16 = rsqrt(epsilon = var_8443_epsilon_0, x = var_8442_cast_fp16)[name = string("op_8443_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_267_cast_fp16 = mul(x = y_305_cast_fp16, y = var_8443_cast_fp16)[name = string("out_267_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109571648)))];
            tensor<fp16, [1, 128, 512]> var_8445_cast_fp16 = mul(x = out_267_cast_fp16, y = tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8445_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_37_cast_fp16 = add(x = input_225_cast_fp16, y = var_8445_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<fp16, [1]> tm_layers_18_layer_scalar_to_fp16 = const()[name = string("tm_layers_18_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.68p-1])];
            tensor<fp16, [1, 128, 512]> x_343_cast_fp16 = mul(x = h_37_cast_fp16, y = tm_layers_18_layer_scalar_to_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8449_cast_fp16 = abs(x = x_343_cast_fp16)[name = string("op_8449_cast_fp16")];
            bool var_8452 = const()[name = string("op_8452"), val = bool(true)];
            tensor<int32, [1]> var_8453_axes_0 = const()[name = string("op_8453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8453_cast_fp16 = reduce_max(axes = var_8453_axes_0, keep_dims = var_8452, x = var_8449_cast_fp16)[name = string("op_8453_cast_fp16")];
            fp16 var_8454_to_fp16 = const()[name = string("op_8454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_307_cast_fp16 = maximum(x = var_8453_cast_fp16, y = var_8454_to_fp16)[name = string("s_307_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_307_cast_fp16 = real_div(x = x_343_cast_fp16, y = s_307_cast_fp16)[name = string("y_307_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8457_cast_fp16 = mul(x = y_307_cast_fp16, y = y_307_cast_fp16)[name = string("op_8457_cast_fp16")];
            tensor<int32, [1]> var_8462_axes_0 = const()[name = string("op_8462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8462_keep_dims_0 = const()[name = string("op_8462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8462_cast_fp16 = reduce_mean(axes = var_8462_axes_0, keep_dims = var_8462_keep_dims_0, x = var_8457_cast_fp16)[name = string("op_8462_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8463_cast_fp16 = mul(x = s_307_cast_fp16, y = s_307_cast_fp16)[name = string("op_8463_cast_fp16")];
            fp32 var_8464_epsilon_0 = const()[name = string("op_8464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8464_cast_fp16 = inverse(epsilon = var_8464_epsilon_0, x = var_8463_cast_fp16)[name = string("op_8464_cast_fp16")];
            fp16 var_8465_to_fp16 = const()[name = string("op_8465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8466_cast_fp16 = mul(x = var_8464_cast_fp16, y = var_8465_to_fp16)[name = string("op_8466_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8468_cast_fp16 = add(x = var_8462_cast_fp16, y = var_8466_cast_fp16)[name = string("op_8468_cast_fp16")];
            fp32 var_8469_epsilon_0 = const()[name = string("op_8469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8469_cast_fp16 = rsqrt(epsilon = var_8469_epsilon_0, x = var_8468_cast_fp16)[name = string("op_8469_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_269_cast_fp16 = mul(x = y_307_cast_fp16, y = var_8469_cast_fp16)[name = string("out_269_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109572736)))];
            tensor<fp16, [1, 128, 512]> input_229_cast_fp16 = mul(x = out_269_cast_fp16, y = tm_layers_19_input_layernorm_weight_to_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109573824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110098176))))[name = string("tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [4]> var_8479 = const()[name = string("op_8479"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_8480_cast_fp16 = reshape(shape = var_8479, x = linear_172_cast_fp16)[name = string("op_8480_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110100288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110362496))))[name = string("tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<int32, [4]> var_8491 = const()[name = string("op_8491"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_8492_cast_fp16 = reshape(shape = var_8491, x = linear_173_cast_fp16)[name = string("op_8492_cast_fp16")];
            tensor<int32, [4]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110363584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110625792))))[name = string("tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<int32, [4]> var_8503 = const()[name = string("op_8503"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_8504_cast_fp16 = reshape(shape = var_8503, x = linear_174_cast_fp16)[name = string("op_8504_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_8480_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 4, 128, 256]> var_8508_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_8508_cast_fp16")];
            bool var_8511 = const()[name = string("op_8511"), val = bool(true)];
            tensor<int32, [1]> var_8512_axes_0 = const()[name = string("op_8512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_8512_cast_fp16 = reduce_max(axes = var_8512_axes_0, keep_dims = var_8511, x = var_8508_cast_fp16)[name = string("op_8512_cast_fp16")];
            fp16 var_8513_to_fp16 = const()[name = string("op_8513_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_309_cast_fp16 = maximum(x = var_8512_cast_fp16, y = var_8513_to_fp16)[name = string("s_309_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_309_cast_fp16 = real_div(x = x_345_cast_fp16, y = s_309_cast_fp16)[name = string("y_309_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8516_cast_fp16 = mul(x = y_309_cast_fp16, y = y_309_cast_fp16)[name = string("op_8516_cast_fp16")];
            tensor<int32, [1]> var_8521_axes_0 = const()[name = string("op_8521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8521_keep_dims_0 = const()[name = string("op_8521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_8521_cast_fp16 = reduce_mean(axes = var_8521_axes_0, keep_dims = var_8521_keep_dims_0, x = var_8516_cast_fp16)[name = string("op_8521_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_8522_cast_fp16 = mul(x = s_309_cast_fp16, y = s_309_cast_fp16)[name = string("op_8522_cast_fp16")];
            fp32 var_8523_epsilon_0 = const()[name = string("op_8523_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_8523_cast_fp16 = inverse(epsilon = var_8523_epsilon_0, x = var_8522_cast_fp16)[name = string("op_8523_cast_fp16")];
            fp16 var_8524_to_fp16 = const()[name = string("op_8524_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_8525_cast_fp16 = mul(x = var_8523_cast_fp16, y = var_8524_to_fp16)[name = string("op_8525_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_8527_cast_fp16 = add(x = var_8521_cast_fp16, y = var_8525_cast_fp16)[name = string("op_8527_cast_fp16")];
            fp32 var_8528_epsilon_0 = const()[name = string("op_8528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_8528_cast_fp16 = rsqrt(epsilon = var_8528_epsilon_0, x = var_8527_cast_fp16)[name = string("op_8528_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_271_cast_fp16 = mul(x = y_309_cast_fp16, y = var_8528_cast_fp16)[name = string("out_271_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_77_cast_fp16 = mul(x = out_271_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_77_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = var_8492_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 128, 256]> var_8531_cast_fp16 = abs(x = x_347_cast_fp16)[name = string("op_8531_cast_fp16")];
            bool var_8534 = const()[name = string("op_8534"), val = bool(true)];
            tensor<int32, [1]> var_8535_axes_0 = const()[name = string("op_8535_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_8535_cast_fp16 = reduce_max(axes = var_8535_axes_0, keep_dims = var_8534, x = var_8531_cast_fp16)[name = string("op_8535_cast_fp16")];
            fp16 var_8536_to_fp16 = const()[name = string("op_8536_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_311_cast_fp16 = maximum(x = var_8535_cast_fp16, y = var_8536_to_fp16)[name = string("s_311_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_311_cast_fp16 = real_div(x = x_347_cast_fp16, y = s_311_cast_fp16)[name = string("y_311_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8539_cast_fp16 = mul(x = y_311_cast_fp16, y = y_311_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<int32, [1]> var_8544_axes_0 = const()[name = string("op_8544_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8544_keep_dims_0 = const()[name = string("op_8544_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_8544_cast_fp16 = reduce_mean(axes = var_8544_axes_0, keep_dims = var_8544_keep_dims_0, x = var_8539_cast_fp16)[name = string("op_8544_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8545_cast_fp16 = mul(x = s_311_cast_fp16, y = s_311_cast_fp16)[name = string("op_8545_cast_fp16")];
            fp32 var_8546_epsilon_0 = const()[name = string("op_8546_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_8546_cast_fp16 = inverse(epsilon = var_8546_epsilon_0, x = var_8545_cast_fp16)[name = string("op_8546_cast_fp16")];
            fp16 var_8547_to_fp16 = const()[name = string("op_8547_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_8548_cast_fp16 = mul(x = var_8546_cast_fp16, y = var_8547_to_fp16)[name = string("op_8548_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8550_cast_fp16 = add(x = var_8544_cast_fp16, y = var_8548_cast_fp16)[name = string("op_8550_cast_fp16")];
            fp32 var_8551_epsilon_0 = const()[name = string("op_8551_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_8551_cast_fp16 = rsqrt(epsilon = var_8551_epsilon_0, x = var_8550_cast_fp16)[name = string("op_8551_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_273_cast_fp16 = mul(x = y_311_cast_fp16, y = var_8551_cast_fp16)[name = string("out_273_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_115_cast_fp16 = mul(x = out_273_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_115_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_8504_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 2, 128, 256]> var_8554_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_8554_cast_fp16")];
            bool var_8557 = const()[name = string("op_8557"), val = bool(true)];
            tensor<int32, [1]> var_8558_axes_0 = const()[name = string("op_8558_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_8558_cast_fp16 = reduce_max(axes = var_8558_axes_0, keep_dims = var_8557, x = var_8554_cast_fp16)[name = string("op_8558_cast_fp16")];
            fp16 var_8559_to_fp16 = const()[name = string("op_8559_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_313_cast_fp16 = maximum(x = var_8558_cast_fp16, y = var_8559_to_fp16)[name = string("s_313_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_313_cast_fp16 = real_div(x = x_349_cast_fp16, y = s_313_cast_fp16)[name = string("y_313_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8562_cast_fp16 = mul(x = y_313_cast_fp16, y = y_313_cast_fp16)[name = string("op_8562_cast_fp16")];
            tensor<int32, [1]> var_8567_axes_0 = const()[name = string("op_8567_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8567_keep_dims_0 = const()[name = string("op_8567_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_8567_cast_fp16 = reduce_mean(axes = var_8567_axes_0, keep_dims = var_8567_keep_dims_0, x = var_8562_cast_fp16)[name = string("op_8567_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8568_cast_fp16 = mul(x = s_313_cast_fp16, y = s_313_cast_fp16)[name = string("op_8568_cast_fp16")];
            fp32 var_8569_epsilon_0 = const()[name = string("op_8569_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_8569_cast_fp16 = inverse(epsilon = var_8569_epsilon_0, x = var_8568_cast_fp16)[name = string("op_8569_cast_fp16")];
            fp16 var_8570_to_fp16 = const()[name = string("op_8570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_8571_cast_fp16 = mul(x = var_8569_cast_fp16, y = var_8570_to_fp16)[name = string("op_8571_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8573_cast_fp16 = add(x = var_8567_cast_fp16, y = var_8571_cast_fp16)[name = string("op_8573_cast_fp16")];
            fp32 var_8574_epsilon_0 = const()[name = string("op_8574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_8574_cast_fp16 = rsqrt(epsilon = var_8574_epsilon_0, x = var_8573_cast_fp16)[name = string("op_8574_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_77_cast_fp16 = mul(x = y_313_cast_fp16, y = var_8574_cast_fp16)[name = string("v_77_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8576_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8576_cast_fp16")];
            tensor<int32, [4]> var_8581_begin_0 = const()[name = string("op_8581_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8581_end_0 = const()[name = string("op_8581_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_8581_end_mask_0 = const()[name = string("op_8581_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_8581_cast_fp16 = slice_by_index(begin = var_8581_begin_0, end = var_8581_end_0, end_mask = var_8581_end_mask_0, x = q_77_cast_fp16)[name = string("op_8581_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_8582_cast_fp16 = mul(x = var_8581_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_8582_cast_fp16")];
            tensor<int32, [4]> var_8587_begin_0 = const()[name = string("op_8587_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8587_end_0 = const()[name = string("op_8587_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_8587_end_mask_0 = const()[name = string("op_8587_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_8587_cast_fp16 = slice_by_index(begin = var_8587_begin_0, end = var_8587_end_0, end_mask = var_8587_end_mask_0, x = q_77_cast_fp16)[name = string("op_8587_cast_fp16")];
            int32 var_8589 = const()[name = string("op_8589"), val = int32(-1)];
            bool var_8590_interleave_0 = const()[name = string("op_8590_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_8590_cast_fp16 = concat(axis = var_8589, interleave = var_8590_interleave_0, values = (var_8582_cast_fp16, var_8587_cast_fp16))[name = string("op_8590_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8591_cast_fp16 = mul(x = var_8590_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8591_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_79_cast_fp16 = add(x = var_8576_cast_fp16, y = var_8591_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8594_cast_fp16 = mul(x = k_115_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8594_cast_fp16")];
            tensor<int32, [4]> var_8599_begin_0 = const()[name = string("op_8599_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8599_end_0 = const()[name = string("op_8599_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_8599_end_mask_0 = const()[name = string("op_8599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_8599_cast_fp16 = slice_by_index(begin = var_8599_begin_0, end = var_8599_end_0, end_mask = var_8599_end_mask_0, x = k_115_cast_fp16)[name = string("op_8599_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_8600_cast_fp16 = mul(x = var_8599_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_8600_cast_fp16")];
            tensor<int32, [4]> var_8605_begin_0 = const()[name = string("op_8605_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8605_end_0 = const()[name = string("op_8605_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_8605_end_mask_0 = const()[name = string("op_8605_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_8605_cast_fp16 = slice_by_index(begin = var_8605_begin_0, end = var_8605_end_0, end_mask = var_8605_end_mask_0, x = k_115_cast_fp16)[name = string("op_8605_cast_fp16")];
            int32 var_8607 = const()[name = string("op_8607"), val = int32(-1)];
            bool var_8608_interleave_0 = const()[name = string("op_8608_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_8608_cast_fp16 = concat(axis = var_8607, interleave = var_8608_interleave_0, values = (var_8600_cast_fp16, var_8605_cast_fp16))[name = string("op_8608_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8609_cast_fp16 = mul(x = var_8608_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8609_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_117_cast_fp16 = add(x = var_8594_cast_fp16, y = var_8609_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_38_reps_0 = const()[name = string("tile_38_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = k_117_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 1, 128, 256]> tile_38_cast_fp16 = tile(reps = tile_38_reps_0, x = transpose_76_cast_fp16)[name = string("tile_38_cast_fp16")];
            tensor<int32, [5]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_76_cast_fp16 = reshape(shape = concat_76, x = tile_38_cast_fp16)[name = string("reshape_76_cast_fp16")];
            tensor<int32, [5]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_77 = const()[name = string("concat_77"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_77_cast_fp16 = transpose(perm = transpose_77_perm_0, x = reshape_76_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 1, 128, 256]> reshape_77_cast_fp16 = reshape(shape = concat_77, x = transpose_77_cast_fp16)[name = string("reshape_77_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_39_reps_0 = const()[name = string("tile_39_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_78_cast_fp16 = transpose(perm = transpose_78_perm_0, x = v_77_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 1, 128, 256]> tile_39_cast_fp16 = tile(reps = tile_39_reps_0, x = transpose_78_cast_fp16)[name = string("tile_39_cast_fp16")];
            tensor<int32, [5]> concat_78 = const()[name = string("concat_78"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_78_cast_fp16 = reshape(shape = concat_78, x = tile_39_cast_fp16)[name = string("reshape_78_cast_fp16")];
            tensor<int32, [5]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = reshape_78_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 1, 128, 256]> reshape_79_cast_fp16 = reshape(shape = concat_79, x = transpose_79_cast_fp16)[name = string("reshape_79_cast_fp16")];
            tensor<int32, [4]> v_79_perm_0 = const()[name = string("v_79_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8623_transpose_x_0 = const()[name = string("op_8623_transpose_x_0"), val = bool(false)];
            bool var_8623_transpose_y_0 = const()[name = string("op_8623_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = reshape_77_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4, 128, 128]> var_8623_cast_fp16 = matmul(transpose_x = var_8623_transpose_x_0, transpose_y = var_8623_transpose_y_0, x = q_79_cast_fp16, y = transpose_115_cast_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8625_cast_fp16")];
            int32 var_8626 = const()[name = string("op_8626"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_313_cast_fp16 = softmax(axis = var_8626, x = var_8625_cast_fp16)[name = string("w_313_cast_fp16")];
            bool var_8629_transpose_x_0 = const()[name = string("op_8629_transpose_x_0"), val = bool(false)];
            bool var_8629_transpose_y_0 = const()[name = string("op_8629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = reshape_79_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 4, 128, 256]> var_8629_cast_fp16 = matmul(transpose_x = var_8629_transpose_x_0, transpose_y = var_8629_transpose_y_0, x = w_313_cast_fp16, y = v_79_cast_fp16)[name = string("op_8629_cast_fp16")];
            tensor<int32, [4]> var_8632_perm_0 = const()[name = string("op_8632_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_8632_cast_fp16 = transpose(perm = var_8632_perm_0, x = var_8629_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 128, 1024]> input_231_cast_fp16 = reshape(shape = var_8636, x = var_8632_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110626880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111151232))))[name = string("tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8641_cast_fp16 = abs(x = linear_175_cast_fp16)[name = string("op_8641_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<int32, [1]> var_8645_axes_0 = const()[name = string("op_8645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8645_cast_fp16 = reduce_max(axes = var_8645_axes_0, keep_dims = var_8644, x = var_8641_cast_fp16)[name = string("op_8645_cast_fp16")];
            fp16 var_8646_to_fp16 = const()[name = string("op_8646_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_315_cast_fp16 = maximum(x = var_8645_cast_fp16, y = var_8646_to_fp16)[name = string("s_315_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_315_cast_fp16 = real_div(x = linear_175_cast_fp16, y = s_315_cast_fp16)[name = string("y_315_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8649_cast_fp16 = mul(x = y_315_cast_fp16, y = y_315_cast_fp16)[name = string("op_8649_cast_fp16")];
            tensor<int32, [1]> var_8654_axes_0 = const()[name = string("op_8654_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8654_keep_dims_0 = const()[name = string("op_8654_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8654_cast_fp16 = reduce_mean(axes = var_8654_axes_0, keep_dims = var_8654_keep_dims_0, x = var_8649_cast_fp16)[name = string("op_8654_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8655_cast_fp16 = mul(x = s_315_cast_fp16, y = s_315_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp32 var_8656_epsilon_0 = const()[name = string("op_8656_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8656_cast_fp16 = inverse(epsilon = var_8656_epsilon_0, x = var_8655_cast_fp16)[name = string("op_8656_cast_fp16")];
            fp16 var_8657_to_fp16 = const()[name = string("op_8657_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8658_cast_fp16 = mul(x = var_8656_cast_fp16, y = var_8657_to_fp16)[name = string("op_8658_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8660_cast_fp16 = add(x = var_8654_cast_fp16, y = var_8658_cast_fp16)[name = string("op_8660_cast_fp16")];
            fp32 var_8661_epsilon_0 = const()[name = string("op_8661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8661_cast_fp16 = rsqrt(epsilon = var_8661_epsilon_0, x = var_8660_cast_fp16)[name = string("op_8661_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_275_cast_fp16 = mul(x = y_315_cast_fp16, y = var_8661_cast_fp16)[name = string("out_275_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111152320)))];
            tensor<fp16, [1, 128, 512]> var_8663_cast_fp16 = mul(x = out_275_cast_fp16, y = tm_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("op_8663_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_353_cast_fp16 = add(x = x_343_cast_fp16, y = var_8663_cast_fp16)[name = string("x_353_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8666_cast_fp16 = abs(x = x_353_cast_fp16)[name = string("op_8666_cast_fp16")];
            bool var_8669 = const()[name = string("op_8669"), val = bool(true)];
            tensor<int32, [1]> var_8670_axes_0 = const()[name = string("op_8670_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8670_cast_fp16 = reduce_max(axes = var_8670_axes_0, keep_dims = var_8669, x = var_8666_cast_fp16)[name = string("op_8670_cast_fp16")];
            fp16 var_8671_to_fp16 = const()[name = string("op_8671_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_317_cast_fp16 = maximum(x = var_8670_cast_fp16, y = var_8671_to_fp16)[name = string("s_317_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_317_cast_fp16 = real_div(x = x_353_cast_fp16, y = s_317_cast_fp16)[name = string("y_317_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8674_cast_fp16 = mul(x = y_317_cast_fp16, y = y_317_cast_fp16)[name = string("op_8674_cast_fp16")];
            tensor<int32, [1]> var_8679_axes_0 = const()[name = string("op_8679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8679_keep_dims_0 = const()[name = string("op_8679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8679_cast_fp16 = reduce_mean(axes = var_8679_axes_0, keep_dims = var_8679_keep_dims_0, x = var_8674_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8680_cast_fp16 = mul(x = s_317_cast_fp16, y = s_317_cast_fp16)[name = string("op_8680_cast_fp16")];
            fp32 var_8681_epsilon_0 = const()[name = string("op_8681_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8681_cast_fp16 = inverse(epsilon = var_8681_epsilon_0, x = var_8680_cast_fp16)[name = string("op_8681_cast_fp16")];
            fp16 var_8682_to_fp16 = const()[name = string("op_8682_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8683_cast_fp16 = mul(x = var_8681_cast_fp16, y = var_8682_to_fp16)[name = string("op_8683_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8685_cast_fp16 = add(x = var_8679_cast_fp16, y = var_8683_cast_fp16)[name = string("op_8685_cast_fp16")];
            fp32 var_8686_epsilon_0 = const()[name = string("op_8686_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8686_cast_fp16 = rsqrt(epsilon = var_8686_epsilon_0, x = var_8685_cast_fp16)[name = string("op_8686_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_277_cast_fp16 = mul(x = y_317_cast_fp16, y = var_8686_cast_fp16)[name = string("out_277_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111153408)))];
            tensor<fp16, [1, 128, 512]> input_233_cast_fp16 = mul(x = out_277_cast_fp16, y = tm_layers_19_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111154496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112203136))))[name = string("tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_176_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_176_cast_fp16")];
            string var_8693_mode_0 = const()[name = string("op_8693_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_8693_cast_fp16 = gelu(mode = var_8693_mode_0, x = linear_176_cast_fp16)[name = string("op_8693_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112207296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113255936))))[name = string("tm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_177_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_235_cast_fp16 = mul(x = var_8693_cast_fp16, y = linear_177_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113260096))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114308736))))[name = string("tm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8701_cast_fp16 = abs(x = linear_178_cast_fp16)[name = string("op_8701_cast_fp16")];
            bool var_8704 = const()[name = string("op_8704"), val = bool(true)];
            tensor<int32, [1]> var_8705_axes_0 = const()[name = string("op_8705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8705_cast_fp16 = reduce_max(axes = var_8705_axes_0, keep_dims = var_8704, x = var_8701_cast_fp16)[name = string("op_8705_cast_fp16")];
            fp16 var_8706_to_fp16 = const()[name = string("op_8706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_319_cast_fp16 = maximum(x = var_8705_cast_fp16, y = var_8706_to_fp16)[name = string("s_319_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_319_cast_fp16 = real_div(x = linear_178_cast_fp16, y = s_319_cast_fp16)[name = string("y_319_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8709_cast_fp16 = mul(x = y_319_cast_fp16, y = y_319_cast_fp16)[name = string("op_8709_cast_fp16")];
            tensor<int32, [1]> var_8714_axes_0 = const()[name = string("op_8714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8714_keep_dims_0 = const()[name = string("op_8714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8714_cast_fp16 = reduce_mean(axes = var_8714_axes_0, keep_dims = var_8714_keep_dims_0, x = var_8709_cast_fp16)[name = string("op_8714_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8715_cast_fp16 = mul(x = s_319_cast_fp16, y = s_319_cast_fp16)[name = string("op_8715_cast_fp16")];
            fp32 var_8716_epsilon_0 = const()[name = string("op_8716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8716_cast_fp16 = inverse(epsilon = var_8716_epsilon_0, x = var_8715_cast_fp16)[name = string("op_8716_cast_fp16")];
            fp16 var_8717_to_fp16 = const()[name = string("op_8717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8718_cast_fp16 = mul(x = var_8716_cast_fp16, y = var_8717_to_fp16)[name = string("op_8718_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8720_cast_fp16 = add(x = var_8714_cast_fp16, y = var_8718_cast_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8721_cast_fp16 = rsqrt(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_279_cast_fp16 = mul(x = y_319_cast_fp16, y = var_8721_cast_fp16)[name = string("out_279_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114309824)))];
            tensor<fp16, [1, 128, 512]> var_8723_cast_fp16 = mul(x = out_279_cast_fp16, y = tm_layers_19_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_237_cast_fp16 = add(x = x_353_cast_fp16, y = var_8723_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114310912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114573120))))[name = string("tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_179_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_237_cast_fp16)[name = string("linear_179_cast_fp16")];
            string x_357_mode_0 = const()[name = string("x_357_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_357_cast_fp16 = gelu(mode = x_357_mode_0, x = linear_179_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [4]> var_8743_begin_0 = const()[name = string("op_8743_begin_0"), val = tensor<int32, [4]>([0, 0, 19, 0])];
            tensor<int32, [4]> var_8743_end_0 = const()[name = string("op_8743_end_0"), val = tensor<int32, [4]>([1, 128, 20, 512])];
            tensor<bool, [4]> var_8743_end_mask_0 = const()[name = string("op_8743_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8743_squeeze_mask_0 = const()[name = string("op_8743_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_8743_cast_fp16 = slice_by_index(begin = var_8743_begin_0, end = var_8743_end_0, end_mask = var_8743_end_mask_0, squeeze_mask = var_8743_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8743_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_239_cast_fp16 = mul(x = x_357_cast_fp16, y = var_8743_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114574208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114836416))))[name = string("tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8753_cast_fp16 = abs(x = linear_180_cast_fp16)[name = string("op_8753_cast_fp16")];
            bool var_8756 = const()[name = string("op_8756"), val = bool(true)];
            tensor<int32, [1]> var_8757_axes_0 = const()[name = string("op_8757_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8757_cast_fp16 = reduce_max(axes = var_8757_axes_0, keep_dims = var_8756, x = var_8753_cast_fp16)[name = string("op_8757_cast_fp16")];
            fp16 var_8758_to_fp16 = const()[name = string("op_8758_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_321_cast_fp16 = maximum(x = var_8757_cast_fp16, y = var_8758_to_fp16)[name = string("s_321_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_321_cast_fp16 = real_div(x = linear_180_cast_fp16, y = s_321_cast_fp16)[name = string("y_321_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8761_cast_fp16 = mul(x = y_321_cast_fp16, y = y_321_cast_fp16)[name = string("op_8761_cast_fp16")];
            tensor<int32, [1]> var_8766_axes_0 = const()[name = string("op_8766_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8766_keep_dims_0 = const()[name = string("op_8766_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8766_cast_fp16 = reduce_mean(axes = var_8766_axes_0, keep_dims = var_8766_keep_dims_0, x = var_8761_cast_fp16)[name = string("op_8766_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8767_cast_fp16 = mul(x = s_321_cast_fp16, y = s_321_cast_fp16)[name = string("op_8767_cast_fp16")];
            fp32 var_8768_epsilon_0 = const()[name = string("op_8768_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8768_cast_fp16 = inverse(epsilon = var_8768_epsilon_0, x = var_8767_cast_fp16)[name = string("op_8768_cast_fp16")];
            fp16 var_8769_to_fp16 = const()[name = string("op_8769_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8770_cast_fp16 = mul(x = var_8768_cast_fp16, y = var_8769_to_fp16)[name = string("op_8770_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8772_cast_fp16 = add(x = var_8766_cast_fp16, y = var_8770_cast_fp16)[name = string("op_8772_cast_fp16")];
            fp32 var_8773_epsilon_0 = const()[name = string("op_8773_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8773_cast_fp16 = rsqrt(epsilon = var_8773_epsilon_0, x = var_8772_cast_fp16)[name = string("op_8773_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_281_cast_fp16 = mul(x = y_321_cast_fp16, y = var_8773_cast_fp16)[name = string("out_281_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114837504)))];
            tensor<fp16, [1, 128, 512]> var_8775_cast_fp16 = mul(x = out_281_cast_fp16, y = tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_39_cast_fp16 = add(x = input_237_cast_fp16, y = var_8775_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<fp16, [1]> tm_layers_19_layer_scalar_to_fp16 = const()[name = string("tm_layers_19_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fp-1])];
            tensor<fp16, [1, 128, 512]> x_361_cast_fp16 = mul(x = h_39_cast_fp16, y = tm_layers_19_layer_scalar_to_fp16)[name = string("x_361_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8779_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_8779_cast_fp16")];
            bool var_8782 = const()[name = string("op_8782"), val = bool(true)];
            tensor<int32, [1]> var_8783_axes_0 = const()[name = string("op_8783_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8783_cast_fp16 = reduce_max(axes = var_8783_axes_0, keep_dims = var_8782, x = var_8779_cast_fp16)[name = string("op_8783_cast_fp16")];
            fp16 var_8784_to_fp16 = const()[name = string("op_8784_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_323_cast_fp16 = maximum(x = var_8783_cast_fp16, y = var_8784_to_fp16)[name = string("s_323_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_323_cast_fp16 = real_div(x = x_361_cast_fp16, y = s_323_cast_fp16)[name = string("y_323_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8787_cast_fp16 = mul(x = y_323_cast_fp16, y = y_323_cast_fp16)[name = string("op_8787_cast_fp16")];
            tensor<int32, [1]> var_8792_axes_0 = const()[name = string("op_8792_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8792_keep_dims_0 = const()[name = string("op_8792_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8792_cast_fp16 = reduce_mean(axes = var_8792_axes_0, keep_dims = var_8792_keep_dims_0, x = var_8787_cast_fp16)[name = string("op_8792_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8793_cast_fp16 = mul(x = s_323_cast_fp16, y = s_323_cast_fp16)[name = string("op_8793_cast_fp16")];
            fp32 var_8794_epsilon_0 = const()[name = string("op_8794_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8794_cast_fp16 = inverse(epsilon = var_8794_epsilon_0, x = var_8793_cast_fp16)[name = string("op_8794_cast_fp16")];
            fp16 var_8795_to_fp16 = const()[name = string("op_8795_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8796_cast_fp16 = mul(x = var_8794_cast_fp16, y = var_8795_to_fp16)[name = string("op_8796_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8798_cast_fp16 = add(x = var_8792_cast_fp16, y = var_8796_cast_fp16)[name = string("op_8798_cast_fp16")];
            fp32 var_8799_epsilon_0 = const()[name = string("op_8799_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8799_cast_fp16 = rsqrt(epsilon = var_8799_epsilon_0, x = var_8798_cast_fp16)[name = string("op_8799_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_283_cast_fp16 = mul(x = y_323_cast_fp16, y = var_8799_cast_fp16)[name = string("out_283_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114838592)))];
            tensor<fp16, [1, 128, 512]> input_241_cast_fp16 = mul(x = out_283_cast_fp16, y = tm_layers_20_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114839680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115364032))))[name = string("tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<int32, [4]> var_8809 = const()[name = string("op_8809"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_8810_cast_fp16 = reshape(shape = var_8809, x = linear_181_cast_fp16)[name = string("op_8810_cast_fp16")];
            tensor<int32, [4]> x_363_perm_0 = const()[name = string("x_363_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115366144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115628352))))[name = string("tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_8821 = const()[name = string("op_8821"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_8822_cast_fp16 = reshape(shape = var_8821, x = linear_182_cast_fp16)[name = string("op_8822_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115629440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115891648))))[name = string("tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_8833 = const()[name = string("op_8833"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_8834_cast_fp16 = reshape(shape = var_8833, x = linear_183_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<int32, [4]> x_367_perm_0 = const()[name = string("x_367_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_363_cast_fp16 = transpose(perm = x_363_perm_0, x = var_8810_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 4, 128, 256]> var_8838_cast_fp16 = abs(x = x_363_cast_fp16)[name = string("op_8838_cast_fp16")];
            bool var_8841 = const()[name = string("op_8841"), val = bool(true)];
            tensor<int32, [1]> var_8842_axes_0 = const()[name = string("op_8842_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_8842_cast_fp16 = reduce_max(axes = var_8842_axes_0, keep_dims = var_8841, x = var_8838_cast_fp16)[name = string("op_8842_cast_fp16")];
            fp16 var_8843_to_fp16 = const()[name = string("op_8843_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_325_cast_fp16 = maximum(x = var_8842_cast_fp16, y = var_8843_to_fp16)[name = string("s_325_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_325_cast_fp16 = real_div(x = x_363_cast_fp16, y = s_325_cast_fp16)[name = string("y_325_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8846_cast_fp16 = mul(x = y_325_cast_fp16, y = y_325_cast_fp16)[name = string("op_8846_cast_fp16")];
            tensor<int32, [1]> var_8851_axes_0 = const()[name = string("op_8851_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8851_keep_dims_0 = const()[name = string("op_8851_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_8851_cast_fp16 = reduce_mean(axes = var_8851_axes_0, keep_dims = var_8851_keep_dims_0, x = var_8846_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_8852_cast_fp16 = mul(x = s_325_cast_fp16, y = s_325_cast_fp16)[name = string("op_8852_cast_fp16")];
            fp32 var_8853_epsilon_0 = const()[name = string("op_8853_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_8853_cast_fp16 = inverse(epsilon = var_8853_epsilon_0, x = var_8852_cast_fp16)[name = string("op_8853_cast_fp16")];
            fp16 var_8854_to_fp16 = const()[name = string("op_8854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_8855_cast_fp16 = mul(x = var_8853_cast_fp16, y = var_8854_to_fp16)[name = string("op_8855_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_8857_cast_fp16 = add(x = var_8851_cast_fp16, y = var_8855_cast_fp16)[name = string("op_8857_cast_fp16")];
            fp32 var_8858_epsilon_0 = const()[name = string("op_8858_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_8858_cast_fp16 = rsqrt(epsilon = var_8858_epsilon_0, x = var_8857_cast_fp16)[name = string("op_8858_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_285_cast_fp16 = mul(x = y_325_cast_fp16, y = var_8858_cast_fp16)[name = string("out_285_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_81_cast_fp16 = mul(x = out_285_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_81_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_8822_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 128, 256]> var_8861_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_8861_cast_fp16")];
            bool var_8864 = const()[name = string("op_8864"), val = bool(true)];
            tensor<int32, [1]> var_8865_axes_0 = const()[name = string("op_8865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_8865_cast_fp16 = reduce_max(axes = var_8865_axes_0, keep_dims = var_8864, x = var_8861_cast_fp16)[name = string("op_8865_cast_fp16")];
            fp16 var_8866_to_fp16 = const()[name = string("op_8866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_327_cast_fp16 = maximum(x = var_8865_cast_fp16, y = var_8866_to_fp16)[name = string("s_327_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_327_cast_fp16 = real_div(x = x_365_cast_fp16, y = s_327_cast_fp16)[name = string("y_327_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8869_cast_fp16 = mul(x = y_327_cast_fp16, y = y_327_cast_fp16)[name = string("op_8869_cast_fp16")];
            tensor<int32, [1]> var_8874_axes_0 = const()[name = string("op_8874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8874_keep_dims_0 = const()[name = string("op_8874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_8874_cast_fp16 = reduce_mean(axes = var_8874_axes_0, keep_dims = var_8874_keep_dims_0, x = var_8869_cast_fp16)[name = string("op_8874_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8875_cast_fp16 = mul(x = s_327_cast_fp16, y = s_327_cast_fp16)[name = string("op_8875_cast_fp16")];
            fp32 var_8876_epsilon_0 = const()[name = string("op_8876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_8876_cast_fp16 = inverse(epsilon = var_8876_epsilon_0, x = var_8875_cast_fp16)[name = string("op_8876_cast_fp16")];
            fp16 var_8877_to_fp16 = const()[name = string("op_8877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_8878_cast_fp16 = mul(x = var_8876_cast_fp16, y = var_8877_to_fp16)[name = string("op_8878_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8880_cast_fp16 = add(x = var_8874_cast_fp16, y = var_8878_cast_fp16)[name = string("op_8880_cast_fp16")];
            fp32 var_8881_epsilon_0 = const()[name = string("op_8881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_8881_cast_fp16 = rsqrt(epsilon = var_8881_epsilon_0, x = var_8880_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_287_cast_fp16 = mul(x = y_327_cast_fp16, y = var_8881_cast_fp16)[name = string("out_287_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_121_cast_fp16 = mul(x = out_287_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_121_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_367_cast_fp16 = transpose(perm = x_367_perm_0, x = var_8834_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 128, 256]> var_8884_cast_fp16 = abs(x = x_367_cast_fp16)[name = string("op_8884_cast_fp16")];
            bool var_8887 = const()[name = string("op_8887"), val = bool(true)];
            tensor<int32, [1]> var_8888_axes_0 = const()[name = string("op_8888_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_8888_cast_fp16 = reduce_max(axes = var_8888_axes_0, keep_dims = var_8887, x = var_8884_cast_fp16)[name = string("op_8888_cast_fp16")];
            fp16 var_8889_to_fp16 = const()[name = string("op_8889_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_329_cast_fp16 = maximum(x = var_8888_cast_fp16, y = var_8889_to_fp16)[name = string("s_329_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_329_cast_fp16 = real_div(x = x_367_cast_fp16, y = s_329_cast_fp16)[name = string("y_329_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8892_cast_fp16 = mul(x = y_329_cast_fp16, y = y_329_cast_fp16)[name = string("op_8892_cast_fp16")];
            tensor<int32, [1]> var_8897_axes_0 = const()[name = string("op_8897_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8897_keep_dims_0 = const()[name = string("op_8897_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_8897_cast_fp16 = reduce_mean(axes = var_8897_axes_0, keep_dims = var_8897_keep_dims_0, x = var_8892_cast_fp16)[name = string("op_8897_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8898_cast_fp16 = mul(x = s_329_cast_fp16, y = s_329_cast_fp16)[name = string("op_8898_cast_fp16")];
            fp32 var_8899_epsilon_0 = const()[name = string("op_8899_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_8899_cast_fp16 = inverse(epsilon = var_8899_epsilon_0, x = var_8898_cast_fp16)[name = string("op_8899_cast_fp16")];
            fp16 var_8900_to_fp16 = const()[name = string("op_8900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_8901_cast_fp16 = mul(x = var_8899_cast_fp16, y = var_8900_to_fp16)[name = string("op_8901_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_8903_cast_fp16 = add(x = var_8897_cast_fp16, y = var_8901_cast_fp16)[name = string("op_8903_cast_fp16")];
            fp32 var_8904_epsilon_0 = const()[name = string("op_8904_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_8904_cast_fp16 = rsqrt(epsilon = var_8904_epsilon_0, x = var_8903_cast_fp16)[name = string("op_8904_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_81_cast_fp16 = mul(x = y_329_cast_fp16, y = var_8904_cast_fp16)[name = string("v_81_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8906_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8906_cast_fp16")];
            tensor<int32, [4]> var_8911_begin_0 = const()[name = string("op_8911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8911_end_0 = const()[name = string("op_8911_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_8911_end_mask_0 = const()[name = string("op_8911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_8911_cast_fp16 = slice_by_index(begin = var_8911_begin_0, end = var_8911_end_0, end_mask = var_8911_end_mask_0, x = q_81_cast_fp16)[name = string("op_8911_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_8912_cast_fp16 = mul(x = var_8911_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_8912_cast_fp16")];
            tensor<int32, [4]> var_8917_begin_0 = const()[name = string("op_8917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8917_end_0 = const()[name = string("op_8917_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_8917_end_mask_0 = const()[name = string("op_8917_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_8917_cast_fp16 = slice_by_index(begin = var_8917_begin_0, end = var_8917_end_0, end_mask = var_8917_end_mask_0, x = q_81_cast_fp16)[name = string("op_8917_cast_fp16")];
            int32 var_8919 = const()[name = string("op_8919"), val = int32(-1)];
            bool var_8920_interleave_0 = const()[name = string("op_8920_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_8920_cast_fp16 = concat(axis = var_8919, interleave = var_8920_interleave_0, values = (var_8912_cast_fp16, var_8917_cast_fp16))[name = string("op_8920_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_8921_cast_fp16 = mul(x = var_8920_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_83_cast_fp16 = add(x = var_8906_cast_fp16, y = var_8921_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8924_cast_fp16 = mul(x = k_121_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8924_cast_fp16")];
            tensor<int32, [4]> var_8929_begin_0 = const()[name = string("op_8929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8929_end_0 = const()[name = string("op_8929_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_8929_end_mask_0 = const()[name = string("op_8929_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, x = k_121_cast_fp16)[name = string("op_8929_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_8930_cast_fp16 = mul(x = var_8929_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_8930_cast_fp16")];
            tensor<int32, [4]> var_8935_begin_0 = const()[name = string("op_8935_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8935_end_0 = const()[name = string("op_8935_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_8935_end_mask_0 = const()[name = string("op_8935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, x = k_121_cast_fp16)[name = string("op_8935_cast_fp16")];
            int32 var_8937 = const()[name = string("op_8937"), val = int32(-1)];
            bool var_8938_interleave_0 = const()[name = string("op_8938_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_8938_cast_fp16 = concat(axis = var_8937, interleave = var_8938_interleave_0, values = (var_8930_cast_fp16, var_8935_cast_fp16))[name = string("op_8938_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_8939_cast_fp16 = mul(x = var_8938_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8939_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_123_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8939_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_40_reps_0 = const()[name = string("tile_40_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = k_123_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 1, 128, 256]> tile_40_cast_fp16 = tile(reps = tile_40_reps_0, x = transpose_80_cast_fp16)[name = string("tile_40_cast_fp16")];
            tensor<int32, [5]> concat_80 = const()[name = string("concat_80"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_80_cast_fp16 = reshape(shape = concat_80, x = tile_40_cast_fp16)[name = string("reshape_80_cast_fp16")];
            tensor<int32, [5]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_81 = const()[name = string("concat_81"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = reshape_80_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 1, 128, 256]> reshape_81_cast_fp16 = reshape(shape = concat_81, x = transpose_81_cast_fp16)[name = string("reshape_81_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_41_reps_0 = const()[name = string("tile_41_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_82_cast_fp16 = transpose(perm = transpose_82_perm_0, x = v_81_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 1, 128, 256]> tile_41_cast_fp16 = tile(reps = tile_41_reps_0, x = transpose_82_cast_fp16)[name = string("tile_41_cast_fp16")];
            tensor<int32, [5]> concat_82 = const()[name = string("concat_82"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_82_cast_fp16 = reshape(shape = concat_82, x = tile_41_cast_fp16)[name = string("reshape_82_cast_fp16")];
            tensor<int32, [5]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_83 = const()[name = string("concat_83"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = reshape_82_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 1, 128, 256]> reshape_83_cast_fp16 = reshape(shape = concat_83, x = transpose_83_cast_fp16)[name = string("reshape_83_cast_fp16")];
            tensor<int32, [4]> v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8953_transpose_x_0 = const()[name = string("op_8953_transpose_x_0"), val = bool(false)];
            bool var_8953_transpose_y_0 = const()[name = string("op_8953_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = reshape_81_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 4, 128, 128]> var_8953_cast_fp16 = matmul(transpose_x = var_8953_transpose_x_0, transpose_y = var_8953_transpose_y_0, x = q_83_cast_fp16, y = transpose_116_cast_fp16)[name = string("op_8953_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_8955_cast_fp16 = add(x = var_8953_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8955_cast_fp16")];
            int32 var_8956 = const()[name = string("op_8956"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_329_cast_fp16 = softmax(axis = var_8956, x = var_8955_cast_fp16)[name = string("w_329_cast_fp16")];
            bool var_8959_transpose_x_0 = const()[name = string("op_8959_transpose_x_0"), val = bool(false)];
            bool var_8959_transpose_y_0 = const()[name = string("op_8959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_83_cast_fp16 = transpose(perm = v_83_perm_0, x = reshape_83_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 4, 128, 256]> var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = w_329_cast_fp16, y = v_83_cast_fp16)[name = string("op_8959_cast_fp16")];
            tensor<int32, [4]> var_8962_perm_0 = const()[name = string("op_8962_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8966 = const()[name = string("op_8966"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_8962_cast_fp16 = transpose(perm = var_8962_perm_0, x = var_8959_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 128, 1024]> input_243_cast_fp16 = reshape(shape = var_8966, x = var_8962_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115892736))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116417088))))[name = string("tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8971_cast_fp16 = abs(x = linear_184_cast_fp16)[name = string("op_8971_cast_fp16")];
            bool var_8974 = const()[name = string("op_8974"), val = bool(true)];
            tensor<int32, [1]> var_8975_axes_0 = const()[name = string("op_8975_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_8975_cast_fp16 = reduce_max(axes = var_8975_axes_0, keep_dims = var_8974, x = var_8971_cast_fp16)[name = string("op_8975_cast_fp16")];
            fp16 var_8976_to_fp16 = const()[name = string("op_8976_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_331_cast_fp16 = maximum(x = var_8975_cast_fp16, y = var_8976_to_fp16)[name = string("s_331_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_331_cast_fp16 = real_div(x = linear_184_cast_fp16, y = s_331_cast_fp16)[name = string("y_331_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8979_cast_fp16 = mul(x = y_331_cast_fp16, y = y_331_cast_fp16)[name = string("op_8979_cast_fp16")];
            tensor<int32, [1]> var_8984_axes_0 = const()[name = string("op_8984_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8984_keep_dims_0 = const()[name = string("op_8984_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_8984_cast_fp16 = reduce_mean(axes = var_8984_axes_0, keep_dims = var_8984_keep_dims_0, x = var_8979_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8985_cast_fp16 = mul(x = s_331_cast_fp16, y = s_331_cast_fp16)[name = string("op_8985_cast_fp16")];
            fp32 var_8986_epsilon_0 = const()[name = string("op_8986_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_8986_cast_fp16 = inverse(epsilon = var_8986_epsilon_0, x = var_8985_cast_fp16)[name = string("op_8986_cast_fp16")];
            fp16 var_8987_to_fp16 = const()[name = string("op_8987_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_8988_cast_fp16 = mul(x = var_8986_cast_fp16, y = var_8987_to_fp16)[name = string("op_8988_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_8990_cast_fp16 = add(x = var_8984_cast_fp16, y = var_8988_cast_fp16)[name = string("op_8990_cast_fp16")];
            fp32 var_8991_epsilon_0 = const()[name = string("op_8991_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_8991_cast_fp16 = rsqrt(epsilon = var_8991_epsilon_0, x = var_8990_cast_fp16)[name = string("op_8991_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_289_cast_fp16 = mul(x = y_331_cast_fp16, y = var_8991_cast_fp16)[name = string("out_289_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116418176)))];
            tensor<fp16, [1, 128, 512]> var_8993_cast_fp16 = mul(x = out_289_cast_fp16, y = tm_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("op_8993_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_371_cast_fp16 = add(x = x_361_cast_fp16, y = var_8993_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_8996_cast_fp16 = abs(x = x_371_cast_fp16)[name = string("op_8996_cast_fp16")];
            bool var_8999 = const()[name = string("op_8999"), val = bool(true)];
            tensor<int32, [1]> var_9000_axes_0 = const()[name = string("op_9000_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9000_cast_fp16 = reduce_max(axes = var_9000_axes_0, keep_dims = var_8999, x = var_8996_cast_fp16)[name = string("op_9000_cast_fp16")];
            fp16 var_9001_to_fp16 = const()[name = string("op_9001_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_333_cast_fp16 = maximum(x = var_9000_cast_fp16, y = var_9001_to_fp16)[name = string("s_333_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_333_cast_fp16 = real_div(x = x_371_cast_fp16, y = s_333_cast_fp16)[name = string("y_333_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9004_cast_fp16 = mul(x = y_333_cast_fp16, y = y_333_cast_fp16)[name = string("op_9004_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9004_cast_fp16)[name = string("op_9009_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9010_cast_fp16 = mul(x = s_333_cast_fp16, y = s_333_cast_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9015_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9015_cast_fp16")];
            fp32 var_9016_epsilon_0 = const()[name = string("op_9016_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9016_cast_fp16 = rsqrt(epsilon = var_9016_epsilon_0, x = var_9015_cast_fp16)[name = string("op_9016_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_291_cast_fp16 = mul(x = y_333_cast_fp16, y = var_9016_cast_fp16)[name = string("out_291_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116419264)))];
            tensor<fp16, [1, 128, 512]> input_245_cast_fp16 = mul(x = out_291_cast_fp16, y = tm_layers_20_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116420352))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117468992))))[name = string("tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_185_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_185_cast_fp16")];
            string var_9023_mode_0 = const()[name = string("op_9023_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_9023_cast_fp16 = gelu(mode = var_9023_mode_0, x = linear_185_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117473152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118521792))))[name = string("tm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_186_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_247_cast_fp16 = mul(x = var_9023_cast_fp16, y = linear_186_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118525952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119574592))))[name = string("tm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9031_cast_fp16 = abs(x = linear_187_cast_fp16)[name = string("op_9031_cast_fp16")];
            bool var_9034 = const()[name = string("op_9034"), val = bool(true)];
            tensor<int32, [1]> var_9035_axes_0 = const()[name = string("op_9035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9035_cast_fp16 = reduce_max(axes = var_9035_axes_0, keep_dims = var_9034, x = var_9031_cast_fp16)[name = string("op_9035_cast_fp16")];
            fp16 var_9036_to_fp16 = const()[name = string("op_9036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_335_cast_fp16 = maximum(x = var_9035_cast_fp16, y = var_9036_to_fp16)[name = string("s_335_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_335_cast_fp16 = real_div(x = linear_187_cast_fp16, y = s_335_cast_fp16)[name = string("y_335_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9039_cast_fp16 = mul(x = y_335_cast_fp16, y = y_335_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [1]> var_9044_axes_0 = const()[name = string("op_9044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9044_keep_dims_0 = const()[name = string("op_9044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9044_cast_fp16 = reduce_mean(axes = var_9044_axes_0, keep_dims = var_9044_keep_dims_0, x = var_9039_cast_fp16)[name = string("op_9044_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9045_cast_fp16 = mul(x = s_335_cast_fp16, y = s_335_cast_fp16)[name = string("op_9045_cast_fp16")];
            fp32 var_9046_epsilon_0 = const()[name = string("op_9046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9046_cast_fp16 = inverse(epsilon = var_9046_epsilon_0, x = var_9045_cast_fp16)[name = string("op_9046_cast_fp16")];
            fp16 var_9047_to_fp16 = const()[name = string("op_9047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9048_cast_fp16 = mul(x = var_9046_cast_fp16, y = var_9047_to_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9050_cast_fp16 = add(x = var_9044_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9050_cast_fp16")];
            fp32 var_9051_epsilon_0 = const()[name = string("op_9051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9051_cast_fp16 = rsqrt(epsilon = var_9051_epsilon_0, x = var_9050_cast_fp16)[name = string("op_9051_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_293_cast_fp16 = mul(x = y_335_cast_fp16, y = var_9051_cast_fp16)[name = string("out_293_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119575680)))];
            tensor<fp16, [1, 128, 512]> var_9053_cast_fp16 = mul(x = out_293_cast_fp16, y = tm_layers_20_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9053_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_249_cast_fp16 = add(x = x_371_cast_fp16, y = var_9053_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119576768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119838976))))[name = string("tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_188_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_188_cast_fp16")];
            string x_375_mode_0 = const()[name = string("x_375_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_375_cast_fp16 = gelu(mode = x_375_mode_0, x = linear_188_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [4]> var_9073_begin_0 = const()[name = string("op_9073_begin_0"), val = tensor<int32, [4]>([0, 0, 20, 0])];
            tensor<int32, [4]> var_9073_end_0 = const()[name = string("op_9073_end_0"), val = tensor<int32, [4]>([1, 128, 21, 512])];
            tensor<bool, [4]> var_9073_end_mask_0 = const()[name = string("op_9073_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9073_squeeze_mask_0 = const()[name = string("op_9073_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_9073_cast_fp16 = slice_by_index(begin = var_9073_begin_0, end = var_9073_end_0, end_mask = var_9073_end_mask_0, squeeze_mask = var_9073_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_251_cast_fp16 = mul(x = x_375_cast_fp16, y = var_9073_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119840064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120102272))))[name = string("tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9083_cast_fp16 = abs(x = linear_189_cast_fp16)[name = string("op_9083_cast_fp16")];
            bool var_9086 = const()[name = string("op_9086"), val = bool(true)];
            tensor<int32, [1]> var_9087_axes_0 = const()[name = string("op_9087_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9087_cast_fp16 = reduce_max(axes = var_9087_axes_0, keep_dims = var_9086, x = var_9083_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_337_cast_fp16 = maximum(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("s_337_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_337_cast_fp16 = real_div(x = linear_189_cast_fp16, y = s_337_cast_fp16)[name = string("y_337_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9091_cast_fp16 = mul(x = y_337_cast_fp16, y = y_337_cast_fp16)[name = string("op_9091_cast_fp16")];
            tensor<int32, [1]> var_9096_axes_0 = const()[name = string("op_9096_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9096_keep_dims_0 = const()[name = string("op_9096_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9096_cast_fp16 = reduce_mean(axes = var_9096_axes_0, keep_dims = var_9096_keep_dims_0, x = var_9091_cast_fp16)[name = string("op_9096_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9097_cast_fp16 = mul(x = s_337_cast_fp16, y = s_337_cast_fp16)[name = string("op_9097_cast_fp16")];
            fp32 var_9098_epsilon_0 = const()[name = string("op_9098_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9098_cast_fp16 = inverse(epsilon = var_9098_epsilon_0, x = var_9097_cast_fp16)[name = string("op_9098_cast_fp16")];
            fp16 var_9099_to_fp16 = const()[name = string("op_9099_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9100_cast_fp16 = mul(x = var_9098_cast_fp16, y = var_9099_to_fp16)[name = string("op_9100_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9102_cast_fp16 = add(x = var_9096_cast_fp16, y = var_9100_cast_fp16)[name = string("op_9102_cast_fp16")];
            fp32 var_9103_epsilon_0 = const()[name = string("op_9103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9103_cast_fp16 = rsqrt(epsilon = var_9103_epsilon_0, x = var_9102_cast_fp16)[name = string("op_9103_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_295_cast_fp16 = mul(x = y_337_cast_fp16, y = var_9103_cast_fp16)[name = string("out_295_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120103360)))];
            tensor<fp16, [1, 128, 512]> var_9105_cast_fp16 = mul(x = out_295_cast_fp16, y = tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9105_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_41_cast_fp16 = add(x = input_249_cast_fp16, y = var_9105_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<fp16, [1]> tm_layers_20_layer_scalar_to_fp16 = const()[name = string("tm_layers_20_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ecp-1])];
            tensor<fp16, [1, 128, 512]> x_379_cast_fp16 = mul(x = h_41_cast_fp16, y = tm_layers_20_layer_scalar_to_fp16)[name = string("x_379_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9109_cast_fp16 = abs(x = x_379_cast_fp16)[name = string("op_9109_cast_fp16")];
            bool var_9112 = const()[name = string("op_9112"), val = bool(true)];
            tensor<int32, [1]> var_9113_axes_0 = const()[name = string("op_9113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9113_cast_fp16 = reduce_max(axes = var_9113_axes_0, keep_dims = var_9112, x = var_9109_cast_fp16)[name = string("op_9113_cast_fp16")];
            fp16 var_9114_to_fp16 = const()[name = string("op_9114_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_339_cast_fp16 = maximum(x = var_9113_cast_fp16, y = var_9114_to_fp16)[name = string("s_339_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_339_cast_fp16 = real_div(x = x_379_cast_fp16, y = s_339_cast_fp16)[name = string("y_339_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9117_cast_fp16 = mul(x = y_339_cast_fp16, y = y_339_cast_fp16)[name = string("op_9117_cast_fp16")];
            tensor<int32, [1]> var_9122_axes_0 = const()[name = string("op_9122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9122_keep_dims_0 = const()[name = string("op_9122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9122_cast_fp16 = reduce_mean(axes = var_9122_axes_0, keep_dims = var_9122_keep_dims_0, x = var_9117_cast_fp16)[name = string("op_9122_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9123_cast_fp16 = mul(x = s_339_cast_fp16, y = s_339_cast_fp16)[name = string("op_9123_cast_fp16")];
            fp32 var_9124_epsilon_0 = const()[name = string("op_9124_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9124_cast_fp16 = inverse(epsilon = var_9124_epsilon_0, x = var_9123_cast_fp16)[name = string("op_9124_cast_fp16")];
            fp16 var_9125_to_fp16 = const()[name = string("op_9125_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9126_cast_fp16 = mul(x = var_9124_cast_fp16, y = var_9125_to_fp16)[name = string("op_9126_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9128_cast_fp16 = add(x = var_9122_cast_fp16, y = var_9126_cast_fp16)[name = string("op_9128_cast_fp16")];
            fp32 var_9129_epsilon_0 = const()[name = string("op_9129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9129_cast_fp16 = rsqrt(epsilon = var_9129_epsilon_0, x = var_9128_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_297_cast_fp16 = mul(x = y_339_cast_fp16, y = var_9129_cast_fp16)[name = string("out_297_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120104448)))];
            tensor<fp16, [1, 128, 512]> input_253_cast_fp16 = mul(x = out_297_cast_fp16, y = tm_layers_21_input_layernorm_weight_to_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120105536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120629888))))[name = string("tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9139 = const()[name = string("op_9139"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_9140_cast_fp16 = reshape(shape = var_9139, x = linear_190_cast_fp16)[name = string("op_9140_cast_fp16")];
            tensor<int32, [4]> x_381_perm_0 = const()[name = string("x_381_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120632000))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120894208))))[name = string("tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9151 = const()[name = string("op_9151"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_9152_cast_fp16 = reshape(shape = var_9151, x = linear_191_cast_fp16)[name = string("op_9152_cast_fp16")];
            tensor<int32, [4]> x_383_perm_0 = const()[name = string("x_383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120895296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121157504))))[name = string("tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_9163 = const()[name = string("op_9163"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_9164_cast_fp16 = reshape(shape = var_9163, x = linear_192_cast_fp16)[name = string("op_9164_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_381_cast_fp16 = transpose(perm = x_381_perm_0, x = var_9140_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 4, 128, 256]> var_9168_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_9168_cast_fp16")];
            bool var_9171 = const()[name = string("op_9171"), val = bool(true)];
            tensor<int32, [1]> var_9172_axes_0 = const()[name = string("op_9172_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_9172_cast_fp16 = reduce_max(axes = var_9172_axes_0, keep_dims = var_9171, x = var_9168_cast_fp16)[name = string("op_9172_cast_fp16")];
            fp16 var_9173_to_fp16 = const()[name = string("op_9173_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_341_cast_fp16 = maximum(x = var_9172_cast_fp16, y = var_9173_to_fp16)[name = string("s_341_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_341_cast_fp16 = real_div(x = x_381_cast_fp16, y = s_341_cast_fp16)[name = string("y_341_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_9176_cast_fp16 = mul(x = y_341_cast_fp16, y = y_341_cast_fp16)[name = string("op_9176_cast_fp16")];
            tensor<int32, [1]> var_9181_axes_0 = const()[name = string("op_9181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9181_keep_dims_0 = const()[name = string("op_9181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_9181_cast_fp16 = reduce_mean(axes = var_9181_axes_0, keep_dims = var_9181_keep_dims_0, x = var_9176_cast_fp16)[name = string("op_9181_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_9182_cast_fp16 = mul(x = s_341_cast_fp16, y = s_341_cast_fp16)[name = string("op_9182_cast_fp16")];
            fp32 var_9183_epsilon_0 = const()[name = string("op_9183_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_9183_cast_fp16 = inverse(epsilon = var_9183_epsilon_0, x = var_9182_cast_fp16)[name = string("op_9183_cast_fp16")];
            fp16 var_9184_to_fp16 = const()[name = string("op_9184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_9185_cast_fp16 = mul(x = var_9183_cast_fp16, y = var_9184_to_fp16)[name = string("op_9185_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_9187_cast_fp16 = add(x = var_9181_cast_fp16, y = var_9185_cast_fp16)[name = string("op_9187_cast_fp16")];
            fp32 var_9188_epsilon_0 = const()[name = string("op_9188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_9188_cast_fp16 = rsqrt(epsilon = var_9188_epsilon_0, x = var_9187_cast_fp16)[name = string("op_9188_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_299_cast_fp16 = mul(x = y_341_cast_fp16, y = var_9188_cast_fp16)[name = string("out_299_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_85_cast_fp16 = mul(x = out_299_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_85_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_383_cast_fp16 = transpose(perm = x_383_perm_0, x = var_9152_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2, 128, 256]> var_9191_cast_fp16 = abs(x = x_383_cast_fp16)[name = string("op_9191_cast_fp16")];
            bool var_9194 = const()[name = string("op_9194"), val = bool(true)];
            tensor<int32, [1]> var_9195_axes_0 = const()[name = string("op_9195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_9195_cast_fp16 = reduce_max(axes = var_9195_axes_0, keep_dims = var_9194, x = var_9191_cast_fp16)[name = string("op_9195_cast_fp16")];
            fp16 var_9196_to_fp16 = const()[name = string("op_9196_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_343_cast_fp16 = maximum(x = var_9195_cast_fp16, y = var_9196_to_fp16)[name = string("s_343_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_343_cast_fp16 = real_div(x = x_383_cast_fp16, y = s_343_cast_fp16)[name = string("y_343_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_9199_cast_fp16 = mul(x = y_343_cast_fp16, y = y_343_cast_fp16)[name = string("op_9199_cast_fp16")];
            tensor<int32, [1]> var_9204_axes_0 = const()[name = string("op_9204_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9204_keep_dims_0 = const()[name = string("op_9204_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_9204_cast_fp16 = reduce_mean(axes = var_9204_axes_0, keep_dims = var_9204_keep_dims_0, x = var_9199_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_9205_cast_fp16 = mul(x = s_343_cast_fp16, y = s_343_cast_fp16)[name = string("op_9205_cast_fp16")];
            fp32 var_9206_epsilon_0 = const()[name = string("op_9206_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_9206_cast_fp16 = inverse(epsilon = var_9206_epsilon_0, x = var_9205_cast_fp16)[name = string("op_9206_cast_fp16")];
            fp16 var_9207_to_fp16 = const()[name = string("op_9207_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_9208_cast_fp16 = mul(x = var_9206_cast_fp16, y = var_9207_to_fp16)[name = string("op_9208_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_9210_cast_fp16 = add(x = var_9204_cast_fp16, y = var_9208_cast_fp16)[name = string("op_9210_cast_fp16")];
            fp32 var_9211_epsilon_0 = const()[name = string("op_9211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_9211_cast_fp16 = rsqrt(epsilon = var_9211_epsilon_0, x = var_9210_cast_fp16)[name = string("op_9211_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_301_cast_fp16 = mul(x = y_343_cast_fp16, y = var_9211_cast_fp16)[name = string("out_301_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_127_cast_fp16 = mul(x = out_301_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_127_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_9164_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2, 128, 256]> var_9214_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_9214_cast_fp16")];
            bool var_9217 = const()[name = string("op_9217"), val = bool(true)];
            tensor<int32, [1]> var_9218_axes_0 = const()[name = string("op_9218_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_9218_cast_fp16 = reduce_max(axes = var_9218_axes_0, keep_dims = var_9217, x = var_9214_cast_fp16)[name = string("op_9218_cast_fp16")];
            fp16 var_9219_to_fp16 = const()[name = string("op_9219_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_345_cast_fp16 = maximum(x = var_9218_cast_fp16, y = var_9219_to_fp16)[name = string("s_345_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_345_cast_fp16 = real_div(x = x_385_cast_fp16, y = s_345_cast_fp16)[name = string("y_345_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_9222_cast_fp16 = mul(x = y_345_cast_fp16, y = y_345_cast_fp16)[name = string("op_9222_cast_fp16")];
            tensor<int32, [1]> var_9227_axes_0 = const()[name = string("op_9227_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9227_keep_dims_0 = const()[name = string("op_9227_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_9227_cast_fp16 = reduce_mean(axes = var_9227_axes_0, keep_dims = var_9227_keep_dims_0, x = var_9222_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_9228_cast_fp16 = mul(x = s_345_cast_fp16, y = s_345_cast_fp16)[name = string("op_9228_cast_fp16")];
            fp32 var_9229_epsilon_0 = const()[name = string("op_9229_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_9229_cast_fp16 = inverse(epsilon = var_9229_epsilon_0, x = var_9228_cast_fp16)[name = string("op_9229_cast_fp16")];
            fp16 var_9230_to_fp16 = const()[name = string("op_9230_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_9231_cast_fp16 = mul(x = var_9229_cast_fp16, y = var_9230_to_fp16)[name = string("op_9231_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_9233_cast_fp16 = add(x = var_9227_cast_fp16, y = var_9231_cast_fp16)[name = string("op_9233_cast_fp16")];
            fp32 var_9234_epsilon_0 = const()[name = string("op_9234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_9234_cast_fp16 = rsqrt(epsilon = var_9234_epsilon_0, x = var_9233_cast_fp16)[name = string("op_9234_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_85_cast_fp16 = mul(x = y_345_cast_fp16, y = var_9234_cast_fp16)[name = string("v_85_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_9236_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9236_cast_fp16")];
            tensor<int32, [4]> var_9241_begin_0 = const()[name = string("op_9241_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9241_end_0 = const()[name = string("op_9241_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_9241_end_mask_0 = const()[name = string("op_9241_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_9241_cast_fp16 = slice_by_index(begin = var_9241_begin_0, end = var_9241_end_0, end_mask = var_9241_end_mask_0, x = q_85_cast_fp16)[name = string("op_9241_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_9242_cast_fp16 = mul(x = var_9241_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_9242_cast_fp16")];
            tensor<int32, [4]> var_9247_begin_0 = const()[name = string("op_9247_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9247_end_0 = const()[name = string("op_9247_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_9247_end_mask_0 = const()[name = string("op_9247_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_9247_cast_fp16 = slice_by_index(begin = var_9247_begin_0, end = var_9247_end_0, end_mask = var_9247_end_mask_0, x = q_85_cast_fp16)[name = string("op_9247_cast_fp16")];
            int32 var_9249 = const()[name = string("op_9249"), val = int32(-1)];
            bool var_9250_interleave_0 = const()[name = string("op_9250_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_9250_cast_fp16 = concat(axis = var_9249, interleave = var_9250_interleave_0, values = (var_9242_cast_fp16, var_9247_cast_fp16))[name = string("op_9250_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_9251_cast_fp16 = mul(x = var_9250_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9251_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_87_cast_fp16 = add(x = var_9236_cast_fp16, y = var_9251_cast_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_9254_cast_fp16 = mul(x = k_127_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9254_cast_fp16")];
            tensor<int32, [4]> var_9259_begin_0 = const()[name = string("op_9259_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9259_end_0 = const()[name = string("op_9259_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_9259_end_mask_0 = const()[name = string("op_9259_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_9259_cast_fp16 = slice_by_index(begin = var_9259_begin_0, end = var_9259_end_0, end_mask = var_9259_end_mask_0, x = k_127_cast_fp16)[name = string("op_9259_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_9260_cast_fp16 = mul(x = var_9259_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_9260_cast_fp16")];
            tensor<int32, [4]> var_9265_begin_0 = const()[name = string("op_9265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9265_end_0 = const()[name = string("op_9265_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_9265_end_mask_0 = const()[name = string("op_9265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_9265_cast_fp16 = slice_by_index(begin = var_9265_begin_0, end = var_9265_end_0, end_mask = var_9265_end_mask_0, x = k_127_cast_fp16)[name = string("op_9265_cast_fp16")];
            int32 var_9267 = const()[name = string("op_9267"), val = int32(-1)];
            bool var_9268_interleave_0 = const()[name = string("op_9268_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_9268_cast_fp16 = concat(axis = var_9267, interleave = var_9268_interleave_0, values = (var_9260_cast_fp16, var_9265_cast_fp16))[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_9269_cast_fp16 = mul(x = var_9268_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_129_cast_fp16 = add(x = var_9254_cast_fp16, y = var_9269_cast_fp16)[name = string("k_129_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_42_reps_0 = const()[name = string("tile_42_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = k_129_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 1, 128, 256]> tile_42_cast_fp16 = tile(reps = tile_42_reps_0, x = transpose_84_cast_fp16)[name = string("tile_42_cast_fp16")];
            tensor<int32, [5]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_84_cast_fp16 = reshape(shape = concat_84, x = tile_42_cast_fp16)[name = string("reshape_84_cast_fp16")];
            tensor<int32, [5]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = reshape_84_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 1, 128, 256]> reshape_85_cast_fp16 = reshape(shape = concat_85, x = transpose_85_cast_fp16)[name = string("reshape_85_cast_fp16")];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_43_reps_0 = const()[name = string("tile_43_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_86_cast_fp16 = transpose(perm = transpose_86_perm_0, x = v_85_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 1, 128, 256]> tile_43_cast_fp16 = tile(reps = tile_43_reps_0, x = transpose_86_cast_fp16)[name = string("tile_43_cast_fp16")];
            tensor<int32, [5]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_86_cast_fp16 = reshape(shape = concat_86, x = tile_43_cast_fp16)[name = string("reshape_86_cast_fp16")];
            tensor<int32, [5]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_87_cast_fp16 = transpose(perm = transpose_87_perm_0, x = reshape_86_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 1, 128, 256]> reshape_87_cast_fp16 = reshape(shape = concat_87, x = transpose_87_cast_fp16)[name = string("reshape_87_cast_fp16")];
            tensor<int32, [4]> v_87_perm_0 = const()[name = string("v_87_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9283_transpose_x_0 = const()[name = string("op_9283_transpose_x_0"), val = bool(false)];
            bool var_9283_transpose_y_0 = const()[name = string("op_9283_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_117_cast_fp16 = transpose(perm = transpose_117_perm_0, x = reshape_85_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 4, 128, 128]> var_9283_cast_fp16 = matmul(transpose_x = var_9283_transpose_x_0, transpose_y = var_9283_transpose_y_0, x = q_87_cast_fp16, y = transpose_117_cast_fp16)[name = string("op_9283_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_9285_cast_fp16 = add(x = var_9283_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9285_cast_fp16")];
            int32 var_9286 = const()[name = string("op_9286"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_345_cast_fp16 = softmax(axis = var_9286, x = var_9285_cast_fp16)[name = string("w_345_cast_fp16")];
            bool var_9289_transpose_x_0 = const()[name = string("op_9289_transpose_x_0"), val = bool(false)];
            bool var_9289_transpose_y_0 = const()[name = string("op_9289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_87_cast_fp16 = transpose(perm = v_87_perm_0, x = reshape_87_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 4, 128, 256]> var_9289_cast_fp16 = matmul(transpose_x = var_9289_transpose_x_0, transpose_y = var_9289_transpose_y_0, x = w_345_cast_fp16, y = v_87_cast_fp16)[name = string("op_9289_cast_fp16")];
            tensor<int32, [4]> var_9292_perm_0 = const()[name = string("op_9292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9296 = const()[name = string("op_9296"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_9292_cast_fp16 = transpose(perm = var_9292_perm_0, x = var_9289_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 128, 1024]> input_255_cast_fp16 = reshape(shape = var_9296, x = var_9292_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121158592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121682944))))[name = string("tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9301_cast_fp16 = abs(x = linear_193_cast_fp16)[name = string("op_9301_cast_fp16")];
            bool var_9304 = const()[name = string("op_9304"), val = bool(true)];
            tensor<int32, [1]> var_9305_axes_0 = const()[name = string("op_9305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9305_cast_fp16 = reduce_max(axes = var_9305_axes_0, keep_dims = var_9304, x = var_9301_cast_fp16)[name = string("op_9305_cast_fp16")];
            fp16 var_9306_to_fp16 = const()[name = string("op_9306_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_347_cast_fp16 = maximum(x = var_9305_cast_fp16, y = var_9306_to_fp16)[name = string("s_347_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_347_cast_fp16 = real_div(x = linear_193_cast_fp16, y = s_347_cast_fp16)[name = string("y_347_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9309_cast_fp16 = mul(x = y_347_cast_fp16, y = y_347_cast_fp16)[name = string("op_9309_cast_fp16")];
            tensor<int32, [1]> var_9314_axes_0 = const()[name = string("op_9314_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9314_keep_dims_0 = const()[name = string("op_9314_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9314_cast_fp16 = reduce_mean(axes = var_9314_axes_0, keep_dims = var_9314_keep_dims_0, x = var_9309_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9315_cast_fp16 = mul(x = s_347_cast_fp16, y = s_347_cast_fp16)[name = string("op_9315_cast_fp16")];
            fp32 var_9316_epsilon_0 = const()[name = string("op_9316_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9316_cast_fp16 = inverse(epsilon = var_9316_epsilon_0, x = var_9315_cast_fp16)[name = string("op_9316_cast_fp16")];
            fp16 var_9317_to_fp16 = const()[name = string("op_9317_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9318_cast_fp16 = mul(x = var_9316_cast_fp16, y = var_9317_to_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9320_cast_fp16 = add(x = var_9314_cast_fp16, y = var_9318_cast_fp16)[name = string("op_9320_cast_fp16")];
            fp32 var_9321_epsilon_0 = const()[name = string("op_9321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9321_cast_fp16 = rsqrt(epsilon = var_9321_epsilon_0, x = var_9320_cast_fp16)[name = string("op_9321_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_303_cast_fp16 = mul(x = y_347_cast_fp16, y = var_9321_cast_fp16)[name = string("out_303_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121684032)))];
            tensor<fp16, [1, 128, 512]> var_9323_cast_fp16 = mul(x = out_303_cast_fp16, y = tm_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("op_9323_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_389_cast_fp16 = add(x = x_379_cast_fp16, y = var_9323_cast_fp16)[name = string("x_389_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9326_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_9326_cast_fp16")];
            bool var_9329 = const()[name = string("op_9329"), val = bool(true)];
            tensor<int32, [1]> var_9330_axes_0 = const()[name = string("op_9330_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9330_cast_fp16 = reduce_max(axes = var_9330_axes_0, keep_dims = var_9329, x = var_9326_cast_fp16)[name = string("op_9330_cast_fp16")];
            fp16 var_9331_to_fp16 = const()[name = string("op_9331_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_349_cast_fp16 = maximum(x = var_9330_cast_fp16, y = var_9331_to_fp16)[name = string("s_349_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_349_cast_fp16 = real_div(x = x_389_cast_fp16, y = s_349_cast_fp16)[name = string("y_349_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9334_cast_fp16 = mul(x = y_349_cast_fp16, y = y_349_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [1]> var_9339_axes_0 = const()[name = string("op_9339_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9339_keep_dims_0 = const()[name = string("op_9339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9339_cast_fp16 = reduce_mean(axes = var_9339_axes_0, keep_dims = var_9339_keep_dims_0, x = var_9334_cast_fp16)[name = string("op_9339_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9340_cast_fp16 = mul(x = s_349_cast_fp16, y = s_349_cast_fp16)[name = string("op_9340_cast_fp16")];
            fp32 var_9341_epsilon_0 = const()[name = string("op_9341_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9341_cast_fp16 = inverse(epsilon = var_9341_epsilon_0, x = var_9340_cast_fp16)[name = string("op_9341_cast_fp16")];
            fp16 var_9342_to_fp16 = const()[name = string("op_9342_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9343_cast_fp16 = mul(x = var_9341_cast_fp16, y = var_9342_to_fp16)[name = string("op_9343_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9345_cast_fp16 = add(x = var_9339_cast_fp16, y = var_9343_cast_fp16)[name = string("op_9345_cast_fp16")];
            fp32 var_9346_epsilon_0 = const()[name = string("op_9346_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9346_cast_fp16 = rsqrt(epsilon = var_9346_epsilon_0, x = var_9345_cast_fp16)[name = string("op_9346_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_305_cast_fp16 = mul(x = y_349_cast_fp16, y = var_9346_cast_fp16)[name = string("out_305_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121685120)))];
            tensor<fp16, [1, 128, 512]> input_257_cast_fp16 = mul(x = out_305_cast_fp16, y = tm_layers_21_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121686208))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122734848))))[name = string("tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_194_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_194_cast_fp16")];
            string var_9353_mode_0 = const()[name = string("op_9353_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_9353_cast_fp16 = gelu(mode = var_9353_mode_0, x = linear_194_cast_fp16)[name = string("op_9353_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122739008))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123787648))))[name = string("tm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_195_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_259_cast_fp16 = mul(x = var_9353_cast_fp16, y = linear_195_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123791808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124840448))))[name = string("tm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9361_cast_fp16 = abs(x = linear_196_cast_fp16)[name = string("op_9361_cast_fp16")];
            bool var_9364 = const()[name = string("op_9364"), val = bool(true)];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9365_cast_fp16 = reduce_max(axes = var_9365_axes_0, keep_dims = var_9364, x = var_9361_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9366_to_fp16 = const()[name = string("op_9366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_351_cast_fp16 = maximum(x = var_9365_cast_fp16, y = var_9366_to_fp16)[name = string("s_351_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_351_cast_fp16 = real_div(x = linear_196_cast_fp16, y = s_351_cast_fp16)[name = string("y_351_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9369_cast_fp16 = mul(x = y_351_cast_fp16, y = y_351_cast_fp16)[name = string("op_9369_cast_fp16")];
            tensor<int32, [1]> var_9374_axes_0 = const()[name = string("op_9374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9374_keep_dims_0 = const()[name = string("op_9374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9374_cast_fp16 = reduce_mean(axes = var_9374_axes_0, keep_dims = var_9374_keep_dims_0, x = var_9369_cast_fp16)[name = string("op_9374_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9375_cast_fp16 = mul(x = s_351_cast_fp16, y = s_351_cast_fp16)[name = string("op_9375_cast_fp16")];
            fp32 var_9376_epsilon_0 = const()[name = string("op_9376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9376_cast_fp16 = inverse(epsilon = var_9376_epsilon_0, x = var_9375_cast_fp16)[name = string("op_9376_cast_fp16")];
            fp16 var_9377_to_fp16 = const()[name = string("op_9377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9378_cast_fp16 = mul(x = var_9376_cast_fp16, y = var_9377_to_fp16)[name = string("op_9378_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9380_cast_fp16 = add(x = var_9374_cast_fp16, y = var_9378_cast_fp16)[name = string("op_9380_cast_fp16")];
            fp32 var_9381_epsilon_0 = const()[name = string("op_9381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9381_cast_fp16 = rsqrt(epsilon = var_9381_epsilon_0, x = var_9380_cast_fp16)[name = string("op_9381_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_307_cast_fp16 = mul(x = y_351_cast_fp16, y = var_9381_cast_fp16)[name = string("out_307_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124841536)))];
            tensor<fp16, [1, 128, 512]> var_9383_cast_fp16 = mul(x = out_307_cast_fp16, y = tm_layers_21_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9383_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_261_cast_fp16 = add(x = x_389_cast_fp16, y = var_9383_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124842624))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125104832))))[name = string("tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_197_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_197_cast_fp16")];
            string x_393_mode_0 = const()[name = string("x_393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_393_cast_fp16 = gelu(mode = x_393_mode_0, x = linear_197_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<int32, [4]> var_9403_begin_0 = const()[name = string("op_9403_begin_0"), val = tensor<int32, [4]>([0, 0, 21, 0])];
            tensor<int32, [4]> var_9403_end_0 = const()[name = string("op_9403_end_0"), val = tensor<int32, [4]>([1, 128, 22, 512])];
            tensor<bool, [4]> var_9403_end_mask_0 = const()[name = string("op_9403_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9403_squeeze_mask_0 = const()[name = string("op_9403_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_9403_cast_fp16 = slice_by_index(begin = var_9403_begin_0, end = var_9403_end_0, end_mask = var_9403_end_mask_0, squeeze_mask = var_9403_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9403_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_263_cast_fp16 = mul(x = x_393_cast_fp16, y = var_9403_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125105920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125368128))))[name = string("tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9413_cast_fp16 = abs(x = linear_198_cast_fp16)[name = string("op_9413_cast_fp16")];
            bool var_9416 = const()[name = string("op_9416"), val = bool(true)];
            tensor<int32, [1]> var_9417_axes_0 = const()[name = string("op_9417_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9417_cast_fp16 = reduce_max(axes = var_9417_axes_0, keep_dims = var_9416, x = var_9413_cast_fp16)[name = string("op_9417_cast_fp16")];
            fp16 var_9418_to_fp16 = const()[name = string("op_9418_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_353_cast_fp16 = maximum(x = var_9417_cast_fp16, y = var_9418_to_fp16)[name = string("s_353_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_353_cast_fp16 = real_div(x = linear_198_cast_fp16, y = s_353_cast_fp16)[name = string("y_353_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9421_cast_fp16 = mul(x = y_353_cast_fp16, y = y_353_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<int32, [1]> var_9426_axes_0 = const()[name = string("op_9426_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9426_keep_dims_0 = const()[name = string("op_9426_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9426_cast_fp16 = reduce_mean(axes = var_9426_axes_0, keep_dims = var_9426_keep_dims_0, x = var_9421_cast_fp16)[name = string("op_9426_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9427_cast_fp16 = mul(x = s_353_cast_fp16, y = s_353_cast_fp16)[name = string("op_9427_cast_fp16")];
            fp32 var_9428_epsilon_0 = const()[name = string("op_9428_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9428_cast_fp16 = inverse(epsilon = var_9428_epsilon_0, x = var_9427_cast_fp16)[name = string("op_9428_cast_fp16")];
            fp16 var_9429_to_fp16 = const()[name = string("op_9429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9430_cast_fp16 = mul(x = var_9428_cast_fp16, y = var_9429_to_fp16)[name = string("op_9430_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9432_cast_fp16 = add(x = var_9426_cast_fp16, y = var_9430_cast_fp16)[name = string("op_9432_cast_fp16")];
            fp32 var_9433_epsilon_0 = const()[name = string("op_9433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9433_cast_fp16 = rsqrt(epsilon = var_9433_epsilon_0, x = var_9432_cast_fp16)[name = string("op_9433_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_309_cast_fp16 = mul(x = y_353_cast_fp16, y = var_9433_cast_fp16)[name = string("out_309_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125369216)))];
            tensor<fp16, [1, 128, 512]> var_9435_cast_fp16 = mul(x = out_309_cast_fp16, y = tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9435_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_43_cast_fp16 = add(x = input_261_cast_fp16, y = var_9435_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<fp16, [1]> tm_layers_21_layer_scalar_to_fp16 = const()[name = string("tm_layers_21_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e8p-1])];
            tensor<fp16, [1, 128, 512]> x_397_cast_fp16 = mul(x = h_43_cast_fp16, y = tm_layers_21_layer_scalar_to_fp16)[name = string("x_397_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9439_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_9439_cast_fp16")];
            bool var_9442 = const()[name = string("op_9442"), val = bool(true)];
            tensor<int32, [1]> var_9443_axes_0 = const()[name = string("op_9443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9443_cast_fp16 = reduce_max(axes = var_9443_axes_0, keep_dims = var_9442, x = var_9439_cast_fp16)[name = string("op_9443_cast_fp16")];
            fp16 var_9444_to_fp16 = const()[name = string("op_9444_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_355_cast_fp16 = maximum(x = var_9443_cast_fp16, y = var_9444_to_fp16)[name = string("s_355_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_355_cast_fp16 = real_div(x = x_397_cast_fp16, y = s_355_cast_fp16)[name = string("y_355_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9447_cast_fp16 = mul(x = y_355_cast_fp16, y = y_355_cast_fp16)[name = string("op_9447_cast_fp16")];
            tensor<int32, [1]> var_9452_axes_0 = const()[name = string("op_9452_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9452_keep_dims_0 = const()[name = string("op_9452_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9452_cast_fp16 = reduce_mean(axes = var_9452_axes_0, keep_dims = var_9452_keep_dims_0, x = var_9447_cast_fp16)[name = string("op_9452_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9453_cast_fp16 = mul(x = s_355_cast_fp16, y = s_355_cast_fp16)[name = string("op_9453_cast_fp16")];
            fp32 var_9454_epsilon_0 = const()[name = string("op_9454_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9454_cast_fp16 = inverse(epsilon = var_9454_epsilon_0, x = var_9453_cast_fp16)[name = string("op_9454_cast_fp16")];
            fp16 var_9455_to_fp16 = const()[name = string("op_9455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9456_cast_fp16 = mul(x = var_9454_cast_fp16, y = var_9455_to_fp16)[name = string("op_9456_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9458_cast_fp16 = add(x = var_9452_cast_fp16, y = var_9456_cast_fp16)[name = string("op_9458_cast_fp16")];
            fp32 var_9459_epsilon_0 = const()[name = string("op_9459_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9459_cast_fp16 = rsqrt(epsilon = var_9459_epsilon_0, x = var_9458_cast_fp16)[name = string("op_9459_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_311_cast_fp16 = mul(x = y_355_cast_fp16, y = var_9459_cast_fp16)[name = string("out_311_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125370304)))];
            tensor<fp16, [1, 128, 512]> input_265_cast_fp16 = mul(x = out_311_cast_fp16, y = tm_layers_22_input_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125371392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125895744))))[name = string("tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 1024]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<int32, [4]> var_9469 = const()[name = string("op_9469"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 128, 4, 256]> var_9470_cast_fp16 = reshape(shape = var_9469, x = linear_199_cast_fp16)[name = string("op_9470_cast_fp16")];
            tensor<int32, [4]> x_399_perm_0 = const()[name = string("x_399_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125897856))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126160064))))[name = string("tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<int32, [4]> var_9481 = const()[name = string("op_9481"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_9482_cast_fp16 = reshape(shape = var_9481, x = linear_200_cast_fp16)[name = string("op_9482_cast_fp16")];
            tensor<int32, [4]> x_401_perm_0 = const()[name = string("x_401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126161152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126423360))))[name = string("tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<int32, [4]> var_9493 = const()[name = string("op_9493"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 128, 2, 256]> var_9494_cast_fp16 = reshape(shape = var_9493, x = linear_201_cast_fp16)[name = string("op_9494_cast_fp16")];
            tensor<int32, [4]> x_403_perm_0 = const()[name = string("x_403_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 256]> x_399_cast_fp16 = transpose(perm = x_399_perm_0, x = var_9470_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 4, 128, 256]> var_9498_cast_fp16 = abs(x = x_399_cast_fp16)[name = string("op_9498_cast_fp16")];
            bool var_9501 = const()[name = string("op_9501"), val = bool(true)];
            tensor<int32, [1]> var_9502_axes_0 = const()[name = string("op_9502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_9502_cast_fp16 = reduce_max(axes = var_9502_axes_0, keep_dims = var_9501, x = var_9498_cast_fp16)[name = string("op_9502_cast_fp16")];
            fp16 var_9503_to_fp16 = const()[name = string("op_9503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_357_cast_fp16 = maximum(x = var_9502_cast_fp16, y = var_9503_to_fp16)[name = string("s_357_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> y_357_cast_fp16 = real_div(x = x_399_cast_fp16, y = s_357_cast_fp16)[name = string("y_357_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_9506_cast_fp16 = mul(x = y_357_cast_fp16, y = y_357_cast_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9511_axes_0 = const()[name = string("op_9511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9511_keep_dims_0 = const()[name = string("op_9511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_9511_cast_fp16 = reduce_mean(axes = var_9511_axes_0, keep_dims = var_9511_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_9512_cast_fp16 = mul(x = s_357_cast_fp16, y = s_357_cast_fp16)[name = string("op_9512_cast_fp16")];
            fp32 var_9513_epsilon_0 = const()[name = string("op_9513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_9513_cast_fp16 = inverse(epsilon = var_9513_epsilon_0, x = var_9512_cast_fp16)[name = string("op_9513_cast_fp16")];
            fp16 var_9514_to_fp16 = const()[name = string("op_9514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_9515_cast_fp16 = mul(x = var_9513_cast_fp16, y = var_9514_to_fp16)[name = string("op_9515_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_9517_cast_fp16 = add(x = var_9511_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9517_cast_fp16")];
            fp32 var_9518_epsilon_0 = const()[name = string("op_9518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_9518_cast_fp16 = rsqrt(epsilon = var_9518_epsilon_0, x = var_9517_cast_fp16)[name = string("op_9518_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> out_313_cast_fp16 = mul(x = y_357_cast_fp16, y = var_9518_cast_fp16)[name = string("out_313_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_89_cast_fp16 = mul(x = out_313_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_89_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_401_cast_fp16 = transpose(perm = x_401_perm_0, x = var_9482_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 128, 256]> var_9521_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_9521_cast_fp16")];
            bool var_9524 = const()[name = string("op_9524"), val = bool(true)];
            tensor<int32, [1]> var_9525_axes_0 = const()[name = string("op_9525_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_9525_cast_fp16 = reduce_max(axes = var_9525_axes_0, keep_dims = var_9524, x = var_9521_cast_fp16)[name = string("op_9525_cast_fp16")];
            fp16 var_9526_to_fp16 = const()[name = string("op_9526_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_359_cast_fp16 = maximum(x = var_9525_cast_fp16, y = var_9526_to_fp16)[name = string("s_359_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_359_cast_fp16 = real_div(x = x_401_cast_fp16, y = s_359_cast_fp16)[name = string("y_359_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_9529_cast_fp16 = mul(x = y_359_cast_fp16, y = y_359_cast_fp16)[name = string("op_9529_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9529_cast_fp16)[name = string("op_9534_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_9535_cast_fp16 = mul(x = s_359_cast_fp16, y = s_359_cast_fp16)[name = string("op_9535_cast_fp16")];
            fp32 var_9536_epsilon_0 = const()[name = string("op_9536_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_9536_cast_fp16 = inverse(epsilon = var_9536_epsilon_0, x = var_9535_cast_fp16)[name = string("op_9536_cast_fp16")];
            fp16 var_9537_to_fp16 = const()[name = string("op_9537_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_9538_cast_fp16 = mul(x = var_9536_cast_fp16, y = var_9537_to_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_9540_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9538_cast_fp16)[name = string("op_9540_cast_fp16")];
            fp32 var_9541_epsilon_0 = const()[name = string("op_9541_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_9541_cast_fp16 = rsqrt(epsilon = var_9541_epsilon_0, x = var_9540_cast_fp16)[name = string("op_9541_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> out_315_cast_fp16 = mul(x = y_359_cast_fp16, y = var_9541_cast_fp16)[name = string("out_315_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_133_cast_fp16 = mul(x = out_315_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_133_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> x_403_cast_fp16 = transpose(perm = x_403_perm_0, x = var_9494_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 128, 256]> var_9544_cast_fp16 = abs(x = x_403_cast_fp16)[name = string("op_9544_cast_fp16")];
            bool var_9547 = const()[name = string("op_9547"), val = bool(true)];
            tensor<int32, [1]> var_9548_axes_0 = const()[name = string("op_9548_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 128, 1]> var_9548_cast_fp16 = reduce_max(axes = var_9548_axes_0, keep_dims = var_9547, x = var_9544_cast_fp16)[name = string("op_9548_cast_fp16")];
            fp16 var_9549_to_fp16 = const()[name = string("op_9549_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 128, 1]> s_361_cast_fp16 = maximum(x = var_9548_cast_fp16, y = var_9549_to_fp16)[name = string("s_361_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> y_361_cast_fp16 = real_div(x = x_403_cast_fp16, y = s_361_cast_fp16)[name = string("y_361_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_9552_cast_fp16 = mul(x = y_361_cast_fp16, y = y_361_cast_fp16)[name = string("op_9552_cast_fp16")];
            tensor<int32, [1]> var_9557_axes_0 = const()[name = string("op_9557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9557_keep_dims_0 = const()[name = string("op_9557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 128, 1]> var_9557_cast_fp16 = reduce_mean(axes = var_9557_axes_0, keep_dims = var_9557_keep_dims_0, x = var_9552_cast_fp16)[name = string("op_9557_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_9558_cast_fp16 = mul(x = s_361_cast_fp16, y = s_361_cast_fp16)[name = string("op_9558_cast_fp16")];
            fp32 var_9559_epsilon_0 = const()[name = string("op_9559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 128, 1]> var_9559_cast_fp16 = inverse(epsilon = var_9559_epsilon_0, x = var_9558_cast_fp16)[name = string("op_9559_cast_fp16")];
            fp16 var_9560_to_fp16 = const()[name = string("op_9560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 128, 1]> var_9561_cast_fp16 = mul(x = var_9559_cast_fp16, y = var_9560_to_fp16)[name = string("op_9561_cast_fp16")];
            tensor<fp16, [1, 2, 128, 1]> var_9563_cast_fp16 = add(x = var_9557_cast_fp16, y = var_9561_cast_fp16)[name = string("op_9563_cast_fp16")];
            fp32 var_9564_epsilon_0 = const()[name = string("op_9564_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 128, 1]> var_9564_cast_fp16 = rsqrt(epsilon = var_9564_epsilon_0, x = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> v_89_cast_fp16 = mul(x = y_361_cast_fp16, y = var_9564_cast_fp16)[name = string("v_89_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_9566_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9566_cast_fp16")];
            tensor<int32, [4]> var_9571_begin_0 = const()[name = string("op_9571_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9571_end_0 = const()[name = string("op_9571_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_9571_end_mask_0 = const()[name = string("op_9571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 128]> var_9571_cast_fp16 = slice_by_index(begin = var_9571_begin_0, end = var_9571_end_0, end_mask = var_9571_end_mask_0, x = q_89_cast_fp16)[name = string("op_9571_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 128]> var_9572_cast_fp16 = mul(x = var_9571_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_9572_cast_fp16")];
            tensor<int32, [4]> var_9577_begin_0 = const()[name = string("op_9577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9577_end_0 = const()[name = string("op_9577_end_0"), val = tensor<int32, [4]>([1, 4, 128, 128])];
            tensor<bool, [4]> var_9577_end_mask_0 = const()[name = string("op_9577_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 128]> var_9577_cast_fp16 = slice_by_index(begin = var_9577_begin_0, end = var_9577_end_0, end_mask = var_9577_end_mask_0, x = q_89_cast_fp16)[name = string("op_9577_cast_fp16")];
            int32 var_9579 = const()[name = string("op_9579"), val = int32(-1)];
            bool var_9580_interleave_0 = const()[name = string("op_9580_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> var_9580_cast_fp16 = concat(axis = var_9579, interleave = var_9580_interleave_0, values = (var_9572_cast_fp16, var_9577_cast_fp16))[name = string("op_9580_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> var_9581_cast_fp16 = mul(x = var_9580_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9581_cast_fp16")];
            tensor<fp16, [1, 4, 128, 256]> q_91_cast_fp16 = add(x = var_9566_cast_fp16, y = var_9581_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_9584_cast_fp16 = mul(x = k_133_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9584_cast_fp16")];
            tensor<int32, [4]> var_9589_begin_0 = const()[name = string("op_9589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9589_end_0 = const()[name = string("op_9589_end_0"), val = tensor<int32, [4]>([1, 2, 128, 256])];
            tensor<bool, [4]> var_9589_end_mask_0 = const()[name = string("op_9589_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 128]> var_9589_cast_fp16 = slice_by_index(begin = var_9589_begin_0, end = var_9589_end_0, end_mask = var_9589_end_mask_0, x = k_133_cast_fp16)[name = string("op_9589_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 128]> var_9590_cast_fp16 = mul(x = var_9589_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_9590_cast_fp16")];
            tensor<int32, [4]> var_9595_begin_0 = const()[name = string("op_9595_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9595_end_0 = const()[name = string("op_9595_end_0"), val = tensor<int32, [4]>([1, 2, 128, 128])];
            tensor<bool, [4]> var_9595_end_mask_0 = const()[name = string("op_9595_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 128]> var_9595_cast_fp16 = slice_by_index(begin = var_9595_begin_0, end = var_9595_end_0, end_mask = var_9595_end_mask_0, x = k_133_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool var_9598_interleave_0 = const()[name = string("op_9598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 256]> var_9598_cast_fp16 = concat(axis = var_9597, interleave = var_9598_interleave_0, values = (var_9590_cast_fp16, var_9595_cast_fp16))[name = string("op_9598_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> var_9599_cast_fp16 = mul(x = var_9598_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9599_cast_fp16")];
            tensor<fp16, [1, 2, 128, 256]> k_135_cast_fp16 = add(x = var_9584_cast_fp16, y = var_9599_cast_fp16)[name = string("k_135_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_44_reps_0 = const()[name = string("tile_44_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = k_135_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 1, 128, 256]> tile_44_cast_fp16 = tile(reps = tile_44_reps_0, x = transpose_88_cast_fp16)[name = string("tile_44_cast_fp16")];
            tensor<int32, [5]> concat_88 = const()[name = string("concat_88"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_88_cast_fp16 = reshape(shape = concat_88, x = tile_44_cast_fp16)[name = string("reshape_88_cast_fp16")];
            tensor<int32, [5]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = reshape_88_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 1, 128, 256]> reshape_89_cast_fp16 = reshape(shape = concat_89, x = transpose_89_cast_fp16)[name = string("reshape_89_cast_fp16")];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_45_reps_0 = const()[name = string("tile_45_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 128, 256]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = v_89_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 1, 128, 256]> tile_45_cast_fp16 = tile(reps = tile_45_reps_0, x = transpose_90_cast_fp16)[name = string("tile_45_cast_fp16")];
            tensor<int32, [5]> concat_90 = const()[name = string("concat_90"), val = tensor<int32, [5]>([2, 2, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> reshape_90_cast_fp16 = reshape(shape = concat_90, x = tile_45_cast_fp16)[name = string("reshape_90_cast_fp16")];
            tensor<int32, [5]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_91 = const()[name = string("concat_91"), val = tensor<int32, [4]>([-1, 1, 128, 256])];
            tensor<fp16, [2, 2, 1, 128, 256]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = reshape_90_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 1, 128, 256]> reshape_91_cast_fp16 = reshape(shape = concat_91, x = transpose_91_cast_fp16)[name = string("reshape_91_cast_fp16")];
            tensor<int32, [4]> v_91_perm_0 = const()[name = string("v_91_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9613_transpose_x_0 = const()[name = string("op_9613_transpose_x_0"), val = bool(false)];
            bool var_9613_transpose_y_0 = const()[name = string("op_9613_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 128]> transpose_118_cast_fp16 = transpose(perm = transpose_118_perm_0, x = reshape_89_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 4, 128, 128]> var_9613_cast_fp16 = matmul(transpose_x = var_9613_transpose_x_0, transpose_y = var_9613_transpose_y_0, x = q_91_cast_fp16, y = transpose_118_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_9615_cast_fp16 = add(x = var_9613_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9615_cast_fp16")];
            int32 var_9616 = const()[name = string("op_9616"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_361_cast_fp16 = softmax(axis = var_9616, x = var_9615_cast_fp16)[name = string("w_361_cast_fp16")];
            bool var_9619_transpose_x_0 = const()[name = string("op_9619_transpose_x_0"), val = bool(false)];
            bool var_9619_transpose_y_0 = const()[name = string("op_9619_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 256]> v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = reshape_91_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 4, 128, 256]> var_9619_cast_fp16 = matmul(transpose_x = var_9619_transpose_x_0, transpose_y = var_9619_transpose_y_0, x = w_361_cast_fp16, y = v_91_cast_fp16)[name = string("op_9619_cast_fp16")];
            tensor<int32, [4]> var_9622_perm_0 = const()[name = string("op_9622_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9626 = const()[name = string("op_9626"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 128, 4, 256]> var_9622_cast_fp16 = transpose(perm = var_9622_perm_0, x = var_9619_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 128, 1024]> input_267_cast_fp16 = reshape(shape = var_9626, x = var_9622_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126424448))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126948800))))[name = string("tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9631_cast_fp16 = abs(x = linear_202_cast_fp16)[name = string("op_9631_cast_fp16")];
            bool var_9634 = const()[name = string("op_9634"), val = bool(true)];
            tensor<int32, [1]> var_9635_axes_0 = const()[name = string("op_9635_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9635_cast_fp16 = reduce_max(axes = var_9635_axes_0, keep_dims = var_9634, x = var_9631_cast_fp16)[name = string("op_9635_cast_fp16")];
            fp16 var_9636_to_fp16 = const()[name = string("op_9636_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_363_cast_fp16 = maximum(x = var_9635_cast_fp16, y = var_9636_to_fp16)[name = string("s_363_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_363_cast_fp16 = real_div(x = linear_202_cast_fp16, y = s_363_cast_fp16)[name = string("y_363_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9639_cast_fp16 = mul(x = y_363_cast_fp16, y = y_363_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [1]> var_9644_axes_0 = const()[name = string("op_9644_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9644_keep_dims_0 = const()[name = string("op_9644_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9644_cast_fp16 = reduce_mean(axes = var_9644_axes_0, keep_dims = var_9644_keep_dims_0, x = var_9639_cast_fp16)[name = string("op_9644_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9645_cast_fp16 = mul(x = s_363_cast_fp16, y = s_363_cast_fp16)[name = string("op_9645_cast_fp16")];
            fp32 var_9646_epsilon_0 = const()[name = string("op_9646_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9646_cast_fp16 = inverse(epsilon = var_9646_epsilon_0, x = var_9645_cast_fp16)[name = string("op_9646_cast_fp16")];
            fp16 var_9647_to_fp16 = const()[name = string("op_9647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9648_cast_fp16 = mul(x = var_9646_cast_fp16, y = var_9647_to_fp16)[name = string("op_9648_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9650_cast_fp16 = add(x = var_9644_cast_fp16, y = var_9648_cast_fp16)[name = string("op_9650_cast_fp16")];
            fp32 var_9651_epsilon_0 = const()[name = string("op_9651_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9651_cast_fp16 = rsqrt(epsilon = var_9651_epsilon_0, x = var_9650_cast_fp16)[name = string("op_9651_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_317_cast_fp16 = mul(x = y_363_cast_fp16, y = var_9651_cast_fp16)[name = string("out_317_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126949888)))];
            tensor<fp16, [1, 128, 512]> var_9653_cast_fp16 = mul(x = out_317_cast_fp16, y = tm_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_407_cast_fp16 = add(x = x_397_cast_fp16, y = var_9653_cast_fp16)[name = string("x_407_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9656_cast_fp16 = abs(x = x_407_cast_fp16)[name = string("op_9656_cast_fp16")];
            bool var_9659 = const()[name = string("op_9659"), val = bool(true)];
            tensor<int32, [1]> var_9660_axes_0 = const()[name = string("op_9660_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9660_cast_fp16 = reduce_max(axes = var_9660_axes_0, keep_dims = var_9659, x = var_9656_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp16 var_9661_to_fp16 = const()[name = string("op_9661_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_365_cast_fp16 = maximum(x = var_9660_cast_fp16, y = var_9661_to_fp16)[name = string("s_365_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_365_cast_fp16 = real_div(x = x_407_cast_fp16, y = s_365_cast_fp16)[name = string("y_365_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9664_cast_fp16 = mul(x = y_365_cast_fp16, y = y_365_cast_fp16)[name = string("op_9664_cast_fp16")];
            tensor<int32, [1]> var_9669_axes_0 = const()[name = string("op_9669_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9669_keep_dims_0 = const()[name = string("op_9669_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9669_cast_fp16 = reduce_mean(axes = var_9669_axes_0, keep_dims = var_9669_keep_dims_0, x = var_9664_cast_fp16)[name = string("op_9669_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9670_cast_fp16 = mul(x = s_365_cast_fp16, y = s_365_cast_fp16)[name = string("op_9670_cast_fp16")];
            fp32 var_9671_epsilon_0 = const()[name = string("op_9671_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9671_cast_fp16 = inverse(epsilon = var_9671_epsilon_0, x = var_9670_cast_fp16)[name = string("op_9671_cast_fp16")];
            fp16 var_9672_to_fp16 = const()[name = string("op_9672_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9673_cast_fp16 = mul(x = var_9671_cast_fp16, y = var_9672_to_fp16)[name = string("op_9673_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9675_cast_fp16 = add(x = var_9669_cast_fp16, y = var_9673_cast_fp16)[name = string("op_9675_cast_fp16")];
            fp32 var_9676_epsilon_0 = const()[name = string("op_9676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9676_cast_fp16 = rsqrt(epsilon = var_9676_epsilon_0, x = var_9675_cast_fp16)[name = string("op_9676_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_319_cast_fp16 = mul(x = y_365_cast_fp16, y = var_9676_cast_fp16)[name = string("out_319_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126950976)))];
            tensor<fp16, [1, 128, 512]> input_269_cast_fp16 = mul(x = out_319_cast_fp16, y = tm_layers_22_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126952064))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128000704))))[name = string("tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_203_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_203_cast_fp16")];
            string var_9683_mode_0 = const()[name = string("op_9683_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_9683_cast_fp16 = gelu(mode = var_9683_mode_0, x = linear_203_cast_fp16)[name = string("op_9683_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128004864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129053504))))[name = string("tm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_204_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_271_cast_fp16 = mul(x = var_9683_cast_fp16, y = linear_204_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129057664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130106304))))[name = string("tm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9691_cast_fp16 = abs(x = linear_205_cast_fp16)[name = string("op_9691_cast_fp16")];
            bool var_9694 = const()[name = string("op_9694"), val = bool(true)];
            tensor<int32, [1]> var_9695_axes_0 = const()[name = string("op_9695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9695_cast_fp16 = reduce_max(axes = var_9695_axes_0, keep_dims = var_9694, x = var_9691_cast_fp16)[name = string("op_9695_cast_fp16")];
            fp16 var_9696_to_fp16 = const()[name = string("op_9696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_367_cast_fp16 = maximum(x = var_9695_cast_fp16, y = var_9696_to_fp16)[name = string("s_367_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_367_cast_fp16 = real_div(x = linear_205_cast_fp16, y = s_367_cast_fp16)[name = string("y_367_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9699_cast_fp16 = mul(x = y_367_cast_fp16, y = y_367_cast_fp16)[name = string("op_9699_cast_fp16")];
            tensor<int32, [1]> var_9704_axes_0 = const()[name = string("op_9704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9704_keep_dims_0 = const()[name = string("op_9704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9704_cast_fp16 = reduce_mean(axes = var_9704_axes_0, keep_dims = var_9704_keep_dims_0, x = var_9699_cast_fp16)[name = string("op_9704_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9705_cast_fp16 = mul(x = s_367_cast_fp16, y = s_367_cast_fp16)[name = string("op_9705_cast_fp16")];
            fp32 var_9706_epsilon_0 = const()[name = string("op_9706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9706_cast_fp16 = inverse(epsilon = var_9706_epsilon_0, x = var_9705_cast_fp16)[name = string("op_9706_cast_fp16")];
            fp16 var_9707_to_fp16 = const()[name = string("op_9707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9708_cast_fp16 = mul(x = var_9706_cast_fp16, y = var_9707_to_fp16)[name = string("op_9708_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9710_cast_fp16 = add(x = var_9704_cast_fp16, y = var_9708_cast_fp16)[name = string("op_9710_cast_fp16")];
            fp32 var_9711_epsilon_0 = const()[name = string("op_9711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9711_cast_fp16 = rsqrt(epsilon = var_9711_epsilon_0, x = var_9710_cast_fp16)[name = string("op_9711_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_321_cast_fp16 = mul(x = y_367_cast_fp16, y = var_9711_cast_fp16)[name = string("out_321_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130107392)))];
            tensor<fp16, [1, 128, 512]> var_9713_cast_fp16 = mul(x = out_321_cast_fp16, y = tm_layers_22_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9713_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_273_cast_fp16 = add(x = x_407_cast_fp16, y = var_9713_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130108480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130370688))))[name = string("tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_206_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_206_cast_fp16")];
            string x_411_mode_0 = const()[name = string("x_411_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_411_cast_fp16 = gelu(mode = x_411_mode_0, x = linear_206_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [4]> var_9733_begin_0 = const()[name = string("op_9733_begin_0"), val = tensor<int32, [4]>([0, 0, 22, 0])];
            tensor<int32, [4]> var_9733_end_0 = const()[name = string("op_9733_end_0"), val = tensor<int32, [4]>([1, 128, 23, 512])];
            tensor<bool, [4]> var_9733_end_mask_0 = const()[name = string("op_9733_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9733_squeeze_mask_0 = const()[name = string("op_9733_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_9733_cast_fp16 = slice_by_index(begin = var_9733_begin_0, end = var_9733_end_0, end_mask = var_9733_end_mask_0, squeeze_mask = var_9733_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9733_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_275_cast_fp16 = mul(x = x_411_cast_fp16, y = var_9733_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130371776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130633984))))[name = string("tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9743_cast_fp16 = abs(x = linear_207_cast_fp16)[name = string("op_9743_cast_fp16")];
            bool var_9746 = const()[name = string("op_9746"), val = bool(true)];
            tensor<int32, [1]> var_9747_axes_0 = const()[name = string("op_9747_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9747_cast_fp16 = reduce_max(axes = var_9747_axes_0, keep_dims = var_9746, x = var_9743_cast_fp16)[name = string("op_9747_cast_fp16")];
            fp16 var_9748_to_fp16 = const()[name = string("op_9748_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_369_cast_fp16 = maximum(x = var_9747_cast_fp16, y = var_9748_to_fp16)[name = string("s_369_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_369_cast_fp16 = real_div(x = linear_207_cast_fp16, y = s_369_cast_fp16)[name = string("y_369_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9751_cast_fp16 = mul(x = y_369_cast_fp16, y = y_369_cast_fp16)[name = string("op_9751_cast_fp16")];
            tensor<int32, [1]> var_9756_axes_0 = const()[name = string("op_9756_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9756_keep_dims_0 = const()[name = string("op_9756_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9756_cast_fp16 = reduce_mean(axes = var_9756_axes_0, keep_dims = var_9756_keep_dims_0, x = var_9751_cast_fp16)[name = string("op_9756_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9757_cast_fp16 = mul(x = s_369_cast_fp16, y = s_369_cast_fp16)[name = string("op_9757_cast_fp16")];
            fp32 var_9758_epsilon_0 = const()[name = string("op_9758_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9758_cast_fp16 = inverse(epsilon = var_9758_epsilon_0, x = var_9757_cast_fp16)[name = string("op_9758_cast_fp16")];
            fp16 var_9759_to_fp16 = const()[name = string("op_9759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9760_cast_fp16 = mul(x = var_9758_cast_fp16, y = var_9759_to_fp16)[name = string("op_9760_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9762_cast_fp16 = add(x = var_9756_cast_fp16, y = var_9760_cast_fp16)[name = string("op_9762_cast_fp16")];
            fp32 var_9763_epsilon_0 = const()[name = string("op_9763_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9763_cast_fp16 = rsqrt(epsilon = var_9763_epsilon_0, x = var_9762_cast_fp16)[name = string("op_9763_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_323_cast_fp16 = mul(x = y_369_cast_fp16, y = var_9763_cast_fp16)[name = string("out_323_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130635072)))];
            tensor<fp16, [1, 128, 512]> var_9765_cast_fp16 = mul(x = out_323_cast_fp16, y = tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9765_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_45_cast_fp16 = add(x = input_273_cast_fp16, y = var_9765_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<fp16, [1]> tm_layers_22_layer_scalar_to_fp16 = const()[name = string("tm_layers_22_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 128, 512]> x_415_cast_fp16 = mul(x = h_45_cast_fp16, y = tm_layers_22_layer_scalar_to_fp16)[name = string("x_415_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9769_cast_fp16 = abs(x = x_415_cast_fp16)[name = string("op_9769_cast_fp16")];
            bool var_9772 = const()[name = string("op_9772"), val = bool(true)];
            tensor<int32, [1]> var_9773_axes_0 = const()[name = string("op_9773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9773_cast_fp16 = reduce_max(axes = var_9773_axes_0, keep_dims = var_9772, x = var_9769_cast_fp16)[name = string("op_9773_cast_fp16")];
            fp16 var_9774_to_fp16 = const()[name = string("op_9774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_371_cast_fp16 = maximum(x = var_9773_cast_fp16, y = var_9774_to_fp16)[name = string("s_371_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_371_cast_fp16 = real_div(x = x_415_cast_fp16, y = s_371_cast_fp16)[name = string("y_371_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9777_cast_fp16 = mul(x = y_371_cast_fp16, y = y_371_cast_fp16)[name = string("op_9777_cast_fp16")];
            tensor<int32, [1]> var_9782_axes_0 = const()[name = string("op_9782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9782_keep_dims_0 = const()[name = string("op_9782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9782_cast_fp16 = reduce_mean(axes = var_9782_axes_0, keep_dims = var_9782_keep_dims_0, x = var_9777_cast_fp16)[name = string("op_9782_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9783_cast_fp16 = mul(x = s_371_cast_fp16, y = s_371_cast_fp16)[name = string("op_9783_cast_fp16")];
            fp32 var_9784_epsilon_0 = const()[name = string("op_9784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9784_cast_fp16 = inverse(epsilon = var_9784_epsilon_0, x = var_9783_cast_fp16)[name = string("op_9784_cast_fp16")];
            fp16 var_9785_to_fp16 = const()[name = string("op_9785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9786_cast_fp16 = mul(x = var_9784_cast_fp16, y = var_9785_to_fp16)[name = string("op_9786_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9788_cast_fp16 = add(x = var_9782_cast_fp16, y = var_9786_cast_fp16)[name = string("op_9788_cast_fp16")];
            fp32 var_9789_epsilon_0 = const()[name = string("op_9789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9789_cast_fp16 = rsqrt(epsilon = var_9789_epsilon_0, x = var_9788_cast_fp16)[name = string("op_9789_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_325_cast_fp16 = mul(x = y_371_cast_fp16, y = var_9789_cast_fp16)[name = string("out_325_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130636160)))];
            tensor<fp16, [1, 128, 512]> input_277_cast_fp16 = mul(x = out_325_cast_fp16, y = tm_layers_23_input_layernorm_weight_to_fp16)[name = string("input_277_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130637248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131685888))))[name = string("tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_208_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<int32, [4]> var_9799 = const()[name = string("op_9799"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 128, 4, 512]> var_9800_cast_fp16 = reshape(shape = var_9799, x = linear_208_cast_fp16)[name = string("op_9800_cast_fp16")];
            tensor<int32, [4]> x_417_perm_0 = const()[name = string("x_417_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131690048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131952256))))[name = string("tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<int32, [4]> var_9811 = const()[name = string("op_9811"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 128, 1, 512]> var_9812_cast_fp16 = reshape(shape = var_9811, x = linear_209_cast_fp16)[name = string("op_9812_cast_fp16")];
            tensor<int32, [4]> x_419_perm_0 = const()[name = string("x_419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131953344))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132215552))))[name = string("tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<int32, [4]> var_9823 = const()[name = string("op_9823"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 128, 1, 512]> var_9824_cast_fp16 = reshape(shape = var_9823, x = linear_210_cast_fp16)[name = string("op_9824_cast_fp16")];
            tensor<int32, [4]> x_421_perm_0 = const()[name = string("x_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 128, 512]> x_417_cast_fp16 = transpose(perm = x_417_perm_0, x = var_9800_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 4, 128, 512]> var_9828_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_9828_cast_fp16")];
            bool var_9831 = const()[name = string("op_9831"), val = bool(true)];
            tensor<int32, [1]> var_9832_axes_0 = const()[name = string("op_9832_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 128, 1]> var_9832_cast_fp16 = reduce_max(axes = var_9832_axes_0, keep_dims = var_9831, x = var_9828_cast_fp16)[name = string("op_9832_cast_fp16")];
            fp16 var_9833_to_fp16 = const()[name = string("op_9833_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 128, 1]> s_373_cast_fp16 = maximum(x = var_9832_cast_fp16, y = var_9833_to_fp16)[name = string("s_373_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> y_373_cast_fp16 = real_div(x = x_417_cast_fp16, y = s_373_cast_fp16)[name = string("y_373_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_9836_cast_fp16 = mul(x = y_373_cast_fp16, y = y_373_cast_fp16)[name = string("op_9836_cast_fp16")];
            tensor<int32, [1]> var_9841_axes_0 = const()[name = string("op_9841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9841_keep_dims_0 = const()[name = string("op_9841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 128, 1]> var_9841_cast_fp16 = reduce_mean(axes = var_9841_axes_0, keep_dims = var_9841_keep_dims_0, x = var_9836_cast_fp16)[name = string("op_9841_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_9842_cast_fp16 = mul(x = s_373_cast_fp16, y = s_373_cast_fp16)[name = string("op_9842_cast_fp16")];
            fp32 var_9843_epsilon_0 = const()[name = string("op_9843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 128, 1]> var_9843_cast_fp16 = inverse(epsilon = var_9843_epsilon_0, x = var_9842_cast_fp16)[name = string("op_9843_cast_fp16")];
            fp16 var_9844_to_fp16 = const()[name = string("op_9844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 128, 1]> var_9845_cast_fp16 = mul(x = var_9843_cast_fp16, y = var_9844_to_fp16)[name = string("op_9845_cast_fp16")];
            tensor<fp16, [1, 4, 128, 1]> var_9847_cast_fp16 = add(x = var_9841_cast_fp16, y = var_9845_cast_fp16)[name = string("op_9847_cast_fp16")];
            fp32 var_9848_epsilon_0 = const()[name = string("op_9848_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 128, 1]> var_9848_cast_fp16 = rsqrt(epsilon = var_9848_epsilon_0, x = var_9847_cast_fp16)[name = string("op_9848_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> out_327_cast_fp16 = mul(x = y_373_cast_fp16, y = var_9848_cast_fp16)[name = string("out_327_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132216640)))];
            tensor<fp16, [1, 4, 128, 512]> q_93_cast_fp16 = mul(x = out_327_cast_fp16, y = tm_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("q_93_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> x_419_cast_fp16 = transpose(perm = x_419_perm_0, x = var_9812_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1, 128, 512]> var_9851_cast_fp16 = abs(x = x_419_cast_fp16)[name = string("op_9851_cast_fp16")];
            bool var_9854 = const()[name = string("op_9854"), val = bool(true)];
            tensor<int32, [1]> var_9855_axes_0 = const()[name = string("op_9855_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 128, 1]> var_9855_cast_fp16 = reduce_max(axes = var_9855_axes_0, keep_dims = var_9854, x = var_9851_cast_fp16)[name = string("op_9855_cast_fp16")];
            fp16 var_9856_to_fp16 = const()[name = string("op_9856_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 128, 1]> s_375_cast_fp16 = maximum(x = var_9855_cast_fp16, y = var_9856_to_fp16)[name = string("s_375_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> y_375_cast_fp16 = real_div(x = x_419_cast_fp16, y = s_375_cast_fp16)[name = string("y_375_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_9859_cast_fp16 = mul(x = y_375_cast_fp16, y = y_375_cast_fp16)[name = string("op_9859_cast_fp16")];
            tensor<int32, [1]> var_9864_axes_0 = const()[name = string("op_9864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9864_keep_dims_0 = const()[name = string("op_9864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_9864_cast_fp16 = reduce_mean(axes = var_9864_axes_0, keep_dims = var_9864_keep_dims_0, x = var_9859_cast_fp16)[name = string("op_9864_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_9865_cast_fp16 = mul(x = s_375_cast_fp16, y = s_375_cast_fp16)[name = string("op_9865_cast_fp16")];
            fp32 var_9866_epsilon_0 = const()[name = string("op_9866_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 128, 1]> var_9866_cast_fp16 = inverse(epsilon = var_9866_epsilon_0, x = var_9865_cast_fp16)[name = string("op_9866_cast_fp16")];
            fp16 var_9867_to_fp16 = const()[name = string("op_9867_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 128, 1]> var_9868_cast_fp16 = mul(x = var_9866_cast_fp16, y = var_9867_to_fp16)[name = string("op_9868_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_9870_cast_fp16 = add(x = var_9864_cast_fp16, y = var_9868_cast_fp16)[name = string("op_9870_cast_fp16")];
            fp32 var_9871_epsilon_0 = const()[name = string("op_9871_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 128, 1]> var_9871_cast_fp16 = rsqrt(epsilon = var_9871_epsilon_0, x = var_9870_cast_fp16)[name = string("op_9871_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> out_329_cast_fp16 = mul(x = y_375_cast_fp16, y = var_9871_cast_fp16)[name = string("out_329_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132217728)))];
            tensor<fp16, [1, 1, 128, 512]> k_139_cast_fp16 = mul(x = out_329_cast_fp16, y = tm_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("k_139_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> x_421_cast_fp16 = transpose(perm = x_421_perm_0, x = var_9824_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 1, 128, 512]> var_9874_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_9874_cast_fp16")];
            bool var_9877 = const()[name = string("op_9877"), val = bool(true)];
            tensor<int32, [1]> var_9878_axes_0 = const()[name = string("op_9878_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 128, 1]> var_9878_cast_fp16 = reduce_max(axes = var_9878_axes_0, keep_dims = var_9877, x = var_9874_cast_fp16)[name = string("op_9878_cast_fp16")];
            fp16 var_9879_to_fp16 = const()[name = string("op_9879_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 128, 1]> s_377_cast_fp16 = maximum(x = var_9878_cast_fp16, y = var_9879_to_fp16)[name = string("s_377_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> y_377_cast_fp16 = real_div(x = x_421_cast_fp16, y = s_377_cast_fp16)[name = string("y_377_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_9882_cast_fp16 = mul(x = y_377_cast_fp16, y = y_377_cast_fp16)[name = string("op_9882_cast_fp16")];
            tensor<int32, [1]> var_9887_axes_0 = const()[name = string("op_9887_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9887_keep_dims_0 = const()[name = string("op_9887_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_9887_cast_fp16 = reduce_mean(axes = var_9887_axes_0, keep_dims = var_9887_keep_dims_0, x = var_9882_cast_fp16)[name = string("op_9887_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_9888_cast_fp16 = mul(x = s_377_cast_fp16, y = s_377_cast_fp16)[name = string("op_9888_cast_fp16")];
            fp32 var_9889_epsilon_0 = const()[name = string("op_9889_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 128, 1]> var_9889_cast_fp16 = inverse(epsilon = var_9889_epsilon_0, x = var_9888_cast_fp16)[name = string("op_9889_cast_fp16")];
            fp16 var_9890_to_fp16 = const()[name = string("op_9890_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 128, 1]> var_9891_cast_fp16 = mul(x = var_9889_cast_fp16, y = var_9890_to_fp16)[name = string("op_9891_cast_fp16")];
            tensor<fp16, [1, 1, 128, 1]> var_9893_cast_fp16 = add(x = var_9887_cast_fp16, y = var_9891_cast_fp16)[name = string("op_9893_cast_fp16")];
            fp32 var_9894_epsilon_0 = const()[name = string("op_9894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 128, 1]> var_9894_cast_fp16 = rsqrt(epsilon = var_9894_epsilon_0, x = var_9893_cast_fp16)[name = string("op_9894_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> v_93_cast_fp16 = mul(x = y_377_cast_fp16, y = var_9894_cast_fp16)[name = string("v_93_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_9896_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9896_cast_fp16")];
            tensor<int32, [4]> var_9901_begin_0 = const()[name = string("op_9901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9901_end_0 = const()[name = string("op_9901_end_0"), val = tensor<int32, [4]>([1, 4, 128, 512])];
            tensor<bool, [4]> var_9901_end_mask_0 = const()[name = string("op_9901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 128, 256]> var_9901_cast_fp16 = slice_by_index(begin = var_9901_begin_0, end = var_9901_end_0, end_mask = var_9901_end_mask_0, x = q_93_cast_fp16)[name = string("op_9901_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 128, 256]> var_9902_cast_fp16 = mul(x = var_9901_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_9902_cast_fp16")];
            tensor<int32, [4]> var_9907_begin_0 = const()[name = string("op_9907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9907_end_0 = const()[name = string("op_9907_end_0"), val = tensor<int32, [4]>([1, 4, 128, 256])];
            tensor<bool, [4]> var_9907_end_mask_0 = const()[name = string("op_9907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 128, 256]> var_9907_cast_fp16 = slice_by_index(begin = var_9907_begin_0, end = var_9907_end_0, end_mask = var_9907_end_mask_0, x = q_93_cast_fp16)[name = string("op_9907_cast_fp16")];
            int32 var_9909 = const()[name = string("op_9909"), val = int32(-1)];
            bool var_9910_interleave_0 = const()[name = string("op_9910_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 512]> var_9910_cast_fp16 = concat(axis = var_9909, interleave = var_9910_interleave_0, values = (var_9902_cast_fp16, var_9907_cast_fp16))[name = string("op_9910_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> var_9911_cast_fp16 = mul(x = var_9910_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9911_cast_fp16")];
            tensor<fp16, [1, 4, 128, 512]> q_cast_fp16 = add(x = var_9896_cast_fp16, y = var_9911_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_9914_cast_fp16 = mul(x = k_139_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9914_cast_fp16")];
            tensor<int32, [4]> var_9919_begin_0 = const()[name = string("op_9919_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9919_end_0 = const()[name = string("op_9919_end_0"), val = tensor<int32, [4]>([1, 1, 128, 512])];
            tensor<bool, [4]> var_9919_end_mask_0 = const()[name = string("op_9919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 128, 256]> var_9919_cast_fp16 = slice_by_index(begin = var_9919_begin_0, end = var_9919_end_0, end_mask = var_9919_end_mask_0, x = k_139_cast_fp16)[name = string("op_9919_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 128, 256]> var_9920_cast_fp16 = mul(x = var_9919_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_9920_cast_fp16")];
            tensor<int32, [4]> var_9925_begin_0 = const()[name = string("op_9925_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9925_end_0 = const()[name = string("op_9925_end_0"), val = tensor<int32, [4]>([1, 1, 128, 256])];
            tensor<bool, [4]> var_9925_end_mask_0 = const()[name = string("op_9925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 128, 256]> var_9925_cast_fp16 = slice_by_index(begin = var_9925_begin_0, end = var_9925_end_0, end_mask = var_9925_end_mask_0, x = k_139_cast_fp16)[name = string("op_9925_cast_fp16")];
            int32 var_9927 = const()[name = string("op_9927"), val = int32(-1)];
            bool var_9928_interleave_0 = const()[name = string("op_9928_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 128, 512]> var_9928_cast_fp16 = concat(axis = var_9927, interleave = var_9928_interleave_0, values = (var_9920_cast_fp16, var_9925_cast_fp16))[name = string("op_9928_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> var_9929_cast_fp16 = mul(x = var_9928_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9929_cast_fp16")];
            tensor<fp16, [1, 1, 128, 512]> k_141_cast_fp16 = add(x = var_9914_cast_fp16, y = var_9929_cast_fp16)[name = string("k_141_cast_fp16")];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_46_reps_0 = const()[name = string("tile_46_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 128, 512]> transpose_92_cast_fp16 = transpose(perm = transpose_92_perm_0, x = k_141_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 1, 128, 512]> tile_46_cast_fp16 = tile(reps = tile_46_reps_0, x = transpose_92_cast_fp16)[name = string("tile_46_cast_fp16")];
            tensor<int32, [5]> concat_92 = const()[name = string("concat_92"), val = tensor<int32, [5]>([4, 1, 1, 128, 512])];
            tensor<fp16, [4, 1, 1, 128, 512]> reshape_92_cast_fp16 = reshape(shape = concat_92, x = tile_46_cast_fp16)[name = string("reshape_92_cast_fp16")];
            tensor<int32, [5]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_93 = const()[name = string("concat_93"), val = tensor<int32, [4]>([-1, 1, 128, 512])];
            tensor<fp16, [1, 4, 1, 128, 512]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = reshape_92_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 1, 128, 512]> reshape_93_cast_fp16 = reshape(shape = concat_93, x = transpose_93_cast_fp16)[name = string("reshape_93_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_47_reps_0 = const()[name = string("tile_47_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 128, 512]> transpose_94_cast_fp16 = transpose(perm = transpose_94_perm_0, x = v_93_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 1, 128, 512]> tile_47_cast_fp16 = tile(reps = tile_47_reps_0, x = transpose_94_cast_fp16)[name = string("tile_47_cast_fp16")];
            tensor<int32, [5]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [5]>([4, 1, 1, 128, 512])];
            tensor<fp16, [4, 1, 1, 128, 512]> reshape_94_cast_fp16 = reshape(shape = concat_94, x = tile_47_cast_fp16)[name = string("reshape_94_cast_fp16")];
            tensor<int32, [5]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [4]>([-1, 1, 128, 512])];
            tensor<fp16, [1, 4, 1, 128, 512]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = reshape_94_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 1, 128, 512]> reshape_95_cast_fp16 = reshape(shape = concat_95, x = transpose_95_cast_fp16)[name = string("reshape_95_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9943_transpose_x_0 = const()[name = string("op_9943_transpose_x_0"), val = bool(false)];
            bool var_9943_transpose_y_0 = const()[name = string("op_9943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 128]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = reshape_93_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 4, 128, 128]> var_9943_cast_fp16 = matmul(transpose_x = var_9943_transpose_x_0, transpose_y = var_9943_transpose_y_0, x = q_cast_fp16, y = transpose_119_cast_fp16)[name = string("op_9943_cast_fp16")];
            tensor<fp16, [1, 4, 128, 128]> var_9945_cast_fp16 = add(x = var_9943_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9945_cast_fp16")];
            int32 var_9946 = const()[name = string("op_9946"), val = int32(-1)];
            tensor<fp16, [1, 4, 128, 128]> w_377_cast_fp16 = softmax(axis = var_9946, x = var_9945_cast_fp16)[name = string("w_377_cast_fp16")];
            bool var_9949_transpose_x_0 = const()[name = string("op_9949_transpose_x_0"), val = bool(false)];
            bool var_9949_transpose_y_0 = const()[name = string("op_9949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 128, 512]> v_cast_fp16 = transpose(perm = v_perm_0, x = reshape_95_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 4, 128, 512]> var_9949_cast_fp16 = matmul(transpose_x = var_9949_transpose_x_0, transpose_y = var_9949_transpose_y_0, x = w_377_cast_fp16, y = v_cast_fp16)[name = string("op_9949_cast_fp16")];
            tensor<int32, [4]> var_9952_perm_0 = const()[name = string("op_9952_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9956 = const()[name = string("op_9956"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 128, 4, 512]> var_9952_cast_fp16 = transpose(perm = var_9952_perm_0, x = var_9949_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 128, 2048]> input_279_cast_fp16 = reshape(shape = var_9956, x = var_9952_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132218816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133267456))))[name = string("tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9961_cast_fp16 = abs(x = linear_211_cast_fp16)[name = string("op_9961_cast_fp16")];
            bool var_9964 = const()[name = string("op_9964"), val = bool(true)];
            tensor<int32, [1]> var_9965_axes_0 = const()[name = string("op_9965_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9965_cast_fp16 = reduce_max(axes = var_9965_axes_0, keep_dims = var_9964, x = var_9961_cast_fp16)[name = string("op_9965_cast_fp16")];
            fp16 var_9966_to_fp16 = const()[name = string("op_9966_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_379_cast_fp16 = maximum(x = var_9965_cast_fp16, y = var_9966_to_fp16)[name = string("s_379_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_379_cast_fp16 = real_div(x = linear_211_cast_fp16, y = s_379_cast_fp16)[name = string("y_379_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9969_cast_fp16 = mul(x = y_379_cast_fp16, y = y_379_cast_fp16)[name = string("op_9969_cast_fp16")];
            tensor<int32, [1]> var_9974_axes_0 = const()[name = string("op_9974_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9974_keep_dims_0 = const()[name = string("op_9974_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9974_cast_fp16 = reduce_mean(axes = var_9974_axes_0, keep_dims = var_9974_keep_dims_0, x = var_9969_cast_fp16)[name = string("op_9974_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9975_cast_fp16 = mul(x = s_379_cast_fp16, y = s_379_cast_fp16)[name = string("op_9975_cast_fp16")];
            fp32 var_9976_epsilon_0 = const()[name = string("op_9976_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_9976_cast_fp16 = inverse(epsilon = var_9976_epsilon_0, x = var_9975_cast_fp16)[name = string("op_9976_cast_fp16")];
            fp16 var_9977_to_fp16 = const()[name = string("op_9977_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_9978_cast_fp16 = mul(x = var_9976_cast_fp16, y = var_9977_to_fp16)[name = string("op_9978_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_9980_cast_fp16 = add(x = var_9974_cast_fp16, y = var_9978_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp32 var_9981_epsilon_0 = const()[name = string("op_9981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_9981_cast_fp16 = rsqrt(epsilon = var_9981_epsilon_0, x = var_9980_cast_fp16)[name = string("op_9981_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_331_cast_fp16 = mul(x = y_379_cast_fp16, y = var_9981_cast_fp16)[name = string("out_331_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133268544)))];
            tensor<fp16, [1, 128, 512]> var_9983_cast_fp16 = mul(x = out_331_cast_fp16, y = tm_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("op_9983_cast_fp16")];
            tensor<fp16, [1, 128, 512]> x_425_cast_fp16 = add(x = x_415_cast_fp16, y = var_9983_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9986_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_9986_cast_fp16")];
            bool var_9989 = const()[name = string("op_9989"), val = bool(true)];
            tensor<int32, [1]> var_9990_axes_0 = const()[name = string("op_9990_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_9990_cast_fp16 = reduce_max(axes = var_9990_axes_0, keep_dims = var_9989, x = var_9986_cast_fp16)[name = string("op_9990_cast_fp16")];
            fp16 var_9991_to_fp16 = const()[name = string("op_9991_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_381_cast_fp16 = maximum(x = var_9990_cast_fp16, y = var_9991_to_fp16)[name = string("s_381_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_381_cast_fp16 = real_div(x = x_425_cast_fp16, y = s_381_cast_fp16)[name = string("y_381_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_9994_cast_fp16 = mul(x = y_381_cast_fp16, y = y_381_cast_fp16)[name = string("op_9994_cast_fp16")];
            tensor<int32, [1]> var_9999_axes_0 = const()[name = string("op_9999_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9999_keep_dims_0 = const()[name = string("op_9999_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9999_cast_fp16 = reduce_mean(axes = var_9999_axes_0, keep_dims = var_9999_keep_dims_0, x = var_9994_cast_fp16)[name = string("op_9999_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_10000_cast_fp16 = mul(x = s_381_cast_fp16, y = s_381_cast_fp16)[name = string("op_10000_cast_fp16")];
            fp32 var_10001_epsilon_0 = const()[name = string("op_10001_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_10001_cast_fp16 = inverse(epsilon = var_10001_epsilon_0, x = var_10000_cast_fp16)[name = string("op_10001_cast_fp16")];
            fp16 var_10002_to_fp16 = const()[name = string("op_10002_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_10003_cast_fp16 = mul(x = var_10001_cast_fp16, y = var_10002_to_fp16)[name = string("op_10003_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_10005_cast_fp16 = add(x = var_9999_cast_fp16, y = var_10003_cast_fp16)[name = string("op_10005_cast_fp16")];
            fp32 var_10006_epsilon_0 = const()[name = string("op_10006_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_10006_cast_fp16 = rsqrt(epsilon = var_10006_epsilon_0, x = var_10005_cast_fp16)[name = string("op_10006_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_333_cast_fp16 = mul(x = y_381_cast_fp16, y = var_10006_cast_fp16)[name = string("out_333_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133269632)))];
            tensor<fp16, [1, 128, 512]> input_281_cast_fp16 = mul(x = out_333_cast_fp16, y = tm_layers_23_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133270720))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134319360))))[name = string("tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_212_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_212_cast_fp16")];
            string var_10013_mode_0 = const()[name = string("op_10013_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 2048]> var_10013_cast_fp16 = gelu(mode = var_10013_mode_0, x = linear_212_cast_fp16)[name = string("op_10013_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134323520))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135372160))))[name = string("tm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 2048]> linear_213_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_213_cast_fp16")];
            tensor<fp16, [1, 128, 2048]> input_283_cast_fp16 = mul(x = var_10013_cast_fp16, y = linear_213_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135376320))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136424960))))[name = string("tm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_283_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_10021_cast_fp16 = abs(x = linear_214_cast_fp16)[name = string("op_10021_cast_fp16")];
            bool var_10024 = const()[name = string("op_10024"), val = bool(true)];
            tensor<int32, [1]> var_10025_axes_0 = const()[name = string("op_10025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_10025_cast_fp16 = reduce_max(axes = var_10025_axes_0, keep_dims = var_10024, x = var_10021_cast_fp16)[name = string("op_10025_cast_fp16")];
            fp16 var_10026_to_fp16 = const()[name = string("op_10026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_383_cast_fp16 = maximum(x = var_10025_cast_fp16, y = var_10026_to_fp16)[name = string("s_383_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_383_cast_fp16 = real_div(x = linear_214_cast_fp16, y = s_383_cast_fp16)[name = string("y_383_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_10029_cast_fp16 = mul(x = y_383_cast_fp16, y = y_383_cast_fp16)[name = string("op_10029_cast_fp16")];
            tensor<int32, [1]> var_10034_axes_0 = const()[name = string("op_10034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10034_keep_dims_0 = const()[name = string("op_10034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_10034_cast_fp16 = reduce_mean(axes = var_10034_axes_0, keep_dims = var_10034_keep_dims_0, x = var_10029_cast_fp16)[name = string("op_10034_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_10035_cast_fp16 = mul(x = s_383_cast_fp16, y = s_383_cast_fp16)[name = string("op_10035_cast_fp16")];
            fp32 var_10036_epsilon_0 = const()[name = string("op_10036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_10036_cast_fp16 = inverse(epsilon = var_10036_epsilon_0, x = var_10035_cast_fp16)[name = string("op_10036_cast_fp16")];
            fp16 var_10037_to_fp16 = const()[name = string("op_10037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_10038_cast_fp16 = mul(x = var_10036_cast_fp16, y = var_10037_to_fp16)[name = string("op_10038_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_10040_cast_fp16 = add(x = var_10034_cast_fp16, y = var_10038_cast_fp16)[name = string("op_10040_cast_fp16")];
            fp32 var_10041_epsilon_0 = const()[name = string("op_10041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_10041_cast_fp16 = rsqrt(epsilon = var_10041_epsilon_0, x = var_10040_cast_fp16)[name = string("op_10041_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_335_cast_fp16 = mul(x = y_383_cast_fp16, y = var_10041_cast_fp16)[name = string("out_335_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136426048)))];
            tensor<fp16, [1, 128, 512]> var_10043_cast_fp16 = mul(x = out_335_cast_fp16, y = tm_layers_23_post_feedforward_layernorm_weight_to_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_285_cast_fp16 = add(x = x_425_cast_fp16, y = var_10043_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136427136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136689344))))[name = string("tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_215_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = string("linear_215_cast_fp16")];
            string x_429_mode_0 = const()[name = string("x_429_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 128, 512]> x_429_cast_fp16 = gelu(mode = x_429_mode_0, x = linear_215_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [4]> var_10063_begin_0 = const()[name = string("op_10063_begin_0"), val = tensor<int32, [4]>([0, 0, 23, 0])];
            tensor<int32, [4]> var_10063_end_0 = const()[name = string("op_10063_end_0"), val = tensor<int32, [4]>([1, 128, 24, 512])];
            tensor<bool, [4]> var_10063_end_mask_0 = const()[name = string("op_10063_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_10063_squeeze_mask_0 = const()[name = string("op_10063_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 128, 512]> var_10063_cast_fp16 = slice_by_index(begin = var_10063_begin_0, end = var_10063_end_0, end_mask = var_10063_end_mask_0, squeeze_mask = var_10063_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 128, 512]> input_287_cast_fp16 = mul(x = x_429_cast_fp16, y = var_10063_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136690432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136952640))))[name = string("tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 512]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_10073_cast_fp16 = abs(x = linear_216_cast_fp16)[name = string("op_10073_cast_fp16")];
            bool var_10076 = const()[name = string("op_10076"), val = bool(true)];
            tensor<int32, [1]> var_10077_axes_0 = const()[name = string("op_10077_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_10077_cast_fp16 = reduce_max(axes = var_10077_axes_0, keep_dims = var_10076, x = var_10073_cast_fp16)[name = string("op_10077_cast_fp16")];
            fp16 var_10078_to_fp16 = const()[name = string("op_10078_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_385_cast_fp16 = maximum(x = var_10077_cast_fp16, y = var_10078_to_fp16)[name = string("s_385_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_385_cast_fp16 = real_div(x = linear_216_cast_fp16, y = s_385_cast_fp16)[name = string("y_385_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_10081_cast_fp16 = mul(x = y_385_cast_fp16, y = y_385_cast_fp16)[name = string("op_10081_cast_fp16")];
            tensor<int32, [1]> var_10086_axes_0 = const()[name = string("op_10086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10086_keep_dims_0 = const()[name = string("op_10086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_10086_cast_fp16 = reduce_mean(axes = var_10086_axes_0, keep_dims = var_10086_keep_dims_0, x = var_10081_cast_fp16)[name = string("op_10086_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_10087_cast_fp16 = mul(x = s_385_cast_fp16, y = s_385_cast_fp16)[name = string("op_10087_cast_fp16")];
            fp32 var_10088_epsilon_0 = const()[name = string("op_10088_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_10088_cast_fp16 = inverse(epsilon = var_10088_epsilon_0, x = var_10087_cast_fp16)[name = string("op_10088_cast_fp16")];
            fp16 var_10089_to_fp16 = const()[name = string("op_10089_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_10090_cast_fp16 = mul(x = var_10088_cast_fp16, y = var_10089_to_fp16)[name = string("op_10090_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_10092_cast_fp16 = add(x = var_10086_cast_fp16, y = var_10090_cast_fp16)[name = string("op_10092_cast_fp16")];
            fp32 var_10093_epsilon_0 = const()[name = string("op_10093_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_10093_cast_fp16 = rsqrt(epsilon = var_10093_epsilon_0, x = var_10092_cast_fp16)[name = string("op_10093_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_337_cast_fp16 = mul(x = y_385_cast_fp16, y = var_10093_cast_fp16)[name = string("out_337_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136953728)))];
            tensor<fp16, [1, 128, 512]> var_10095_cast_fp16 = mul(x = out_337_cast_fp16, y = tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_10095_cast_fp16")];
            tensor<fp16, [1, 128, 512]> h_47_cast_fp16 = add(x = input_285_cast_fp16, y = var_10095_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<fp16, [1]> tm_layers_23_layer_scalar_to_fp16 = const()[name = string("tm_layers_23_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d6p-4])];
            tensor<fp16, [1, 128, 512]> x_cast_fp16 = mul(x = h_47_cast_fp16, y = tm_layers_23_layer_scalar_to_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_10099_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_10099_cast_fp16")];
            bool var_10102 = const()[name = string("op_10102"), val = bool(true)];
            tensor<int32, [1]> var_10103_axes_0 = const()[name = string("op_10103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_10103_cast_fp16 = reduce_max(axes = var_10103_axes_0, keep_dims = var_10102, x = var_10099_cast_fp16)[name = string("op_10103_cast_fp16")];
            fp16 var_10104_to_fp16 = const()[name = string("op_10104_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 128, 1]> s_387_cast_fp16 = maximum(x = var_10103_cast_fp16, y = var_10104_to_fp16)[name = string("s_387_cast_fp16")];
            tensor<fp16, [1, 128, 512]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_387_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 128, 512]> var_10107_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_10107_cast_fp16")];
            tensor<int32, [1]> var_10112_axes_0 = const()[name = string("op_10112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10112_keep_dims_0 = const()[name = string("op_10112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_10112_cast_fp16 = reduce_mean(axes = var_10112_axes_0, keep_dims = var_10112_keep_dims_0, x = var_10107_cast_fp16)[name = string("op_10112_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_10113_cast_fp16 = mul(x = s_387_cast_fp16, y = s_387_cast_fp16)[name = string("op_10113_cast_fp16")];
            fp32 var_10114_epsilon_0 = const()[name = string("op_10114_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_10114_cast_fp16 = inverse(epsilon = var_10114_epsilon_0, x = var_10113_cast_fp16)[name = string("op_10114_cast_fp16")];
            fp16 var_10115_to_fp16 = const()[name = string("op_10115_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_10116_cast_fp16 = mul(x = var_10114_cast_fp16, y = var_10115_to_fp16)[name = string("op_10116_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_10118_cast_fp16 = add(x = var_10112_cast_fp16, y = var_10116_cast_fp16)[name = string("op_10118_cast_fp16")];
            fp32 var_10119_epsilon_0 = const()[name = string("op_10119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_10119_cast_fp16 = rsqrt(epsilon = var_10119_epsilon_0, x = var_10118_cast_fp16)[name = string("op_10119_cast_fp16")];
            tensor<fp16, [1, 128, 512]> out_cast_fp16 = mul(x = y_cast_fp16, y = var_10119_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [512]> tm_norm_weight_to_fp16 = const()[name = string("tm_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136954816)))];
            tensor<fp16, [1, 128, 512]> h_cast_fp16 = mul(x = out_cast_fp16, y = tm_norm_weight_to_fp16)[name = string("h_cast_fp16")];
            tensor<int32, [1]> var_10126_axes_0 = const()[name = string("op_10126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10126_keep_dims_0 = const()[name = string("op_10126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10126_cast_fp16 = reduce_sum(axes = var_10126_axes_0, keep_dims = var_10126_keep_dims_0, x = attention_mask)[name = string("op_10126_cast_fp16")];
            tensor<fp16, [1, 128]> var_10127_cast_fp16 = real_div(x = attention_mask, y = var_10126_cast_fp16)[name = string("op_10127_cast_fp16")];
            tensor<int32, [1]> var_10134_axes_0 = const()[name = string("op_10134_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 128]> var_10134_cast_fp16 = expand_dims(axes = var_10134_axes_0, x = var_10127_cast_fp16)[name = string("op_10134_cast_fp16")];
            bool input_transpose_x_0 = const()[name = string("input_transpose_x_0"), val = bool(false)];
            bool input_transpose_y_0 = const()[name = string("input_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512]> input_cast_fp16 = matmul(transpose_x = input_transpose_x_0, transpose_y = input_transpose_y_0, x = var_10134_cast_fp16, y = h_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [768, 512]> tm_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136955904))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137349184))))[name = string("tm_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_217_bias_0_to_fp16 = const()[name = string("linear_217_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137350784)))];
            tensor<fp16, [1, 1, 768]> linear_217_cast_fp16 = linear(bias = linear_217_bias_0_to_fp16, weight = tm_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<int32, [3]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_10151_squeeze_mask_0 = const()[name = string("op_10151_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, squeeze_mask = var_10151_squeeze_mask_0, x = linear_217_cast_fp16)[name = string("op_10151_cast_fp16")];
            tensor<fp16, [1, 768]> var_10157_cast_fp16 = abs(x = var_10151_cast_fp16)[name = string("op_10157_cast_fp16")];
            bool var_10160 = const()[name = string("op_10160"), val = bool(true)];
            tensor<int32, [1]> s_axes_0 = const()[name = string("s_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> s_cast_fp16 = reduce_max(axes = s_axes_0, keep_dims = var_10160, x = var_10157_cast_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [1, 768]> pooled_cast_fp16 = real_div(x = var_10151_cast_fp16, y = s_cast_fp16)[name = string("pooled_cast_fp16")];
            tensor<fp16, [1, 768]> var_10163_cast_fp16 = mul(x = pooled_cast_fp16, y = pooled_cast_fp16)[name = string("op_10163_cast_fp16")];
            tensor<int32, [1]> var_10168_axes_0 = const()[name = string("op_10168_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10168_keep_dims_0 = const()[name = string("op_10168_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10168_cast_fp16 = reduce_sum(axes = var_10168_axes_0, keep_dims = var_10168_keep_dims_0, x = var_10163_cast_fp16)[name = string("op_10168_cast_fp16")];
            fp32 var_10169_epsilon_0 = const()[name = string("op_10169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_10169_cast_fp16 = rsqrt(epsilon = var_10169_epsilon_0, x = var_10168_cast_fp16)[name = string("op_10169_cast_fp16")];
            tensor<fp16, [1, 768]> var_10170_cast_fp16 = mul(x = pooled_cast_fp16, y = var_10169_cast_fp16)[name = string("op_10170_cast_fp16")];
            string var_10170_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_10170_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 768]> embedding = cast(dtype = var_10170_cast_fp16_to_fp32_dtype_0, x = var_10170_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func embed_256<ios18>(tensor<fp16, [1, 256]> attention_mask, tensor<fp16, [1, 256, 512]> inputs_embeds) {
            tensor<fp16, [12288, 512]> tm_ple_per_layer_model_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [12288, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1728))), scale = tensor<fp16, [12288, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6293248))))[name = string("tm_ple_per_layer_model_projection_weight_to_fp16_quantized")];
            tensor<fp16, [12288]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [12288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6317888)))];
            tensor<fp16, [1, 256, 12288]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = tm_ple_per_layer_model_projection_weight_to_fp16_quantized, x = inputs_embeds)[name = string("linear_0_cast_fp16")];
            fp16 var_2129_to_fp16 = const()[name = string("op_2129_to_fp16"), val = fp16(0x1.6ap-5)];
            tensor<fp16, [1, 256, 12288]> ple_1_cast_fp16 = mul(x = linear_0_cast_fp16, y = var_2129_to_fp16)[name = string("ple_1_cast_fp16")];
            tensor<int32, [4]> var_2135 = const()[name = string("op_2135"), val = tensor<int32, [4]>([1, -1, 24, 512])];
            tensor<fp16, [1, 256, 24, 512]> x_1_cast_fp16 = reshape(shape = var_2135, x = ple_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 256, 24, 512]> var_2137_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_2137_cast_fp16")];
            bool var_2140 = const()[name = string("op_2140"), val = bool(true)];
            tensor<int32, [1]> var_2141_axes_0 = const()[name = string("op_2141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 24, 1]> var_2141_cast_fp16 = reduce_max(axes = var_2141_axes_0, keep_dims = var_2140, x = var_2137_cast_fp16)[name = string("op_2141_cast_fp16")];
            fp16 var_2142_to_fp16 = const()[name = string("op_2142_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 24, 1]> s_1_cast_fp16 = maximum(x = var_2141_cast_fp16, y = var_2142_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 256, 24, 512]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 256, 24, 512]> var_2145_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [1]> var_2150_axes_0 = const()[name = string("op_2150_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2150_keep_dims_0 = const()[name = string("op_2150_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 24, 1]> var_2150_cast_fp16 = reduce_mean(axes = var_2150_axes_0, keep_dims = var_2150_keep_dims_0, x = var_2145_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 256, 24, 1]> var_2151_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 24, 1]> var_2152_cast_fp16 = inverse(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            fp16 var_2153_to_fp16 = const()[name = string("op_2153_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 24, 1]> var_2154_cast_fp16 = mul(x = var_2152_cast_fp16, y = var_2153_to_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 256, 24, 1]> var_2156_cast_fp16 = add(x = var_2150_cast_fp16, y = var_2154_cast_fp16)[name = string("op_2156_cast_fp16")];
            fp32 var_2157_epsilon_0 = const()[name = string("op_2157_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 24, 1]> var_2157_cast_fp16 = rsqrt(epsilon = var_2157_epsilon_0, x = var_2156_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<fp16, [1, 256, 24, 512]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_2157_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [512]> tm_ple_per_layer_projection_norm_weight_to_fp16 = const()[name = string("tm_ple_per_layer_projection_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6342528)))];
            tensor<fp16, [1, 256, 24, 512]> ple_3_cast_fp16 = mul(x = out_1_cast_fp16, y = tm_ple_per_layer_projection_norm_weight_to_fp16)[name = string("ple_3_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 256]> var_2162_cast_fp16 = sub(x = var_2160_to_fp16, y = attention_mask)[name = string("op_2162_cast_fp16")];
            fp16 var_2163_to_fp16 = const()[name = string("op_2163_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 256]> var_2164_cast_fp16 = mul(x = var_2162_cast_fp16, y = var_2163_to_fp16)[name = string("op_2164_cast_fp16")];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 256]> var_2171_cast_fp16 = expand_dims(axes = var_2171_axes_0, x = var_2164_cast_fp16)[name = string("op_2171_cast_fp16")];
            tensor<int32, [1]> var_2173_axes_0 = const()[name = string("op_2173_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 256]> var_2173_cast_fp16 = expand_dims(axes = var_2173_axes_0, x = var_2171_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2179_cast_fp16 = abs(x = inputs_embeds)[name = string("op_2179_cast_fp16")];
            bool var_2182 = const()[name = string("op_2182"), val = bool(true)];
            tensor<int32, [1]> var_2183_axes_0 = const()[name = string("op_2183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2183_cast_fp16 = reduce_max(axes = var_2183_axes_0, keep_dims = var_2182, x = var_2179_cast_fp16)[name = string("op_2183_cast_fp16")];
            fp16 var_2184_to_fp16 = const()[name = string("op_2184_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_3_cast_fp16 = maximum(x = var_2183_cast_fp16, y = var_2184_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_3_cast_fp16 = real_div(x = inputs_embeds, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2187_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> var_2192_axes_0 = const()[name = string("op_2192_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2192_keep_dims_0 = const()[name = string("op_2192_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2192_cast_fp16 = reduce_mean(axes = var_2192_axes_0, keep_dims = var_2192_keep_dims_0, x = var_2187_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2193_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp32 var_2194_epsilon_0 = const()[name = string("op_2194_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2194_cast_fp16 = inverse(epsilon = var_2194_epsilon_0, x = var_2193_cast_fp16)[name = string("op_2194_cast_fp16")];
            fp16 var_2195_to_fp16 = const()[name = string("op_2195_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2196_cast_fp16 = mul(x = var_2194_cast_fp16, y = var_2195_to_fp16)[name = string("op_2196_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2198_cast_fp16 = add(x = var_2192_cast_fp16, y = var_2196_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_2199_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6343616)))];
            tensor<fp16, [1, 256, 512]> input_1_cast_fp16 = mul(x = out_3_cast_fp16, y = tm_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6344704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6869056))))[name = string("tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6871168)))];
            tensor<fp16, [1, 256, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_1_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6873280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7135488))))[name = string("tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7136576)))];
            tensor<fp16, [1, 256, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_2221 = const()[name = string("op_2221"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2222_cast_fp16 = reshape(shape = var_2221, x = linear_2_cast_fp16)[name = string("op_2222_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7137664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7399872))))[name = string("tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2234_cast_fp16 = reshape(shape = var_2233, x = linear_3_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<int32, [4]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_2210_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 4, 256, 256]> var_2238_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_2238_cast_fp16")];
            bool var_2241 = const()[name = string("op_2241"), val = bool(true)];
            tensor<int32, [1]> var_2242_axes_0 = const()[name = string("op_2242_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_2242_cast_fp16 = reduce_max(axes = var_2242_axes_0, keep_dims = var_2241, x = var_2238_cast_fp16)[name = string("op_2242_cast_fp16")];
            fp16 var_2243_to_fp16 = const()[name = string("op_2243_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_5_cast_fp16 = maximum(x = var_2242_cast_fp16, y = var_2243_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2246_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<int32, [1]> var_2251_axes_0 = const()[name = string("op_2251_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2251_keep_dims_0 = const()[name = string("op_2251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_2251_cast_fp16 = reduce_mean(axes = var_2251_axes_0, keep_dims = var_2251_keep_dims_0, x = var_2246_cast_fp16)[name = string("op_2251_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2252_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_2252_cast_fp16")];
            fp32 var_2253_epsilon_0 = const()[name = string("op_2253_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_2253_cast_fp16 = inverse(epsilon = var_2253_epsilon_0, x = var_2252_cast_fp16)[name = string("op_2253_cast_fp16")];
            fp16 var_2254_to_fp16 = const()[name = string("op_2254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_2255_cast_fp16 = mul(x = var_2253_cast_fp16, y = var_2254_to_fp16)[name = string("op_2255_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2257_cast_fp16 = add(x = var_2251_cast_fp16, y = var_2255_cast_fp16)[name = string("op_2257_cast_fp16")];
            fp32 var_2258_epsilon_0 = const()[name = string("op_2258_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_2258_cast_fp16 = rsqrt(epsilon = var_2258_epsilon_0, x = var_2257_cast_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_2258_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7400960)))];
            tensor<fp16, [1, 4, 256, 256]> q_1_cast_fp16 = mul(x = out_5_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_2222_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 2, 256, 256]> var_2261_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_2261_cast_fp16")];
            bool var_2264 = const()[name = string("op_2264"), val = bool(true)];
            tensor<int32, [1]> var_2265_axes_0 = const()[name = string("op_2265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2265_cast_fp16 = reduce_max(axes = var_2265_axes_0, keep_dims = var_2264, x = var_2261_cast_fp16)[name = string("op_2265_cast_fp16")];
            fp16 var_2266_to_fp16 = const()[name = string("op_2266_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_7_cast_fp16 = maximum(x = var_2265_cast_fp16, y = var_2266_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2269_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> var_2274_axes_0 = const()[name = string("op_2274_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2274_keep_dims_0 = const()[name = string("op_2274_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2274_cast_fp16 = reduce_mean(axes = var_2274_axes_0, keep_dims = var_2274_keep_dims_0, x = var_2269_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2275_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2276_cast_fp16 = inverse(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            fp16 var_2277_to_fp16 = const()[name = string("op_2277_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2278_cast_fp16 = mul(x = var_2276_cast_fp16, y = var_2277_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2280_cast_fp16 = add(x = var_2274_cast_fp16, y = var_2278_cast_fp16)[name = string("op_2280_cast_fp16")];
            fp32 var_2281_epsilon_0 = const()[name = string("op_2281_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2281_cast_fp16 = rsqrt(epsilon = var_2281_epsilon_0, x = var_2280_cast_fp16)[name = string("op_2281_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_7_cast_fp16 = mul(x = y_7_cast_fp16, y = var_2281_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7401536)))];
            tensor<fp16, [1, 2, 256, 256]> k_1_cast_fp16 = mul(x = out_7_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = var_2234_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2, 256, 256]> var_2284_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_2284_cast_fp16")];
            bool var_2287 = const()[name = string("op_2287"), val = bool(true)];
            tensor<int32, [1]> var_2288_axes_0 = const()[name = string("op_2288_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2288_cast_fp16 = reduce_max(axes = var_2288_axes_0, keep_dims = var_2287, x = var_2284_cast_fp16)[name = string("op_2288_cast_fp16")];
            fp16 var_2289_to_fp16 = const()[name = string("op_2289_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_9_cast_fp16 = maximum(x = var_2288_cast_fp16, y = var_2289_to_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_9_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2292_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_2292_cast_fp16")];
            tensor<int32, [1]> var_2297_axes_0 = const()[name = string("op_2297_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2297_keep_dims_0 = const()[name = string("op_2297_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2297_cast_fp16 = reduce_mean(axes = var_2297_axes_0, keep_dims = var_2297_keep_dims_0, x = var_2292_cast_fp16)[name = string("op_2297_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2298_cast_fp16 = mul(x = s_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_2298_cast_fp16")];
            fp32 var_2299_epsilon_0 = const()[name = string("op_2299_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2299_cast_fp16 = inverse(epsilon = var_2299_epsilon_0, x = var_2298_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 var_2300_to_fp16 = const()[name = string("op_2300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2301_cast_fp16 = mul(x = var_2299_cast_fp16, y = var_2300_to_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2303_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2301_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp32 var_2304_epsilon_0 = const()[name = string("op_2304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2304_cast_fp16 = rsqrt(epsilon = var_2304_epsilon_0, x = var_2303_cast_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_1_cast_fp16 = mul(x = y_9_cast_fp16, y = var_2304_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<fp16, [256, 256]> cos_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137575872))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137641472))))[name = string("cos_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 256, 256]> var_2306_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2306_cast_fp16")];
            tensor<int32, [4]> var_2311_begin_0 = const()[name = string("op_2311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2311_end_0 = const()[name = string("op_2311_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_2311_end_mask_0 = const()[name = string("op_2311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_2311_cast_fp16 = slice_by_index(begin = var_2311_begin_0, end = var_2311_end_0, end_mask = var_2311_end_mask_0, x = q_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = q_1_cast_fp16)[name = string("op_2317_cast_fp16")];
            int32 var_2319 = const()[name = string("op_2319"), val = int32(-1)];
            bool var_2320_interleave_0 = const()[name = string("op_2320_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_2320_cast_fp16 = concat(axis = var_2319, interleave = var_2320_interleave_0, values = (var_2312_cast_fp16, var_2317_cast_fp16))[name = string("op_2320_cast_fp16")];
            tensor<fp16, [256, 256]> sin_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137642048))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137707648))))[name = string("sin_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 256, 256]> var_2321_cast_fp16 = mul(x = var_2320_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_3_cast_fp16 = add(x = var_2306_cast_fp16, y = var_2321_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2324_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2329_begin_0 = const()[name = string("op_2329_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2329_end_0 = const()[name = string("op_2329_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_2329_end_mask_0 = const()[name = string("op_2329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, x = k_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_2330_cast_fp16 = mul(x = var_2329_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_2330_cast_fp16")];
            tensor<int32, [4]> var_2335_begin_0 = const()[name = string("op_2335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2335_end_0 = const()[name = string("op_2335_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_2335_end_mask_0 = const()[name = string("op_2335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_2335_cast_fp16 = slice_by_index(begin = var_2335_begin_0, end = var_2335_end_0, end_mask = var_2335_end_mask_0, x = k_1_cast_fp16)[name = string("op_2335_cast_fp16")];
            int32 var_2337 = const()[name = string("op_2337"), val = int32(-1)];
            bool var_2338_interleave_0 = const()[name = string("op_2338_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_2338_cast_fp16 = concat(axis = var_2337, interleave = var_2338_interleave_0, values = (var_2330_cast_fp16, var_2335_cast_fp16))[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2339_cast_fp16 = mul(x = var_2338_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2339_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_3_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2339_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = k_3_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [4, 1, 256, 256]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
            tensor<int32, [5]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_0_cast_fp16 = reshape(shape = concat_0, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [4, 1, 256, 256]> reshape_1_cast_fp16 = reshape(shape = concat_1, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = v_1_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [4, 1, 256, 256]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
            tensor<int32, [5]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_2_cast_fp16 = reshape(shape = concat_2, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [4, 1, 256, 256]> reshape_3_cast_fp16 = reshape(shape = concat_3, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2353_transpose_x_0 = const()[name = string("op_2353_transpose_x_0"), val = bool(false)];
            bool var_2353_transpose_y_0 = const()[name = string("op_2353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 4, 256, 256]> var_2353_cast_fp16 = matmul(transpose_x = var_2353_transpose_x_0, transpose_y = var_2353_transpose_y_0, x = q_3_cast_fp16, y = transpose_96_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2355_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2356 = const()[name = string("op_2356"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_9_cast_fp16 = softmax(axis = var_2356, x = var_2355_cast_fp16)[name = string("w_9_cast_fp16")];
            bool var_2359_transpose_x_0 = const()[name = string("op_2359_transpose_x_0"), val = bool(false)];
            bool var_2359_transpose_y_0 = const()[name = string("op_2359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 4, 256, 256]> var_2359_cast_fp16 = matmul(transpose_x = var_2359_transpose_x_0, transpose_y = var_2359_transpose_y_0, x = w_9_cast_fp16, y = v_3_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<int32, [4]> var_2362_perm_0 = const()[name = string("op_2362_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2366 = const()[name = string("op_2366"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_2362_cast_fp16 = transpose(perm = var_2362_perm_0, x = var_2359_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 256, 1024]> input_3_cast_fp16 = reshape(shape = var_2366, x = var_2362_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7402112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7926464))))[name = string("tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2371_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_2371_cast_fp16")];
            bool var_2374 = const()[name = string("op_2374"), val = bool(true)];
            tensor<int32, [1]> var_2375_axes_0 = const()[name = string("op_2375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2375_cast_fp16 = reduce_max(axes = var_2375_axes_0, keep_dims = var_2374, x = var_2371_cast_fp16)[name = string("op_2375_cast_fp16")];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_11_cast_fp16 = maximum(x = var_2375_cast_fp16, y = var_2376_to_fp16)[name = string("s_11_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_11_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_11_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2379_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_2379_cast_fp16")];
            tensor<int32, [1]> var_2384_axes_0 = const()[name = string("op_2384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2384_keep_dims_0 = const()[name = string("op_2384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2384_cast_fp16 = reduce_mean(axes = var_2384_axes_0, keep_dims = var_2384_keep_dims_0, x = var_2379_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2385_cast_fp16 = mul(x = s_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2385_cast_fp16")];
            fp32 var_2386_epsilon_0 = const()[name = string("op_2386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2386_cast_fp16 = inverse(epsilon = var_2386_epsilon_0, x = var_2385_cast_fp16)[name = string("op_2386_cast_fp16")];
            fp16 var_2387_to_fp16 = const()[name = string("op_2387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2388_cast_fp16 = mul(x = var_2386_cast_fp16, y = var_2387_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2390_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2388_cast_fp16)[name = string("op_2390_cast_fp16")];
            fp32 var_2391_epsilon_0 = const()[name = string("op_2391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2391_cast_fp16 = rsqrt(epsilon = var_2391_epsilon_0, x = var_2390_cast_fp16)[name = string("op_2391_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_2391_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7927552)))];
            tensor<fp16, [1, 256, 512]> var_2393_cast_fp16 = mul(x = out_9_cast_fp16, y = tm_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_11_cast_fp16 = add(x = inputs_embeds, y = var_2393_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2396_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_2396_cast_fp16")];
            bool var_2399 = const()[name = string("op_2399"), val = bool(true)];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_2399, x = var_2396_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 var_2401_to_fp16 = const()[name = string("op_2401_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_13_cast_fp16 = maximum(x = var_2400_cast_fp16, y = var_2401_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = s_13_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2404_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [1]> var_2409_axes_0 = const()[name = string("op_2409_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2409_keep_dims_0 = const()[name = string("op_2409_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2409_cast_fp16 = reduce_mean(axes = var_2409_axes_0, keep_dims = var_2409_keep_dims_0, x = var_2404_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2410_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2411_cast_fp16 = inverse(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2413_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("op_2413_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2415_cast_fp16 = add(x = var_2409_cast_fp16, y = var_2413_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp32 var_2416_epsilon_0 = const()[name = string("op_2416_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2416_cast_fp16 = rsqrt(epsilon = var_2416_epsilon_0, x = var_2415_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_2416_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7928640)))];
            tensor<fp16, [1, 256, 512]> input_5_cast_fp16 = mul(x = out_11_cast_fp16, y = tm_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7929728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8978368))))[name = string("tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8982528)))];
            tensor<fp16, [1, 256, 2048]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_2423_mode_0 = const()[name = string("op_2423_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_2423_cast_fp16 = gelu(mode = var_2423_mode_0, x = linear_5_cast_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8986688))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10035328))))[name = string("tm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_7_cast_fp16 = mul(x = var_2423_cast_fp16, y = linear_6_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10039488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11088128))))[name = string("tm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2431_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_2431_cast_fp16")];
            bool var_2434 = const()[name = string("op_2434"), val = bool(true)];
            tensor<int32, [1]> var_2435_axes_0 = const()[name = string("op_2435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2435_cast_fp16 = reduce_max(axes = var_2435_axes_0, keep_dims = var_2434, x = var_2431_cast_fp16)[name = string("op_2435_cast_fp16")];
            fp16 var_2436_to_fp16 = const()[name = string("op_2436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_15_cast_fp16 = maximum(x = var_2435_cast_fp16, y = var_2436_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_15_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_15_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2439_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_2439_cast_fp16")];
            tensor<int32, [1]> var_2444_axes_0 = const()[name = string("op_2444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2444_keep_dims_0 = const()[name = string("op_2444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2444_cast_fp16 = reduce_mean(axes = var_2444_axes_0, keep_dims = var_2444_keep_dims_0, x = var_2439_cast_fp16)[name = string("op_2444_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2445_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_2445_cast_fp16")];
            fp32 var_2446_epsilon_0 = const()[name = string("op_2446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2446_cast_fp16 = inverse(epsilon = var_2446_epsilon_0, x = var_2445_cast_fp16)[name = string("op_2446_cast_fp16")];
            fp16 var_2447_to_fp16 = const()[name = string("op_2447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2448_cast_fp16 = mul(x = var_2446_cast_fp16, y = var_2447_to_fp16)[name = string("op_2448_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2450_cast_fp16 = add(x = var_2444_cast_fp16, y = var_2448_cast_fp16)[name = string("op_2450_cast_fp16")];
            fp32 var_2451_epsilon_0 = const()[name = string("op_2451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2451_cast_fp16 = rsqrt(epsilon = var_2451_epsilon_0, x = var_2450_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_2451_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11089216)))];
            tensor<fp16, [1, 256, 512]> var_2453_cast_fp16 = mul(x = out_13_cast_fp16, y = tm_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_9_cast_fp16 = add(x = x_11_cast_fp16, y = var_2453_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11090304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11352512))))[name = string("tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_8_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_8_cast_fp16")];
            string x_15_mode_0 = const()[name = string("x_15_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_15_cast_fp16 = gelu(mode = x_15_mode_0, x = linear_8_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([1, 256, 1, 512])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2473_squeeze_mask_0 = const()[name = string("op_2473_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, squeeze_mask = var_2473_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2473_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_11_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2473_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11353600))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11615808))))[name = string("tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2483_cast_fp16 = abs(x = linear_9_cast_fp16)[name = string("op_2483_cast_fp16")];
            bool var_2486 = const()[name = string("op_2486"), val = bool(true)];
            tensor<int32, [1]> var_2487_axes_0 = const()[name = string("op_2487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2487_cast_fp16 = reduce_max(axes = var_2487_axes_0, keep_dims = var_2486, x = var_2483_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp16 var_2488_to_fp16 = const()[name = string("op_2488_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_17_cast_fp16 = maximum(x = var_2487_cast_fp16, y = var_2488_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_17_cast_fp16 = real_div(x = linear_9_cast_fp16, y = s_17_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2491_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_2491_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2496_keep_dims_0 = const()[name = string("op_2496_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2496_cast_fp16 = reduce_mean(axes = var_2496_axes_0, keep_dims = var_2496_keep_dims_0, x = var_2491_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2497_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_2497_cast_fp16")];
            fp32 var_2498_epsilon_0 = const()[name = string("op_2498_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2498_cast_fp16 = inverse(epsilon = var_2498_epsilon_0, x = var_2497_cast_fp16)[name = string("op_2498_cast_fp16")];
            fp16 var_2499_to_fp16 = const()[name = string("op_2499_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2500_cast_fp16 = mul(x = var_2498_cast_fp16, y = var_2499_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2502_cast_fp16 = add(x = var_2496_cast_fp16, y = var_2500_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp32 var_2503_epsilon_0 = const()[name = string("op_2503_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2503_cast_fp16 = rsqrt(epsilon = var_2503_epsilon_0, x = var_2502_cast_fp16)[name = string("op_2503_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_2503_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11616896)))];
            tensor<fp16, [1, 256, 512]> var_2505_cast_fp16 = mul(x = out_15_cast_fp16, y = tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_1_cast_fp16 = add(x = input_9_cast_fp16, y = var_2505_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<fp16, [1]> tm_layers_0_layer_scalar_to_fp16 = const()[name = string("tm_layers_0_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9p-2])];
            tensor<fp16, [1, 256, 512]> x_19_cast_fp16 = mul(x = h_1_cast_fp16, y = tm_layers_0_layer_scalar_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2509_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool var_2512 = const()[name = string("op_2512"), val = bool(true)];
            tensor<int32, [1]> var_2513_axes_0 = const()[name = string("op_2513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2513_cast_fp16 = reduce_max(axes = var_2513_axes_0, keep_dims = var_2512, x = var_2509_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_19_cast_fp16 = maximum(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_19_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2517_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [1]> var_2522_axes_0 = const()[name = string("op_2522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2522_keep_dims_0 = const()[name = string("op_2522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2522_cast_fp16 = reduce_mean(axes = var_2522_axes_0, keep_dims = var_2522_keep_dims_0, x = var_2517_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2523_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_2523_cast_fp16")];
            fp32 var_2524_epsilon_0 = const()[name = string("op_2524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2524_cast_fp16 = inverse(epsilon = var_2524_epsilon_0, x = var_2523_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp16 var_2525_to_fp16 = const()[name = string("op_2525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2526_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_2525_to_fp16)[name = string("op_2526_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2528_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2526_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_2529_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11617984)))];
            tensor<fp16, [1, 256, 512]> input_13_cast_fp16 = mul(x = out_17_cast_fp16, y = tm_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11619072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12143424))))[name = string("tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_2540_cast_fp16 = reshape(shape = var_2539, x = linear_10_cast_fp16)[name = string("op_2540_cast_fp16")];
            tensor<int32, [4]> x_21_perm_0 = const()[name = string("x_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12145536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12407744))))[name = string("tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<int32, [4]> var_2551 = const()[name = string("op_2551"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2552_cast_fp16 = reshape(shape = var_2551, x = linear_11_cast_fp16)[name = string("op_2552_cast_fp16")];
            tensor<int32, [4]> x_23_perm_0 = const()[name = string("x_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12408832))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671040))))[name = string("tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [4]> var_2563 = const()[name = string("op_2563"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2564_cast_fp16 = reshape(shape = var_2563, x = linear_12_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = var_2540_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 4, 256, 256]> var_2568_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_2568_cast_fp16")];
            bool var_2571 = const()[name = string("op_2571"), val = bool(true)];
            tensor<int32, [1]> var_2572_axes_0 = const()[name = string("op_2572_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_2572_cast_fp16 = reduce_max(axes = var_2572_axes_0, keep_dims = var_2571, x = var_2568_cast_fp16)[name = string("op_2572_cast_fp16")];
            fp16 var_2573_to_fp16 = const()[name = string("op_2573_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_21_cast_fp16 = maximum(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_21_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2576_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> var_2581_axes_0 = const()[name = string("op_2581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2581_keep_dims_0 = const()[name = string("op_2581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_2581_cast_fp16 = reduce_mean(axes = var_2581_axes_0, keep_dims = var_2581_keep_dims_0, x = var_2576_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2582_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp32 var_2583_epsilon_0 = const()[name = string("op_2583_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_2583_cast_fp16 = inverse(epsilon = var_2583_epsilon_0, x = var_2582_cast_fp16)[name = string("op_2583_cast_fp16")];
            fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_2585_cast_fp16 = mul(x = var_2583_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2587_cast_fp16 = add(x = var_2581_cast_fp16, y = var_2585_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_19_cast_fp16 = mul(x = y_21_cast_fp16, y = var_2588_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672128)))];
            tensor<fp16, [1, 4, 256, 256]> q_5_cast_fp16 = mul(x = out_19_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_23_cast_fp16 = transpose(perm = x_23_perm_0, x = var_2552_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2, 256, 256]> var_2591_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2594 = const()[name = string("op_2594"), val = bool(true)];
            tensor<int32, [1]> var_2595_axes_0 = const()[name = string("op_2595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2595_cast_fp16 = reduce_max(axes = var_2595_axes_0, keep_dims = var_2594, x = var_2591_cast_fp16)[name = string("op_2595_cast_fp16")];
            fp16 var_2596_to_fp16 = const()[name = string("op_2596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_23_cast_fp16 = maximum(x = var_2595_cast_fp16, y = var_2596_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_23_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2599_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2599_cast_fp16")];
            tensor<int32, [1]> var_2604_axes_0 = const()[name = string("op_2604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2604_keep_dims_0 = const()[name = string("op_2604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2604_cast_fp16 = reduce_mean(axes = var_2604_axes_0, keep_dims = var_2604_keep_dims_0, x = var_2599_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2605_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_2605_cast_fp16")];
            fp32 var_2606_epsilon_0 = const()[name = string("op_2606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2606_cast_fp16 = inverse(epsilon = var_2606_epsilon_0, x = var_2605_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp16 var_2607_to_fp16 = const()[name = string("op_2607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2608_cast_fp16 = mul(x = var_2606_cast_fp16, y = var_2607_to_fp16)[name = string("op_2608_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2610_cast_fp16 = add(x = var_2604_cast_fp16, y = var_2608_cast_fp16)[name = string("op_2610_cast_fp16")];
            fp32 var_2611_epsilon_0 = const()[name = string("op_2611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2611_cast_fp16 = rsqrt(epsilon = var_2611_epsilon_0, x = var_2610_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_21_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2611_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672704)))];
            tensor<fp16, [1, 2, 256, 256]> k_7_cast_fp16 = mul(x = out_21_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_2564_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 2, 256, 256]> var_2614_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_2614_cast_fp16")];
            bool var_2617 = const()[name = string("op_2617"), val = bool(true)];
            tensor<int32, [1]> var_2618_axes_0 = const()[name = string("op_2618_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2618_cast_fp16 = reduce_max(axes = var_2618_axes_0, keep_dims = var_2617, x = var_2614_cast_fp16)[name = string("op_2618_cast_fp16")];
            fp16 var_2619_to_fp16 = const()[name = string("op_2619_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_25_cast_fp16 = maximum(x = var_2618_cast_fp16, y = var_2619_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_25_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2622_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2622_cast_fp16")];
            tensor<int32, [1]> var_2627_axes_0 = const()[name = string("op_2627_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2627_keep_dims_0 = const()[name = string("op_2627_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2627_cast_fp16 = reduce_mean(axes = var_2627_axes_0, keep_dims = var_2627_keep_dims_0, x = var_2622_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2628_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_2628_cast_fp16")];
            fp32 var_2629_epsilon_0 = const()[name = string("op_2629_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2629_cast_fp16 = inverse(epsilon = var_2629_epsilon_0, x = var_2628_cast_fp16)[name = string("op_2629_cast_fp16")];
            fp16 var_2630_to_fp16 = const()[name = string("op_2630_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2631_cast_fp16 = mul(x = var_2629_cast_fp16, y = var_2630_to_fp16)[name = string("op_2631_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2633_cast_fp16 = add(x = var_2627_cast_fp16, y = var_2631_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2634_cast_fp16 = rsqrt(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_5_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2634_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2636_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2636_cast_fp16")];
            tensor<int32, [4]> var_2641_begin_0 = const()[name = string("op_2641_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2641_end_0 = const()[name = string("op_2641_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_2641_end_mask_0 = const()[name = string("op_2641_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_2641_cast_fp16 = slice_by_index(begin = var_2641_begin_0, end = var_2641_end_0, end_mask = var_2641_end_mask_0, x = q_5_cast_fp16)[name = string("op_2641_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_2642_cast_fp16 = mul(x = var_2641_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_2642_cast_fp16")];
            tensor<int32, [4]> var_2647_begin_0 = const()[name = string("op_2647_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2647_end_0 = const()[name = string("op_2647_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_2647_end_mask_0 = const()[name = string("op_2647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_2647_cast_fp16 = slice_by_index(begin = var_2647_begin_0, end = var_2647_end_0, end_mask = var_2647_end_mask_0, x = q_5_cast_fp16)[name = string("op_2647_cast_fp16")];
            int32 var_2649 = const()[name = string("op_2649"), val = int32(-1)];
            bool var_2650_interleave_0 = const()[name = string("op_2650_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_2650_cast_fp16 = concat(axis = var_2649, interleave = var_2650_interleave_0, values = (var_2642_cast_fp16, var_2647_cast_fp16))[name = string("op_2650_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2651_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_7_cast_fp16 = add(x = var_2636_cast_fp16, y = var_2651_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2654_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = k_7_cast_fp16)[name = string("op_2659_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_2660_cast_fp16 = mul(x = var_2659_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2660_cast_fp16")];
            tensor<int32, [4]> var_2665_begin_0 = const()[name = string("op_2665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2665_end_0 = const()[name = string("op_2665_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_2665_end_mask_0 = const()[name = string("op_2665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_2665_cast_fp16 = slice_by_index(begin = var_2665_begin_0, end = var_2665_end_0, end_mask = var_2665_end_mask_0, x = k_7_cast_fp16)[name = string("op_2665_cast_fp16")];
            int32 var_2667 = const()[name = string("op_2667"), val = int32(-1)];
            bool var_2668_interleave_0 = const()[name = string("op_2668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_2668_cast_fp16 = concat(axis = var_2667, interleave = var_2668_interleave_0, values = (var_2660_cast_fp16, var_2665_cast_fp16))[name = string("op_2668_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2669_cast_fp16 = mul(x = var_2668_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_9_cast_fp16 = add(x = var_2654_cast_fp16, y = var_2669_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = k_9_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [4, 1, 256, 256]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_4_cast_fp16 = reshape(shape = concat_4, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [4, 1, 256, 256]> reshape_5_cast_fp16 = reshape(shape = concat_5, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = v_5_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [4, 1, 256, 256]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_6_cast_fp16 = reshape(shape = concat_6, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [4, 1, 256, 256]> reshape_7_cast_fp16 = reshape(shape = concat_7, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2683_transpose_x_0 = const()[name = string("op_2683_transpose_x_0"), val = bool(false)];
            bool var_2683_transpose_y_0 = const()[name = string("op_2683_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_97_cast_fp16 = transpose(perm = transpose_97_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 4, 256, 256]> var_2683_cast_fp16 = matmul(transpose_x = var_2683_transpose_x_0, transpose_y = var_2683_transpose_y_0, x = q_7_cast_fp16, y = transpose_97_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2685_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2685_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_25_cast_fp16 = softmax(axis = var_2686, x = var_2685_cast_fp16)[name = string("w_25_cast_fp16")];
            bool var_2689_transpose_x_0 = const()[name = string("op_2689_transpose_x_0"), val = bool(false)];
            bool var_2689_transpose_y_0 = const()[name = string("op_2689_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 4, 256, 256]> var_2689_cast_fp16 = matmul(transpose_x = var_2689_transpose_x_0, transpose_y = var_2689_transpose_y_0, x = w_25_cast_fp16, y = v_7_cast_fp16)[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> var_2692_perm_0 = const()[name = string("op_2692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2696 = const()[name = string("op_2696"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_2692_cast_fp16 = transpose(perm = var_2692_perm_0, x = var_2689_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 256, 1024]> input_15_cast_fp16 = reshape(shape = var_2696, x = var_2692_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13197632))))[name = string("tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2701_cast_fp16 = abs(x = linear_13_cast_fp16)[name = string("op_2701_cast_fp16")];
            bool var_2704 = const()[name = string("op_2704"), val = bool(true)];
            tensor<int32, [1]> var_2705_axes_0 = const()[name = string("op_2705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2705_cast_fp16 = reduce_max(axes = var_2705_axes_0, keep_dims = var_2704, x = var_2701_cast_fp16)[name = string("op_2705_cast_fp16")];
            fp16 var_2706_to_fp16 = const()[name = string("op_2706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_27_cast_fp16 = maximum(x = var_2705_cast_fp16, y = var_2706_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_27_cast_fp16 = real_div(x = linear_13_cast_fp16, y = s_27_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2709_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2709_cast_fp16")];
            tensor<int32, [1]> var_2714_axes_0 = const()[name = string("op_2714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2714_keep_dims_0 = const()[name = string("op_2714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2714_cast_fp16 = reduce_mean(axes = var_2714_axes_0, keep_dims = var_2714_keep_dims_0, x = var_2709_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2715_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2715_cast_fp16")];
            fp32 var_2716_epsilon_0 = const()[name = string("op_2716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2716_cast_fp16 = inverse(epsilon = var_2716_epsilon_0, x = var_2715_cast_fp16)[name = string("op_2716_cast_fp16")];
            fp16 var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2718_cast_fp16 = mul(x = var_2716_cast_fp16, y = var_2717_to_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2720_cast_fp16 = add(x = var_2714_cast_fp16, y = var_2718_cast_fp16)[name = string("op_2720_cast_fp16")];
            fp32 var_2721_epsilon_0 = const()[name = string("op_2721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2721_cast_fp16 = rsqrt(epsilon = var_2721_epsilon_0, x = var_2720_cast_fp16)[name = string("op_2721_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2721_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13198720)))];
            tensor<fp16, [1, 256, 512]> var_2723_cast_fp16 = mul(x = out_23_cast_fp16, y = tm_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2723_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_29_cast_fp16 = add(x = x_19_cast_fp16, y = var_2723_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2726_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_2726_cast_fp16")];
            bool var_2729 = const()[name = string("op_2729"), val = bool(true)];
            tensor<int32, [1]> var_2730_axes_0 = const()[name = string("op_2730_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2730_cast_fp16 = reduce_max(axes = var_2730_axes_0, keep_dims = var_2729, x = var_2726_cast_fp16)[name = string("op_2730_cast_fp16")];
            fp16 var_2731_to_fp16 = const()[name = string("op_2731_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_29_cast_fp16 = maximum(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_29_cast_fp16 = real_div(x = x_29_cast_fp16, y = s_29_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2734_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2734_cast_fp16")];
            tensor<int32, [1]> var_2739_axes_0 = const()[name = string("op_2739_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2739_keep_dims_0 = const()[name = string("op_2739_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2739_cast_fp16 = reduce_mean(axes = var_2739_axes_0, keep_dims = var_2739_keep_dims_0, x = var_2734_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2740_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2740_cast_fp16")];
            fp32 var_2741_epsilon_0 = const()[name = string("op_2741_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2741_cast_fp16 = inverse(epsilon = var_2741_epsilon_0, x = var_2740_cast_fp16)[name = string("op_2741_cast_fp16")];
            fp16 var_2742_to_fp16 = const()[name = string("op_2742_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2743_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_2742_to_fp16)[name = string("op_2743_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2745_cast_fp16 = add(x = var_2739_cast_fp16, y = var_2743_cast_fp16)[name = string("op_2745_cast_fp16")];
            fp32 var_2746_epsilon_0 = const()[name = string("op_2746_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2746_cast_fp16 = rsqrt(epsilon = var_2746_epsilon_0, x = var_2745_cast_fp16)[name = string("op_2746_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2746_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13199808)))];
            tensor<fp16, [1, 256, 512]> input_17_cast_fp16 = mul(x = out_25_cast_fp16, y = tm_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13200896))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14249536))))[name = string("tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_14_cast_fp16")];
            string var_2753_mode_0 = const()[name = string("op_2753_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_2753_cast_fp16 = gelu(mode = var_2753_mode_0, x = linear_14_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14253696))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15302336))))[name = string("tm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_15_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_19_cast_fp16 = mul(x = var_2753_cast_fp16, y = linear_15_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15306496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16355136))))[name = string("tm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2761_cast_fp16 = abs(x = linear_16_cast_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2764 = const()[name = string("op_2764"), val = bool(true)];
            tensor<int32, [1]> var_2765_axes_0 = const()[name = string("op_2765_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2765_cast_fp16 = reduce_max(axes = var_2765_axes_0, keep_dims = var_2764, x = var_2761_cast_fp16)[name = string("op_2765_cast_fp16")];
            fp16 var_2766_to_fp16 = const()[name = string("op_2766_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_31_cast_fp16 = maximum(x = var_2765_cast_fp16, y = var_2766_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_31_cast_fp16 = real_div(x = linear_16_cast_fp16, y = s_31_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2769_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [1]> var_2774_axes_0 = const()[name = string("op_2774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2774_keep_dims_0 = const()[name = string("op_2774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2774_cast_fp16 = reduce_mean(axes = var_2774_axes_0, keep_dims = var_2774_keep_dims_0, x = var_2769_cast_fp16)[name = string("op_2774_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2775_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp32 var_2776_epsilon_0 = const()[name = string("op_2776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2776_cast_fp16 = inverse(epsilon = var_2776_epsilon_0, x = var_2775_cast_fp16)[name = string("op_2776_cast_fp16")];
            fp16 var_2777_to_fp16 = const()[name = string("op_2777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2778_cast_fp16 = mul(x = var_2776_cast_fp16, y = var_2777_to_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2780_cast_fp16 = add(x = var_2774_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2780_cast_fp16")];
            fp32 var_2781_epsilon_0 = const()[name = string("op_2781_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2781_cast_fp16 = rsqrt(epsilon = var_2781_epsilon_0, x = var_2780_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2781_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16356224)))];
            tensor<fp16, [1, 256, 512]> var_2783_cast_fp16 = mul(x = out_27_cast_fp16, y = tm_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2783_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_21_cast_fp16 = add(x = x_29_cast_fp16, y = var_2783_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16357312))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16619520))))[name = string("tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_17_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            string x_33_mode_0 = const()[name = string("x_33_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_33_cast_fp16 = gelu(mode = x_33_mode_0, x = linear_17_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> var_2803_begin_0 = const()[name = string("op_2803_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2803_end_0 = const()[name = string("op_2803_end_0"), val = tensor<int32, [4]>([1, 256, 2, 512])];
            tensor<bool, [4]> var_2803_end_mask_0 = const()[name = string("op_2803_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2803_squeeze_mask_0 = const()[name = string("op_2803_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_2803_cast_fp16 = slice_by_index(begin = var_2803_begin_0, end = var_2803_end_0, end_mask = var_2803_end_mask_0, squeeze_mask = var_2803_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_23_cast_fp16 = mul(x = x_33_cast_fp16, y = var_2803_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16620608))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16882816))))[name = string("tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2813_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2813_cast_fp16")];
            bool var_2816 = const()[name = string("op_2816"), val = bool(true)];
            tensor<int32, [1]> var_2817_axes_0 = const()[name = string("op_2817_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2817_cast_fp16 = reduce_max(axes = var_2817_axes_0, keep_dims = var_2816, x = var_2813_cast_fp16)[name = string("op_2817_cast_fp16")];
            fp16 var_2818_to_fp16 = const()[name = string("op_2818_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_33_cast_fp16 = maximum(x = var_2817_cast_fp16, y = var_2818_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_33_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_33_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2821_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<int32, [1]> var_2826_axes_0 = const()[name = string("op_2826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2826_keep_dims_0 = const()[name = string("op_2826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2826_cast_fp16 = reduce_mean(axes = var_2826_axes_0, keep_dims = var_2826_keep_dims_0, x = var_2821_cast_fp16)[name = string("op_2826_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2827_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp32 var_2828_epsilon_0 = const()[name = string("op_2828_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2828_cast_fp16 = inverse(epsilon = var_2828_epsilon_0, x = var_2827_cast_fp16)[name = string("op_2828_cast_fp16")];
            fp16 var_2829_to_fp16 = const()[name = string("op_2829_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2830_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_2829_to_fp16)[name = string("op_2830_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2832_cast_fp16 = add(x = var_2826_cast_fp16, y = var_2830_cast_fp16)[name = string("op_2832_cast_fp16")];
            fp32 var_2833_epsilon_0 = const()[name = string("op_2833_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2833_cast_fp16 = rsqrt(epsilon = var_2833_epsilon_0, x = var_2832_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2833_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16883904)))];
            tensor<fp16, [1, 256, 512]> var_2835_cast_fp16 = mul(x = out_29_cast_fp16, y = tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_3_cast_fp16 = add(x = input_21_cast_fp16, y = var_2835_cast_fp16)[name = string("h_3_cast_fp16")];
            tensor<fp16, [1]> tm_layers_1_layer_scalar_to_fp16 = const()[name = string("tm_layers_1_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.5ap-3])];
            tensor<fp16, [1, 256, 512]> x_37_cast_fp16 = mul(x = h_3_cast_fp16, y = tm_layers_1_layer_scalar_to_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2839_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2839_cast_fp16")];
            bool var_2842 = const()[name = string("op_2842"), val = bool(true)];
            tensor<int32, [1]> var_2843_axes_0 = const()[name = string("op_2843_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2843_cast_fp16 = reduce_max(axes = var_2843_axes_0, keep_dims = var_2842, x = var_2839_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp16 var_2844_to_fp16 = const()[name = string("op_2844_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_35_cast_fp16 = maximum(x = var_2843_cast_fp16, y = var_2844_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_35_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_35_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2847_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2847_cast_fp16")];
            tensor<int32, [1]> var_2852_axes_0 = const()[name = string("op_2852_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2852_keep_dims_0 = const()[name = string("op_2852_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2852_cast_fp16 = reduce_mean(axes = var_2852_axes_0, keep_dims = var_2852_keep_dims_0, x = var_2847_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2853_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp32 var_2854_epsilon_0 = const()[name = string("op_2854_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2854_cast_fp16 = inverse(epsilon = var_2854_epsilon_0, x = var_2853_cast_fp16)[name = string("op_2854_cast_fp16")];
            fp16 var_2855_to_fp16 = const()[name = string("op_2855_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2856_cast_fp16 = mul(x = var_2854_cast_fp16, y = var_2855_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2858_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2856_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp32 var_2859_epsilon_0 = const()[name = string("op_2859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2859_cast_fp16 = rsqrt(epsilon = var_2859_epsilon_0, x = var_2858_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_31_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2859_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16884992)))];
            tensor<fp16, [1, 256, 512]> input_25_cast_fp16 = mul(x = out_31_cast_fp16, y = tm_layers_2_input_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16886080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17410432))))[name = string("tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [4]> var_2869 = const()[name = string("op_2869"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_2870_cast_fp16 = reshape(shape = var_2869, x = linear_19_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17412544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17674752))))[name = string("tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [4]> var_2881 = const()[name = string("op_2881"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2882_cast_fp16 = reshape(shape = var_2881, x = linear_20_cast_fp16)[name = string("op_2882_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17675840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17938048))))[name = string("tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2894_cast_fp16 = reshape(shape = var_2893, x = linear_21_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [4]> x_43_perm_0 = const()[name = string("x_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2870_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 4, 256, 256]> var_2898_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2898_cast_fp16")];
            bool var_2901 = const()[name = string("op_2901"), val = bool(true)];
            tensor<int32, [1]> var_2902_axes_0 = const()[name = string("op_2902_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_2902_cast_fp16 = reduce_max(axes = var_2902_axes_0, keep_dims = var_2901, x = var_2898_cast_fp16)[name = string("op_2902_cast_fp16")];
            fp16 var_2903_to_fp16 = const()[name = string("op_2903_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_37_cast_fp16 = maximum(x = var_2902_cast_fp16, y = var_2903_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_37_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_37_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2906_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<int32, [1]> var_2911_axes_0 = const()[name = string("op_2911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2911_keep_dims_0 = const()[name = string("op_2911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_2911_cast_fp16 = reduce_mean(axes = var_2911_axes_0, keep_dims = var_2911_keep_dims_0, x = var_2906_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2912_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2912_cast_fp16")];
            fp32 var_2913_epsilon_0 = const()[name = string("op_2913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_2913_cast_fp16 = inverse(epsilon = var_2913_epsilon_0, x = var_2912_cast_fp16)[name = string("op_2913_cast_fp16")];
            fp16 var_2914_to_fp16 = const()[name = string("op_2914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_2915_cast_fp16 = mul(x = var_2913_cast_fp16, y = var_2914_to_fp16)[name = string("op_2915_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2917_cast_fp16 = add(x = var_2911_cast_fp16, y = var_2915_cast_fp16)[name = string("op_2917_cast_fp16")];
            fp32 var_2918_epsilon_0 = const()[name = string("op_2918_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_2918_cast_fp16 = rsqrt(epsilon = var_2918_epsilon_0, x = var_2917_cast_fp16)[name = string("op_2918_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_33_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2918_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_9_cast_fp16 = mul(x = out_33_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2882_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 2, 256, 256]> var_2921_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2921_cast_fp16")];
            bool var_2924 = const()[name = string("op_2924"), val = bool(true)];
            tensor<int32, [1]> var_2925_axes_0 = const()[name = string("op_2925_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2925_cast_fp16 = reduce_max(axes = var_2925_axes_0, keep_dims = var_2924, x = var_2921_cast_fp16)[name = string("op_2925_cast_fp16")];
            fp16 var_2926_to_fp16 = const()[name = string("op_2926_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_39_cast_fp16 = maximum(x = var_2925_cast_fp16, y = var_2926_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_39_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_39_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2929_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2929_cast_fp16")];
            tensor<int32, [1]> var_2934_axes_0 = const()[name = string("op_2934_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2934_keep_dims_0 = const()[name = string("op_2934_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2934_cast_fp16 = reduce_mean(axes = var_2934_axes_0, keep_dims = var_2934_keep_dims_0, x = var_2929_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2935_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2935_cast_fp16")];
            fp32 var_2936_epsilon_0 = const()[name = string("op_2936_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2936_cast_fp16 = inverse(epsilon = var_2936_epsilon_0, x = var_2935_cast_fp16)[name = string("op_2936_cast_fp16")];
            fp16 var_2937_to_fp16 = const()[name = string("op_2937_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2938_cast_fp16 = mul(x = var_2936_cast_fp16, y = var_2937_to_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2940_cast_fp16 = add(x = var_2934_cast_fp16, y = var_2938_cast_fp16)[name = string("op_2940_cast_fp16")];
            fp32 var_2941_epsilon_0 = const()[name = string("op_2941_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2941_cast_fp16 = rsqrt(epsilon = var_2941_epsilon_0, x = var_2940_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_35_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2941_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_13_cast_fp16 = mul(x = out_35_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_43_cast_fp16 = transpose(perm = x_43_perm_0, x = var_2894_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 2, 256, 256]> var_2944_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2944_cast_fp16")];
            bool var_2947 = const()[name = string("op_2947"), val = bool(true)];
            tensor<int32, [1]> var_2948_axes_0 = const()[name = string("op_2948_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2948_cast_fp16 = reduce_max(axes = var_2948_axes_0, keep_dims = var_2947, x = var_2944_cast_fp16)[name = string("op_2948_cast_fp16")];
            fp16 var_2949_to_fp16 = const()[name = string("op_2949_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_41_cast_fp16 = maximum(x = var_2948_cast_fp16, y = var_2949_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_41_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_41_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2952_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2952_cast_fp16")];
            tensor<int32, [1]> var_2957_axes_0 = const()[name = string("op_2957_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2957_keep_dims_0 = const()[name = string("op_2957_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2957_cast_fp16 = reduce_mean(axes = var_2957_axes_0, keep_dims = var_2957_keep_dims_0, x = var_2952_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2958_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2958_cast_fp16")];
            fp32 var_2959_epsilon_0 = const()[name = string("op_2959_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2959_cast_fp16 = inverse(epsilon = var_2959_epsilon_0, x = var_2958_cast_fp16)[name = string("op_2959_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2961_cast_fp16 = mul(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2963_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2961_cast_fp16)[name = string("op_2963_cast_fp16")];
            fp32 var_2964_epsilon_0 = const()[name = string("op_2964_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2964_cast_fp16 = rsqrt(epsilon = var_2964_epsilon_0, x = var_2963_cast_fp16)[name = string("op_2964_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_9_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2964_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2966_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2966_cast_fp16")];
            tensor<int32, [4]> var_2971_begin_0 = const()[name = string("op_2971_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2971_end_0 = const()[name = string("op_2971_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_2971_end_mask_0 = const()[name = string("op_2971_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_2971_cast_fp16 = slice_by_index(begin = var_2971_begin_0, end = var_2971_end_0, end_mask = var_2971_end_mask_0, x = q_9_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_2972_cast_fp16 = mul(x = var_2971_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2972_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            int32 var_2979 = const()[name = string("op_2979"), val = int32(-1)];
            bool var_2980_interleave_0 = const()[name = string("op_2980_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_2980_cast_fp16 = concat(axis = var_2979, interleave = var_2980_interleave_0, values = (var_2972_cast_fp16, var_2977_cast_fp16))[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2981_cast_fp16 = mul(x = var_2980_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2981_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_11_cast_fp16 = add(x = var_2966_cast_fp16, y = var_2981_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2984_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2989_begin_0 = const()[name = string("op_2989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2989_end_0 = const()[name = string("op_2989_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_2989_end_mask_0 = const()[name = string("op_2989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_2989_cast_fp16 = slice_by_index(begin = var_2989_begin_0, end = var_2989_end_0, end_mask = var_2989_end_mask_0, x = k_13_cast_fp16)[name = string("op_2989_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_2990_cast_fp16 = mul(x = var_2989_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2990_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = k_13_cast_fp16)[name = string("op_2995_cast_fp16")];
            int32 var_2997 = const()[name = string("op_2997"), val = int32(-1)];
            bool var_2998_interleave_0 = const()[name = string("op_2998_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_2998_cast_fp16 = concat(axis = var_2997, interleave = var_2998_interleave_0, values = (var_2990_cast_fp16, var_2995_cast_fp16))[name = string("op_2998_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2999_cast_fp16 = mul(x = var_2998_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2999_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_15_cast_fp16 = add(x = var_2984_cast_fp16, y = var_2999_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = k_15_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [4, 1, 256, 256]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
            tensor<int32, [5]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_8_cast_fp16 = reshape(shape = concat_8, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [4, 1, 256, 256]> reshape_9_cast_fp16 = reshape(shape = concat_9, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = v_9_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [4, 1, 256, 256]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
            tensor<int32, [5]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_10_cast_fp16 = reshape(shape = concat_10, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [4, 1, 256, 256]> reshape_11_cast_fp16 = reshape(shape = concat_11, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3013_transpose_x_0 = const()[name = string("op_3013_transpose_x_0"), val = bool(false)];
            bool var_3013_transpose_y_0 = const()[name = string("op_3013_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 4, 256, 256]> var_3013_cast_fp16 = matmul(transpose_x = var_3013_transpose_x_0, transpose_y = var_3013_transpose_y_0, x = q_11_cast_fp16, y = transpose_98_cast_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3015_cast_fp16 = add(x = var_3013_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3015_cast_fp16")];
            int32 var_3016 = const()[name = string("op_3016"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_41_cast_fp16 = softmax(axis = var_3016, x = var_3015_cast_fp16)[name = string("w_41_cast_fp16")];
            bool var_3019_transpose_x_0 = const()[name = string("op_3019_transpose_x_0"), val = bool(false)];
            bool var_3019_transpose_y_0 = const()[name = string("op_3019_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 4, 256, 256]> var_3019_cast_fp16 = matmul(transpose_x = var_3019_transpose_x_0, transpose_y = var_3019_transpose_y_0, x = w_41_cast_fp16, y = v_11_cast_fp16)[name = string("op_3019_cast_fp16")];
            tensor<int32, [4]> var_3022_perm_0 = const()[name = string("op_3022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3026 = const()[name = string("op_3026"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_3022_cast_fp16 = transpose(perm = var_3022_perm_0, x = var_3019_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 256, 1024]> input_27_cast_fp16 = reshape(shape = var_3026, x = var_3022_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17939136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18463488))))[name = string("tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3031_cast_fp16 = abs(x = linear_22_cast_fp16)[name = string("op_3031_cast_fp16")];
            bool var_3034 = const()[name = string("op_3034"), val = bool(true)];
            tensor<int32, [1]> var_3035_axes_0 = const()[name = string("op_3035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3035_cast_fp16 = reduce_max(axes = var_3035_axes_0, keep_dims = var_3034, x = var_3031_cast_fp16)[name = string("op_3035_cast_fp16")];
            fp16 var_3036_to_fp16 = const()[name = string("op_3036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_43_cast_fp16 = maximum(x = var_3035_cast_fp16, y = var_3036_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_43_cast_fp16 = real_div(x = linear_22_cast_fp16, y = s_43_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3039_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3039_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3045_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_3045_cast_fp16")];
            fp32 var_3046_epsilon_0 = const()[name = string("op_3046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3046_cast_fp16 = inverse(epsilon = var_3046_epsilon_0, x = var_3045_cast_fp16)[name = string("op_3046_cast_fp16")];
            fp16 var_3047_to_fp16 = const()[name = string("op_3047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3048_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3047_to_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3050_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3048_cast_fp16)[name = string("op_3050_cast_fp16")];
            fp32 var_3051_epsilon_0 = const()[name = string("op_3051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3051_cast_fp16 = rsqrt(epsilon = var_3051_epsilon_0, x = var_3050_cast_fp16)[name = string("op_3051_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_3051_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18464576)))];
            tensor<fp16, [1, 256, 512]> var_3053_cast_fp16 = mul(x = out_37_cast_fp16, y = tm_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = var_3053_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3056_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_3056_cast_fp16")];
            bool var_3059 = const()[name = string("op_3059"), val = bool(true)];
            tensor<int32, [1]> var_3060_axes_0 = const()[name = string("op_3060_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3060_cast_fp16 = reduce_max(axes = var_3060_axes_0, keep_dims = var_3059, x = var_3056_cast_fp16)[name = string("op_3060_cast_fp16")];
            fp16 var_3061_to_fp16 = const()[name = string("op_3061_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_45_cast_fp16 = maximum(x = var_3060_cast_fp16, y = var_3061_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_45_cast_fp16 = real_div(x = x_47_cast_fp16, y = s_45_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3064_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<int32, [1]> var_3069_axes_0 = const()[name = string("op_3069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3069_keep_dims_0 = const()[name = string("op_3069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3069_cast_fp16 = reduce_mean(axes = var_3069_axes_0, keep_dims = var_3069_keep_dims_0, x = var_3064_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3070_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp32 var_3071_epsilon_0 = const()[name = string("op_3071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3071_cast_fp16 = inverse(epsilon = var_3071_epsilon_0, x = var_3070_cast_fp16)[name = string("op_3071_cast_fp16")];
            fp16 var_3072_to_fp16 = const()[name = string("op_3072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3073_cast_fp16 = mul(x = var_3071_cast_fp16, y = var_3072_to_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3075_cast_fp16 = add(x = var_3069_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3075_cast_fp16")];
            fp32 var_3076_epsilon_0 = const()[name = string("op_3076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3076_cast_fp16 = rsqrt(epsilon = var_3076_epsilon_0, x = var_3075_cast_fp16)[name = string("op_3076_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_3076_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18465664)))];
            tensor<fp16, [1, 256, 512]> input_29_cast_fp16 = mul(x = out_39_cast_fp16, y = tm_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18466752))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19515392))))[name = string("tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_23_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_23_cast_fp16")];
            string var_3083_mode_0 = const()[name = string("op_3083_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_3083_cast_fp16 = gelu(mode = var_3083_mode_0, x = linear_23_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19519552))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20568192))))[name = string("tm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_24_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_31_cast_fp16 = mul(x = var_3083_cast_fp16, y = linear_24_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20572352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21620992))))[name = string("tm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3091_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_3091_cast_fp16")];
            bool var_3094 = const()[name = string("op_3094"), val = bool(true)];
            tensor<int32, [1]> var_3095_axes_0 = const()[name = string("op_3095_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3095_cast_fp16 = reduce_max(axes = var_3095_axes_0, keep_dims = var_3094, x = var_3091_cast_fp16)[name = string("op_3095_cast_fp16")];
            fp16 var_3096_to_fp16 = const()[name = string("op_3096_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_47_cast_fp16 = maximum(x = var_3095_cast_fp16, y = var_3096_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_47_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_47_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3099_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3104_keep_dims_0 = const()[name = string("op_3104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3104_cast_fp16 = reduce_mean(axes = var_3104_axes_0, keep_dims = var_3104_keep_dims_0, x = var_3099_cast_fp16)[name = string("op_3104_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3105_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_3105_cast_fp16")];
            fp32 var_3106_epsilon_0 = const()[name = string("op_3106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3106_cast_fp16 = inverse(epsilon = var_3106_epsilon_0, x = var_3105_cast_fp16)[name = string("op_3106_cast_fp16")];
            fp16 var_3107_to_fp16 = const()[name = string("op_3107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3108_cast_fp16 = mul(x = var_3106_cast_fp16, y = var_3107_to_fp16)[name = string("op_3108_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3110_cast_fp16 = add(x = var_3104_cast_fp16, y = var_3108_cast_fp16)[name = string("op_3110_cast_fp16")];
            fp32 var_3111_epsilon_0 = const()[name = string("op_3111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3111_cast_fp16 = rsqrt(epsilon = var_3111_epsilon_0, x = var_3110_cast_fp16)[name = string("op_3111_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_3111_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21622080)))];
            tensor<fp16, [1, 256, 512]> var_3113_cast_fp16 = mul(x = out_41_cast_fp16, y = tm_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_33_cast_fp16 = add(x = x_47_cast_fp16, y = var_3113_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21623168))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21885376))))[name = string("tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_26_cast_fp16")];
            string x_51_mode_0 = const()[name = string("x_51_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_51_cast_fp16 = gelu(mode = x_51_mode_0, x = linear_26_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([1, 256, 3, 512])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3133_squeeze_mask_0 = const()[name = string("op_3133_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, squeeze_mask = var_3133_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_35_cast_fp16 = mul(x = x_51_cast_fp16, y = var_3133_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21886464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22148672))))[name = string("tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3143_cast_fp16 = abs(x = linear_27_cast_fp16)[name = string("op_3143_cast_fp16")];
            bool var_3146 = const()[name = string("op_3146"), val = bool(true)];
            tensor<int32, [1]> var_3147_axes_0 = const()[name = string("op_3147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3147_cast_fp16 = reduce_max(axes = var_3147_axes_0, keep_dims = var_3146, x = var_3143_cast_fp16)[name = string("op_3147_cast_fp16")];
            fp16 var_3148_to_fp16 = const()[name = string("op_3148_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_49_cast_fp16 = maximum(x = var_3147_cast_fp16, y = var_3148_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_49_cast_fp16 = real_div(x = linear_27_cast_fp16, y = s_49_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3151_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_3151_cast_fp16")];
            tensor<int32, [1]> var_3156_axes_0 = const()[name = string("op_3156_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3156_keep_dims_0 = const()[name = string("op_3156_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3156_cast_fp16 = reduce_mean(axes = var_3156_axes_0, keep_dims = var_3156_keep_dims_0, x = var_3151_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3157_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_3157_cast_fp16")];
            fp32 var_3158_epsilon_0 = const()[name = string("op_3158_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3158_cast_fp16 = inverse(epsilon = var_3158_epsilon_0, x = var_3157_cast_fp16)[name = string("op_3158_cast_fp16")];
            fp16 var_3159_to_fp16 = const()[name = string("op_3159_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3160_cast_fp16 = mul(x = var_3158_cast_fp16, y = var_3159_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3162_cast_fp16 = add(x = var_3156_cast_fp16, y = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp32 var_3163_epsilon_0 = const()[name = string("op_3163_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3163_cast_fp16 = rsqrt(epsilon = var_3163_epsilon_0, x = var_3162_cast_fp16)[name = string("op_3163_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_43_cast_fp16 = mul(x = y_49_cast_fp16, y = var_3163_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22149760)))];
            tensor<fp16, [1, 256, 512]> var_3165_cast_fp16 = mul(x = out_43_cast_fp16, y = tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_5_cast_fp16 = add(x = input_33_cast_fp16, y = var_3165_cast_fp16)[name = string("h_5_cast_fp16")];
            tensor<fp16, [1]> tm_layers_2_layer_scalar_to_fp16 = const()[name = string("tm_layers_2_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ep-1])];
            tensor<fp16, [1, 256, 512]> x_55_cast_fp16 = mul(x = h_5_cast_fp16, y = tm_layers_2_layer_scalar_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3169_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_3169_cast_fp16")];
            bool var_3172 = const()[name = string("op_3172"), val = bool(true)];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3173_cast_fp16 = reduce_max(axes = var_3173_axes_0, keep_dims = var_3172, x = var_3169_cast_fp16)[name = string("op_3173_cast_fp16")];
            fp16 var_3174_to_fp16 = const()[name = string("op_3174_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_51_cast_fp16 = maximum(x = var_3173_cast_fp16, y = var_3174_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_51_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_51_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3177_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_3177_cast_fp16")];
            tensor<int32, [1]> var_3182_axes_0 = const()[name = string("op_3182_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3182_keep_dims_0 = const()[name = string("op_3182_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3182_cast_fp16 = reduce_mean(axes = var_3182_axes_0, keep_dims = var_3182_keep_dims_0, x = var_3177_cast_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3183_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_3183_cast_fp16")];
            fp32 var_3184_epsilon_0 = const()[name = string("op_3184_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3184_cast_fp16 = inverse(epsilon = var_3184_epsilon_0, x = var_3183_cast_fp16)[name = string("op_3184_cast_fp16")];
            fp16 var_3185_to_fp16 = const()[name = string("op_3185_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3186_cast_fp16 = mul(x = var_3184_cast_fp16, y = var_3185_to_fp16)[name = string("op_3186_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3188_cast_fp16 = add(x = var_3182_cast_fp16, y = var_3186_cast_fp16)[name = string("op_3188_cast_fp16")];
            fp32 var_3189_epsilon_0 = const()[name = string("op_3189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3189_cast_fp16 = rsqrt(epsilon = var_3189_epsilon_0, x = var_3188_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_45_cast_fp16 = mul(x = y_51_cast_fp16, y = var_3189_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22150848)))];
            tensor<fp16, [1, 256, 512]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = tm_layers_3_input_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22151936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22676288))))[name = string("tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_3200_cast_fp16 = reshape(shape = var_3199, x = linear_28_cast_fp16)[name = string("op_3200_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22678400))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22940608))))[name = string("tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_29_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_59_perm_0 = const()[name = string("x_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22941696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23203904))))[name = string("tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_3223 = const()[name = string("op_3223"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_3224_cast_fp16 = reshape(shape = var_3223, x = linear_30_cast_fp16)[name = string("op_3224_cast_fp16")];
            tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_3200_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 4, 256, 256]> var_3228_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3231 = const()[name = string("op_3231"), val = bool(true)];
            tensor<int32, [1]> var_3232_axes_0 = const()[name = string("op_3232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_3232_cast_fp16 = reduce_max(axes = var_3232_axes_0, keep_dims = var_3231, x = var_3228_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_53_cast_fp16 = maximum(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_53_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_53_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3236_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_3236_cast_fp16")];
            tensor<int32, [1]> var_3241_axes_0 = const()[name = string("op_3241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3241_keep_dims_0 = const()[name = string("op_3241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_3241_cast_fp16 = reduce_mean(axes = var_3241_axes_0, keep_dims = var_3241_keep_dims_0, x = var_3236_cast_fp16)[name = string("op_3241_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3242_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_3242_cast_fp16")];
            fp32 var_3243_epsilon_0 = const()[name = string("op_3243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_3243_cast_fp16 = inverse(epsilon = var_3243_epsilon_0, x = var_3242_cast_fp16)[name = string("op_3243_cast_fp16")];
            fp16 var_3244_to_fp16 = const()[name = string("op_3244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_3245_cast_fp16 = mul(x = var_3243_cast_fp16, y = var_3244_to_fp16)[name = string("op_3245_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3247_cast_fp16 = add(x = var_3241_cast_fp16, y = var_3245_cast_fp16)[name = string("op_3247_cast_fp16")];
            fp32 var_3248_epsilon_0 = const()[name = string("op_3248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_3248_cast_fp16 = rsqrt(epsilon = var_3248_epsilon_0, x = var_3247_cast_fp16)[name = string("op_3248_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_47_cast_fp16 = mul(x = y_53_cast_fp16, y = var_3248_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23204992)))];
            tensor<fp16, [1, 4, 256, 256]> q_13_cast_fp16 = mul(x = out_47_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_59_cast_fp16 = transpose(perm = x_59_perm_0, x = var_3212_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 2, 256, 256]> var_3251_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_3251_cast_fp16")];
            bool var_3254 = const()[name = string("op_3254"), val = bool(true)];
            tensor<int32, [1]> var_3255_axes_0 = const()[name = string("op_3255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_3255_cast_fp16 = reduce_max(axes = var_3255_axes_0, keep_dims = var_3254, x = var_3251_cast_fp16)[name = string("op_3255_cast_fp16")];
            fp16 var_3256_to_fp16 = const()[name = string("op_3256_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_55_cast_fp16 = maximum(x = var_3255_cast_fp16, y = var_3256_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_55_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_55_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3259_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<int32, [1]> var_3264_axes_0 = const()[name = string("op_3264_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3264_keep_dims_0 = const()[name = string("op_3264_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_3264_cast_fp16 = reduce_mean(axes = var_3264_axes_0, keep_dims = var_3264_keep_dims_0, x = var_3259_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3265_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_3265_cast_fp16")];
            fp32 var_3266_epsilon_0 = const()[name = string("op_3266_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_3266_cast_fp16 = inverse(epsilon = var_3266_epsilon_0, x = var_3265_cast_fp16)[name = string("op_3266_cast_fp16")];
            fp16 var_3267_to_fp16 = const()[name = string("op_3267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_3268_cast_fp16 = mul(x = var_3266_cast_fp16, y = var_3267_to_fp16)[name = string("op_3268_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3270_cast_fp16 = add(x = var_3264_cast_fp16, y = var_3268_cast_fp16)[name = string("op_3270_cast_fp16")];
            fp32 var_3271_epsilon_0 = const()[name = string("op_3271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_3271_cast_fp16 = rsqrt(epsilon = var_3271_epsilon_0, x = var_3270_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_49_cast_fp16 = mul(x = y_55_cast_fp16, y = var_3271_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23205568)))];
            tensor<fp16, [1, 2, 256, 256]> k_19_cast_fp16 = mul(x = out_49_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_3224_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2, 256, 256]> var_3274_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_3274_cast_fp16")];
            bool var_3277 = const()[name = string("op_3277"), val = bool(true)];
            tensor<int32, [1]> var_3278_axes_0 = const()[name = string("op_3278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_3278_cast_fp16 = reduce_max(axes = var_3278_axes_0, keep_dims = var_3277, x = var_3274_cast_fp16)[name = string("op_3278_cast_fp16")];
            fp16 var_3279_to_fp16 = const()[name = string("op_3279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_57_cast_fp16 = maximum(x = var_3278_cast_fp16, y = var_3279_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_57_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_57_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3282_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_3282_cast_fp16")];
            tensor<int32, [1]> var_3287_axes_0 = const()[name = string("op_3287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3287_keep_dims_0 = const()[name = string("op_3287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_3287_cast_fp16 = reduce_mean(axes = var_3287_axes_0, keep_dims = var_3287_keep_dims_0, x = var_3282_cast_fp16)[name = string("op_3287_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3288_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_3289_cast_fp16 = inverse(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            fp16 var_3290_to_fp16 = const()[name = string("op_3290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_3291_cast_fp16 = mul(x = var_3289_cast_fp16, y = var_3290_to_fp16)[name = string("op_3291_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3293_cast_fp16 = add(x = var_3287_cast_fp16, y = var_3291_cast_fp16)[name = string("op_3293_cast_fp16")];
            fp32 var_3294_epsilon_0 = const()[name = string("op_3294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_3294_cast_fp16 = rsqrt(epsilon = var_3294_epsilon_0, x = var_3293_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_13_cast_fp16 = mul(x = y_57_cast_fp16, y = var_3294_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3296_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3296_cast_fp16")];
            tensor<int32, [4]> var_3301_begin_0 = const()[name = string("op_3301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3301_end_0 = const()[name = string("op_3301_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_3301_end_mask_0 = const()[name = string("op_3301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_3301_cast_fp16 = slice_by_index(begin = var_3301_begin_0, end = var_3301_end_0, end_mask = var_3301_end_mask_0, x = q_13_cast_fp16)[name = string("op_3301_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_3302_cast_fp16 = mul(x = var_3301_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3307_begin_0 = const()[name = string("op_3307_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3307_end_0 = const()[name = string("op_3307_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_3307_end_mask_0 = const()[name = string("op_3307_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_3307_cast_fp16 = slice_by_index(begin = var_3307_begin_0, end = var_3307_end_0, end_mask = var_3307_end_mask_0, x = q_13_cast_fp16)[name = string("op_3307_cast_fp16")];
            int32 var_3309 = const()[name = string("op_3309"), val = int32(-1)];
            bool var_3310_interleave_0 = const()[name = string("op_3310_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_3310_cast_fp16 = concat(axis = var_3309, interleave = var_3310_interleave_0, values = (var_3302_cast_fp16, var_3307_cast_fp16))[name = string("op_3310_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3311_cast_fp16 = mul(x = var_3310_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3311_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_15_cast_fp16 = add(x = var_3296_cast_fp16, y = var_3311_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3314_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3314_cast_fp16")];
            tensor<int32, [4]> var_3319_begin_0 = const()[name = string("op_3319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3319_end_0 = const()[name = string("op_3319_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_3319_end_mask_0 = const()[name = string("op_3319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_3319_cast_fp16 = slice_by_index(begin = var_3319_begin_0, end = var_3319_end_0, end_mask = var_3319_end_mask_0, x = k_19_cast_fp16)[name = string("op_3319_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_3320_cast_fp16 = mul(x = var_3319_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_3320_cast_fp16")];
            tensor<int32, [4]> var_3325_begin_0 = const()[name = string("op_3325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3325_end_0 = const()[name = string("op_3325_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_3325_end_mask_0 = const()[name = string("op_3325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_3325_cast_fp16 = slice_by_index(begin = var_3325_begin_0, end = var_3325_end_0, end_mask = var_3325_end_mask_0, x = k_19_cast_fp16)[name = string("op_3325_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), val = int32(-1)];
            bool var_3328_interleave_0 = const()[name = string("op_3328_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_3328_cast_fp16 = concat(axis = var_3327, interleave = var_3328_interleave_0, values = (var_3320_cast_fp16, var_3325_cast_fp16))[name = string("op_3328_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3329_cast_fp16 = mul(x = var_3328_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3329_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_21_cast_fp16 = add(x = var_3314_cast_fp16, y = var_3329_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = k_21_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [4, 1, 256, 256]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_12_cast_fp16 = reshape(shape = concat_12, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [4, 1, 256, 256]> reshape_13_cast_fp16 = reshape(shape = concat_13, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = v_13_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [4, 1, 256, 256]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_14_cast_fp16 = reshape(shape = concat_14, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [4, 1, 256, 256]> reshape_15_cast_fp16 = reshape(shape = concat_15, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3343_transpose_x_0 = const()[name = string("op_3343_transpose_x_0"), val = bool(false)];
            bool var_3343_transpose_y_0 = const()[name = string("op_3343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 4, 256, 256]> var_3343_cast_fp16 = matmul(transpose_x = var_3343_transpose_x_0, transpose_y = var_3343_transpose_y_0, x = q_15_cast_fp16, y = transpose_99_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3345_cast_fp16 = add(x = var_3343_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3345_cast_fp16")];
            int32 var_3346 = const()[name = string("op_3346"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_57_cast_fp16 = softmax(axis = var_3346, x = var_3345_cast_fp16)[name = string("w_57_cast_fp16")];
            bool var_3349_transpose_x_0 = const()[name = string("op_3349_transpose_x_0"), val = bool(false)];
            bool var_3349_transpose_y_0 = const()[name = string("op_3349_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 4, 256, 256]> var_3349_cast_fp16 = matmul(transpose_x = var_3349_transpose_x_0, transpose_y = var_3349_transpose_y_0, x = w_57_cast_fp16, y = v_15_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<int32, [4]> var_3352_perm_0 = const()[name = string("op_3352_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3356 = const()[name = string("op_3356"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_3352_cast_fp16 = transpose(perm = var_3352_perm_0, x = var_3349_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 256, 1024]> input_39_cast_fp16 = reshape(shape = var_3356, x = var_3352_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23206144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23730496))))[name = string("tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3361_cast_fp16 = abs(x = linear_31_cast_fp16)[name = string("op_3361_cast_fp16")];
            bool var_3364 = const()[name = string("op_3364"), val = bool(true)];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3365_cast_fp16 = reduce_max(axes = var_3365_axes_0, keep_dims = var_3364, x = var_3361_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_59_cast_fp16 = maximum(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_59_cast_fp16 = real_div(x = linear_31_cast_fp16, y = s_59_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3369_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<int32, [1]> var_3374_axes_0 = const()[name = string("op_3374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3374_keep_dims_0 = const()[name = string("op_3374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3374_cast_fp16 = reduce_mean(axes = var_3374_axes_0, keep_dims = var_3374_keep_dims_0, x = var_3369_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3375_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_3375_cast_fp16")];
            fp32 var_3376_epsilon_0 = const()[name = string("op_3376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3376_cast_fp16 = inverse(epsilon = var_3376_epsilon_0, x = var_3375_cast_fp16)[name = string("op_3376_cast_fp16")];
            fp16 var_3377_to_fp16 = const()[name = string("op_3377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3378_cast_fp16 = mul(x = var_3376_cast_fp16, y = var_3377_to_fp16)[name = string("op_3378_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3380_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3380_cast_fp16")];
            fp32 var_3381_epsilon_0 = const()[name = string("op_3381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3381_cast_fp16 = rsqrt(epsilon = var_3381_epsilon_0, x = var_3380_cast_fp16)[name = string("op_3381_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_3381_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23731584)))];
            tensor<fp16, [1, 256, 512]> var_3383_cast_fp16 = mul(x = out_51_cast_fp16, y = tm_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_65_cast_fp16 = add(x = x_55_cast_fp16, y = var_3383_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3386_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_3386_cast_fp16")];
            bool var_3389 = const()[name = string("op_3389"), val = bool(true)];
            tensor<int32, [1]> var_3390_axes_0 = const()[name = string("op_3390_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3390_cast_fp16 = reduce_max(axes = var_3390_axes_0, keep_dims = var_3389, x = var_3386_cast_fp16)[name = string("op_3390_cast_fp16")];
            fp16 var_3391_to_fp16 = const()[name = string("op_3391_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_61_cast_fp16 = maximum(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_61_cast_fp16 = real_div(x = x_65_cast_fp16, y = s_61_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3394_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<int32, [1]> var_3399_axes_0 = const()[name = string("op_3399_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3399_keep_dims_0 = const()[name = string("op_3399_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3399_cast_fp16 = reduce_mean(axes = var_3399_axes_0, keep_dims = var_3399_keep_dims_0, x = var_3394_cast_fp16)[name = string("op_3399_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3400_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_3400_cast_fp16")];
            fp32 var_3401_epsilon_0 = const()[name = string("op_3401_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3401_cast_fp16 = inverse(epsilon = var_3401_epsilon_0, x = var_3400_cast_fp16)[name = string("op_3401_cast_fp16")];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3403_cast_fp16 = mul(x = var_3401_cast_fp16, y = var_3402_to_fp16)[name = string("op_3403_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3405_cast_fp16 = add(x = var_3399_cast_fp16, y = var_3403_cast_fp16)[name = string("op_3405_cast_fp16")];
            fp32 var_3406_epsilon_0 = const()[name = string("op_3406_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3406_cast_fp16 = rsqrt(epsilon = var_3406_epsilon_0, x = var_3405_cast_fp16)[name = string("op_3406_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_3406_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23732672)))];
            tensor<fp16, [1, 256, 512]> input_41_cast_fp16 = mul(x = out_53_cast_fp16, y = tm_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23733760))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24782400))))[name = string("tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_32_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_32_cast_fp16")];
            string var_3413_mode_0 = const()[name = string("op_3413_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_3413_cast_fp16 = gelu(mode = var_3413_mode_0, x = linear_32_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24786560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25835200))))[name = string("tm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_43_cast_fp16 = mul(x = var_3413_cast_fp16, y = linear_33_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25839360))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26888000))))[name = string("tm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3421_cast_fp16 = abs(x = linear_34_cast_fp16)[name = string("op_3421_cast_fp16")];
            bool var_3424 = const()[name = string("op_3424"), val = bool(true)];
            tensor<int32, [1]> var_3425_axes_0 = const()[name = string("op_3425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3425_cast_fp16 = reduce_max(axes = var_3425_axes_0, keep_dims = var_3424, x = var_3421_cast_fp16)[name = string("op_3425_cast_fp16")];
            fp16 var_3426_to_fp16 = const()[name = string("op_3426_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_63_cast_fp16 = maximum(x = var_3425_cast_fp16, y = var_3426_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_63_cast_fp16 = real_div(x = linear_34_cast_fp16, y = s_63_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3429_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_3429_cast_fp16")];
            tensor<int32, [1]> var_3434_axes_0 = const()[name = string("op_3434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3434_keep_dims_0 = const()[name = string("op_3434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3434_cast_fp16 = reduce_mean(axes = var_3434_axes_0, keep_dims = var_3434_keep_dims_0, x = var_3429_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3435_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp32 var_3436_epsilon_0 = const()[name = string("op_3436_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3436_cast_fp16 = inverse(epsilon = var_3436_epsilon_0, x = var_3435_cast_fp16)[name = string("op_3436_cast_fp16")];
            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3438_cast_fp16 = mul(x = var_3436_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3440_cast_fp16 = add(x = var_3434_cast_fp16, y = var_3438_cast_fp16)[name = string("op_3440_cast_fp16")];
            fp32 var_3441_epsilon_0 = const()[name = string("op_3441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3441_cast_fp16 = rsqrt(epsilon = var_3441_epsilon_0, x = var_3440_cast_fp16)[name = string("op_3441_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_55_cast_fp16 = mul(x = y_63_cast_fp16, y = var_3441_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26889088)))];
            tensor<fp16, [1, 256, 512]> var_3443_cast_fp16 = mul(x = out_55_cast_fp16, y = tm_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_45_cast_fp16 = add(x = x_65_cast_fp16, y = var_3443_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26890176))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27152384))))[name = string("tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_35_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_35_cast_fp16")];
            string x_69_mode_0 = const()[name = string("x_69_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_69_cast_fp16 = gelu(mode = x_69_mode_0, x = linear_35_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [4]> var_3463_begin_0 = const()[name = string("op_3463_begin_0"), val = tensor<int32, [4]>([0, 0, 3, 0])];
            tensor<int32, [4]> var_3463_end_0 = const()[name = string("op_3463_end_0"), val = tensor<int32, [4]>([1, 256, 4, 512])];
            tensor<bool, [4]> var_3463_end_mask_0 = const()[name = string("op_3463_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3463_squeeze_mask_0 = const()[name = string("op_3463_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_3463_cast_fp16 = slice_by_index(begin = var_3463_begin_0, end = var_3463_end_0, end_mask = var_3463_end_mask_0, squeeze_mask = var_3463_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3463_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_47_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3463_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27153472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27415680))))[name = string("tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3473_cast_fp16 = abs(x = linear_36_cast_fp16)[name = string("op_3473_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<int32, [1]> var_3477_axes_0 = const()[name = string("op_3477_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3477_cast_fp16 = reduce_max(axes = var_3477_axes_0, keep_dims = var_3476, x = var_3473_cast_fp16)[name = string("op_3477_cast_fp16")];
            fp16 var_3478_to_fp16 = const()[name = string("op_3478_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_65_cast_fp16 = maximum(x = var_3477_cast_fp16, y = var_3478_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_65_cast_fp16 = real_div(x = linear_36_cast_fp16, y = s_65_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3481_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [1]> var_3486_axes_0 = const()[name = string("op_3486_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3486_keep_dims_0 = const()[name = string("op_3486_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3486_cast_fp16 = reduce_mean(axes = var_3486_axes_0, keep_dims = var_3486_keep_dims_0, x = var_3481_cast_fp16)[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3487_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp32 var_3488_epsilon_0 = const()[name = string("op_3488_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3488_cast_fp16 = inverse(epsilon = var_3488_epsilon_0, x = var_3487_cast_fp16)[name = string("op_3488_cast_fp16")];
            fp16 var_3489_to_fp16 = const()[name = string("op_3489_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3490_cast_fp16 = mul(x = var_3488_cast_fp16, y = var_3489_to_fp16)[name = string("op_3490_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3492_cast_fp16 = add(x = var_3486_cast_fp16, y = var_3490_cast_fp16)[name = string("op_3492_cast_fp16")];
            fp32 var_3493_epsilon_0 = const()[name = string("op_3493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3493_cast_fp16 = rsqrt(epsilon = var_3493_epsilon_0, x = var_3492_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_57_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3493_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27416768)))];
            tensor<fp16, [1, 256, 512]> var_3495_cast_fp16 = mul(x = out_57_cast_fp16, y = tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3495_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_7_cast_fp16 = add(x = input_45_cast_fp16, y = var_3495_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<fp16, [1]> tm_layers_3_layer_scalar_to_fp16 = const()[name = string("tm_layers_3_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 256, 512]> x_73_cast_fp16 = mul(x = h_7_cast_fp16, y = tm_layers_3_layer_scalar_to_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3499_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_3499_cast_fp16")];
            bool var_3502 = const()[name = string("op_3502"), val = bool(true)];
            tensor<int32, [1]> var_3503_axes_0 = const()[name = string("op_3503_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3503_cast_fp16 = reduce_max(axes = var_3503_axes_0, keep_dims = var_3502, x = var_3499_cast_fp16)[name = string("op_3503_cast_fp16")];
            fp16 var_3504_to_fp16 = const()[name = string("op_3504_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_67_cast_fp16 = maximum(x = var_3503_cast_fp16, y = var_3504_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_67_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_67_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3507_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<int32, [1]> var_3512_axes_0 = const()[name = string("op_3512_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3512_keep_dims_0 = const()[name = string("op_3512_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3512_cast_fp16 = reduce_mean(axes = var_3512_axes_0, keep_dims = var_3512_keep_dims_0, x = var_3507_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3513_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_3513_cast_fp16")];
            fp32 var_3514_epsilon_0 = const()[name = string("op_3514_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3514_cast_fp16 = inverse(epsilon = var_3514_epsilon_0, x = var_3513_cast_fp16)[name = string("op_3514_cast_fp16")];
            fp16 var_3515_to_fp16 = const()[name = string("op_3515_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3516_cast_fp16 = mul(x = var_3514_cast_fp16, y = var_3515_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3518_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp32 var_3519_epsilon_0 = const()[name = string("op_3519_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3519_cast_fp16 = rsqrt(epsilon = var_3519_epsilon_0, x = var_3518_cast_fp16)[name = string("op_3519_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_59_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3519_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27417856)))];
            tensor<fp16, [1, 256, 512]> input_49_cast_fp16 = mul(x = out_59_cast_fp16, y = tm_layers_4_input_layernorm_weight_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27418944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27943296))))[name = string("tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3529 = const()[name = string("op_3529"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_3530_cast_fp16 = reshape(shape = var_3529, x = linear_37_cast_fp16)[name = string("op_3530_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27945408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28207616))))[name = string("tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3541 = const()[name = string("op_3541"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_3542_cast_fp16 = reshape(shape = var_3541, x = linear_38_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28208704))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28470912))))[name = string("tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_3553 = const()[name = string("op_3553"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_3554_cast_fp16 = reshape(shape = var_3553, x = linear_39_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<int32, [4]> x_79_perm_0 = const()[name = string("x_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3530_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 4, 256, 256]> var_3558_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3558_cast_fp16")];
            bool var_3561 = const()[name = string("op_3561"), val = bool(true)];
            tensor<int32, [1]> var_3562_axes_0 = const()[name = string("op_3562_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_3562_cast_fp16 = reduce_max(axes = var_3562_axes_0, keep_dims = var_3561, x = var_3558_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp16 var_3563_to_fp16 = const()[name = string("op_3563_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_69_cast_fp16 = maximum(x = var_3562_cast_fp16, y = var_3563_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_69_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_69_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3566_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3566_cast_fp16")];
            tensor<int32, [1]> var_3571_axes_0 = const()[name = string("op_3571_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3571_keep_dims_0 = const()[name = string("op_3571_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_3571_cast_fp16 = reduce_mean(axes = var_3571_axes_0, keep_dims = var_3571_keep_dims_0, x = var_3566_cast_fp16)[name = string("op_3571_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3572_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3572_cast_fp16")];
            fp32 var_3573_epsilon_0 = const()[name = string("op_3573_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_3573_cast_fp16 = inverse(epsilon = var_3573_epsilon_0, x = var_3572_cast_fp16)[name = string("op_3573_cast_fp16")];
            fp16 var_3574_to_fp16 = const()[name = string("op_3574_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_3575_cast_fp16 = mul(x = var_3573_cast_fp16, y = var_3574_to_fp16)[name = string("op_3575_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3577_cast_fp16 = add(x = var_3571_cast_fp16, y = var_3575_cast_fp16)[name = string("op_3577_cast_fp16")];
            fp32 var_3578_epsilon_0 = const()[name = string("op_3578_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_3578_cast_fp16 = rsqrt(epsilon = var_3578_epsilon_0, x = var_3577_cast_fp16)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_61_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3578_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [256]> tm_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472000)))];
            tensor<fp16, [1, 4, 256, 256]> q_17_cast_fp16 = mul(x = out_61_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3542_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2, 256, 256]> var_3581_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3581_cast_fp16")];
            bool var_3584 = const()[name = string("op_3584"), val = bool(true)];
            tensor<int32, [1]> var_3585_axes_0 = const()[name = string("op_3585_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_3585_cast_fp16 = reduce_max(axes = var_3585_axes_0, keep_dims = var_3584, x = var_3581_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp16 var_3586_to_fp16 = const()[name = string("op_3586_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_71_cast_fp16 = maximum(x = var_3585_cast_fp16, y = var_3586_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_71_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_71_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3589_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3589_cast_fp16")];
            tensor<int32, [1]> var_3594_axes_0 = const()[name = string("op_3594_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3594_keep_dims_0 = const()[name = string("op_3594_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_3594_cast_fp16 = reduce_mean(axes = var_3594_axes_0, keep_dims = var_3594_keep_dims_0, x = var_3589_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3595_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp32 var_3596_epsilon_0 = const()[name = string("op_3596_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_3596_cast_fp16 = inverse(epsilon = var_3596_epsilon_0, x = var_3595_cast_fp16)[name = string("op_3596_cast_fp16")];
            fp16 var_3597_to_fp16 = const()[name = string("op_3597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_3598_cast_fp16 = mul(x = var_3596_cast_fp16, y = var_3597_to_fp16)[name = string("op_3598_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3600_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3598_cast_fp16)[name = string("op_3600_cast_fp16")];
            fp32 var_3601_epsilon_0 = const()[name = string("op_3601_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_3601_cast_fp16 = rsqrt(epsilon = var_3601_epsilon_0, x = var_3600_cast_fp16)[name = string("op_3601_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_63_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3601_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472576)))];
            tensor<fp16, [1, 2, 256, 256]> k_25_cast_fp16 = mul(x = out_63_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_79_cast_fp16 = transpose(perm = x_79_perm_0, x = var_3554_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 2, 256, 256]> var_3604_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3604_cast_fp16")];
            bool var_3607 = const()[name = string("op_3607"), val = bool(true)];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_3607, x = var_3604_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_73_cast_fp16 = maximum(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_73_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_73_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3612_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<int32, [1]> var_3617_axes_0 = const()[name = string("op_3617_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3617_keep_dims_0 = const()[name = string("op_3617_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_3617_cast_fp16 = reduce_mean(axes = var_3617_axes_0, keep_dims = var_3617_keep_dims_0, x = var_3612_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3618_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_3619_cast_fp16 = inverse(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            fp16 var_3620_to_fp16 = const()[name = string("op_3620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_3621_cast_fp16 = mul(x = var_3619_cast_fp16, y = var_3620_to_fp16)[name = string("op_3621_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3623_cast_fp16 = add(x = var_3617_cast_fp16, y = var_3621_cast_fp16)[name = string("op_3623_cast_fp16")];
            fp32 var_3624_epsilon_0 = const()[name = string("op_3624_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_3624_cast_fp16 = rsqrt(epsilon = var_3624_epsilon_0, x = var_3623_cast_fp16)[name = string("op_3624_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_17_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3624_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3626_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3626_cast_fp16")];
            tensor<int32, [4]> var_3631_begin_0 = const()[name = string("op_3631_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3631_end_0 = const()[name = string("op_3631_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_3631_end_mask_0 = const()[name = string("op_3631_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_3631_cast_fp16 = slice_by_index(begin = var_3631_begin_0, end = var_3631_end_0, end_mask = var_3631_end_mask_0, x = q_17_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_3632_cast_fp16 = mul(x = var_3631_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> var_3637_begin_0 = const()[name = string("op_3637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3637_end_0 = const()[name = string("op_3637_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_3637_end_mask_0 = const()[name = string("op_3637_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_3637_cast_fp16 = slice_by_index(begin = var_3637_begin_0, end = var_3637_end_0, end_mask = var_3637_end_mask_0, x = q_17_cast_fp16)[name = string("op_3637_cast_fp16")];
            int32 var_3639 = const()[name = string("op_3639"), val = int32(-1)];
            bool var_3640_interleave_0 = const()[name = string("op_3640_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_3640_cast_fp16 = concat(axis = var_3639, interleave = var_3640_interleave_0, values = (var_3632_cast_fp16, var_3637_cast_fp16))[name = string("op_3640_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3641_cast_fp16 = mul(x = var_3640_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3641_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_19_cast_fp16 = add(x = var_3626_cast_fp16, y = var_3641_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3644_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = k_25_cast_fp16)[name = string("op_3649_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_3650_cast_fp16 = mul(x = var_3649_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            tensor<int32, [4]> var_3655_begin_0 = const()[name = string("op_3655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3655_end_0 = const()[name = string("op_3655_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_3655_end_mask_0 = const()[name = string("op_3655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_3655_cast_fp16 = slice_by_index(begin = var_3655_begin_0, end = var_3655_end_0, end_mask = var_3655_end_mask_0, x = k_25_cast_fp16)[name = string("op_3655_cast_fp16")];
            int32 var_3657 = const()[name = string("op_3657"), val = int32(-1)];
            bool var_3658_interleave_0 = const()[name = string("op_3658_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_3658_cast_fp16 = concat(axis = var_3657, interleave = var_3658_interleave_0, values = (var_3650_cast_fp16, var_3655_cast_fp16))[name = string("op_3658_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3659_cast_fp16 = mul(x = var_3658_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3659_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_27_cast_fp16 = add(x = var_3644_cast_fp16, y = var_3659_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = k_27_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [4, 1, 256, 256]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
            tensor<int32, [5]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_16_cast_fp16 = reshape(shape = concat_16, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [4, 1, 256, 256]> reshape_17_cast_fp16 = reshape(shape = concat_17, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = v_17_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [4, 1, 256, 256]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
            tensor<int32, [5]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_18_cast_fp16 = reshape(shape = concat_18, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [4, 1, 256, 256]> reshape_19_cast_fp16 = reshape(shape = concat_19, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3673_transpose_x_0 = const()[name = string("op_3673_transpose_x_0"), val = bool(false)];
            bool var_3673_transpose_y_0 = const()[name = string("op_3673_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 4, 256, 256]> var_3673_cast_fp16 = matmul(transpose_x = var_3673_transpose_x_0, transpose_y = var_3673_transpose_y_0, x = q_19_cast_fp16, y = transpose_100_cast_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3675_cast_fp16 = add(x = var_3673_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3675_cast_fp16")];
            int32 var_3676 = const()[name = string("op_3676"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_73_cast_fp16 = softmax(axis = var_3676, x = var_3675_cast_fp16)[name = string("w_73_cast_fp16")];
            bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)];
            bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 4, 256, 256]> var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = w_73_cast_fp16, y = v_19_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<int32, [4]> var_3682_perm_0 = const()[name = string("op_3682_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3686 = const()[name = string("op_3686"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_3682_cast_fp16 = transpose(perm = var_3682_perm_0, x = var_3679_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 256, 1024]> input_51_cast_fp16 = reshape(shape = var_3686, x = var_3682_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28473152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28997504))))[name = string("tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3691_cast_fp16 = abs(x = linear_40_cast_fp16)[name = string("op_3691_cast_fp16")];
            bool var_3694 = const()[name = string("op_3694"), val = bool(true)];
            tensor<int32, [1]> var_3695_axes_0 = const()[name = string("op_3695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3695_cast_fp16 = reduce_max(axes = var_3695_axes_0, keep_dims = var_3694, x = var_3691_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 var_3696_to_fp16 = const()[name = string("op_3696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_75_cast_fp16 = maximum(x = var_3695_cast_fp16, y = var_3696_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_75_cast_fp16 = real_div(x = linear_40_cast_fp16, y = s_75_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3699_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3699_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3704_keep_dims_0 = const()[name = string("op_3704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3704_cast_fp16 = reduce_mean(axes = var_3704_axes_0, keep_dims = var_3704_keep_dims_0, x = var_3699_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3705_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3705_cast_fp16")];
            fp32 var_3706_epsilon_0 = const()[name = string("op_3706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3706_cast_fp16 = inverse(epsilon = var_3706_epsilon_0, x = var_3705_cast_fp16)[name = string("op_3706_cast_fp16")];
            fp16 var_3707_to_fp16 = const()[name = string("op_3707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3708_cast_fp16 = mul(x = var_3706_cast_fp16, y = var_3707_to_fp16)[name = string("op_3708_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3710_cast_fp16 = add(x = var_3704_cast_fp16, y = var_3708_cast_fp16)[name = string("op_3710_cast_fp16")];
            fp32 var_3711_epsilon_0 = const()[name = string("op_3711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3711_cast_fp16 = rsqrt(epsilon = var_3711_epsilon_0, x = var_3710_cast_fp16)[name = string("op_3711_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3711_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28998592)))];
            tensor<fp16, [1, 256, 512]> var_3713_cast_fp16 = mul(x = out_65_cast_fp16, y = tm_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = var_3713_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3716_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3716_cast_fp16")];
            bool var_3719 = const()[name = string("op_3719"), val = bool(true)];
            tensor<int32, [1]> var_3720_axes_0 = const()[name = string("op_3720_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3720_cast_fp16 = reduce_max(axes = var_3720_axes_0, keep_dims = var_3719, x = var_3716_cast_fp16)[name = string("op_3720_cast_fp16")];
            fp16 var_3721_to_fp16 = const()[name = string("op_3721_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_77_cast_fp16 = maximum(x = var_3720_cast_fp16, y = var_3721_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_77_cast_fp16 = real_div(x = x_83_cast_fp16, y = s_77_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3724_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3729_axes_0 = const()[name = string("op_3729_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3729_keep_dims_0 = const()[name = string("op_3729_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3729_cast_fp16 = reduce_mean(axes = var_3729_axes_0, keep_dims = var_3729_keep_dims_0, x = var_3724_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3730_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3730_cast_fp16")];
            fp32 var_3731_epsilon_0 = const()[name = string("op_3731_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3731_cast_fp16 = inverse(epsilon = var_3731_epsilon_0, x = var_3730_cast_fp16)[name = string("op_3731_cast_fp16")];
            fp16 var_3732_to_fp16 = const()[name = string("op_3732_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3733_cast_fp16 = mul(x = var_3731_cast_fp16, y = var_3732_to_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3735_cast_fp16 = add(x = var_3729_cast_fp16, y = var_3733_cast_fp16)[name = string("op_3735_cast_fp16")];
            fp32 var_3736_epsilon_0 = const()[name = string("op_3736_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3736_cast_fp16 = rsqrt(epsilon = var_3736_epsilon_0, x = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_67_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3736_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28999680)))];
            tensor<fp16, [1, 256, 512]> input_53_cast_fp16 = mul(x = out_67_cast_fp16, y = tm_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29000768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30049408))))[name = string("tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_41_cast_fp16")];
            string var_3743_mode_0 = const()[name = string("op_3743_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_3743_cast_fp16 = gelu(mode = var_3743_mode_0, x = linear_41_cast_fp16)[name = string("op_3743_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30053568))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31102208))))[name = string("tm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_42_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_55_cast_fp16 = mul(x = var_3743_cast_fp16, y = linear_42_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31106368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32155008))))[name = string("tm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3751_cast_fp16 = abs(x = linear_43_cast_fp16)[name = string("op_3751_cast_fp16")];
            bool var_3754 = const()[name = string("op_3754"), val = bool(true)];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_3754, x = var_3751_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 var_3756_to_fp16 = const()[name = string("op_3756_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_79_cast_fp16 = maximum(x = var_3755_cast_fp16, y = var_3756_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_79_cast_fp16 = real_div(x = linear_43_cast_fp16, y = s_79_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3759_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<int32, [1]> var_3764_axes_0 = const()[name = string("op_3764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3764_keep_dims_0 = const()[name = string("op_3764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3764_cast_fp16 = reduce_mean(axes = var_3764_axes_0, keep_dims = var_3764_keep_dims_0, x = var_3759_cast_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3765_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3766_cast_fp16 = inverse(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3768_cast_fp16 = mul(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("op_3768_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3770_cast_fp16 = add(x = var_3764_cast_fp16, y = var_3768_cast_fp16)[name = string("op_3770_cast_fp16")];
            fp32 var_3771_epsilon_0 = const()[name = string("op_3771_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3771_cast_fp16 = rsqrt(epsilon = var_3771_epsilon_0, x = var_3770_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_69_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3771_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32156096)))];
            tensor<fp16, [1, 256, 512]> var_3773_cast_fp16 = mul(x = out_69_cast_fp16, y = tm_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3773_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_57_cast_fp16 = add(x = x_83_cast_fp16, y = var_3773_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32157184))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32419392))))[name = string("tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_44_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_44_cast_fp16")];
            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_87_cast_fp16 = gelu(mode = x_87_mode_0, x = linear_44_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [4]> var_3793_begin_0 = const()[name = string("op_3793_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_3793_end_0 = const()[name = string("op_3793_end_0"), val = tensor<int32, [4]>([1, 256, 5, 512])];
            tensor<bool, [4]> var_3793_end_mask_0 = const()[name = string("op_3793_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3793_squeeze_mask_0 = const()[name = string("op_3793_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_3793_cast_fp16 = slice_by_index(begin = var_3793_begin_0, end = var_3793_end_0, end_mask = var_3793_end_mask_0, squeeze_mask = var_3793_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_59_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3793_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32420480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32682688))))[name = string("tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3803_cast_fp16 = abs(x = linear_45_cast_fp16)[name = string("op_3803_cast_fp16")];
            bool var_3806 = const()[name = string("op_3806"), val = bool(true)];
            tensor<int32, [1]> var_3807_axes_0 = const()[name = string("op_3807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3807_cast_fp16 = reduce_max(axes = var_3807_axes_0, keep_dims = var_3806, x = var_3803_cast_fp16)[name = string("op_3807_cast_fp16")];
            fp16 var_3808_to_fp16 = const()[name = string("op_3808_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_81_cast_fp16 = maximum(x = var_3807_cast_fp16, y = var_3808_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_81_cast_fp16 = real_div(x = linear_45_cast_fp16, y = s_81_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3811_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [1]> var_3816_axes_0 = const()[name = string("op_3816_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3816_keep_dims_0 = const()[name = string("op_3816_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3816_cast_fp16 = reduce_mean(axes = var_3816_axes_0, keep_dims = var_3816_keep_dims_0, x = var_3811_cast_fp16)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3817_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3817_cast_fp16")];
            fp32 var_3818_epsilon_0 = const()[name = string("op_3818_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3818_cast_fp16 = inverse(epsilon = var_3818_epsilon_0, x = var_3817_cast_fp16)[name = string("op_3818_cast_fp16")];
            fp16 var_3819_to_fp16 = const()[name = string("op_3819_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3820_cast_fp16 = mul(x = var_3818_cast_fp16, y = var_3819_to_fp16)[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3822_cast_fp16 = add(x = var_3816_cast_fp16, y = var_3820_cast_fp16)[name = string("op_3822_cast_fp16")];
            fp32 var_3823_epsilon_0 = const()[name = string("op_3823_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3823_cast_fp16 = rsqrt(epsilon = var_3823_epsilon_0, x = var_3822_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_71_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3823_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32683776)))];
            tensor<fp16, [1, 256, 512]> var_3825_cast_fp16 = mul(x = out_71_cast_fp16, y = tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3825_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_9_cast_fp16 = add(x = input_57_cast_fp16, y = var_3825_cast_fp16)[name = string("h_9_cast_fp16")];
            tensor<fp16, [1]> tm_layers_4_layer_scalar_to_fp16 = const()[name = string("tm_layers_4_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.62p-1])];
            tensor<fp16, [1, 256, 512]> x_91_cast_fp16 = mul(x = h_9_cast_fp16, y = tm_layers_4_layer_scalar_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3829_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3829_cast_fp16")];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<int32, [1]> var_3833_axes_0 = const()[name = string("op_3833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3833_cast_fp16 = reduce_max(axes = var_3833_axes_0, keep_dims = var_3832, x = var_3829_cast_fp16)[name = string("op_3833_cast_fp16")];
            fp16 var_3834_to_fp16 = const()[name = string("op_3834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_83_cast_fp16 = maximum(x = var_3833_cast_fp16, y = var_3834_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_83_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_83_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3837_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3837_cast_fp16")];
            tensor<int32, [1]> var_3842_axes_0 = const()[name = string("op_3842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3842_keep_dims_0 = const()[name = string("op_3842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3842_cast_fp16 = reduce_mean(axes = var_3842_axes_0, keep_dims = var_3842_keep_dims_0, x = var_3837_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3843_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp32 var_3844_epsilon_0 = const()[name = string("op_3844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3844_cast_fp16 = inverse(epsilon = var_3844_epsilon_0, x = var_3843_cast_fp16)[name = string("op_3844_cast_fp16")];
            fp16 var_3845_to_fp16 = const()[name = string("op_3845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3846_cast_fp16 = mul(x = var_3844_cast_fp16, y = var_3845_to_fp16)[name = string("op_3846_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3848_cast_fp16 = add(x = var_3842_cast_fp16, y = var_3846_cast_fp16)[name = string("op_3848_cast_fp16")];
            fp32 var_3849_epsilon_0 = const()[name = string("op_3849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3849_cast_fp16 = rsqrt(epsilon = var_3849_epsilon_0, x = var_3848_cast_fp16)[name = string("op_3849_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_73_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3849_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32684864)))];
            tensor<fp16, [1, 256, 512]> input_61_cast_fp16 = mul(x = out_73_cast_fp16, y = tm_layers_5_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32685952))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33734592))))[name = string("tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_46_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [4]> var_3859 = const()[name = string("op_3859"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 256, 4, 512]> var_3860_cast_fp16 = reshape(shape = var_3859, x = linear_46_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<int32, [4]> x_93_perm_0 = const()[name = string("x_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33738752))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34000960))))[name = string("tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<int32, [4]> var_3871 = const()[name = string("op_3871"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_3872_cast_fp16 = reshape(shape = var_3871, x = linear_47_cast_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [4]> x_95_perm_0 = const()[name = string("x_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34002048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34264256))))[name = string("tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [4]> var_3883 = const()[name = string("op_3883"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_3884_cast_fp16 = reshape(shape = var_3883, x = linear_48_cast_fp16)[name = string("op_3884_cast_fp16")];
            tensor<int32, [4]> x_97_perm_0 = const()[name = string("x_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 512]> x_93_cast_fp16 = transpose(perm = x_93_perm_0, x = var_3860_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 4, 256, 512]> var_3888_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3888_cast_fp16")];
            bool var_3891 = const()[name = string("op_3891"), val = bool(true)];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_3892_cast_fp16 = reduce_max(axes = var_3892_axes_0, keep_dims = var_3891, x = var_3888_cast_fp16)[name = string("op_3892_cast_fp16")];
            fp16 var_3893_to_fp16 = const()[name = string("op_3893_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_85_cast_fp16 = maximum(x = var_3892_cast_fp16, y = var_3893_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> y_85_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_85_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_3896_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3896_cast_fp16")];
            tensor<int32, [1]> var_3901_axes_0 = const()[name = string("op_3901_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3901_keep_dims_0 = const()[name = string("op_3901_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_3901_cast_fp16 = reduce_mean(axes = var_3901_axes_0, keep_dims = var_3901_keep_dims_0, x = var_3896_cast_fp16)[name = string("op_3901_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3902_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3902_cast_fp16")];
            fp32 var_3903_epsilon_0 = const()[name = string("op_3903_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_3903_cast_fp16 = inverse(epsilon = var_3903_epsilon_0, x = var_3902_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_3905_cast_fp16 = mul(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3907_cast_fp16 = add(x = var_3901_cast_fp16, y = var_3905_cast_fp16)[name = string("op_3907_cast_fp16")];
            fp32 var_3908_epsilon_0 = const()[name = string("op_3908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_3908_cast_fp16 = rsqrt(epsilon = var_3908_epsilon_0, x = var_3907_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> out_75_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3908_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34265344)))];
            tensor<fp16, [1, 4, 256, 512]> q_21_cast_fp16 = mul(x = out_75_cast_fp16, y = tm_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_95_cast_fp16 = transpose(perm = x_95_perm_0, x = var_3872_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1, 256, 512]> var_3911_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_87_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_87_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_87_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_3919_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_3925_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> out_77_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3931_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34266432)))];
            tensor<fp16, [1, 1, 256, 512]> k_31_cast_fp16 = mul(x = out_77_cast_fp16, y = tm_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = var_3884_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 1, 256, 512]> var_3934_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3934_cast_fp16")];
            bool var_3937 = const()[name = string("op_3937"), val = bool(true)];
            tensor<int32, [1]> var_3938_axes_0 = const()[name = string("op_3938_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_3938_cast_fp16 = reduce_max(axes = var_3938_axes_0, keep_dims = var_3937, x = var_3934_cast_fp16)[name = string("op_3938_cast_fp16")];
            fp16 var_3939_to_fp16 = const()[name = string("op_3939_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_89_cast_fp16 = maximum(x = var_3938_cast_fp16, y = var_3939_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_89_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_89_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_3942_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3942_cast_fp16")];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3947_keep_dims_0 = const()[name = string("op_3947_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_3947_cast_fp16 = reduce_mean(axes = var_3947_axes_0, keep_dims = var_3947_keep_dims_0, x = var_3942_cast_fp16)[name = string("op_3947_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_3948_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3948_cast_fp16")];
            fp32 var_3949_epsilon_0 = const()[name = string("op_3949_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_3949_cast_fp16 = inverse(epsilon = var_3949_epsilon_0, x = var_3948_cast_fp16)[name = string("op_3949_cast_fp16")];
            fp16 var_3950_to_fp16 = const()[name = string("op_3950_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_3951_cast_fp16 = mul(x = var_3949_cast_fp16, y = var_3950_to_fp16)[name = string("op_3951_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_3953_cast_fp16 = add(x = var_3947_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3953_cast_fp16")];
            fp32 var_3954_epsilon_0 = const()[name = string("op_3954_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_3954_cast_fp16 = rsqrt(epsilon = var_3954_epsilon_0, x = var_3953_cast_fp16)[name = string("op_3954_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> v_21_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3954_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<fp16, [256, 512]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137708224))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137641472))))[name = string("cos_to_fp16_quantized")];
            tensor<fp16, [1, 4, 256, 512]> var_3956_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3956_cast_fp16")];
            tensor<int32, [4]> var_3961_begin_0 = const()[name = string("op_3961_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3961_end_0 = const()[name = string("op_3961_end_0"), val = tensor<int32, [4]>([1, 4, 256, 512])];
            tensor<bool, [4]> var_3961_end_mask_0 = const()[name = string("op_3961_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 256]> var_3961_cast_fp16 = slice_by_index(begin = var_3961_begin_0, end = var_3961_end_0, end_mask = var_3961_end_mask_0, x = q_21_cast_fp16)[name = string("op_3961_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 256]> var_3962_cast_fp16 = mul(x = var_3961_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_3962_cast_fp16")];
            tensor<int32, [4]> var_3967_begin_0 = const()[name = string("op_3967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3967_end_0 = const()[name = string("op_3967_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_3967_end_mask_0 = const()[name = string("op_3967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 256]> var_3967_cast_fp16 = slice_by_index(begin = var_3967_begin_0, end = var_3967_end_0, end_mask = var_3967_end_mask_0, x = q_21_cast_fp16)[name = string("op_3967_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(-1)];
            bool var_3970_interleave_0 = const()[name = string("op_3970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> var_3970_cast_fp16 = concat(axis = var_3969, interleave = var_3970_interleave_0, values = (var_3962_cast_fp16, var_3967_cast_fp16))[name = string("op_3970_cast_fp16")];
            tensor<fp16, [256, 512]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137839360))), scale = tensor<fp16, [256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137970496))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [1, 4, 256, 512]> var_3971_cast_fp16 = mul(x = var_3970_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3971_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> q_23_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3971_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_3974_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3974_cast_fp16")];
            tensor<int32, [4]> var_3979_begin_0 = const()[name = string("op_3979_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3979_end_0 = const()[name = string("op_3979_end_0"), val = tensor<int32, [4]>([1, 1, 256, 512])];
            tensor<bool, [4]> var_3979_end_mask_0 = const()[name = string("op_3979_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 256, 256]> var_3979_cast_fp16 = slice_by_index(begin = var_3979_begin_0, end = var_3979_end_0, end_mask = var_3979_end_mask_0, x = k_31_cast_fp16)[name = string("op_3979_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 256, 256]> var_3980_cast_fp16 = mul(x = var_3979_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_3980_cast_fp16")];
            tensor<int32, [4]> var_3985_begin_0 = const()[name = string("op_3985_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3985_end_0 = const()[name = string("op_3985_end_0"), val = tensor<int32, [4]>([1, 1, 256, 256])];
            tensor<bool, [4]> var_3985_end_mask_0 = const()[name = string("op_3985_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 256, 256]> var_3985_cast_fp16 = slice_by_index(begin = var_3985_begin_0, end = var_3985_end_0, end_mask = var_3985_end_mask_0, x = k_31_cast_fp16)[name = string("op_3985_cast_fp16")];
            int32 var_3987 = const()[name = string("op_3987"), val = int32(-1)];
            bool var_3988_interleave_0 = const()[name = string("op_3988_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 256, 512]> var_3988_cast_fp16 = concat(axis = var_3987, interleave = var_3988_interleave_0, values = (var_3980_cast_fp16, var_3985_cast_fp16))[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_3989_cast_fp16 = mul(x = var_3988_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> k_33_cast_fp16 = add(x = var_3974_cast_fp16, y = var_3989_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = k_33_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [4, 1, 256, 512]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_20_cast_fp16 = reshape(shape = concat_20, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [4, 1, 256, 512]> reshape_21_cast_fp16 = reshape(shape = concat_21, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = v_21_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [4, 1, 256, 512]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_22_cast_fp16 = reshape(shape = concat_22, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [4, 1, 256, 512]> reshape_23_cast_fp16 = reshape(shape = concat_23, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4003_transpose_x_0 = const()[name = string("op_4003_transpose_x_0"), val = bool(false)];
            bool var_4003_transpose_y_0 = const()[name = string("op_4003_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 4, 256, 256]> var_4003_cast_fp16 = matmul(transpose_x = var_4003_transpose_x_0, transpose_y = var_4003_transpose_y_0, x = q_23_cast_fp16, y = transpose_101_cast_fp16)[name = string("op_4003_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4005_cast_fp16 = add(x = var_4003_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4005_cast_fp16")];
            int32 var_4006 = const()[name = string("op_4006"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_89_cast_fp16 = softmax(axis = var_4006, x = var_4005_cast_fp16)[name = string("w_89_cast_fp16")];
            bool var_4009_transpose_x_0 = const()[name = string("op_4009_transpose_x_0"), val = bool(false)];
            bool var_4009_transpose_y_0 = const()[name = string("op_4009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 4, 256, 512]> var_4009_cast_fp16 = matmul(transpose_x = var_4009_transpose_x_0, transpose_y = var_4009_transpose_y_0, x = w_89_cast_fp16, y = v_23_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4012_perm_0 = const()[name = string("op_4012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4016 = const()[name = string("op_4016"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 256, 4, 512]> var_4012_cast_fp16 = transpose(perm = var_4012_perm_0, x = var_4009_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 256, 2048]> input_63_cast_fp16 = reshape(shape = var_4016, x = var_4012_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34267520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35316160))))[name = string("tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4021_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_4021_cast_fp16")];
            bool var_4024 = const()[name = string("op_4024"), val = bool(true)];
            tensor<int32, [1]> var_4025_axes_0 = const()[name = string("op_4025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4025_cast_fp16 = reduce_max(axes = var_4025_axes_0, keep_dims = var_4024, x = var_4021_cast_fp16)[name = string("op_4025_cast_fp16")];
            fp16 var_4026_to_fp16 = const()[name = string("op_4026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_91_cast_fp16 = maximum(x = var_4025_cast_fp16, y = var_4026_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_91_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_91_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4029_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<int32, [1]> var_4034_axes_0 = const()[name = string("op_4034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4034_keep_dims_0 = const()[name = string("op_4034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4034_cast_fp16 = reduce_mean(axes = var_4034_axes_0, keep_dims = var_4034_keep_dims_0, x = var_4029_cast_fp16)[name = string("op_4034_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4035_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_4035_cast_fp16")];
            fp32 var_4036_epsilon_0 = const()[name = string("op_4036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4036_cast_fp16 = inverse(epsilon = var_4036_epsilon_0, x = var_4035_cast_fp16)[name = string("op_4036_cast_fp16")];
            fp16 var_4037_to_fp16 = const()[name = string("op_4037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4038_cast_fp16 = mul(x = var_4036_cast_fp16, y = var_4037_to_fp16)[name = string("op_4038_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4040_cast_fp16 = add(x = var_4034_cast_fp16, y = var_4038_cast_fp16)[name = string("op_4040_cast_fp16")];
            fp32 var_4041_epsilon_0 = const()[name = string("op_4041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4041_cast_fp16 = rsqrt(epsilon = var_4041_epsilon_0, x = var_4040_cast_fp16)[name = string("op_4041_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_79_cast_fp16 = mul(x = y_91_cast_fp16, y = var_4041_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35317248)))];
            tensor<fp16, [1, 256, 512]> var_4043_cast_fp16 = mul(x = out_79_cast_fp16, y = tm_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_4043_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_101_cast_fp16 = add(x = x_91_cast_fp16, y = var_4043_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4046_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_4046_cast_fp16")];
            bool var_4049 = const()[name = string("op_4049"), val = bool(true)];
            tensor<int32, [1]> var_4050_axes_0 = const()[name = string("op_4050_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4050_cast_fp16 = reduce_max(axes = var_4050_axes_0, keep_dims = var_4049, x = var_4046_cast_fp16)[name = string("op_4050_cast_fp16")];
            fp16 var_4051_to_fp16 = const()[name = string("op_4051_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_93_cast_fp16 = maximum(x = var_4050_cast_fp16, y = var_4051_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_93_cast_fp16 = real_div(x = x_101_cast_fp16, y = s_93_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4054_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [1]> var_4059_axes_0 = const()[name = string("op_4059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4059_keep_dims_0 = const()[name = string("op_4059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4059_cast_fp16 = reduce_mean(axes = var_4059_axes_0, keep_dims = var_4059_keep_dims_0, x = var_4054_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4060_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_4060_cast_fp16")];
            fp32 var_4061_epsilon_0 = const()[name = string("op_4061_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4061_cast_fp16 = inverse(epsilon = var_4061_epsilon_0, x = var_4060_cast_fp16)[name = string("op_4061_cast_fp16")];
            fp16 var_4062_to_fp16 = const()[name = string("op_4062_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4063_cast_fp16 = mul(x = var_4061_cast_fp16, y = var_4062_to_fp16)[name = string("op_4063_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4065_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4063_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp32 var_4066_epsilon_0 = const()[name = string("op_4066_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4066_cast_fp16 = rsqrt(epsilon = var_4066_epsilon_0, x = var_4065_cast_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_81_cast_fp16 = mul(x = y_93_cast_fp16, y = var_4066_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35318336)))];
            tensor<fp16, [1, 256, 512]> input_65_cast_fp16 = mul(x = out_81_cast_fp16, y = tm_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35319424))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36368064))))[name = string("tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_50_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_50_cast_fp16")];
            string var_4073_mode_0 = const()[name = string("op_4073_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_4073_cast_fp16 = gelu(mode = var_4073_mode_0, x = linear_50_cast_fp16)[name = string("op_4073_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36372224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37420864))))[name = string("tm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_51_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_67_cast_fp16 = mul(x = var_4073_cast_fp16, y = linear_51_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37425024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38473664))))[name = string("tm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4081_cast_fp16 = abs(x = linear_52_cast_fp16)[name = string("op_4081_cast_fp16")];
            bool var_4084 = const()[name = string("op_4084"), val = bool(true)];
            tensor<int32, [1]> var_4085_axes_0 = const()[name = string("op_4085_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4085_cast_fp16 = reduce_max(axes = var_4085_axes_0, keep_dims = var_4084, x = var_4081_cast_fp16)[name = string("op_4085_cast_fp16")];
            fp16 var_4086_to_fp16 = const()[name = string("op_4086_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_95_cast_fp16 = maximum(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_95_cast_fp16 = real_div(x = linear_52_cast_fp16, y = s_95_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4089_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_4089_cast_fp16")];
            tensor<int32, [1]> var_4094_axes_0 = const()[name = string("op_4094_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4094_keep_dims_0 = const()[name = string("op_4094_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4094_cast_fp16 = reduce_mean(axes = var_4094_axes_0, keep_dims = var_4094_keep_dims_0, x = var_4089_cast_fp16)[name = string("op_4094_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4095_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_4095_cast_fp16")];
            fp32 var_4096_epsilon_0 = const()[name = string("op_4096_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4096_cast_fp16 = inverse(epsilon = var_4096_epsilon_0, x = var_4095_cast_fp16)[name = string("op_4096_cast_fp16")];
            fp16 var_4097_to_fp16 = const()[name = string("op_4097_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4098_cast_fp16 = mul(x = var_4096_cast_fp16, y = var_4097_to_fp16)[name = string("op_4098_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4100_cast_fp16 = add(x = var_4094_cast_fp16, y = var_4098_cast_fp16)[name = string("op_4100_cast_fp16")];
            fp32 var_4101_epsilon_0 = const()[name = string("op_4101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4101_cast_fp16 = rsqrt(epsilon = var_4101_epsilon_0, x = var_4100_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_83_cast_fp16 = mul(x = y_95_cast_fp16, y = var_4101_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38474752)))];
            tensor<fp16, [1, 256, 512]> var_4103_cast_fp16 = mul(x = out_83_cast_fp16, y = tm_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_69_cast_fp16 = add(x = x_101_cast_fp16, y = var_4103_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38475840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38738048))))[name = string("tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_53_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_53_cast_fp16")];
            string x_105_mode_0 = const()[name = string("x_105_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_105_cast_fp16 = gelu(mode = x_105_mode_0, x = linear_53_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [4]> var_4123_begin_0 = const()[name = string("op_4123_begin_0"), val = tensor<int32, [4]>([0, 0, 5, 0])];
            tensor<int32, [4]> var_4123_end_0 = const()[name = string("op_4123_end_0"), val = tensor<int32, [4]>([1, 256, 6, 512])];
            tensor<bool, [4]> var_4123_end_mask_0 = const()[name = string("op_4123_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4123_squeeze_mask_0 = const()[name = string("op_4123_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_4123_cast_fp16 = slice_by_index(begin = var_4123_begin_0, end = var_4123_end_0, end_mask = var_4123_end_mask_0, squeeze_mask = var_4123_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_71_cast_fp16 = mul(x = x_105_cast_fp16, y = var_4123_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38739136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39001344))))[name = string("tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4133_cast_fp16 = abs(x = linear_54_cast_fp16)[name = string("op_4133_cast_fp16")];
            bool var_4136 = const()[name = string("op_4136"), val = bool(true)];
            tensor<int32, [1]> var_4137_axes_0 = const()[name = string("op_4137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4137_cast_fp16 = reduce_max(axes = var_4137_axes_0, keep_dims = var_4136, x = var_4133_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 var_4138_to_fp16 = const()[name = string("op_4138_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_97_cast_fp16 = maximum(x = var_4137_cast_fp16, y = var_4138_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_97_cast_fp16 = real_div(x = linear_54_cast_fp16, y = s_97_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4141_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_4141_cast_fp16")];
            tensor<int32, [1]> var_4146_axes_0 = const()[name = string("op_4146_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4146_keep_dims_0 = const()[name = string("op_4146_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4146_cast_fp16 = reduce_mean(axes = var_4146_axes_0, keep_dims = var_4146_keep_dims_0, x = var_4141_cast_fp16)[name = string("op_4146_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4147_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_4147_cast_fp16")];
            fp32 var_4148_epsilon_0 = const()[name = string("op_4148_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4148_cast_fp16 = inverse(epsilon = var_4148_epsilon_0, x = var_4147_cast_fp16)[name = string("op_4148_cast_fp16")];
            fp16 var_4149_to_fp16 = const()[name = string("op_4149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4150_cast_fp16 = mul(x = var_4148_cast_fp16, y = var_4149_to_fp16)[name = string("op_4150_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4152_cast_fp16 = add(x = var_4146_cast_fp16, y = var_4150_cast_fp16)[name = string("op_4152_cast_fp16")];
            fp32 var_4153_epsilon_0 = const()[name = string("op_4153_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4153_cast_fp16 = rsqrt(epsilon = var_4153_epsilon_0, x = var_4152_cast_fp16)[name = string("op_4153_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_85_cast_fp16 = mul(x = y_97_cast_fp16, y = var_4153_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39002432)))];
            tensor<fp16, [1, 256, 512]> var_4155_cast_fp16 = mul(x = out_85_cast_fp16, y = tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4155_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_11_cast_fp16 = add(x = input_69_cast_fp16, y = var_4155_cast_fp16)[name = string("h_11_cast_fp16")];
            tensor<fp16, [1]> tm_layers_5_layer_scalar_to_fp16 = const()[name = string("tm_layers_5_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-2])];
            tensor<fp16, [1, 256, 512]> x_109_cast_fp16 = mul(x = h_11_cast_fp16, y = tm_layers_5_layer_scalar_to_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4159_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_4159_cast_fp16")];
            bool var_4162 = const()[name = string("op_4162"), val = bool(true)];
            tensor<int32, [1]> var_4163_axes_0 = const()[name = string("op_4163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4163_cast_fp16 = reduce_max(axes = var_4163_axes_0, keep_dims = var_4162, x = var_4159_cast_fp16)[name = string("op_4163_cast_fp16")];
            fp16 var_4164_to_fp16 = const()[name = string("op_4164_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_99_cast_fp16 = maximum(x = var_4163_cast_fp16, y = var_4164_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_99_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_99_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4167_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_4167_cast_fp16")];
            tensor<int32, [1]> var_4172_axes_0 = const()[name = string("op_4172_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4172_keep_dims_0 = const()[name = string("op_4172_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4172_cast_fp16 = reduce_mean(axes = var_4172_axes_0, keep_dims = var_4172_keep_dims_0, x = var_4167_cast_fp16)[name = string("op_4172_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4173_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4174_cast_fp16 = inverse(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            fp16 var_4175_to_fp16 = const()[name = string("op_4175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4176_cast_fp16 = mul(x = var_4174_cast_fp16, y = var_4175_to_fp16)[name = string("op_4176_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4178_cast_fp16 = add(x = var_4172_cast_fp16, y = var_4176_cast_fp16)[name = string("op_4178_cast_fp16")];
            fp32 var_4179_epsilon_0 = const()[name = string("op_4179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4179_cast_fp16 = rsqrt(epsilon = var_4179_epsilon_0, x = var_4178_cast_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_87_cast_fp16 = mul(x = y_99_cast_fp16, y = var_4179_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39003520)))];
            tensor<fp16, [1, 256, 512]> input_73_cast_fp16 = mul(x = out_87_cast_fp16, y = tm_layers_6_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39004608))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39528960))))[name = string("tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [4]> var_4189 = const()[name = string("op_4189"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_4190_cast_fp16 = reshape(shape = var_4189, x = linear_55_cast_fp16)[name = string("op_4190_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39531072))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39793280))))[name = string("tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_4201 = const()[name = string("op_4201"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4202_cast_fp16 = reshape(shape = var_4201, x = linear_56_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39794368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40056576))))[name = string("tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_57_cast_fp16)[name = string("op_4214_cast_fp16")];
            tensor<int32, [4]> x_115_perm_0 = const()[name = string("x_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4190_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 4, 256, 256]> var_4218_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_4218_cast_fp16")];
            bool var_4221 = const()[name = string("op_4221"), val = bool(true)];
            tensor<int32, [1]> var_4222_axes_0 = const()[name = string("op_4222_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_4222_cast_fp16 = reduce_max(axes = var_4222_axes_0, keep_dims = var_4221, x = var_4218_cast_fp16)[name = string("op_4222_cast_fp16")];
            fp16 var_4223_to_fp16 = const()[name = string("op_4223_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_101_cast_fp16 = maximum(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_101_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_101_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4226_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [1]> var_4231_axes_0 = const()[name = string("op_4231_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4231_keep_dims_0 = const()[name = string("op_4231_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_4231_cast_fp16 = reduce_mean(axes = var_4231_axes_0, keep_dims = var_4231_keep_dims_0, x = var_4226_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4232_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_4232_cast_fp16")];
            fp32 var_4233_epsilon_0 = const()[name = string("op_4233_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_4233_cast_fp16 = inverse(epsilon = var_4233_epsilon_0, x = var_4232_cast_fp16)[name = string("op_4233_cast_fp16")];
            fp16 var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_4235_cast_fp16 = mul(x = var_4233_cast_fp16, y = var_4234_to_fp16)[name = string("op_4235_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4237_cast_fp16 = add(x = var_4231_cast_fp16, y = var_4235_cast_fp16)[name = string("op_4237_cast_fp16")];
            fp32 var_4238_epsilon_0 = const()[name = string("op_4238_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_4238_cast_fp16 = rsqrt(epsilon = var_4238_epsilon_0, x = var_4237_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_89_cast_fp16 = mul(x = y_101_cast_fp16, y = var_4238_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40057664)))];
            tensor<fp16, [1, 4, 256, 256]> q_25_cast_fp16 = mul(x = out_89_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4202_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 2, 256, 256]> var_4241_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_4241_cast_fp16")];
            bool var_4244 = const()[name = string("op_4244"), val = bool(true)];
            tensor<int32, [1]> var_4245_axes_0 = const()[name = string("op_4245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4245_cast_fp16 = reduce_max(axes = var_4245_axes_0, keep_dims = var_4244, x = var_4241_cast_fp16)[name = string("op_4245_cast_fp16")];
            fp16 var_4246_to_fp16 = const()[name = string("op_4246_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_103_cast_fp16 = maximum(x = var_4245_cast_fp16, y = var_4246_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_103_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_103_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4249_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<int32, [1]> var_4254_axes_0 = const()[name = string("op_4254_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4254_keep_dims_0 = const()[name = string("op_4254_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4254_cast_fp16 = reduce_mean(axes = var_4254_axes_0, keep_dims = var_4254_keep_dims_0, x = var_4249_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4255_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_4255_cast_fp16")];
            fp32 var_4256_epsilon_0 = const()[name = string("op_4256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4256_cast_fp16 = inverse(epsilon = var_4256_epsilon_0, x = var_4255_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp16 var_4257_to_fp16 = const()[name = string("op_4257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4258_cast_fp16 = mul(x = var_4256_cast_fp16, y = var_4257_to_fp16)[name = string("op_4258_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4260_cast_fp16 = add(x = var_4254_cast_fp16, y = var_4258_cast_fp16)[name = string("op_4260_cast_fp16")];
            fp32 var_4261_epsilon_0 = const()[name = string("op_4261_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4261_cast_fp16 = rsqrt(epsilon = var_4261_epsilon_0, x = var_4260_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_91_cast_fp16 = mul(x = y_103_cast_fp16, y = var_4261_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058240)))];
            tensor<fp16, [1, 2, 256, 256]> k_37_cast_fp16 = mul(x = out_91_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_37_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_115_cast_fp16 = transpose(perm = x_115_perm_0, x = var_4214_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 2, 256, 256]> var_4264_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_4264_cast_fp16")];
            bool var_4267 = const()[name = string("op_4267"), val = bool(true)];
            tensor<int32, [1]> var_4268_axes_0 = const()[name = string("op_4268_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4268_cast_fp16 = reduce_max(axes = var_4268_axes_0, keep_dims = var_4267, x = var_4264_cast_fp16)[name = string("op_4268_cast_fp16")];
            fp16 var_4269_to_fp16 = const()[name = string("op_4269_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_105_cast_fp16 = maximum(x = var_4268_cast_fp16, y = var_4269_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_105_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_105_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4272_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_4272_cast_fp16")];
            tensor<int32, [1]> var_4277_axes_0 = const()[name = string("op_4277_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4277_keep_dims_0 = const()[name = string("op_4277_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4277_cast_fp16 = reduce_mean(axes = var_4277_axes_0, keep_dims = var_4277_keep_dims_0, x = var_4272_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4278_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp32 var_4279_epsilon_0 = const()[name = string("op_4279_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4279_cast_fp16 = inverse(epsilon = var_4279_epsilon_0, x = var_4278_cast_fp16)[name = string("op_4279_cast_fp16")];
            fp16 var_4280_to_fp16 = const()[name = string("op_4280_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4281_cast_fp16 = mul(x = var_4279_cast_fp16, y = var_4280_to_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4283_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4281_cast_fp16)[name = string("op_4283_cast_fp16")];
            fp32 var_4284_epsilon_0 = const()[name = string("op_4284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4284_cast_fp16 = rsqrt(epsilon = var_4284_epsilon_0, x = var_4283_cast_fp16)[name = string("op_4284_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_25_cast_fp16 = mul(x = y_105_cast_fp16, y = var_4284_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4286_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4286_cast_fp16")];
            tensor<int32, [4]> var_4291_begin_0 = const()[name = string("op_4291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4291_end_0 = const()[name = string("op_4291_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_4291_end_mask_0 = const()[name = string("op_4291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_4291_cast_fp16 = slice_by_index(begin = var_4291_begin_0, end = var_4291_end_0, end_mask = var_4291_end_mask_0, x = q_25_cast_fp16)[name = string("op_4291_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_4292_cast_fp16 = mul(x = var_4291_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = q_25_cast_fp16)[name = string("op_4297_cast_fp16")];
            int32 var_4299 = const()[name = string("op_4299"), val = int32(-1)];
            bool var_4300_interleave_0 = const()[name = string("op_4300_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_4300_cast_fp16 = concat(axis = var_4299, interleave = var_4300_interleave_0, values = (var_4292_cast_fp16, var_4297_cast_fp16))[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4301_cast_fp16 = mul(x = var_4300_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4301_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_27_cast_fp16 = add(x = var_4286_cast_fp16, y = var_4301_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4304_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4304_cast_fp16")];
            tensor<int32, [4]> var_4309_begin_0 = const()[name = string("op_4309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4309_end_0 = const()[name = string("op_4309_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_4309_end_mask_0 = const()[name = string("op_4309_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_4309_cast_fp16 = slice_by_index(begin = var_4309_begin_0, end = var_4309_end_0, end_mask = var_4309_end_mask_0, x = k_37_cast_fp16)[name = string("op_4309_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_4310_cast_fp16 = mul(x = var_4309_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_4310_cast_fp16")];
            tensor<int32, [4]> var_4315_begin_0 = const()[name = string("op_4315_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4315_end_0 = const()[name = string("op_4315_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_4315_end_mask_0 = const()[name = string("op_4315_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_4315_cast_fp16 = slice_by_index(begin = var_4315_begin_0, end = var_4315_end_0, end_mask = var_4315_end_mask_0, x = k_37_cast_fp16)[name = string("op_4315_cast_fp16")];
            int32 var_4317 = const()[name = string("op_4317"), val = int32(-1)];
            bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_4318_cast_fp16 = concat(axis = var_4317, interleave = var_4318_interleave_0, values = (var_4310_cast_fp16, var_4315_cast_fp16))[name = string("op_4318_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4319_cast_fp16 = mul(x = var_4318_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_39_cast_fp16 = add(x = var_4304_cast_fp16, y = var_4319_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = k_39_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [4, 1, 256, 256]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
            tensor<int32, [5]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_24_cast_fp16 = reshape(shape = concat_24, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [4, 1, 256, 256]> reshape_25_cast_fp16 = reshape(shape = concat_25, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = v_25_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [4, 1, 256, 256]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
            tensor<int32, [5]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_26_cast_fp16 = reshape(shape = concat_26, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [4, 1, 256, 256]> reshape_27_cast_fp16 = reshape(shape = concat_27, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4333_transpose_x_0 = const()[name = string("op_4333_transpose_x_0"), val = bool(false)];
            bool var_4333_transpose_y_0 = const()[name = string("op_4333_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_102_cast_fp16 = transpose(perm = transpose_102_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 4, 256, 256]> var_4333_cast_fp16 = matmul(transpose_x = var_4333_transpose_x_0, transpose_y = var_4333_transpose_y_0, x = q_27_cast_fp16, y = transpose_102_cast_fp16)[name = string("op_4333_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4335_cast_fp16 = add(x = var_4333_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4335_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_105_cast_fp16 = softmax(axis = var_4336, x = var_4335_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4339_transpose_x_0 = const()[name = string("op_4339_transpose_x_0"), val = bool(false)];
            bool var_4339_transpose_y_0 = const()[name = string("op_4339_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 4, 256, 256]> var_4339_cast_fp16 = matmul(transpose_x = var_4339_transpose_x_0, transpose_y = var_4339_transpose_y_0, x = w_105_cast_fp16, y = v_27_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<int32, [4]> var_4342_perm_0 = const()[name = string("op_4342_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4346 = const()[name = string("op_4346"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_4342_cast_fp16 = transpose(perm = var_4342_perm_0, x = var_4339_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 256, 1024]> input_75_cast_fp16 = reshape(shape = var_4346, x = var_4342_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40583168))))[name = string("tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4351_cast_fp16 = abs(x = linear_58_cast_fp16)[name = string("op_4351_cast_fp16")];
            bool var_4354 = const()[name = string("op_4354"), val = bool(true)];
            tensor<int32, [1]> var_4355_axes_0 = const()[name = string("op_4355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4355_cast_fp16 = reduce_max(axes = var_4355_axes_0, keep_dims = var_4354, x = var_4351_cast_fp16)[name = string("op_4355_cast_fp16")];
            fp16 var_4356_to_fp16 = const()[name = string("op_4356_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_107_cast_fp16 = maximum(x = var_4355_cast_fp16, y = var_4356_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_107_cast_fp16 = real_div(x = linear_58_cast_fp16, y = s_107_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4359_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<int32, [1]> var_4364_axes_0 = const()[name = string("op_4364_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4364_keep_dims_0 = const()[name = string("op_4364_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4364_cast_fp16 = reduce_mean(axes = var_4364_axes_0, keep_dims = var_4364_keep_dims_0, x = var_4359_cast_fp16)[name = string("op_4364_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4365_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_4365_cast_fp16")];
            fp32 var_4366_epsilon_0 = const()[name = string("op_4366_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4366_cast_fp16 = inverse(epsilon = var_4366_epsilon_0, x = var_4365_cast_fp16)[name = string("op_4366_cast_fp16")];
            fp16 var_4367_to_fp16 = const()[name = string("op_4367_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4368_cast_fp16 = mul(x = var_4366_cast_fp16, y = var_4367_to_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4370_cast_fp16 = add(x = var_4364_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            fp32 var_4371_epsilon_0 = const()[name = string("op_4371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4371_cast_fp16 = rsqrt(epsilon = var_4371_epsilon_0, x = var_4370_cast_fp16)[name = string("op_4371_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_93_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4371_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40584256)))];
            tensor<fp16, [1, 256, 512]> var_4373_cast_fp16 = mul(x = out_93_cast_fp16, y = tm_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_4373_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = var_4373_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4376_cast_fp16 = abs(x = x_119_cast_fp16)[name = string("op_4376_cast_fp16")];
            bool var_4379 = const()[name = string("op_4379"), val = bool(true)];
            tensor<int32, [1]> var_4380_axes_0 = const()[name = string("op_4380_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4380_cast_fp16 = reduce_max(axes = var_4380_axes_0, keep_dims = var_4379, x = var_4376_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 var_4381_to_fp16 = const()[name = string("op_4381_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_109_cast_fp16 = maximum(x = var_4380_cast_fp16, y = var_4381_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_109_cast_fp16 = real_div(x = x_119_cast_fp16, y = s_109_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4384_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4384_cast_fp16")];
            tensor<int32, [1]> var_4389_axes_0 = const()[name = string("op_4389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4389_keep_dims_0 = const()[name = string("op_4389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4389_cast_fp16 = reduce_mean(axes = var_4389_axes_0, keep_dims = var_4389_keep_dims_0, x = var_4384_cast_fp16)[name = string("op_4389_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4390_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_4390_cast_fp16")];
            fp32 var_4391_epsilon_0 = const()[name = string("op_4391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4391_cast_fp16 = inverse(epsilon = var_4391_epsilon_0, x = var_4390_cast_fp16)[name = string("op_4391_cast_fp16")];
            fp16 var_4392_to_fp16 = const()[name = string("op_4392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = string("op_4393_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4395_cast_fp16 = add(x = var_4389_cast_fp16, y = var_4393_cast_fp16)[name = string("op_4395_cast_fp16")];
            fp32 var_4396_epsilon_0 = const()[name = string("op_4396_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4396_cast_fp16 = rsqrt(epsilon = var_4396_epsilon_0, x = var_4395_cast_fp16)[name = string("op_4396_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_95_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4396_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40585344)))];
            tensor<fp16, [1, 256, 512]> input_77_cast_fp16 = mul(x = out_95_cast_fp16, y = tm_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40586432))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41635072))))[name = string("tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_59_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_59_cast_fp16")];
            string var_4403_mode_0 = const()[name = string("op_4403_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_4403_cast_fp16 = gelu(mode = var_4403_mode_0, x = linear_59_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41639232))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42687872))))[name = string("tm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_60_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_79_cast_fp16 = mul(x = var_4403_cast_fp16, y = linear_60_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42692032))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43740672))))[name = string("tm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4411_cast_fp16 = abs(x = linear_61_cast_fp16)[name = string("op_4411_cast_fp16")];
            bool var_4414 = const()[name = string("op_4414"), val = bool(true)];
            tensor<int32, [1]> var_4415_axes_0 = const()[name = string("op_4415_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4415_cast_fp16 = reduce_max(axes = var_4415_axes_0, keep_dims = var_4414, x = var_4411_cast_fp16)[name = string("op_4415_cast_fp16")];
            fp16 var_4416_to_fp16 = const()[name = string("op_4416_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_111_cast_fp16 = maximum(x = var_4415_cast_fp16, y = var_4416_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_111_cast_fp16 = real_div(x = linear_61_cast_fp16, y = s_111_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4419_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4419_cast_fp16")];
            tensor<int32, [1]> var_4424_axes_0 = const()[name = string("op_4424_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4424_keep_dims_0 = const()[name = string("op_4424_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4424_cast_fp16 = reduce_mean(axes = var_4424_axes_0, keep_dims = var_4424_keep_dims_0, x = var_4419_cast_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4425_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4426_cast_fp16 = inverse(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            fp16 var_4427_to_fp16 = const()[name = string("op_4427_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4428_cast_fp16 = mul(x = var_4426_cast_fp16, y = var_4427_to_fp16)[name = string("op_4428_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4430_cast_fp16 = add(x = var_4424_cast_fp16, y = var_4428_cast_fp16)[name = string("op_4430_cast_fp16")];
            fp32 var_4431_epsilon_0 = const()[name = string("op_4431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4431_cast_fp16 = rsqrt(epsilon = var_4431_epsilon_0, x = var_4430_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_97_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4431_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43741760)))];
            tensor<fp16, [1, 256, 512]> var_4433_cast_fp16 = mul(x = out_97_cast_fp16, y = tm_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4433_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_81_cast_fp16 = add(x = x_119_cast_fp16, y = var_4433_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43742848))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44005056))))[name = string("tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_62_cast_fp16")];
            string x_123_mode_0 = const()[name = string("x_123_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_123_cast_fp16 = gelu(mode = x_123_mode_0, x = linear_62_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [4]> var_4453_begin_0 = const()[name = string("op_4453_begin_0"), val = tensor<int32, [4]>([0, 0, 6, 0])];
            tensor<int32, [4]> var_4453_end_0 = const()[name = string("op_4453_end_0"), val = tensor<int32, [4]>([1, 256, 7, 512])];
            tensor<bool, [4]> var_4453_end_mask_0 = const()[name = string("op_4453_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4453_squeeze_mask_0 = const()[name = string("op_4453_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_4453_cast_fp16 = slice_by_index(begin = var_4453_begin_0, end = var_4453_end_0, end_mask = var_4453_end_mask_0, squeeze_mask = var_4453_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_83_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4453_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44006144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44268352))))[name = string("tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4463_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4463_cast_fp16")];
            bool var_4466 = const()[name = string("op_4466"), val = bool(true)];
            tensor<int32, [1]> var_4467_axes_0 = const()[name = string("op_4467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4467_cast_fp16 = reduce_max(axes = var_4467_axes_0, keep_dims = var_4466, x = var_4463_cast_fp16)[name = string("op_4467_cast_fp16")];
            fp16 var_4468_to_fp16 = const()[name = string("op_4468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_113_cast_fp16 = maximum(x = var_4467_cast_fp16, y = var_4468_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_113_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_113_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4471_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [1]> var_4476_axes_0 = const()[name = string("op_4476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4476_keep_dims_0 = const()[name = string("op_4476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4476_cast_fp16 = reduce_mean(axes = var_4476_axes_0, keep_dims = var_4476_keep_dims_0, x = var_4471_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4477_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_fp16)[name = string("op_4477_cast_fp16")];
            fp32 var_4478_epsilon_0 = const()[name = string("op_4478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4478_cast_fp16 = inverse(epsilon = var_4478_epsilon_0, x = var_4477_cast_fp16)[name = string("op_4478_cast_fp16")];
            fp16 var_4479_to_fp16 = const()[name = string("op_4479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4480_cast_fp16 = mul(x = var_4478_cast_fp16, y = var_4479_to_fp16)[name = string("op_4480_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4482_cast_fp16 = add(x = var_4476_cast_fp16, y = var_4480_cast_fp16)[name = string("op_4482_cast_fp16")];
            fp32 var_4483_epsilon_0 = const()[name = string("op_4483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4483_cast_fp16 = rsqrt(epsilon = var_4483_epsilon_0, x = var_4482_cast_fp16)[name = string("op_4483_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_99_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4483_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44269440)))];
            tensor<fp16, [1, 256, 512]> var_4485_cast_fp16 = mul(x = out_99_cast_fp16, y = tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_13_cast_fp16 = add(x = input_81_cast_fp16, y = var_4485_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<fp16, [1]> tm_layers_6_layer_scalar_to_fp16 = const()[name = string("tm_layers_6_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.cep-3])];
            tensor<fp16, [1, 256, 512]> x_127_cast_fp16 = mul(x = h_13_cast_fp16, y = tm_layers_6_layer_scalar_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4489_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_4489_cast_fp16")];
            bool var_4492 = const()[name = string("op_4492"), val = bool(true)];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4493_cast_fp16 = reduce_max(axes = var_4493_axes_0, keep_dims = var_4492, x = var_4489_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_115_cast_fp16 = maximum(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_115_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_115_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4497_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<int32, [1]> var_4502_axes_0 = const()[name = string("op_4502_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4502_keep_dims_0 = const()[name = string("op_4502_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4502_cast_fp16 = reduce_mean(axes = var_4502_axes_0, keep_dims = var_4502_keep_dims_0, x = var_4497_cast_fp16)[name = string("op_4502_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4503_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4504_cast_fp16 = inverse(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            fp16 var_4505_to_fp16 = const()[name = string("op_4505_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4506_cast_fp16 = mul(x = var_4504_cast_fp16, y = var_4505_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4508_cast_fp16 = add(x = var_4502_cast_fp16, y = var_4506_cast_fp16)[name = string("op_4508_cast_fp16")];
            fp32 var_4509_epsilon_0 = const()[name = string("op_4509_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4509_cast_fp16 = rsqrt(epsilon = var_4509_epsilon_0, x = var_4508_cast_fp16)[name = string("op_4509_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_101_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4509_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44270528)))];
            tensor<fp16, [1, 256, 512]> input_85_cast_fp16 = mul(x = out_101_cast_fp16, y = tm_layers_7_input_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44271616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44795968))))[name = string("tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4519 = const()[name = string("op_4519"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_4520_cast_fp16 = reshape(shape = var_4519, x = linear_64_cast_fp16)[name = string("op_4520_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44798080))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45060288))))[name = string("tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4531 = const()[name = string("op_4531"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4532_cast_fp16 = reshape(shape = var_4531, x = linear_65_cast_fp16)[name = string("op_4532_cast_fp16")];
            tensor<int32, [4]> x_131_perm_0 = const()[name = string("x_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45061376))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45323584))))[name = string("tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4543 = const()[name = string("op_4543"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4544_cast_fp16 = reshape(shape = var_4543, x = linear_66_cast_fp16)[name = string("op_4544_cast_fp16")];
            tensor<int32, [4]> x_133_perm_0 = const()[name = string("x_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_4520_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 4, 256, 256]> var_4548_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_4548_cast_fp16")];
            bool var_4551 = const()[name = string("op_4551"), val = bool(true)];
            tensor<int32, [1]> var_4552_axes_0 = const()[name = string("op_4552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_4552_cast_fp16 = reduce_max(axes = var_4552_axes_0, keep_dims = var_4551, x = var_4548_cast_fp16)[name = string("op_4552_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_117_cast_fp16 = maximum(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_117_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_117_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4556_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<int32, [1]> var_4561_axes_0 = const()[name = string("op_4561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4561_keep_dims_0 = const()[name = string("op_4561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_4561_cast_fp16 = reduce_mean(axes = var_4561_axes_0, keep_dims = var_4561_keep_dims_0, x = var_4556_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4562_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4562_cast_fp16")];
            fp32 var_4563_epsilon_0 = const()[name = string("op_4563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_4563_cast_fp16 = inverse(epsilon = var_4563_epsilon_0, x = var_4562_cast_fp16)[name = string("op_4563_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_4565_cast_fp16 = mul(x = var_4563_cast_fp16, y = var_4564_to_fp16)[name = string("op_4565_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4567_cast_fp16 = add(x = var_4561_cast_fp16, y = var_4565_cast_fp16)[name = string("op_4567_cast_fp16")];
            fp32 var_4568_epsilon_0 = const()[name = string("op_4568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_4568_cast_fp16 = rsqrt(epsilon = var_4568_epsilon_0, x = var_4567_cast_fp16)[name = string("op_4568_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_103_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4568_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_29_cast_fp16 = mul(x = out_103_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_131_cast_fp16 = transpose(perm = x_131_perm_0, x = var_4532_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2, 256, 256]> var_4571_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_4571_cast_fp16")];
            bool var_4574 = const()[name = string("op_4574"), val = bool(true)];
            tensor<int32, [1]> var_4575_axes_0 = const()[name = string("op_4575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4575_cast_fp16 = reduce_max(axes = var_4575_axes_0, keep_dims = var_4574, x = var_4571_cast_fp16)[name = string("op_4575_cast_fp16")];
            fp16 var_4576_to_fp16 = const()[name = string("op_4576_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_119_cast_fp16 = maximum(x = var_4575_cast_fp16, y = var_4576_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_119_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_119_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4579_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4579_cast_fp16")];
            tensor<int32, [1]> var_4584_axes_0 = const()[name = string("op_4584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4584_keep_dims_0 = const()[name = string("op_4584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4584_cast_fp16 = reduce_mean(axes = var_4584_axes_0, keep_dims = var_4584_keep_dims_0, x = var_4579_cast_fp16)[name = string("op_4584_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4585_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4585_cast_fp16")];
            fp32 var_4586_epsilon_0 = const()[name = string("op_4586_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4586_cast_fp16 = inverse(epsilon = var_4586_epsilon_0, x = var_4585_cast_fp16)[name = string("op_4586_cast_fp16")];
            fp16 var_4587_to_fp16 = const()[name = string("op_4587_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4588_cast_fp16 = mul(x = var_4586_cast_fp16, y = var_4587_to_fp16)[name = string("op_4588_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4590_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4588_cast_fp16)[name = string("op_4590_cast_fp16")];
            fp32 var_4591_epsilon_0 = const()[name = string("op_4591_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4591_cast_fp16 = rsqrt(epsilon = var_4591_epsilon_0, x = var_4590_cast_fp16)[name = string("op_4591_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_105_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4591_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [256]> tm_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45324672)))];
            tensor<fp16, [1, 2, 256, 256]> k_43_cast_fp16 = mul(x = out_105_cast_fp16, y = tm_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_133_cast_fp16 = transpose(perm = x_133_perm_0, x = var_4544_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 2, 256, 256]> var_4594_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_4594_cast_fp16")];
            bool var_4597 = const()[name = string("op_4597"), val = bool(true)];
            tensor<int32, [1]> var_4598_axes_0 = const()[name = string("op_4598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4598_cast_fp16 = reduce_max(axes = var_4598_axes_0, keep_dims = var_4597, x = var_4594_cast_fp16)[name = string("op_4598_cast_fp16")];
            fp16 var_4599_to_fp16 = const()[name = string("op_4599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_121_cast_fp16 = maximum(x = var_4598_cast_fp16, y = var_4599_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_121_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_121_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4602_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4602_cast_fp16")];
            tensor<int32, [1]> var_4607_axes_0 = const()[name = string("op_4607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4607_keep_dims_0 = const()[name = string("op_4607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4607_cast_fp16 = reduce_mean(axes = var_4607_axes_0, keep_dims = var_4607_keep_dims_0, x = var_4602_cast_fp16)[name = string("op_4607_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4608_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4608_cast_fp16")];
            fp32 var_4609_epsilon_0 = const()[name = string("op_4609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4609_cast_fp16 = inverse(epsilon = var_4609_epsilon_0, x = var_4608_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4611_cast_fp16 = mul(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("op_4611_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4613_cast_fp16 = add(x = var_4607_cast_fp16, y = var_4611_cast_fp16)[name = string("op_4613_cast_fp16")];
            fp32 var_4614_epsilon_0 = const()[name = string("op_4614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4614_cast_fp16 = rsqrt(epsilon = var_4614_epsilon_0, x = var_4613_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_29_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4614_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4616_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4616_cast_fp16")];
            tensor<int32, [4]> var_4621_begin_0 = const()[name = string("op_4621_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4621_end_0 = const()[name = string("op_4621_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_4621_end_mask_0 = const()[name = string("op_4621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_4621_cast_fp16 = slice_by_index(begin = var_4621_begin_0, end = var_4621_end_0, end_mask = var_4621_end_mask_0, x = q_29_cast_fp16)[name = string("op_4621_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_4622_cast_fp16 = mul(x = var_4621_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<int32, [4]> var_4627_begin_0 = const()[name = string("op_4627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4627_end_0 = const()[name = string("op_4627_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_4627_end_mask_0 = const()[name = string("op_4627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_4627_cast_fp16 = slice_by_index(begin = var_4627_begin_0, end = var_4627_end_0, end_mask = var_4627_end_mask_0, x = q_29_cast_fp16)[name = string("op_4627_cast_fp16")];
            int32 var_4629 = const()[name = string("op_4629"), val = int32(-1)];
            bool var_4630_interleave_0 = const()[name = string("op_4630_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_4630_cast_fp16 = concat(axis = var_4629, interleave = var_4630_interleave_0, values = (var_4622_cast_fp16, var_4627_cast_fp16))[name = string("op_4630_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4631_cast_fp16 = mul(x = var_4630_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4631_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_31_cast_fp16 = add(x = var_4616_cast_fp16, y = var_4631_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4634_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4634_cast_fp16")];
            tensor<int32, [4]> var_4639_begin_0 = const()[name = string("op_4639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4639_end_0 = const()[name = string("op_4639_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_4639_end_mask_0 = const()[name = string("op_4639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_4639_cast_fp16 = slice_by_index(begin = var_4639_begin_0, end = var_4639_end_0, end_mask = var_4639_end_mask_0, x = k_43_cast_fp16)[name = string("op_4639_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_4640_cast_fp16 = mul(x = var_4639_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4640_cast_fp16")];
            tensor<int32, [4]> var_4645_begin_0 = const()[name = string("op_4645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4645_end_0 = const()[name = string("op_4645_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_4645_end_mask_0 = const()[name = string("op_4645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_4645_cast_fp16 = slice_by_index(begin = var_4645_begin_0, end = var_4645_end_0, end_mask = var_4645_end_mask_0, x = k_43_cast_fp16)[name = string("op_4645_cast_fp16")];
            int32 var_4647 = const()[name = string("op_4647"), val = int32(-1)];
            bool var_4648_interleave_0 = const()[name = string("op_4648_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_4648_cast_fp16 = concat(axis = var_4647, interleave = var_4648_interleave_0, values = (var_4640_cast_fp16, var_4645_cast_fp16))[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4649_cast_fp16 = mul(x = var_4648_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4649_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_45_cast_fp16 = add(x = var_4634_cast_fp16, y = var_4649_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_14_reps_0 = const()[name = string("tile_14_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = k_45_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [4, 1, 256, 256]> tile_14_cast_fp16 = tile(reps = tile_14_reps_0, x = transpose_28_cast_fp16)[name = string("tile_14_cast_fp16")];
            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_28_cast_fp16 = reshape(shape = concat_28, x = tile_14_cast_fp16)[name = string("reshape_28_cast_fp16")];
            tensor<int32, [5]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_28_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [4, 1, 256, 256]> reshape_29_cast_fp16 = reshape(shape = concat_29, x = transpose_29_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_15_reps_0 = const()[name = string("tile_15_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = v_29_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [4, 1, 256, 256]> tile_15_cast_fp16 = tile(reps = tile_15_reps_0, x = transpose_30_cast_fp16)[name = string("tile_15_cast_fp16")];
            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_30_cast_fp16 = reshape(shape = concat_30, x = tile_15_cast_fp16)[name = string("reshape_30_cast_fp16")];
            tensor<int32, [5]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_30_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [4, 1, 256, 256]> reshape_31_cast_fp16 = reshape(shape = concat_31, x = transpose_31_cast_fp16)[name = string("reshape_31_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4663_transpose_x_0 = const()[name = string("op_4663_transpose_x_0"), val = bool(false)];
            bool var_4663_transpose_y_0 = const()[name = string("op_4663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 4, 256, 256]> var_4663_cast_fp16 = matmul(transpose_x = var_4663_transpose_x_0, transpose_y = var_4663_transpose_y_0, x = q_31_cast_fp16, y = transpose_103_cast_fp16)[name = string("op_4663_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4665_cast_fp16 = add(x = var_4663_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4666 = const()[name = string("op_4666"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_121_cast_fp16 = softmax(axis = var_4666, x = var_4665_cast_fp16)[name = string("w_121_cast_fp16")];
            bool var_4669_transpose_x_0 = const()[name = string("op_4669_transpose_x_0"), val = bool(false)];
            bool var_4669_transpose_y_0 = const()[name = string("op_4669_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = reshape_31_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 4, 256, 256]> var_4669_cast_fp16 = matmul(transpose_x = var_4669_transpose_x_0, transpose_y = var_4669_transpose_y_0, x = w_121_cast_fp16, y = v_31_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<int32, [4]> var_4672_perm_0 = const()[name = string("op_4672_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4676 = const()[name = string("op_4676"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_4672_cast_fp16 = transpose(perm = var_4672_perm_0, x = var_4669_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 256, 1024]> input_87_cast_fp16 = reshape(shape = var_4676, x = var_4672_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45325248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45849600))))[name = string("tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4681_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4684 = const()[name = string("op_4684"), val = bool(true)];
            tensor<int32, [1]> var_4685_axes_0 = const()[name = string("op_4685_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4685_cast_fp16 = reduce_max(axes = var_4685_axes_0, keep_dims = var_4684, x = var_4681_cast_fp16)[name = string("op_4685_cast_fp16")];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_123_cast_fp16 = maximum(x = var_4685_cast_fp16, y = var_4686_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_123_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_123_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4689_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [1]> var_4694_axes_0 = const()[name = string("op_4694_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4694_keep_dims_0 = const()[name = string("op_4694_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4694_cast_fp16 = reduce_mean(axes = var_4694_axes_0, keep_dims = var_4694_keep_dims_0, x = var_4689_cast_fp16)[name = string("op_4694_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4695_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4695_cast_fp16")];
            fp32 var_4696_epsilon_0 = const()[name = string("op_4696_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4696_cast_fp16 = inverse(epsilon = var_4696_epsilon_0, x = var_4695_cast_fp16)[name = string("op_4696_cast_fp16")];
            fp16 var_4697_to_fp16 = const()[name = string("op_4697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4698_cast_fp16 = mul(x = var_4696_cast_fp16, y = var_4697_to_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4700_cast_fp16 = add(x = var_4694_cast_fp16, y = var_4698_cast_fp16)[name = string("op_4700_cast_fp16")];
            fp32 var_4701_epsilon_0 = const()[name = string("op_4701_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4701_cast_fp16 = rsqrt(epsilon = var_4701_epsilon_0, x = var_4700_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_107_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4701_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45850688)))];
            tensor<fp16, [1, 256, 512]> var_4703_cast_fp16 = mul(x = out_107_cast_fp16, y = tm_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4703_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_137_cast_fp16 = add(x = x_127_cast_fp16, y = var_4703_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4706_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_4706_cast_fp16")];
            bool var_4709 = const()[name = string("op_4709"), val = bool(true)];
            tensor<int32, [1]> var_4710_axes_0 = const()[name = string("op_4710_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4710_cast_fp16 = reduce_max(axes = var_4710_axes_0, keep_dims = var_4709, x = var_4706_cast_fp16)[name = string("op_4710_cast_fp16")];
            fp16 var_4711_to_fp16 = const()[name = string("op_4711_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_125_cast_fp16 = maximum(x = var_4710_cast_fp16, y = var_4711_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_125_cast_fp16 = real_div(x = x_137_cast_fp16, y = s_125_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4714_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4714_cast_fp16")];
            tensor<int32, [1]> var_4719_axes_0 = const()[name = string("op_4719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4719_keep_dims_0 = const()[name = string("op_4719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4719_cast_fp16 = reduce_mean(axes = var_4719_axes_0, keep_dims = var_4719_keep_dims_0, x = var_4714_cast_fp16)[name = string("op_4719_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4720_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4720_cast_fp16")];
            fp32 var_4721_epsilon_0 = const()[name = string("op_4721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4721_cast_fp16 = inverse(epsilon = var_4721_epsilon_0, x = var_4720_cast_fp16)[name = string("op_4721_cast_fp16")];
            fp16 var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4723_cast_fp16 = mul(x = var_4721_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4725_cast_fp16 = add(x = var_4719_cast_fp16, y = var_4723_cast_fp16)[name = string("op_4725_cast_fp16")];
            fp32 var_4726_epsilon_0 = const()[name = string("op_4726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4726_cast_fp16 = rsqrt(epsilon = var_4726_epsilon_0, x = var_4725_cast_fp16)[name = string("op_4726_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_109_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4726_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45851776)))];
            tensor<fp16, [1, 256, 512]> input_89_cast_fp16 = mul(x = out_109_cast_fp16, y = tm_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45852864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46901504))))[name = string("tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4733_mode_0 = const()[name = string("op_4733_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_4733_cast_fp16 = gelu(mode = var_4733_mode_0, x = linear_68_cast_fp16)[name = string("op_4733_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46905664))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47954304))))[name = string("tm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_91_cast_fp16 = mul(x = var_4733_cast_fp16, y = linear_69_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47958464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49007104))))[name = string("tm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4741_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4741_cast_fp16")];
            bool var_4744 = const()[name = string("op_4744"), val = bool(true)];
            tensor<int32, [1]> var_4745_axes_0 = const()[name = string("op_4745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4745_cast_fp16 = reduce_max(axes = var_4745_axes_0, keep_dims = var_4744, x = var_4741_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_127_cast_fp16 = maximum(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_127_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_127_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4749_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<int32, [1]> var_4754_axes_0 = const()[name = string("op_4754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4754_keep_dims_0 = const()[name = string("op_4754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4754_cast_fp16 = reduce_mean(axes = var_4754_axes_0, keep_dims = var_4754_keep_dims_0, x = var_4749_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4755_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4755_cast_fp16")];
            fp32 var_4756_epsilon_0 = const()[name = string("op_4756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4756_cast_fp16 = inverse(epsilon = var_4756_epsilon_0, x = var_4755_cast_fp16)[name = string("op_4756_cast_fp16")];
            fp16 var_4757_to_fp16 = const()[name = string("op_4757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4758_cast_fp16 = mul(x = var_4756_cast_fp16, y = var_4757_to_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4760_cast_fp16 = add(x = var_4754_cast_fp16, y = var_4758_cast_fp16)[name = string("op_4760_cast_fp16")];
            fp32 var_4761_epsilon_0 = const()[name = string("op_4761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4761_cast_fp16 = rsqrt(epsilon = var_4761_epsilon_0, x = var_4760_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_111_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4761_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49008192)))];
            tensor<fp16, [1, 256, 512]> var_4763_cast_fp16 = mul(x = out_111_cast_fp16, y = tm_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4763_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_93_cast_fp16 = add(x = x_137_cast_fp16, y = var_4763_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49009280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49271488))))[name = string("tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_71_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_71_cast_fp16")];
            string x_141_mode_0 = const()[name = string("x_141_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_141_cast_fp16 = gelu(mode = x_141_mode_0, x = linear_71_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_4783_begin_0 = const()[name = string("op_4783_begin_0"), val = tensor<int32, [4]>([0, 0, 7, 0])];
            tensor<int32, [4]> var_4783_end_0 = const()[name = string("op_4783_end_0"), val = tensor<int32, [4]>([1, 256, 8, 512])];
            tensor<bool, [4]> var_4783_end_mask_0 = const()[name = string("op_4783_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4783_squeeze_mask_0 = const()[name = string("op_4783_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_4783_cast_fp16 = slice_by_index(begin = var_4783_begin_0, end = var_4783_end_0, end_mask = var_4783_end_mask_0, squeeze_mask = var_4783_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_95_cast_fp16 = mul(x = x_141_cast_fp16, y = var_4783_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49272576))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49534784))))[name = string("tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4793_cast_fp16 = abs(x = linear_72_cast_fp16)[name = string("op_4793_cast_fp16")];
            bool var_4796 = const()[name = string("op_4796"), val = bool(true)];
            tensor<int32, [1]> var_4797_axes_0 = const()[name = string("op_4797_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4797_cast_fp16 = reduce_max(axes = var_4797_axes_0, keep_dims = var_4796, x = var_4793_cast_fp16)[name = string("op_4797_cast_fp16")];
            fp16 var_4798_to_fp16 = const()[name = string("op_4798_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_129_cast_fp16 = maximum(x = var_4797_cast_fp16, y = var_4798_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_129_cast_fp16 = real_div(x = linear_72_cast_fp16, y = s_129_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4801_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4801_cast_fp16")];
            tensor<int32, [1]> var_4806_axes_0 = const()[name = string("op_4806_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4806_keep_dims_0 = const()[name = string("op_4806_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4806_cast_fp16 = reduce_mean(axes = var_4806_axes_0, keep_dims = var_4806_keep_dims_0, x = var_4801_cast_fp16)[name = string("op_4806_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4807_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4807_cast_fp16")];
            fp32 var_4808_epsilon_0 = const()[name = string("op_4808_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4808_cast_fp16 = inverse(epsilon = var_4808_epsilon_0, x = var_4807_cast_fp16)[name = string("op_4808_cast_fp16")];
            fp16 var_4809_to_fp16 = const()[name = string("op_4809_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4810_cast_fp16 = mul(x = var_4808_cast_fp16, y = var_4809_to_fp16)[name = string("op_4810_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4812_cast_fp16 = add(x = var_4806_cast_fp16, y = var_4810_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp32 var_4813_epsilon_0 = const()[name = string("op_4813_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4813_cast_fp16 = rsqrt(epsilon = var_4813_epsilon_0, x = var_4812_cast_fp16)[name = string("op_4813_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_113_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4813_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49535872)))];
            tensor<fp16, [1, 256, 512]> var_4815_cast_fp16 = mul(x = out_113_cast_fp16, y = tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4815_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_15_cast_fp16 = add(x = input_93_cast_fp16, y = var_4815_cast_fp16)[name = string("h_15_cast_fp16")];
            tensor<fp16, [1]> tm_layers_7_layer_scalar_to_fp16 = const()[name = string("tm_layers_7_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-1])];
            tensor<fp16, [1, 256, 512]> x_145_cast_fp16 = mul(x = h_15_cast_fp16, y = tm_layers_7_layer_scalar_to_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4819_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4819_cast_fp16")];
            bool var_4822 = const()[name = string("op_4822"), val = bool(true)];
            tensor<int32, [1]> var_4823_axes_0 = const()[name = string("op_4823_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4823_cast_fp16 = reduce_max(axes = var_4823_axes_0, keep_dims = var_4822, x = var_4819_cast_fp16)[name = string("op_4823_cast_fp16")];
            fp16 var_4824_to_fp16 = const()[name = string("op_4824_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_131_cast_fp16 = maximum(x = var_4823_cast_fp16, y = var_4824_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_131_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_131_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4827_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [1]> var_4832_axes_0 = const()[name = string("op_4832_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4832_keep_dims_0 = const()[name = string("op_4832_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4832_cast_fp16 = reduce_mean(axes = var_4832_axes_0, keep_dims = var_4832_keep_dims_0, x = var_4827_cast_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4833_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4833_cast_fp16")];
            fp32 var_4834_epsilon_0 = const()[name = string("op_4834_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4834_cast_fp16 = inverse(epsilon = var_4834_epsilon_0, x = var_4833_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp16 var_4835_to_fp16 = const()[name = string("op_4835_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4836_cast_fp16 = mul(x = var_4834_cast_fp16, y = var_4835_to_fp16)[name = string("op_4836_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4838_cast_fp16 = add(x = var_4832_cast_fp16, y = var_4836_cast_fp16)[name = string("op_4838_cast_fp16")];
            fp32 var_4839_epsilon_0 = const()[name = string("op_4839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4839_cast_fp16 = rsqrt(epsilon = var_4839_epsilon_0, x = var_4838_cast_fp16)[name = string("op_4839_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_115_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4839_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49536960)))];
            tensor<fp16, [1, 256, 512]> input_97_cast_fp16 = mul(x = out_115_cast_fp16, y = tm_layers_8_input_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49538048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50062400))))[name = string("tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4849 = const()[name = string("op_4849"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_4850_cast_fp16 = reshape(shape = var_4849, x = linear_73_cast_fp16)[name = string("op_4850_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50064512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50326720))))[name = string("tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [4]> var_4861 = const()[name = string("op_4861"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4862_cast_fp16 = reshape(shape = var_4861, x = linear_74_cast_fp16)[name = string("op_4862_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50327808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50590016))))[name = string("tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [4]> var_4873 = const()[name = string("op_4873"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4874_cast_fp16 = reshape(shape = var_4873, x = linear_75_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<int32, [4]> x_151_perm_0 = const()[name = string("x_151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_4850_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 4, 256, 256]> var_4878_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4878_cast_fp16")];
            bool var_4881 = const()[name = string("op_4881"), val = bool(true)];
            tensor<int32, [1]> var_4882_axes_0 = const()[name = string("op_4882_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_4882_cast_fp16 = reduce_max(axes = var_4882_axes_0, keep_dims = var_4881, x = var_4878_cast_fp16)[name = string("op_4882_cast_fp16")];
            fp16 var_4883_to_fp16 = const()[name = string("op_4883_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_133_cast_fp16 = maximum(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_133_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_133_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4886_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4886_cast_fp16")];
            tensor<int32, [1]> var_4891_axes_0 = const()[name = string("op_4891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4891_keep_dims_0 = const()[name = string("op_4891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_4891_cast_fp16 = reduce_mean(axes = var_4891_axes_0, keep_dims = var_4891_keep_dims_0, x = var_4886_cast_fp16)[name = string("op_4891_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4892_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4892_cast_fp16")];
            fp32 var_4893_epsilon_0 = const()[name = string("op_4893_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_4893_cast_fp16 = inverse(epsilon = var_4893_epsilon_0, x = var_4892_cast_fp16)[name = string("op_4893_cast_fp16")];
            fp16 var_4894_to_fp16 = const()[name = string("op_4894_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_4895_cast_fp16 = mul(x = var_4893_cast_fp16, y = var_4894_to_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4897_cast_fp16 = add(x = var_4891_cast_fp16, y = var_4895_cast_fp16)[name = string("op_4897_cast_fp16")];
            fp32 var_4898_epsilon_0 = const()[name = string("op_4898_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_4898_cast_fp16 = rsqrt(epsilon = var_4898_epsilon_0, x = var_4897_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_117_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4898_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591104)))];
            tensor<fp16, [1, 4, 256, 256]> q_33_cast_fp16 = mul(x = out_117_cast_fp16, y = tm_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_4862_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 2, 256, 256]> var_4901_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4901_cast_fp16")];
            bool var_4904 = const()[name = string("op_4904"), val = bool(true)];
            tensor<int32, [1]> var_4905_axes_0 = const()[name = string("op_4905_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4905_cast_fp16 = reduce_max(axes = var_4905_axes_0, keep_dims = var_4904, x = var_4901_cast_fp16)[name = string("op_4905_cast_fp16")];
            fp16 var_4906_to_fp16 = const()[name = string("op_4906_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_135_cast_fp16 = maximum(x = var_4905_cast_fp16, y = var_4906_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_135_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_135_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4909_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [1]> var_4914_axes_0 = const()[name = string("op_4914_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4914_keep_dims_0 = const()[name = string("op_4914_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4914_cast_fp16 = reduce_mean(axes = var_4914_axes_0, keep_dims = var_4914_keep_dims_0, x = var_4909_cast_fp16)[name = string("op_4914_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4915_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4915_cast_fp16")];
            fp32 var_4916_epsilon_0 = const()[name = string("op_4916_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4916_cast_fp16 = inverse(epsilon = var_4916_epsilon_0, x = var_4915_cast_fp16)[name = string("op_4916_cast_fp16")];
            fp16 var_4917_to_fp16 = const()[name = string("op_4917_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4918_cast_fp16 = mul(x = var_4916_cast_fp16, y = var_4917_to_fp16)[name = string("op_4918_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4920_cast_fp16 = add(x = var_4914_cast_fp16, y = var_4918_cast_fp16)[name = string("op_4920_cast_fp16")];
            fp32 var_4921_epsilon_0 = const()[name = string("op_4921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4921_cast_fp16 = rsqrt(epsilon = var_4921_epsilon_0, x = var_4920_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_119_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4921_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591680)))];
            tensor<fp16, [1, 2, 256, 256]> k_49_cast_fp16 = mul(x = out_119_cast_fp16, y = tm_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_151_cast_fp16 = transpose(perm = x_151_perm_0, x = var_4874_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 2, 256, 256]> var_4924_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4924_cast_fp16")];
            bool var_4927 = const()[name = string("op_4927"), val = bool(true)];
            tensor<int32, [1]> var_4928_axes_0 = const()[name = string("op_4928_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4928_cast_fp16 = reduce_max(axes = var_4928_axes_0, keep_dims = var_4927, x = var_4924_cast_fp16)[name = string("op_4928_cast_fp16")];
            fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_137_cast_fp16 = maximum(x = var_4928_cast_fp16, y = var_4929_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_137_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_137_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4932_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4932_cast_fp16")];
            tensor<int32, [1]> var_4937_axes_0 = const()[name = string("op_4937_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4937_keep_dims_0 = const()[name = string("op_4937_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4937_cast_fp16 = reduce_mean(axes = var_4937_axes_0, keep_dims = var_4937_keep_dims_0, x = var_4932_cast_fp16)[name = string("op_4937_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4938_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4938_cast_fp16")];
            fp32 var_4939_epsilon_0 = const()[name = string("op_4939_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4939_cast_fp16 = inverse(epsilon = var_4939_epsilon_0, x = var_4938_cast_fp16)[name = string("op_4939_cast_fp16")];
            fp16 var_4940_to_fp16 = const()[name = string("op_4940_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4941_cast_fp16 = mul(x = var_4939_cast_fp16, y = var_4940_to_fp16)[name = string("op_4941_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4943_cast_fp16 = add(x = var_4937_cast_fp16, y = var_4941_cast_fp16)[name = string("op_4943_cast_fp16")];
            fp32 var_4944_epsilon_0 = const()[name = string("op_4944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4944_cast_fp16 = rsqrt(epsilon = var_4944_epsilon_0, x = var_4943_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_33_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4944_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4946_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = q_33_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_4952_cast_fp16 = mul(x = var_4951_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4957_begin_0 = const()[name = string("op_4957_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4957_end_0 = const()[name = string("op_4957_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_4957_end_mask_0 = const()[name = string("op_4957_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_4957_cast_fp16 = slice_by_index(begin = var_4957_begin_0, end = var_4957_end_0, end_mask = var_4957_end_mask_0, x = q_33_cast_fp16)[name = string("op_4957_cast_fp16")];
            int32 var_4959 = const()[name = string("op_4959"), val = int32(-1)];
            bool var_4960_interleave_0 = const()[name = string("op_4960_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_4960_cast_fp16 = concat(axis = var_4959, interleave = var_4960_interleave_0, values = (var_4952_cast_fp16, var_4957_cast_fp16))[name = string("op_4960_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4961_cast_fp16 = mul(x = var_4960_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4961_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_35_cast_fp16 = add(x = var_4946_cast_fp16, y = var_4961_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4964_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4964_cast_fp16")];
            tensor<int32, [4]> var_4969_begin_0 = const()[name = string("op_4969_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4969_end_0 = const()[name = string("op_4969_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_4969_end_mask_0 = const()[name = string("op_4969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_4969_cast_fp16 = slice_by_index(begin = var_4969_begin_0, end = var_4969_end_0, end_mask = var_4969_end_mask_0, x = k_49_cast_fp16)[name = string("op_4969_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_4970_cast_fp16 = mul(x = var_4969_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_4970_cast_fp16")];
            tensor<int32, [4]> var_4975_begin_0 = const()[name = string("op_4975_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4975_end_0 = const()[name = string("op_4975_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_4975_end_mask_0 = const()[name = string("op_4975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_4975_cast_fp16 = slice_by_index(begin = var_4975_begin_0, end = var_4975_end_0, end_mask = var_4975_end_mask_0, x = k_49_cast_fp16)[name = string("op_4975_cast_fp16")];
            int32 var_4977 = const()[name = string("op_4977"), val = int32(-1)];
            bool var_4978_interleave_0 = const()[name = string("op_4978_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_4978_cast_fp16 = concat(axis = var_4977, interleave = var_4978_interleave_0, values = (var_4970_cast_fp16, var_4975_cast_fp16))[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4979_cast_fp16 = mul(x = var_4978_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_51_cast_fp16 = add(x = var_4964_cast_fp16, y = var_4979_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_16_reps_0 = const()[name = string("tile_16_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = k_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [4, 1, 256, 256]> tile_16_cast_fp16 = tile(reps = tile_16_reps_0, x = transpose_32_cast_fp16)[name = string("tile_16_cast_fp16")];
            tensor<int32, [5]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_32_cast_fp16 = reshape(shape = concat_32, x = tile_16_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [5]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [4, 1, 256, 256]> reshape_33_cast_fp16 = reshape(shape = concat_33, x = transpose_33_cast_fp16)[name = string("reshape_33_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_17_reps_0 = const()[name = string("tile_17_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = v_33_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [4, 1, 256, 256]> tile_17_cast_fp16 = tile(reps = tile_17_reps_0, x = transpose_34_cast_fp16)[name = string("tile_17_cast_fp16")];
            tensor<int32, [5]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_34_cast_fp16 = reshape(shape = concat_34, x = tile_17_cast_fp16)[name = string("reshape_34_cast_fp16")];
            tensor<int32, [5]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = reshape_34_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [4, 1, 256, 256]> reshape_35_cast_fp16 = reshape(shape = concat_35, x = transpose_35_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = reshape_33_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 4, 256, 256]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = q_35_cast_fp16, y = transpose_104_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4995_cast_fp16 = add(x = var_4993_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4995_cast_fp16")];
            int32 var_4996 = const()[name = string("op_4996"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_137_cast_fp16 = softmax(axis = var_4996, x = var_4995_cast_fp16)[name = string("w_137_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 4, 256, 256]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = w_137_cast_fp16, y = v_35_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5002_perm_0 = const()[name = string("op_5002_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5006 = const()[name = string("op_5006"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_5002_cast_fp16 = transpose(perm = var_5002_perm_0, x = var_4999_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 256, 1024]> input_99_cast_fp16 = reshape(shape = var_5006, x = var_5002_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50592256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51116608))))[name = string("tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5011_cast_fp16 = abs(x = linear_76_cast_fp16)[name = string("op_5011_cast_fp16")];
            bool var_5014 = const()[name = string("op_5014"), val = bool(true)];
            tensor<int32, [1]> var_5015_axes_0 = const()[name = string("op_5015_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5015_cast_fp16 = reduce_max(axes = var_5015_axes_0, keep_dims = var_5014, x = var_5011_cast_fp16)[name = string("op_5015_cast_fp16")];
            fp16 var_5016_to_fp16 = const()[name = string("op_5016_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_139_cast_fp16 = maximum(x = var_5015_cast_fp16, y = var_5016_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_139_cast_fp16 = real_div(x = linear_76_cast_fp16, y = s_139_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5019_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_5019_cast_fp16")];
            tensor<int32, [1]> var_5024_axes_0 = const()[name = string("op_5024_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5024_keep_dims_0 = const()[name = string("op_5024_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5024_cast_fp16 = reduce_mean(axes = var_5024_axes_0, keep_dims = var_5024_keep_dims_0, x = var_5019_cast_fp16)[name = string("op_5024_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5025_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_5025_cast_fp16")];
            fp32 var_5026_epsilon_0 = const()[name = string("op_5026_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5026_cast_fp16 = inverse(epsilon = var_5026_epsilon_0, x = var_5025_cast_fp16)[name = string("op_5026_cast_fp16")];
            fp16 var_5027_to_fp16 = const()[name = string("op_5027_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5028_cast_fp16 = mul(x = var_5026_cast_fp16, y = var_5027_to_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5030_cast_fp16 = add(x = var_5024_cast_fp16, y = var_5028_cast_fp16)[name = string("op_5030_cast_fp16")];
            fp32 var_5031_epsilon_0 = const()[name = string("op_5031_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5031_cast_fp16 = rsqrt(epsilon = var_5031_epsilon_0, x = var_5030_cast_fp16)[name = string("op_5031_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_121_cast_fp16 = mul(x = y_139_cast_fp16, y = var_5031_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51117696)))];
            tensor<fp16, [1, 256, 512]> var_5033_cast_fp16 = mul(x = out_121_cast_fp16, y = tm_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_5033_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = var_5033_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5036_cast_fp16 = abs(x = x_155_cast_fp16)[name = string("op_5036_cast_fp16")];
            bool var_5039 = const()[name = string("op_5039"), val = bool(true)];
            tensor<int32, [1]> var_5040_axes_0 = const()[name = string("op_5040_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5040_cast_fp16 = reduce_max(axes = var_5040_axes_0, keep_dims = var_5039, x = var_5036_cast_fp16)[name = string("op_5040_cast_fp16")];
            fp16 var_5041_to_fp16 = const()[name = string("op_5041_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_141_cast_fp16 = maximum(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_141_cast_fp16 = real_div(x = x_155_cast_fp16, y = s_141_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5044_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<int32, [1]> var_5049_axes_0 = const()[name = string("op_5049_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5049_keep_dims_0 = const()[name = string("op_5049_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5049_cast_fp16 = reduce_mean(axes = var_5049_axes_0, keep_dims = var_5049_keep_dims_0, x = var_5044_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5050_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_5050_cast_fp16")];
            fp32 var_5051_epsilon_0 = const()[name = string("op_5051_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5051_cast_fp16 = inverse(epsilon = var_5051_epsilon_0, x = var_5050_cast_fp16)[name = string("op_5051_cast_fp16")];
            fp16 var_5052_to_fp16 = const()[name = string("op_5052_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5053_cast_fp16 = mul(x = var_5051_cast_fp16, y = var_5052_to_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5055_cast_fp16 = add(x = var_5049_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            fp32 var_5056_epsilon_0 = const()[name = string("op_5056_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5056_cast_fp16 = rsqrt(epsilon = var_5056_epsilon_0, x = var_5055_cast_fp16)[name = string("op_5056_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_123_cast_fp16 = mul(x = y_141_cast_fp16, y = var_5056_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51118784)))];
            tensor<fp16, [1, 256, 512]> input_101_cast_fp16 = mul(x = out_123_cast_fp16, y = tm_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51119872))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52168512))))[name = string("tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_77_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_77_cast_fp16")];
            string var_5063_mode_0 = const()[name = string("op_5063_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_5063_cast_fp16 = gelu(mode = var_5063_mode_0, x = linear_77_cast_fp16)[name = string("op_5063_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52172672))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53221312))))[name = string("tm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_78_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_103_cast_fp16 = mul(x = var_5063_cast_fp16, y = linear_78_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53225472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54274112))))[name = string("tm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5071_cast_fp16 = abs(x = linear_79_cast_fp16)[name = string("op_5071_cast_fp16")];
            bool var_5074 = const()[name = string("op_5074"), val = bool(true)];
            tensor<int32, [1]> var_5075_axes_0 = const()[name = string("op_5075_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5075_cast_fp16 = reduce_max(axes = var_5075_axes_0, keep_dims = var_5074, x = var_5071_cast_fp16)[name = string("op_5075_cast_fp16")];
            fp16 var_5076_to_fp16 = const()[name = string("op_5076_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_143_cast_fp16 = maximum(x = var_5075_cast_fp16, y = var_5076_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_143_cast_fp16 = real_div(x = linear_79_cast_fp16, y = s_143_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5079_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_5079_cast_fp16")];
            tensor<int32, [1]> var_5084_axes_0 = const()[name = string("op_5084_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5084_keep_dims_0 = const()[name = string("op_5084_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5084_cast_fp16 = reduce_mean(axes = var_5084_axes_0, keep_dims = var_5084_keep_dims_0, x = var_5079_cast_fp16)[name = string("op_5084_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5085_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5086_cast_fp16 = inverse(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            fp16 var_5087_to_fp16 = const()[name = string("op_5087_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5088_cast_fp16 = mul(x = var_5086_cast_fp16, y = var_5087_to_fp16)[name = string("op_5088_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5090_cast_fp16 = add(x = var_5084_cast_fp16, y = var_5088_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp32 var_5091_epsilon_0 = const()[name = string("op_5091_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5091_cast_fp16 = rsqrt(epsilon = var_5091_epsilon_0, x = var_5090_cast_fp16)[name = string("op_5091_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_125_cast_fp16 = mul(x = y_143_cast_fp16, y = var_5091_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54275200)))];
            tensor<fp16, [1, 256, 512]> var_5093_cast_fp16 = mul(x = out_125_cast_fp16, y = tm_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5093_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_105_cast_fp16 = add(x = x_155_cast_fp16, y = var_5093_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54276288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54538496))))[name = string("tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_80_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_80_cast_fp16")];
            string x_159_mode_0 = const()[name = string("x_159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_159_cast_fp16 = gelu(mode = x_159_mode_0, x = linear_80_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [4]> var_5113_begin_0 = const()[name = string("op_5113_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_5113_end_0 = const()[name = string("op_5113_end_0"), val = tensor<int32, [4]>([1, 256, 9, 512])];
            tensor<bool, [4]> var_5113_end_mask_0 = const()[name = string("op_5113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5113_squeeze_mask_0 = const()[name = string("op_5113_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_5113_cast_fp16 = slice_by_index(begin = var_5113_begin_0, end = var_5113_end_0, end_mask = var_5113_end_mask_0, squeeze_mask = var_5113_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_107_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5113_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54539584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54801792))))[name = string("tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5123_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5123_cast_fp16")];
            bool var_5126 = const()[name = string("op_5126"), val = bool(true)];
            tensor<int32, [1]> var_5127_axes_0 = const()[name = string("op_5127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5127_cast_fp16 = reduce_max(axes = var_5127_axes_0, keep_dims = var_5126, x = var_5123_cast_fp16)[name = string("op_5127_cast_fp16")];
            fp16 var_5128_to_fp16 = const()[name = string("op_5128_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_145_cast_fp16 = maximum(x = var_5127_cast_fp16, y = var_5128_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_145_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_145_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5131_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5136_axes_0 = const()[name = string("op_5136_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5136_keep_dims_0 = const()[name = string("op_5136_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5136_cast_fp16 = reduce_mean(axes = var_5136_axes_0, keep_dims = var_5136_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5136_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5137_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5138_cast_fp16 = inverse(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp16 var_5139_to_fp16 = const()[name = string("op_5139_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5140_cast_fp16 = mul(x = var_5138_cast_fp16, y = var_5139_to_fp16)[name = string("op_5140_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5142_cast_fp16 = add(x = var_5136_cast_fp16, y = var_5140_cast_fp16)[name = string("op_5142_cast_fp16")];
            fp32 var_5143_epsilon_0 = const()[name = string("op_5143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5143_cast_fp16 = rsqrt(epsilon = var_5143_epsilon_0, x = var_5142_cast_fp16)[name = string("op_5143_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_127_cast_fp16 = mul(x = y_145_cast_fp16, y = var_5143_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54802880)))];
            tensor<fp16, [1, 256, 512]> var_5145_cast_fp16 = mul(x = out_127_cast_fp16, y = tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5145_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_17_cast_fp16 = add(x = input_105_cast_fp16, y = var_5145_cast_fp16)[name = string("h_17_cast_fp16")];
            tensor<fp16, [1]> tm_layers_8_layer_scalar_to_fp16 = const()[name = string("tm_layers_8_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bcp-1])];
            tensor<fp16, [1, 256, 512]> x_163_cast_fp16 = mul(x = h_17_cast_fp16, y = tm_layers_8_layer_scalar_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5149_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_5149_cast_fp16")];
            bool var_5152 = const()[name = string("op_5152"), val = bool(true)];
            tensor<int32, [1]> var_5153_axes_0 = const()[name = string("op_5153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5153_cast_fp16 = reduce_max(axes = var_5153_axes_0, keep_dims = var_5152, x = var_5149_cast_fp16)[name = string("op_5153_cast_fp16")];
            fp16 var_5154_to_fp16 = const()[name = string("op_5154_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_147_cast_fp16 = maximum(x = var_5153_cast_fp16, y = var_5154_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_147_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_147_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5157_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_5157_cast_fp16")];
            tensor<int32, [1]> var_5162_axes_0 = const()[name = string("op_5162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5162_keep_dims_0 = const()[name = string("op_5162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5162_cast_fp16 = reduce_mean(axes = var_5162_axes_0, keep_dims = var_5162_keep_dims_0, x = var_5157_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5163_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_5163_cast_fp16")];
            fp32 var_5164_epsilon_0 = const()[name = string("op_5164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5164_cast_fp16 = inverse(epsilon = var_5164_epsilon_0, x = var_5163_cast_fp16)[name = string("op_5164_cast_fp16")];
            fp16 var_5165_to_fp16 = const()[name = string("op_5165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5166_cast_fp16 = mul(x = var_5164_cast_fp16, y = var_5165_to_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5168_cast_fp16 = add(x = var_5162_cast_fp16, y = var_5166_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp32 var_5169_epsilon_0 = const()[name = string("op_5169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5169_cast_fp16 = rsqrt(epsilon = var_5169_epsilon_0, x = var_5168_cast_fp16)[name = string("op_5169_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_129_cast_fp16 = mul(x = y_147_cast_fp16, y = var_5169_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54803968)))];
            tensor<fp16, [1, 256, 512]> input_109_cast_fp16 = mul(x = out_129_cast_fp16, y = tm_layers_9_input_layernorm_weight_to_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54805056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55329408))))[name = string("tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [4]> var_5179 = const()[name = string("op_5179"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_5180_cast_fp16 = reshape(shape = var_5179, x = linear_82_cast_fp16)[name = string("op_5180_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55331520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55593728))))[name = string("tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [4]> var_5191 = const()[name = string("op_5191"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_5192_cast_fp16 = reshape(shape = var_5191, x = linear_83_cast_fp16)[name = string("op_5192_cast_fp16")];
            tensor<int32, [4]> x_167_perm_0 = const()[name = string("x_167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55594816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55857024))))[name = string("tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5203 = const()[name = string("op_5203"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_5204_cast_fp16 = reshape(shape = var_5203, x = linear_84_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_5180_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 4, 256, 256]> var_5208_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_5208_cast_fp16")];
            bool var_5211 = const()[name = string("op_5211"), val = bool(true)];
            tensor<int32, [1]> var_5212_axes_0 = const()[name = string("op_5212_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_5212_cast_fp16 = reduce_max(axes = var_5212_axes_0, keep_dims = var_5211, x = var_5208_cast_fp16)[name = string("op_5212_cast_fp16")];
            fp16 var_5213_to_fp16 = const()[name = string("op_5213_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_149_cast_fp16 = maximum(x = var_5212_cast_fp16, y = var_5213_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_149_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_149_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5216_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<int32, [1]> var_5221_axes_0 = const()[name = string("op_5221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5221_keep_dims_0 = const()[name = string("op_5221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_5221_cast_fp16 = reduce_mean(axes = var_5221_axes_0, keep_dims = var_5221_keep_dims_0, x = var_5216_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5222_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_5223_cast_fp16 = inverse(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            fp16 var_5224_to_fp16 = const()[name = string("op_5224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_5225_cast_fp16 = mul(x = var_5223_cast_fp16, y = var_5224_to_fp16)[name = string("op_5225_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5227_cast_fp16 = add(x = var_5221_cast_fp16, y = var_5225_cast_fp16)[name = string("op_5227_cast_fp16")];
            fp32 var_5228_epsilon_0 = const()[name = string("op_5228_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_5228_cast_fp16 = rsqrt(epsilon = var_5228_epsilon_0, x = var_5227_cast_fp16)[name = string("op_5228_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_131_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5228_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = var_5192_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 2, 256, 256]> var_5231_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_5231_cast_fp16")];
            bool var_5234 = const()[name = string("op_5234"), val = bool(true)];
            tensor<int32, [1]> var_5235_axes_0 = const()[name = string("op_5235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_5235_cast_fp16 = reduce_max(axes = var_5235_axes_0, keep_dims = var_5234, x = var_5231_cast_fp16)[name = string("op_5235_cast_fp16")];
            fp16 var_5236_to_fp16 = const()[name = string("op_5236_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_151_cast_fp16 = maximum(x = var_5235_cast_fp16, y = var_5236_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_151_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_151_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5239_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5239_cast_fp16")];
            tensor<int32, [1]> var_5244_axes_0 = const()[name = string("op_5244_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5244_keep_dims_0 = const()[name = string("op_5244_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_5244_cast_fp16 = reduce_mean(axes = var_5244_axes_0, keep_dims = var_5244_keep_dims_0, x = var_5239_cast_fp16)[name = string("op_5244_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5245_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_5245_cast_fp16")];
            fp32 var_5246_epsilon_0 = const()[name = string("op_5246_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_5246_cast_fp16 = inverse(epsilon = var_5246_epsilon_0, x = var_5245_cast_fp16)[name = string("op_5246_cast_fp16")];
            fp16 var_5247_to_fp16 = const()[name = string("op_5247_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_5248_cast_fp16 = mul(x = var_5246_cast_fp16, y = var_5247_to_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5250_cast_fp16 = add(x = var_5244_cast_fp16, y = var_5248_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp32 var_5251_epsilon_0 = const()[name = string("op_5251_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_5251_cast_fp16 = rsqrt(epsilon = var_5251_epsilon_0, x = var_5250_cast_fp16)[name = string("op_5251_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_133_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5251_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [256]> tm_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858112)))];
            tensor<fp16, [1, 2, 256, 256]> k_55_cast_fp16 = mul(x = out_133_cast_fp16, y = tm_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_5204_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2, 256, 256]> var_5254_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_5254_cast_fp16")];
            bool var_5257 = const()[name = string("op_5257"), val = bool(true)];
            tensor<int32, [1]> var_5258_axes_0 = const()[name = string("op_5258_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_5258_cast_fp16 = reduce_max(axes = var_5258_axes_0, keep_dims = var_5257, x = var_5254_cast_fp16)[name = string("op_5258_cast_fp16")];
            fp16 var_5259_to_fp16 = const()[name = string("op_5259_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_153_cast_fp16 = maximum(x = var_5258_cast_fp16, y = var_5259_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_153_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_153_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5262_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5262_cast_fp16")];
            tensor<int32, [1]> var_5267_axes_0 = const()[name = string("op_5267_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5267_keep_dims_0 = const()[name = string("op_5267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_5267_cast_fp16 = reduce_mean(axes = var_5267_axes_0, keep_dims = var_5267_keep_dims_0, x = var_5262_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5268_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5268_cast_fp16")];
            fp32 var_5269_epsilon_0 = const()[name = string("op_5269_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_5269_cast_fp16 = inverse(epsilon = var_5269_epsilon_0, x = var_5268_cast_fp16)[name = string("op_5269_cast_fp16")];
            fp16 var_5270_to_fp16 = const()[name = string("op_5270_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_5271_cast_fp16 = mul(x = var_5269_cast_fp16, y = var_5270_to_fp16)[name = string("op_5271_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5273_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5271_cast_fp16)[name = string("op_5273_cast_fp16")];
            fp32 var_5274_epsilon_0 = const()[name = string("op_5274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_5274_cast_fp16 = rsqrt(epsilon = var_5274_epsilon_0, x = var_5273_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_37_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5274_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5276_cast_fp16 = mul(x = out_131_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5276_cast_fp16")];
            tensor<int32, [4]> var_5281_begin_0 = const()[name = string("op_5281_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5281_end_0 = const()[name = string("op_5281_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_5281_end_mask_0 = const()[name = string("op_5281_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_5281_cast_fp16 = slice_by_index(begin = var_5281_begin_0, end = var_5281_end_0, end_mask = var_5281_end_mask_0, x = out_131_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_5282_cast_fp16 = mul(x = var_5281_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_5282_cast_fp16")];
            tensor<int32, [4]> var_5287_begin_0 = const()[name = string("op_5287_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5287_end_0 = const()[name = string("op_5287_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_5287_end_mask_0 = const()[name = string("op_5287_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_5287_cast_fp16 = slice_by_index(begin = var_5287_begin_0, end = var_5287_end_0, end_mask = var_5287_end_mask_0, x = out_131_cast_fp16)[name = string("op_5287_cast_fp16")];
            int32 var_5289 = const()[name = string("op_5289"), val = int32(-1)];
            bool var_5290_interleave_0 = const()[name = string("op_5290_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_5290_cast_fp16 = concat(axis = var_5289, interleave = var_5290_interleave_0, values = (var_5282_cast_fp16, var_5287_cast_fp16))[name = string("op_5290_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5291_cast_fp16 = mul(x = var_5290_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5291_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_39_cast_fp16 = add(x = var_5276_cast_fp16, y = var_5291_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5294_cast_fp16 = mul(x = k_55_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5294_cast_fp16")];
            tensor<int32, [4]> var_5299_begin_0 = const()[name = string("op_5299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5299_end_0 = const()[name = string("op_5299_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_5299_end_mask_0 = const()[name = string("op_5299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_5299_cast_fp16 = slice_by_index(begin = var_5299_begin_0, end = var_5299_end_0, end_mask = var_5299_end_mask_0, x = k_55_cast_fp16)[name = string("op_5299_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_5300_cast_fp16 = mul(x = var_5299_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_5300_cast_fp16")];
            tensor<int32, [4]> var_5305_begin_0 = const()[name = string("op_5305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5305_end_0 = const()[name = string("op_5305_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_5305_end_mask_0 = const()[name = string("op_5305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_5305_cast_fp16 = slice_by_index(begin = var_5305_begin_0, end = var_5305_end_0, end_mask = var_5305_end_mask_0, x = k_55_cast_fp16)[name = string("op_5305_cast_fp16")];
            int32 var_5307 = const()[name = string("op_5307"), val = int32(-1)];
            bool var_5308_interleave_0 = const()[name = string("op_5308_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_5308_cast_fp16 = concat(axis = var_5307, interleave = var_5308_interleave_0, values = (var_5300_cast_fp16, var_5305_cast_fp16))[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5309_cast_fp16 = mul(x = var_5308_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_57_cast_fp16 = add(x = var_5294_cast_fp16, y = var_5309_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_18_reps_0 = const()[name = string("tile_18_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = k_57_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [4, 1, 256, 256]> tile_18_cast_fp16 = tile(reps = tile_18_reps_0, x = transpose_36_cast_fp16)[name = string("tile_18_cast_fp16")];
            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_36_cast_fp16 = reshape(shape = concat_36, x = tile_18_cast_fp16)[name = string("reshape_36_cast_fp16")];
            tensor<int32, [5]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = reshape_36_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [4, 1, 256, 256]> reshape_37_cast_fp16 = reshape(shape = concat_37, x = transpose_37_cast_fp16)[name = string("reshape_37_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_19_reps_0 = const()[name = string("tile_19_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = v_37_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [4, 1, 256, 256]> tile_19_cast_fp16 = tile(reps = tile_19_reps_0, x = transpose_38_cast_fp16)[name = string("tile_19_cast_fp16")];
            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_38_cast_fp16 = reshape(shape = concat_38, x = tile_19_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [5]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [4, 1, 256, 256]> reshape_39_cast_fp16 = reshape(shape = concat_39, x = transpose_39_cast_fp16)[name = string("reshape_39_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5323_transpose_x_0 = const()[name = string("op_5323_transpose_x_0"), val = bool(false)];
            bool var_5323_transpose_y_0 = const()[name = string("op_5323_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = reshape_37_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 4, 256, 256]> var_5323_cast_fp16 = matmul(transpose_x = var_5323_transpose_x_0, transpose_y = var_5323_transpose_y_0, x = q_39_cast_fp16, y = transpose_105_cast_fp16)[name = string("op_5323_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5325_cast_fp16 = add(x = var_5323_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5325_cast_fp16")];
            int32 var_5326 = const()[name = string("op_5326"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_153_cast_fp16 = softmax(axis = var_5326, x = var_5325_cast_fp16)[name = string("w_153_cast_fp16")];
            bool var_5329_transpose_x_0 = const()[name = string("op_5329_transpose_x_0"), val = bool(false)];
            bool var_5329_transpose_y_0 = const()[name = string("op_5329_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = reshape_39_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 4, 256, 256]> var_5329_cast_fp16 = matmul(transpose_x = var_5329_transpose_x_0, transpose_y = var_5329_transpose_y_0, x = w_153_cast_fp16, y = v_39_cast_fp16)[name = string("op_5329_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = var_5329_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 256, 1024]> input_111_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56383040))))[name = string("tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5341_cast_fp16 = abs(x = linear_85_cast_fp16)[name = string("op_5341_cast_fp16")];
            bool var_5344 = const()[name = string("op_5344"), val = bool(true)];
            tensor<int32, [1]> var_5345_axes_0 = const()[name = string("op_5345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5345_cast_fp16 = reduce_max(axes = var_5345_axes_0, keep_dims = var_5344, x = var_5341_cast_fp16)[name = string("op_5345_cast_fp16")];
            fp16 var_5346_to_fp16 = const()[name = string("op_5346_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_155_cast_fp16 = maximum(x = var_5345_cast_fp16, y = var_5346_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_155_cast_fp16 = real_div(x = linear_85_cast_fp16, y = s_155_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5349_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5354_axes_0 = const()[name = string("op_5354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5354_keep_dims_0 = const()[name = string("op_5354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5354_cast_fp16 = reduce_mean(axes = var_5354_axes_0, keep_dims = var_5354_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5355_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5355_cast_fp16")];
            fp32 var_5356_epsilon_0 = const()[name = string("op_5356_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5356_cast_fp16 = inverse(epsilon = var_5356_epsilon_0, x = var_5355_cast_fp16)[name = string("op_5356_cast_fp16")];
            fp16 var_5357_to_fp16 = const()[name = string("op_5357_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5358_cast_fp16 = mul(x = var_5356_cast_fp16, y = var_5357_to_fp16)[name = string("op_5358_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5360_cast_fp16 = add(x = var_5354_cast_fp16, y = var_5358_cast_fp16)[name = string("op_5360_cast_fp16")];
            fp32 var_5361_epsilon_0 = const()[name = string("op_5361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5361_cast_fp16 = rsqrt(epsilon = var_5361_epsilon_0, x = var_5360_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_135_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5361_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56384128)))];
            tensor<fp16, [1, 256, 512]> var_5363_cast_fp16 = mul(x = out_135_cast_fp16, y = tm_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_5363_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_173_cast_fp16 = add(x = x_163_cast_fp16, y = var_5363_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5366_cast_fp16 = abs(x = x_173_cast_fp16)[name = string("op_5366_cast_fp16")];
            bool var_5369 = const()[name = string("op_5369"), val = bool(true)];
            tensor<int32, [1]> var_5370_axes_0 = const()[name = string("op_5370_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5370_cast_fp16 = reduce_max(axes = var_5370_axes_0, keep_dims = var_5369, x = var_5366_cast_fp16)[name = string("op_5370_cast_fp16")];
            fp16 var_5371_to_fp16 = const()[name = string("op_5371_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_157_cast_fp16 = maximum(x = var_5370_cast_fp16, y = var_5371_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_157_cast_fp16 = real_div(x = x_173_cast_fp16, y = s_157_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5374_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [1]> var_5379_axes_0 = const()[name = string("op_5379_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5379_keep_dims_0 = const()[name = string("op_5379_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5379_cast_fp16 = reduce_mean(axes = var_5379_axes_0, keep_dims = var_5379_keep_dims_0, x = var_5374_cast_fp16)[name = string("op_5379_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5380_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5381_cast_fp16 = inverse(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            fp16 var_5382_to_fp16 = const()[name = string("op_5382_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5383_cast_fp16 = mul(x = var_5381_cast_fp16, y = var_5382_to_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5385_cast_fp16 = add(x = var_5379_cast_fp16, y = var_5383_cast_fp16)[name = string("op_5385_cast_fp16")];
            fp32 var_5386_epsilon_0 = const()[name = string("op_5386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5386_cast_fp16 = rsqrt(epsilon = var_5386_epsilon_0, x = var_5385_cast_fp16)[name = string("op_5386_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_137_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5386_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56385216)))];
            tensor<fp16, [1, 256, 512]> input_113_cast_fp16 = mul(x = out_137_cast_fp16, y = tm_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56386304))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57434944))))[name = string("tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_86_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_86_cast_fp16")];
            string var_5393_mode_0 = const()[name = string("op_5393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_5393_cast_fp16 = gelu(mode = var_5393_mode_0, x = linear_86_cast_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57439104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58487744))))[name = string("tm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_87_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_115_cast_fp16 = mul(x = var_5393_cast_fp16, y = linear_87_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58491904))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59540544))))[name = string("tm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5401_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5401_cast_fp16")];
            bool var_5404 = const()[name = string("op_5404"), val = bool(true)];
            tensor<int32, [1]> var_5405_axes_0 = const()[name = string("op_5405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5405_cast_fp16 = reduce_max(axes = var_5405_axes_0, keep_dims = var_5404, x = var_5401_cast_fp16)[name = string("op_5405_cast_fp16")];
            fp16 var_5406_to_fp16 = const()[name = string("op_5406_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_159_cast_fp16 = maximum(x = var_5405_cast_fp16, y = var_5406_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_159_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_159_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5409_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5409_cast_fp16")];
            tensor<int32, [1]> var_5414_axes_0 = const()[name = string("op_5414_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5414_keep_dims_0 = const()[name = string("op_5414_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5414_cast_fp16 = reduce_mean(axes = var_5414_axes_0, keep_dims = var_5414_keep_dims_0, x = var_5409_cast_fp16)[name = string("op_5414_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5415_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5415_cast_fp16")];
            fp32 var_5416_epsilon_0 = const()[name = string("op_5416_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5416_cast_fp16 = inverse(epsilon = var_5416_epsilon_0, x = var_5415_cast_fp16)[name = string("op_5416_cast_fp16")];
            fp16 var_5417_to_fp16 = const()[name = string("op_5417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5418_cast_fp16 = mul(x = var_5416_cast_fp16, y = var_5417_to_fp16)[name = string("op_5418_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5420_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5418_cast_fp16)[name = string("op_5420_cast_fp16")];
            fp32 var_5421_epsilon_0 = const()[name = string("op_5421_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5421_cast_fp16 = rsqrt(epsilon = var_5421_epsilon_0, x = var_5420_cast_fp16)[name = string("op_5421_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_139_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5421_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59541632)))];
            tensor<fp16, [1, 256, 512]> var_5423_cast_fp16 = mul(x = out_139_cast_fp16, y = tm_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5423_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_117_cast_fp16 = add(x = x_173_cast_fp16, y = var_5423_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59542720))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59804928))))[name = string("tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_89_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_89_cast_fp16")];
            string x_177_mode_0 = const()[name = string("x_177_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_177_cast_fp16 = gelu(mode = x_177_mode_0, x = linear_89_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [4]> var_5443_begin_0 = const()[name = string("op_5443_begin_0"), val = tensor<int32, [4]>([0, 0, 9, 0])];
            tensor<int32, [4]> var_5443_end_0 = const()[name = string("op_5443_end_0"), val = tensor<int32, [4]>([1, 256, 10, 512])];
            tensor<bool, [4]> var_5443_end_mask_0 = const()[name = string("op_5443_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5443_squeeze_mask_0 = const()[name = string("op_5443_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_5443_cast_fp16 = slice_by_index(begin = var_5443_begin_0, end = var_5443_end_0, end_mask = var_5443_end_mask_0, squeeze_mask = var_5443_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5443_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_119_cast_fp16 = mul(x = x_177_cast_fp16, y = var_5443_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59806016))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60068224))))[name = string("tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5453_cast_fp16 = abs(x = linear_90_cast_fp16)[name = string("op_5453_cast_fp16")];
            bool var_5456 = const()[name = string("op_5456"), val = bool(true)];
            tensor<int32, [1]> var_5457_axes_0 = const()[name = string("op_5457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5457_cast_fp16 = reduce_max(axes = var_5457_axes_0, keep_dims = var_5456, x = var_5453_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_161_cast_fp16 = maximum(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_161_cast_fp16 = real_div(x = linear_90_cast_fp16, y = s_161_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5461_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<int32, [1]> var_5466_axes_0 = const()[name = string("op_5466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5466_keep_dims_0 = const()[name = string("op_5466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5466_cast_fp16 = reduce_mean(axes = var_5466_axes_0, keep_dims = var_5466_keep_dims_0, x = var_5461_cast_fp16)[name = string("op_5466_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5467_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5467_cast_fp16")];
            fp32 var_5468_epsilon_0 = const()[name = string("op_5468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5468_cast_fp16 = inverse(epsilon = var_5468_epsilon_0, x = var_5467_cast_fp16)[name = string("op_5468_cast_fp16")];
            fp16 var_5469_to_fp16 = const()[name = string("op_5469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5470_cast_fp16 = mul(x = var_5468_cast_fp16, y = var_5469_to_fp16)[name = string("op_5470_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5472_cast_fp16 = add(x = var_5466_cast_fp16, y = var_5470_cast_fp16)[name = string("op_5472_cast_fp16")];
            fp32 var_5473_epsilon_0 = const()[name = string("op_5473_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5473_cast_fp16 = rsqrt(epsilon = var_5473_epsilon_0, x = var_5472_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_141_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5473_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60069312)))];
            tensor<fp16, [1, 256, 512]> var_5475_cast_fp16 = mul(x = out_141_cast_fp16, y = tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5475_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_19_cast_fp16 = add(x = input_117_cast_fp16, y = var_5475_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<fp16, [1]> tm_layers_9_layer_scalar_to_fp16 = const()[name = string("tm_layers_9_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9ep-1])];
            tensor<fp16, [1, 256, 512]> x_181_cast_fp16 = mul(x = h_19_cast_fp16, y = tm_layers_9_layer_scalar_to_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5479_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_5479_cast_fp16")];
            bool var_5482 = const()[name = string("op_5482"), val = bool(true)];
            tensor<int32, [1]> var_5483_axes_0 = const()[name = string("op_5483_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5483_cast_fp16 = reduce_max(axes = var_5483_axes_0, keep_dims = var_5482, x = var_5479_cast_fp16)[name = string("op_5483_cast_fp16")];
            fp16 var_5484_to_fp16 = const()[name = string("op_5484_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_163_cast_fp16 = maximum(x = var_5483_cast_fp16, y = var_5484_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_163_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_163_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5487_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5492_axes_0 = const()[name = string("op_5492_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5492_keep_dims_0 = const()[name = string("op_5492_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5492_cast_fp16 = reduce_mean(axes = var_5492_axes_0, keep_dims = var_5492_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5493_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5493_cast_fp16")];
            fp32 var_5494_epsilon_0 = const()[name = string("op_5494_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5494_cast_fp16 = inverse(epsilon = var_5494_epsilon_0, x = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp16 var_5495_to_fp16 = const()[name = string("op_5495_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5496_cast_fp16 = mul(x = var_5494_cast_fp16, y = var_5495_to_fp16)[name = string("op_5496_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5498_cast_fp16 = add(x = var_5492_cast_fp16, y = var_5496_cast_fp16)[name = string("op_5498_cast_fp16")];
            fp32 var_5499_epsilon_0 = const()[name = string("op_5499_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5499_cast_fp16 = rsqrt(epsilon = var_5499_epsilon_0, x = var_5498_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_143_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5499_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60070400)))];
            tensor<fp16, [1, 256, 512]> input_121_cast_fp16 = mul(x = out_143_cast_fp16, y = tm_layers_10_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60071488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60595840))))[name = string("tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5509 = const()[name = string("op_5509"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_5510_cast_fp16 = reshape(shape = var_5509, x = linear_91_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60597952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60860160))))[name = string("tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5521 = const()[name = string("op_5521"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_5522_cast_fp16 = reshape(shape = var_5521, x = linear_92_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60861248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61123456))))[name = string("tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5533 = const()[name = string("op_5533"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_5534_cast_fp16 = reshape(shape = var_5533, x = linear_93_cast_fp16)[name = string("op_5534_cast_fp16")];
            tensor<int32, [4]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5510_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 4, 256, 256]> var_5538_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_5538_cast_fp16")];
            bool var_5541 = const()[name = string("op_5541"), val = bool(true)];
            tensor<int32, [1]> var_5542_axes_0 = const()[name = string("op_5542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_5542_cast_fp16 = reduce_max(axes = var_5542_axes_0, keep_dims = var_5541, x = var_5538_cast_fp16)[name = string("op_5542_cast_fp16")];
            fp16 var_5543_to_fp16 = const()[name = string("op_5543_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_165_cast_fp16 = maximum(x = var_5542_cast_fp16, y = var_5543_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_165_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_165_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5546_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<int32, [1]> var_5551_axes_0 = const()[name = string("op_5551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5551_keep_dims_0 = const()[name = string("op_5551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_5551_cast_fp16 = reduce_mean(axes = var_5551_axes_0, keep_dims = var_5551_keep_dims_0, x = var_5546_cast_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5552_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5552_cast_fp16")];
            fp32 var_5553_epsilon_0 = const()[name = string("op_5553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_5553_cast_fp16 = inverse(epsilon = var_5553_epsilon_0, x = var_5552_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp16 var_5554_to_fp16 = const()[name = string("op_5554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_5555_cast_fp16 = mul(x = var_5553_cast_fp16, y = var_5554_to_fp16)[name = string("op_5555_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5557_cast_fp16 = add(x = var_5551_cast_fp16, y = var_5555_cast_fp16)[name = string("op_5557_cast_fp16")];
            fp32 var_5558_epsilon_0 = const()[name = string("op_5558_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_5558_cast_fp16 = rsqrt(epsilon = var_5558_epsilon_0, x = var_5557_cast_fp16)[name = string("op_5558_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_145_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5558_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61124544)))];
            tensor<fp16, [1, 4, 256, 256]> q_41_cast_fp16 = mul(x = out_145_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_41_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_5522_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2, 256, 256]> var_5561_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_5561_cast_fp16")];
            bool var_5564 = const()[name = string("op_5564"), val = bool(true)];
            tensor<int32, [1]> var_5565_axes_0 = const()[name = string("op_5565_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_5565_cast_fp16 = reduce_max(axes = var_5565_axes_0, keep_dims = var_5564, x = var_5561_cast_fp16)[name = string("op_5565_cast_fp16")];
            fp16 var_5566_to_fp16 = const()[name = string("op_5566_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_167_cast_fp16 = maximum(x = var_5565_cast_fp16, y = var_5566_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_167_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_167_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5569_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5569_cast_fp16")];
            tensor<int32, [1]> var_5574_axes_0 = const()[name = string("op_5574_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5574_keep_dims_0 = const()[name = string("op_5574_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_5574_cast_fp16 = reduce_mean(axes = var_5574_axes_0, keep_dims = var_5574_keep_dims_0, x = var_5569_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5575_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5575_cast_fp16")];
            fp32 var_5576_epsilon_0 = const()[name = string("op_5576_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_5576_cast_fp16 = inverse(epsilon = var_5576_epsilon_0, x = var_5575_cast_fp16)[name = string("op_5576_cast_fp16")];
            fp16 var_5577_to_fp16 = const()[name = string("op_5577_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_5578_cast_fp16 = mul(x = var_5576_cast_fp16, y = var_5577_to_fp16)[name = string("op_5578_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5580_cast_fp16 = add(x = var_5574_cast_fp16, y = var_5578_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_5581_cast_fp16 = rsqrt(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_147_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5581_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [256]> tm_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125120)))];
            tensor<fp16, [1, 2, 256, 256]> k_61_cast_fp16 = mul(x = out_147_cast_fp16, y = tm_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = var_5534_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 2, 256, 256]> var_5584_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_5584_cast_fp16")];
            bool var_5587 = const()[name = string("op_5587"), val = bool(true)];
            tensor<int32, [1]> var_5588_axes_0 = const()[name = string("op_5588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_5588_cast_fp16 = reduce_max(axes = var_5588_axes_0, keep_dims = var_5587, x = var_5584_cast_fp16)[name = string("op_5588_cast_fp16")];
            fp16 var_5589_to_fp16 = const()[name = string("op_5589_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_169_cast_fp16 = maximum(x = var_5588_cast_fp16, y = var_5589_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_169_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_169_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5592_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5592_cast_fp16")];
            tensor<int32, [1]> var_5597_axes_0 = const()[name = string("op_5597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5597_keep_dims_0 = const()[name = string("op_5597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_5597_cast_fp16 = reduce_mean(axes = var_5597_axes_0, keep_dims = var_5597_keep_dims_0, x = var_5592_cast_fp16)[name = string("op_5597_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5598_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5598_cast_fp16")];
            fp32 var_5599_epsilon_0 = const()[name = string("op_5599_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_5599_cast_fp16 = inverse(epsilon = var_5599_epsilon_0, x = var_5598_cast_fp16)[name = string("op_5599_cast_fp16")];
            fp16 var_5600_to_fp16 = const()[name = string("op_5600_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_5601_cast_fp16 = mul(x = var_5599_cast_fp16, y = var_5600_to_fp16)[name = string("op_5601_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5603_cast_fp16 = add(x = var_5597_cast_fp16, y = var_5601_cast_fp16)[name = string("op_5603_cast_fp16")];
            fp32 var_5604_epsilon_0 = const()[name = string("op_5604_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_5604_cast_fp16 = rsqrt(epsilon = var_5604_epsilon_0, x = var_5603_cast_fp16)[name = string("op_5604_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_41_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5604_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5606_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5611_begin_0 = const()[name = string("op_5611_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5611_end_0 = const()[name = string("op_5611_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_5611_end_mask_0 = const()[name = string("op_5611_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_5611_cast_fp16 = slice_by_index(begin = var_5611_begin_0, end = var_5611_end_0, end_mask = var_5611_end_mask_0, x = q_41_cast_fp16)[name = string("op_5611_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_5612_cast_fp16 = mul(x = var_5611_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_5612_cast_fp16")];
            tensor<int32, [4]> var_5617_begin_0 = const()[name = string("op_5617_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5617_end_0 = const()[name = string("op_5617_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_5617_end_mask_0 = const()[name = string("op_5617_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_5617_cast_fp16 = slice_by_index(begin = var_5617_begin_0, end = var_5617_end_0, end_mask = var_5617_end_mask_0, x = q_41_cast_fp16)[name = string("op_5617_cast_fp16")];
            int32 var_5619 = const()[name = string("op_5619"), val = int32(-1)];
            bool var_5620_interleave_0 = const()[name = string("op_5620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_5620_cast_fp16 = concat(axis = var_5619, interleave = var_5620_interleave_0, values = (var_5612_cast_fp16, var_5617_cast_fp16))[name = string("op_5620_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5621_cast_fp16 = mul(x = var_5620_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5621_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_43_cast_fp16 = add(x = var_5606_cast_fp16, y = var_5621_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5624_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = k_61_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = k_61_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool var_5638_interleave_0 = const()[name = string("op_5638_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_5638_cast_fp16 = concat(axis = var_5637, interleave = var_5638_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("op_5638_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5639_cast_fp16 = mul(x = var_5638_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_63_cast_fp16 = add(x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_20_reps_0 = const()[name = string("tile_20_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = k_63_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [4, 1, 256, 256]> tile_20_cast_fp16 = tile(reps = tile_20_reps_0, x = transpose_40_cast_fp16)[name = string("tile_20_cast_fp16")];
            tensor<int32, [5]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_40_cast_fp16 = reshape(shape = concat_40, x = tile_20_cast_fp16)[name = string("reshape_40_cast_fp16")];
            tensor<int32, [5]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = reshape_40_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [4, 1, 256, 256]> reshape_41_cast_fp16 = reshape(shape = concat_41, x = transpose_41_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_21_reps_0 = const()[name = string("tile_21_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = v_41_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [4, 1, 256, 256]> tile_21_cast_fp16 = tile(reps = tile_21_reps_0, x = transpose_42_cast_fp16)[name = string("tile_21_cast_fp16")];
            tensor<int32, [5]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_42_cast_fp16 = reshape(shape = concat_42, x = tile_21_cast_fp16)[name = string("reshape_42_cast_fp16")];
            tensor<int32, [5]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = reshape_42_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [4, 1, 256, 256]> reshape_43_cast_fp16 = reshape(shape = concat_43, x = transpose_43_cast_fp16)[name = string("reshape_43_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5653_transpose_x_0 = const()[name = string("op_5653_transpose_x_0"), val = bool(false)];
            bool var_5653_transpose_y_0 = const()[name = string("op_5653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 4, 256, 256]> var_5653_cast_fp16 = matmul(transpose_x = var_5653_transpose_x_0, transpose_y = var_5653_transpose_y_0, x = q_43_cast_fp16, y = transpose_106_cast_fp16)[name = string("op_5653_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5655_cast_fp16 = add(x = var_5653_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5655_cast_fp16")];
            int32 var_5656 = const()[name = string("op_5656"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_169_cast_fp16 = softmax(axis = var_5656, x = var_5655_cast_fp16)[name = string("w_169_cast_fp16")];
            bool var_5659_transpose_x_0 = const()[name = string("op_5659_transpose_x_0"), val = bool(false)];
            bool var_5659_transpose_y_0 = const()[name = string("op_5659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = reshape_43_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 4, 256, 256]> var_5659_cast_fp16 = matmul(transpose_x = var_5659_transpose_x_0, transpose_y = var_5659_transpose_y_0, x = w_169_cast_fp16, y = v_43_cast_fp16)[name = string("op_5659_cast_fp16")];
            tensor<int32, [4]> var_5662_perm_0 = const()[name = string("op_5662_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5666 = const()[name = string("op_5666"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_5662_cast_fp16 = transpose(perm = var_5662_perm_0, x = var_5659_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 256, 1024]> input_123_cast_fp16 = reshape(shape = var_5666, x = var_5662_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61650048))))[name = string("tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5671_cast_fp16 = abs(x = linear_94_cast_fp16)[name = string("op_5671_cast_fp16")];
            bool var_5674 = const()[name = string("op_5674"), val = bool(true)];
            tensor<int32, [1]> var_5675_axes_0 = const()[name = string("op_5675_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5675_cast_fp16 = reduce_max(axes = var_5675_axes_0, keep_dims = var_5674, x = var_5671_cast_fp16)[name = string("op_5675_cast_fp16")];
            fp16 var_5676_to_fp16 = const()[name = string("op_5676_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_171_cast_fp16 = maximum(x = var_5675_cast_fp16, y = var_5676_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_171_cast_fp16 = real_div(x = linear_94_cast_fp16, y = s_171_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5679_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5679_cast_fp16")];
            tensor<int32, [1]> var_5684_axes_0 = const()[name = string("op_5684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5684_keep_dims_0 = const()[name = string("op_5684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5684_cast_fp16 = reduce_mean(axes = var_5684_axes_0, keep_dims = var_5684_keep_dims_0, x = var_5679_cast_fp16)[name = string("op_5684_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5685_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5685_cast_fp16")];
            fp32 var_5686_epsilon_0 = const()[name = string("op_5686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5686_cast_fp16 = inverse(epsilon = var_5686_epsilon_0, x = var_5685_cast_fp16)[name = string("op_5686_cast_fp16")];
            fp16 var_5687_to_fp16 = const()[name = string("op_5687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5688_cast_fp16 = mul(x = var_5686_cast_fp16, y = var_5687_to_fp16)[name = string("op_5688_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5690_cast_fp16 = add(x = var_5684_cast_fp16, y = var_5688_cast_fp16)[name = string("op_5690_cast_fp16")];
            fp32 var_5691_epsilon_0 = const()[name = string("op_5691_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5691_cast_fp16 = rsqrt(epsilon = var_5691_epsilon_0, x = var_5690_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_149_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5691_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61651136)))];
            tensor<fp16, [1, 256, 512]> var_5693_cast_fp16 = mul(x = out_149_cast_fp16, y = tm_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5693_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = var_5693_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5696_cast_fp16 = abs(x = x_191_cast_fp16)[name = string("op_5696_cast_fp16")];
            bool var_5699 = const()[name = string("op_5699"), val = bool(true)];
            tensor<int32, [1]> var_5700_axes_0 = const()[name = string("op_5700_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5700_cast_fp16 = reduce_max(axes = var_5700_axes_0, keep_dims = var_5699, x = var_5696_cast_fp16)[name = string("op_5700_cast_fp16")];
            fp16 var_5701_to_fp16 = const()[name = string("op_5701_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_173_cast_fp16 = maximum(x = var_5700_cast_fp16, y = var_5701_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_173_cast_fp16 = real_div(x = x_191_cast_fp16, y = s_173_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5704_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5704_cast_fp16")];
            tensor<int32, [1]> var_5709_axes_0 = const()[name = string("op_5709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5709_keep_dims_0 = const()[name = string("op_5709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5709_cast_fp16 = reduce_mean(axes = var_5709_axes_0, keep_dims = var_5709_keep_dims_0, x = var_5704_cast_fp16)[name = string("op_5709_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5710_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5710_cast_fp16")];
            fp32 var_5711_epsilon_0 = const()[name = string("op_5711_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5711_cast_fp16 = inverse(epsilon = var_5711_epsilon_0, x = var_5710_cast_fp16)[name = string("op_5711_cast_fp16")];
            fp16 var_5712_to_fp16 = const()[name = string("op_5712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = string("op_5713_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5715_cast_fp16 = add(x = var_5709_cast_fp16, y = var_5713_cast_fp16)[name = string("op_5715_cast_fp16")];
            fp32 var_5716_epsilon_0 = const()[name = string("op_5716_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5716_cast_fp16 = rsqrt(epsilon = var_5716_epsilon_0, x = var_5715_cast_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_151_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5716_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61652224)))];
            tensor<fp16, [1, 256, 512]> input_125_cast_fp16 = mul(x = out_151_cast_fp16, y = tm_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61653312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62701952))))[name = string("tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_95_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_95_cast_fp16")];
            string var_5723_mode_0 = const()[name = string("op_5723_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_5723_cast_fp16 = gelu(mode = var_5723_mode_0, x = linear_95_cast_fp16)[name = string("op_5723_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62706112))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63754752))))[name = string("tm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_127_cast_fp16 = mul(x = var_5723_cast_fp16, y = linear_96_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63758912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64807552))))[name = string("tm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5731_cast_fp16 = abs(x = linear_97_cast_fp16)[name = string("op_5731_cast_fp16")];
            bool var_5734 = const()[name = string("op_5734"), val = bool(true)];
            tensor<int32, [1]> var_5735_axes_0 = const()[name = string("op_5735_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5735_cast_fp16 = reduce_max(axes = var_5735_axes_0, keep_dims = var_5734, x = var_5731_cast_fp16)[name = string("op_5735_cast_fp16")];
            fp16 var_5736_to_fp16 = const()[name = string("op_5736_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_175_cast_fp16 = maximum(x = var_5735_cast_fp16, y = var_5736_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_175_cast_fp16 = real_div(x = linear_97_cast_fp16, y = s_175_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5739_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [1]> var_5744_axes_0 = const()[name = string("op_5744_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5744_keep_dims_0 = const()[name = string("op_5744_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5744_cast_fp16 = reduce_mean(axes = var_5744_axes_0, keep_dims = var_5744_keep_dims_0, x = var_5739_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5745_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5745_cast_fp16")];
            fp32 var_5746_epsilon_0 = const()[name = string("op_5746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5746_cast_fp16 = inverse(epsilon = var_5746_epsilon_0, x = var_5745_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5747_to_fp16 = const()[name = string("op_5747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5748_cast_fp16 = mul(x = var_5746_cast_fp16, y = var_5747_to_fp16)[name = string("op_5748_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5750_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5748_cast_fp16)[name = string("op_5750_cast_fp16")];
            fp32 var_5751_epsilon_0 = const()[name = string("op_5751_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5751_cast_fp16 = rsqrt(epsilon = var_5751_epsilon_0, x = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_153_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5751_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64808640)))];
            tensor<fp16, [1, 256, 512]> var_5753_cast_fp16 = mul(x = out_153_cast_fp16, y = tm_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_129_cast_fp16 = add(x = x_191_cast_fp16, y = var_5753_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64809728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65071936))))[name = string("tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_98_cast_fp16")];
            string x_195_mode_0 = const()[name = string("x_195_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_195_cast_fp16 = gelu(mode = x_195_mode_0, x = linear_98_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [4]> var_5773_begin_0 = const()[name = string("op_5773_begin_0"), val = tensor<int32, [4]>([0, 0, 10, 0])];
            tensor<int32, [4]> var_5773_end_0 = const()[name = string("op_5773_end_0"), val = tensor<int32, [4]>([1, 256, 11, 512])];
            tensor<bool, [4]> var_5773_end_mask_0 = const()[name = string("op_5773_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5773_squeeze_mask_0 = const()[name = string("op_5773_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_5773_cast_fp16 = slice_by_index(begin = var_5773_begin_0, end = var_5773_end_0, end_mask = var_5773_end_mask_0, squeeze_mask = var_5773_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_131_cast_fp16 = mul(x = x_195_cast_fp16, y = var_5773_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65073024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65335232))))[name = string("tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5783_cast_fp16 = abs(x = linear_99_cast_fp16)[name = string("op_5783_cast_fp16")];
            bool var_5786 = const()[name = string("op_5786"), val = bool(true)];
            tensor<int32, [1]> var_5787_axes_0 = const()[name = string("op_5787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5787_cast_fp16 = reduce_max(axes = var_5787_axes_0, keep_dims = var_5786, x = var_5783_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp16 var_5788_to_fp16 = const()[name = string("op_5788_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_177_cast_fp16 = maximum(x = var_5787_cast_fp16, y = var_5788_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_177_cast_fp16 = real_div(x = linear_99_cast_fp16, y = s_177_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5791_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5791_cast_fp16")];
            tensor<int32, [1]> var_5796_axes_0 = const()[name = string("op_5796_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5796_keep_dims_0 = const()[name = string("op_5796_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5796_cast_fp16 = reduce_mean(axes = var_5796_axes_0, keep_dims = var_5796_keep_dims_0, x = var_5791_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5797_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5797_cast_fp16")];
            fp32 var_5798_epsilon_0 = const()[name = string("op_5798_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5798_cast_fp16 = inverse(epsilon = var_5798_epsilon_0, x = var_5797_cast_fp16)[name = string("op_5798_cast_fp16")];
            fp16 var_5799_to_fp16 = const()[name = string("op_5799_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5800_cast_fp16 = mul(x = var_5798_cast_fp16, y = var_5799_to_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5802_cast_fp16 = add(x = var_5796_cast_fp16, y = var_5800_cast_fp16)[name = string("op_5802_cast_fp16")];
            fp32 var_5803_epsilon_0 = const()[name = string("op_5803_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5803_cast_fp16 = rsqrt(epsilon = var_5803_epsilon_0, x = var_5802_cast_fp16)[name = string("op_5803_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_155_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5803_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65336320)))];
            tensor<fp16, [1, 256, 512]> var_5805_cast_fp16 = mul(x = out_155_cast_fp16, y = tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5805_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_21_cast_fp16 = add(x = input_129_cast_fp16, y = var_5805_cast_fp16)[name = string("h_21_cast_fp16")];
            tensor<fp16, [1]> tm_layers_10_layer_scalar_to_fp16 = const()[name = string("tm_layers_10_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.1cp-1])];
            tensor<fp16, [1, 256, 512]> x_199_cast_fp16 = mul(x = h_21_cast_fp16, y = tm_layers_10_layer_scalar_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5809_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5809_cast_fp16")];
            bool var_5812 = const()[name = string("op_5812"), val = bool(true)];
            tensor<int32, [1]> var_5813_axes_0 = const()[name = string("op_5813_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5813_cast_fp16 = reduce_max(axes = var_5813_axes_0, keep_dims = var_5812, x = var_5809_cast_fp16)[name = string("op_5813_cast_fp16")];
            fp16 var_5814_to_fp16 = const()[name = string("op_5814_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_179_cast_fp16 = maximum(x = var_5813_cast_fp16, y = var_5814_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_179_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_179_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5817_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5817_cast_fp16")];
            tensor<int32, [1]> var_5822_axes_0 = const()[name = string("op_5822_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5822_keep_dims_0 = const()[name = string("op_5822_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5822_cast_fp16 = reduce_mean(axes = var_5822_axes_0, keep_dims = var_5822_keep_dims_0, x = var_5817_cast_fp16)[name = string("op_5822_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5823_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp32 var_5824_epsilon_0 = const()[name = string("op_5824_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5824_cast_fp16 = inverse(epsilon = var_5824_epsilon_0, x = var_5823_cast_fp16)[name = string("op_5824_cast_fp16")];
            fp16 var_5825_to_fp16 = const()[name = string("op_5825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5826_cast_fp16 = mul(x = var_5824_cast_fp16, y = var_5825_to_fp16)[name = string("op_5826_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5828_cast_fp16 = add(x = var_5822_cast_fp16, y = var_5826_cast_fp16)[name = string("op_5828_cast_fp16")];
            fp32 var_5829_epsilon_0 = const()[name = string("op_5829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5829_cast_fp16 = rsqrt(epsilon = var_5829_epsilon_0, x = var_5828_cast_fp16)[name = string("op_5829_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_157_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5829_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65337408)))];
            tensor<fp16, [1, 256, 512]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = tm_layers_11_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65338496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66387136))))[name = string("tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_100_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5839 = const()[name = string("op_5839"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 256, 4, 512]> var_5840_cast_fp16 = reshape(shape = var_5839, x = linear_100_cast_fp16)[name = string("op_5840_cast_fp16")];
            tensor<int32, [4]> x_201_perm_0 = const()[name = string("x_201_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66391296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66653504))))[name = string("tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_5851 = const()[name = string("op_5851"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_5852_cast_fp16 = reshape(shape = var_5851, x = linear_101_cast_fp16)[name = string("op_5852_cast_fp16")];
            tensor<int32, [4]> x_203_perm_0 = const()[name = string("x_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66654592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66916800))))[name = string("tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_5864_cast_fp16 = reshape(shape = var_5863, x = linear_102_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 512]> x_201_cast_fp16 = transpose(perm = x_201_perm_0, x = var_5840_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 4, 256, 512]> var_5868_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5868_cast_fp16")];
            bool var_5871 = const()[name = string("op_5871"), val = bool(true)];
            tensor<int32, [1]> var_5872_axes_0 = const()[name = string("op_5872_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_5872_cast_fp16 = reduce_max(axes = var_5872_axes_0, keep_dims = var_5871, x = var_5868_cast_fp16)[name = string("op_5872_cast_fp16")];
            fp16 var_5873_to_fp16 = const()[name = string("op_5873_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_181_cast_fp16 = maximum(x = var_5872_cast_fp16, y = var_5873_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> y_181_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_181_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_5876_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5876_cast_fp16")];
            tensor<int32, [1]> var_5881_axes_0 = const()[name = string("op_5881_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5881_keep_dims_0 = const()[name = string("op_5881_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_5881_cast_fp16 = reduce_mean(axes = var_5881_axes_0, keep_dims = var_5881_keep_dims_0, x = var_5876_cast_fp16)[name = string("op_5881_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5882_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5882_cast_fp16")];
            fp32 var_5883_epsilon_0 = const()[name = string("op_5883_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_5883_cast_fp16 = inverse(epsilon = var_5883_epsilon_0, x = var_5882_cast_fp16)[name = string("op_5883_cast_fp16")];
            fp16 var_5884_to_fp16 = const()[name = string("op_5884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_5885_cast_fp16 = mul(x = var_5883_cast_fp16, y = var_5884_to_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5887_cast_fp16 = add(x = var_5881_cast_fp16, y = var_5885_cast_fp16)[name = string("op_5887_cast_fp16")];
            fp32 var_5888_epsilon_0 = const()[name = string("op_5888_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_5888_cast_fp16 = rsqrt(epsilon = var_5888_epsilon_0, x = var_5887_cast_fp16)[name = string("op_5888_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> out_159_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5888_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66917888)))];
            tensor<fp16, [1, 4, 256, 512]> q_45_cast_fp16 = mul(x = out_159_cast_fp16, y = tm_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_203_cast_fp16 = transpose(perm = x_203_perm_0, x = var_5852_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1, 256, 512]> var_5891_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5891_cast_fp16")];
            bool var_5894 = const()[name = string("op_5894"), val = bool(true)];
            tensor<int32, [1]> var_5895_axes_0 = const()[name = string("op_5895_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_5895_cast_fp16 = reduce_max(axes = var_5895_axes_0, keep_dims = var_5894, x = var_5891_cast_fp16)[name = string("op_5895_cast_fp16")];
            fp16 var_5896_to_fp16 = const()[name = string("op_5896_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_183_cast_fp16 = maximum(x = var_5895_cast_fp16, y = var_5896_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_183_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_183_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_5899_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5899_cast_fp16")];
            tensor<int32, [1]> var_5904_axes_0 = const()[name = string("op_5904_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5904_keep_dims_0 = const()[name = string("op_5904_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_5904_cast_fp16 = reduce_mean(axes = var_5904_axes_0, keep_dims = var_5904_keep_dims_0, x = var_5899_cast_fp16)[name = string("op_5904_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_5905_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5905_cast_fp16")];
            fp32 var_5906_epsilon_0 = const()[name = string("op_5906_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_5906_cast_fp16 = inverse(epsilon = var_5906_epsilon_0, x = var_5905_cast_fp16)[name = string("op_5906_cast_fp16")];
            fp16 var_5907_to_fp16 = const()[name = string("op_5907_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_5908_cast_fp16 = mul(x = var_5906_cast_fp16, y = var_5907_to_fp16)[name = string("op_5908_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_5910_cast_fp16 = add(x = var_5904_cast_fp16, y = var_5908_cast_fp16)[name = string("op_5910_cast_fp16")];
            fp32 var_5911_epsilon_0 = const()[name = string("op_5911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_5911_cast_fp16 = rsqrt(epsilon = var_5911_epsilon_0, x = var_5910_cast_fp16)[name = string("op_5911_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> out_161_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5911_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66918976)))];
            tensor<fp16, [1, 1, 256, 512]> k_67_cast_fp16 = mul(x = out_161_cast_fp16, y = tm_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_5864_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1, 256, 512]> var_5914_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5914_cast_fp16")];
            bool var_5917 = const()[name = string("op_5917"), val = bool(true)];
            tensor<int32, [1]> var_5918_axes_0 = const()[name = string("op_5918_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_5918_cast_fp16 = reduce_max(axes = var_5918_axes_0, keep_dims = var_5917, x = var_5914_cast_fp16)[name = string("op_5918_cast_fp16")];
            fp16 var_5919_to_fp16 = const()[name = string("op_5919_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_185_cast_fp16 = maximum(x = var_5918_cast_fp16, y = var_5919_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_185_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_185_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_5922_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5922_cast_fp16")];
            tensor<int32, [1]> var_5927_axes_0 = const()[name = string("op_5927_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5927_keep_dims_0 = const()[name = string("op_5927_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_5927_cast_fp16 = reduce_mean(axes = var_5927_axes_0, keep_dims = var_5927_keep_dims_0, x = var_5922_cast_fp16)[name = string("op_5927_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_5928_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5928_cast_fp16")];
            fp32 var_5929_epsilon_0 = const()[name = string("op_5929_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_5929_cast_fp16 = inverse(epsilon = var_5929_epsilon_0, x = var_5928_cast_fp16)[name = string("op_5929_cast_fp16")];
            fp16 var_5930_to_fp16 = const()[name = string("op_5930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_5931_cast_fp16 = mul(x = var_5929_cast_fp16, y = var_5930_to_fp16)[name = string("op_5931_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_5933_cast_fp16 = add(x = var_5927_cast_fp16, y = var_5931_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp32 var_5934_epsilon_0 = const()[name = string("op_5934_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_5934_cast_fp16 = rsqrt(epsilon = var_5934_epsilon_0, x = var_5933_cast_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> v_45_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5934_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_5936_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5936_cast_fp16")];
            tensor<int32, [4]> var_5941_begin_0 = const()[name = string("op_5941_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5941_end_0 = const()[name = string("op_5941_end_0"), val = tensor<int32, [4]>([1, 4, 256, 512])];
            tensor<bool, [4]> var_5941_end_mask_0 = const()[name = string("op_5941_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 256]> var_5941_cast_fp16 = slice_by_index(begin = var_5941_begin_0, end = var_5941_end_0, end_mask = var_5941_end_mask_0, x = q_45_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 256]> var_5942_cast_fp16 = mul(x = var_5941_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_5942_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 256]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = q_45_cast_fp16)[name = string("op_5947_cast_fp16")];
            int32 var_5949 = const()[name = string("op_5949"), val = int32(-1)];
            bool var_5950_interleave_0 = const()[name = string("op_5950_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> var_5950_cast_fp16 = concat(axis = var_5949, interleave = var_5950_interleave_0, values = (var_5942_cast_fp16, var_5947_cast_fp16))[name = string("op_5950_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_5951_cast_fp16 = mul(x = var_5950_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5951_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> q_47_cast_fp16 = add(x = var_5936_cast_fp16, y = var_5951_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_5954_cast_fp16 = mul(x = k_67_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5954_cast_fp16")];
            tensor<int32, [4]> var_5959_begin_0 = const()[name = string("op_5959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5959_end_0 = const()[name = string("op_5959_end_0"), val = tensor<int32, [4]>([1, 1, 256, 512])];
            tensor<bool, [4]> var_5959_end_mask_0 = const()[name = string("op_5959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 256, 256]> var_5959_cast_fp16 = slice_by_index(begin = var_5959_begin_0, end = var_5959_end_0, end_mask = var_5959_end_mask_0, x = k_67_cast_fp16)[name = string("op_5959_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 256, 256]> var_5960_cast_fp16 = mul(x = var_5959_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_5960_cast_fp16")];
            tensor<int32, [4]> var_5965_begin_0 = const()[name = string("op_5965_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5965_end_0 = const()[name = string("op_5965_end_0"), val = tensor<int32, [4]>([1, 1, 256, 256])];
            tensor<bool, [4]> var_5965_end_mask_0 = const()[name = string("op_5965_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 256, 256]> var_5965_cast_fp16 = slice_by_index(begin = var_5965_begin_0, end = var_5965_end_0, end_mask = var_5965_end_mask_0, x = k_67_cast_fp16)[name = string("op_5965_cast_fp16")];
            int32 var_5967 = const()[name = string("op_5967"), val = int32(-1)];
            bool var_5968_interleave_0 = const()[name = string("op_5968_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 256, 512]> var_5968_cast_fp16 = concat(axis = var_5967, interleave = var_5968_interleave_0, values = (var_5960_cast_fp16, var_5965_cast_fp16))[name = string("op_5968_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_5969_cast_fp16 = mul(x = var_5968_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5969_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> k_69_cast_fp16 = add(x = var_5954_cast_fp16, y = var_5969_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_22_reps_0 = const()[name = string("tile_22_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = k_69_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [4, 1, 256, 512]> tile_22_cast_fp16 = tile(reps = tile_22_reps_0, x = transpose_44_cast_fp16)[name = string("tile_22_cast_fp16")];
            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_44_cast_fp16 = reshape(shape = concat_44, x = tile_22_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [5]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [4, 1, 256, 512]> reshape_45_cast_fp16 = reshape(shape = concat_45, x = transpose_45_cast_fp16)[name = string("reshape_45_cast_fp16")];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_23_reps_0 = const()[name = string("tile_23_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = v_45_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [4, 1, 256, 512]> tile_23_cast_fp16 = tile(reps = tile_23_reps_0, x = transpose_46_cast_fp16)[name = string("tile_23_cast_fp16")];
            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_46_cast_fp16 = reshape(shape = concat_46, x = tile_23_cast_fp16)[name = string("reshape_46_cast_fp16")];
            tensor<int32, [5]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = reshape_46_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [4, 1, 256, 512]> reshape_47_cast_fp16 = reshape(shape = concat_47, x = transpose_47_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5983_transpose_x_0 = const()[name = string("op_5983_transpose_x_0"), val = bool(false)];
            bool var_5983_transpose_y_0 = const()[name = string("op_5983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = reshape_45_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 4, 256, 256]> var_5983_cast_fp16 = matmul(transpose_x = var_5983_transpose_x_0, transpose_y = var_5983_transpose_y_0, x = q_47_cast_fp16, y = transpose_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5985_cast_fp16 = add(x = var_5983_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5985_cast_fp16")];
            int32 var_5986 = const()[name = string("op_5986"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_185_cast_fp16 = softmax(axis = var_5986, x = var_5985_cast_fp16)[name = string("w_185_cast_fp16")];
            bool var_5989_transpose_x_0 = const()[name = string("op_5989_transpose_x_0"), val = bool(false)];
            bool var_5989_transpose_y_0 = const()[name = string("op_5989_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 4, 256, 512]> var_5989_cast_fp16 = matmul(transpose_x = var_5989_transpose_x_0, transpose_y = var_5989_transpose_y_0, x = w_185_cast_fp16, y = v_47_cast_fp16)[name = string("op_5989_cast_fp16")];
            tensor<int32, [4]> var_5992_perm_0 = const()[name = string("op_5992_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5996 = const()[name = string("op_5996"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 256, 4, 512]> var_5992_cast_fp16 = transpose(perm = var_5992_perm_0, x = var_5989_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 256, 2048]> input_135_cast_fp16 = reshape(shape = var_5996, x = var_5992_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66920064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67968704))))[name = string("tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6001_cast_fp16 = abs(x = linear_103_cast_fp16)[name = string("op_6001_cast_fp16")];
            bool var_6004 = const()[name = string("op_6004"), val = bool(true)];
            tensor<int32, [1]> var_6005_axes_0 = const()[name = string("op_6005_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6005_cast_fp16 = reduce_max(axes = var_6005_axes_0, keep_dims = var_6004, x = var_6001_cast_fp16)[name = string("op_6005_cast_fp16")];
            fp16 var_6006_to_fp16 = const()[name = string("op_6006_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_187_cast_fp16 = maximum(x = var_6005_cast_fp16, y = var_6006_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_187_cast_fp16 = real_div(x = linear_103_cast_fp16, y = s_187_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6009_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_6009_cast_fp16")];
            tensor<int32, [1]> var_6014_axes_0 = const()[name = string("op_6014_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6014_keep_dims_0 = const()[name = string("op_6014_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6014_cast_fp16 = reduce_mean(axes = var_6014_axes_0, keep_dims = var_6014_keep_dims_0, x = var_6009_cast_fp16)[name = string("op_6014_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6015_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_6015_cast_fp16")];
            fp32 var_6016_epsilon_0 = const()[name = string("op_6016_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6016_cast_fp16 = inverse(epsilon = var_6016_epsilon_0, x = var_6015_cast_fp16)[name = string("op_6016_cast_fp16")];
            fp16 var_6017_to_fp16 = const()[name = string("op_6017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6018_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6020_cast_fp16 = add(x = var_6014_cast_fp16, y = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp32 var_6021_epsilon_0 = const()[name = string("op_6021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6021_cast_fp16 = rsqrt(epsilon = var_6021_epsilon_0, x = var_6020_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_163_cast_fp16 = mul(x = y_187_cast_fp16, y = var_6021_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969792)))];
            tensor<fp16, [1, 256, 512]> var_6023_cast_fp16 = mul(x = out_163_cast_fp16, y = tm_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_209_cast_fp16 = add(x = x_199_cast_fp16, y = var_6023_cast_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6026_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_6026_cast_fp16")];
            bool var_6029 = const()[name = string("op_6029"), val = bool(true)];
            tensor<int32, [1]> var_6030_axes_0 = const()[name = string("op_6030_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6030_cast_fp16 = reduce_max(axes = var_6030_axes_0, keep_dims = var_6029, x = var_6026_cast_fp16)[name = string("op_6030_cast_fp16")];
            fp16 var_6031_to_fp16 = const()[name = string("op_6031_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_189_cast_fp16 = maximum(x = var_6030_cast_fp16, y = var_6031_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_189_cast_fp16 = real_div(x = x_209_cast_fp16, y = s_189_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6034_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [1]> var_6039_axes_0 = const()[name = string("op_6039_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6039_keep_dims_0 = const()[name = string("op_6039_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6039_cast_fp16 = reduce_mean(axes = var_6039_axes_0, keep_dims = var_6039_keep_dims_0, x = var_6034_cast_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6040_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6041_cast_fp16 = inverse(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            fp16 var_6042_to_fp16 = const()[name = string("op_6042_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6043_cast_fp16 = mul(x = var_6041_cast_fp16, y = var_6042_to_fp16)[name = string("op_6043_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6045_cast_fp16 = add(x = var_6039_cast_fp16, y = var_6043_cast_fp16)[name = string("op_6045_cast_fp16")];
            fp32 var_6046_epsilon_0 = const()[name = string("op_6046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6046_cast_fp16 = rsqrt(epsilon = var_6046_epsilon_0, x = var_6045_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_165_cast_fp16 = mul(x = y_189_cast_fp16, y = var_6046_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67970880)))];
            tensor<fp16, [1, 256, 512]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = tm_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971968))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69020608))))[name = string("tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_104_cast_fp16")];
            string var_6053_mode_0 = const()[name = string("op_6053_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_6053_cast_fp16 = gelu(mode = var_6053_mode_0, x = linear_104_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70073408))))[name = string("tm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_105_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_139_cast_fp16 = mul(x = var_6053_cast_fp16, y = linear_105_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70077568))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71126208))))[name = string("tm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6061_cast_fp16 = abs(x = linear_106_cast_fp16)[name = string("op_6061_cast_fp16")];
            bool var_6064 = const()[name = string("op_6064"), val = bool(true)];
            tensor<int32, [1]> var_6065_axes_0 = const()[name = string("op_6065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6065_cast_fp16 = reduce_max(axes = var_6065_axes_0, keep_dims = var_6064, x = var_6061_cast_fp16)[name = string("op_6065_cast_fp16")];
            fp16 var_6066_to_fp16 = const()[name = string("op_6066_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_191_cast_fp16 = maximum(x = var_6065_cast_fp16, y = var_6066_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_191_cast_fp16 = real_div(x = linear_106_cast_fp16, y = s_191_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6069_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<int32, [1]> var_6074_axes_0 = const()[name = string("op_6074_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6074_keep_dims_0 = const()[name = string("op_6074_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6074_cast_fp16 = reduce_mean(axes = var_6074_axes_0, keep_dims = var_6074_keep_dims_0, x = var_6069_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6075_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_6075_cast_fp16")];
            fp32 var_6076_epsilon_0 = const()[name = string("op_6076_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6076_cast_fp16 = inverse(epsilon = var_6076_epsilon_0, x = var_6075_cast_fp16)[name = string("op_6076_cast_fp16")];
            fp16 var_6077_to_fp16 = const()[name = string("op_6077_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6078_cast_fp16 = mul(x = var_6076_cast_fp16, y = var_6077_to_fp16)[name = string("op_6078_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6080_cast_fp16 = add(x = var_6074_cast_fp16, y = var_6078_cast_fp16)[name = string("op_6080_cast_fp16")];
            fp32 var_6081_epsilon_0 = const()[name = string("op_6081_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6081_cast_fp16 = rsqrt(epsilon = var_6081_epsilon_0, x = var_6080_cast_fp16)[name = string("op_6081_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_167_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6081_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71127296)))];
            tensor<fp16, [1, 256, 512]> var_6083_cast_fp16 = mul(x = out_167_cast_fp16, y = tm_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6083_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_141_cast_fp16 = add(x = x_209_cast_fp16, y = var_6083_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71128384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71390592))))[name = string("tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_107_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_107_cast_fp16")];
            string x_213_mode_0 = const()[name = string("x_213_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_213_cast_fp16 = gelu(mode = x_213_mode_0, x = linear_107_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [4]> var_6103_begin_0 = const()[name = string("op_6103_begin_0"), val = tensor<int32, [4]>([0, 0, 11, 0])];
            tensor<int32, [4]> var_6103_end_0 = const()[name = string("op_6103_end_0"), val = tensor<int32, [4]>([1, 256, 12, 512])];
            tensor<bool, [4]> var_6103_end_mask_0 = const()[name = string("op_6103_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6103_squeeze_mask_0 = const()[name = string("op_6103_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_6103_cast_fp16 = slice_by_index(begin = var_6103_begin_0, end = var_6103_end_0, end_mask = var_6103_end_mask_0, squeeze_mask = var_6103_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6103_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_143_cast_fp16 = mul(x = x_213_cast_fp16, y = var_6103_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71391680))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71653888))))[name = string("tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6113_cast_fp16 = abs(x = linear_108_cast_fp16)[name = string("op_6113_cast_fp16")];
            bool var_6116 = const()[name = string("op_6116"), val = bool(true)];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6117_cast_fp16 = reduce_max(axes = var_6117_axes_0, keep_dims = var_6116, x = var_6113_cast_fp16)[name = string("op_6117_cast_fp16")];
            fp16 var_6118_to_fp16 = const()[name = string("op_6118_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_193_cast_fp16 = maximum(x = var_6117_cast_fp16, y = var_6118_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_193_cast_fp16 = real_div(x = linear_108_cast_fp16, y = s_193_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6121_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6121_cast_fp16")];
            tensor<int32, [1]> var_6126_axes_0 = const()[name = string("op_6126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6126_keep_dims_0 = const()[name = string("op_6126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6126_cast_fp16 = reduce_mean(axes = var_6126_axes_0, keep_dims = var_6126_keep_dims_0, x = var_6121_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6127_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_6127_cast_fp16")];
            fp32 var_6128_epsilon_0 = const()[name = string("op_6128_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6128_cast_fp16 = inverse(epsilon = var_6128_epsilon_0, x = var_6127_cast_fp16)[name = string("op_6128_cast_fp16")];
            fp16 var_6129_to_fp16 = const()[name = string("op_6129_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6130_cast_fp16 = mul(x = var_6128_cast_fp16, y = var_6129_to_fp16)[name = string("op_6130_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6132_cast_fp16 = add(x = var_6126_cast_fp16, y = var_6130_cast_fp16)[name = string("op_6132_cast_fp16")];
            fp32 var_6133_epsilon_0 = const()[name = string("op_6133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6133_cast_fp16 = rsqrt(epsilon = var_6133_epsilon_0, x = var_6132_cast_fp16)[name = string("op_6133_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_169_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6133_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71654976)))];
            tensor<fp16, [1, 256, 512]> var_6135_cast_fp16 = mul(x = out_169_cast_fp16, y = tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6135_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_23_cast_fp16 = add(x = input_141_cast_fp16, y = var_6135_cast_fp16)[name = string("h_23_cast_fp16")];
            tensor<fp16, [1]> tm_layers_11_layer_scalar_to_fp16 = const()[name = string("tm_layers_11_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.86p-2])];
            tensor<fp16, [1, 256, 512]> x_217_cast_fp16 = mul(x = h_23_cast_fp16, y = tm_layers_11_layer_scalar_to_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6139_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_195_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_195_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_195_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6153_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_171_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71656064)))];
            tensor<fp16, [1, 256, 512]> input_145_cast_fp16 = mul(x = out_171_cast_fp16, y = tm_layers_12_input_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71657152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72181504))))[name = string("tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<int32, [4]> var_6169 = const()[name = string("op_6169"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_6170_cast_fp16 = reshape(shape = var_6169, x = linear_109_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72183616))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72445824))))[name = string("tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [4]> var_6181 = const()[name = string("op_6181"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6182_cast_fp16 = reshape(shape = var_6181, x = linear_110_cast_fp16)[name = string("op_6182_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72446912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72709120))))[name = string("tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<int32, [4]> var_6193 = const()[name = string("op_6193"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6194_cast_fp16 = reshape(shape = var_6193, x = linear_111_cast_fp16)[name = string("op_6194_cast_fp16")];
            tensor<int32, [4]> x_223_perm_0 = const()[name = string("x_223_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6170_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 4, 256, 256]> var_6198_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_6198_cast_fp16")];
            bool var_6201 = const()[name = string("op_6201"), val = bool(true)];
            tensor<int32, [1]> var_6202_axes_0 = const()[name = string("op_6202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_6202_cast_fp16 = reduce_max(axes = var_6202_axes_0, keep_dims = var_6201, x = var_6198_cast_fp16)[name = string("op_6202_cast_fp16")];
            fp16 var_6203_to_fp16 = const()[name = string("op_6203_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_197_cast_fp16 = maximum(x = var_6202_cast_fp16, y = var_6203_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_197_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_197_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6206_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6206_cast_fp16")];
            tensor<int32, [1]> var_6211_axes_0 = const()[name = string("op_6211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6211_keep_dims_0 = const()[name = string("op_6211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_6211_cast_fp16 = reduce_mean(axes = var_6211_axes_0, keep_dims = var_6211_keep_dims_0, x = var_6206_cast_fp16)[name = string("op_6211_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6212_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6212_cast_fp16")];
            fp32 var_6213_epsilon_0 = const()[name = string("op_6213_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_6213_cast_fp16 = inverse(epsilon = var_6213_epsilon_0, x = var_6212_cast_fp16)[name = string("op_6213_cast_fp16")];
            fp16 var_6214_to_fp16 = const()[name = string("op_6214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_6215_cast_fp16 = mul(x = var_6213_cast_fp16, y = var_6214_to_fp16)[name = string("op_6215_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6217_cast_fp16 = add(x = var_6211_cast_fp16, y = var_6215_cast_fp16)[name = string("op_6217_cast_fp16")];
            fp32 var_6218_epsilon_0 = const()[name = string("op_6218_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_6218_cast_fp16 = rsqrt(epsilon = var_6218_epsilon_0, x = var_6217_cast_fp16)[name = string("op_6218_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_173_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6218_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_49_cast_fp16 = mul(x = out_173_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_49_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6182_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 2, 256, 256]> var_6221_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_6221_cast_fp16")];
            bool var_6224 = const()[name = string("op_6224"), val = bool(true)];
            tensor<int32, [1]> var_6225_axes_0 = const()[name = string("op_6225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6225_cast_fp16 = reduce_max(axes = var_6225_axes_0, keep_dims = var_6224, x = var_6221_cast_fp16)[name = string("op_6225_cast_fp16")];
            fp16 var_6226_to_fp16 = const()[name = string("op_6226_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_199_cast_fp16 = maximum(x = var_6225_cast_fp16, y = var_6226_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_199_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_199_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6229_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6229_cast_fp16")];
            tensor<int32, [1]> var_6234_axes_0 = const()[name = string("op_6234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6234_keep_dims_0 = const()[name = string("op_6234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6234_cast_fp16 = reduce_mean(axes = var_6234_axes_0, keep_dims = var_6234_keep_dims_0, x = var_6229_cast_fp16)[name = string("op_6234_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6235_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6235_cast_fp16")];
            fp32 var_6236_epsilon_0 = const()[name = string("op_6236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6236_cast_fp16 = inverse(epsilon = var_6236_epsilon_0, x = var_6235_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6238_cast_fp16 = mul(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6240_cast_fp16 = add(x = var_6234_cast_fp16, y = var_6238_cast_fp16)[name = string("op_6240_cast_fp16")];
            fp32 var_6241_epsilon_0 = const()[name = string("op_6241_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6241_cast_fp16 = rsqrt(epsilon = var_6241_epsilon_0, x = var_6240_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_175_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6241_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_73_cast_fp16 = mul(x = out_175_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_223_cast_fp16 = transpose(perm = x_223_perm_0, x = var_6194_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2, 256, 256]> var_6244_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_6244_cast_fp16")];
            bool var_6247 = const()[name = string("op_6247"), val = bool(true)];
            tensor<int32, [1]> var_6248_axes_0 = const()[name = string("op_6248_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6248_cast_fp16 = reduce_max(axes = var_6248_axes_0, keep_dims = var_6247, x = var_6244_cast_fp16)[name = string("op_6248_cast_fp16")];
            fp16 var_6249_to_fp16 = const()[name = string("op_6249_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_201_cast_fp16 = maximum(x = var_6248_cast_fp16, y = var_6249_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_201_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_201_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6252_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6252_cast_fp16")];
            tensor<int32, [1]> var_6257_axes_0 = const()[name = string("op_6257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6257_keep_dims_0 = const()[name = string("op_6257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6257_cast_fp16 = reduce_mean(axes = var_6257_axes_0, keep_dims = var_6257_keep_dims_0, x = var_6252_cast_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6258_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6258_cast_fp16")];
            fp32 var_6259_epsilon_0 = const()[name = string("op_6259_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6259_cast_fp16 = inverse(epsilon = var_6259_epsilon_0, x = var_6258_cast_fp16)[name = string("op_6259_cast_fp16")];
            fp16 var_6260_to_fp16 = const()[name = string("op_6260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6261_cast_fp16 = mul(x = var_6259_cast_fp16, y = var_6260_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6263_cast_fp16 = add(x = var_6257_cast_fp16, y = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp32 var_6264_epsilon_0 = const()[name = string("op_6264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6264_cast_fp16 = rsqrt(epsilon = var_6264_epsilon_0, x = var_6263_cast_fp16)[name = string("op_6264_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_49_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6264_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6266_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6266_cast_fp16")];
            tensor<int32, [4]> var_6271_begin_0 = const()[name = string("op_6271_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6271_end_0 = const()[name = string("op_6271_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_6271_end_mask_0 = const()[name = string("op_6271_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_6271_cast_fp16 = slice_by_index(begin = var_6271_begin_0, end = var_6271_end_0, end_mask = var_6271_end_mask_0, x = q_49_cast_fp16)[name = string("op_6271_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_6272_cast_fp16 = mul(x = var_6271_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_6272_cast_fp16")];
            tensor<int32, [4]> var_6277_begin_0 = const()[name = string("op_6277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6277_end_0 = const()[name = string("op_6277_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_6277_end_mask_0 = const()[name = string("op_6277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_6277_cast_fp16 = slice_by_index(begin = var_6277_begin_0, end = var_6277_end_0, end_mask = var_6277_end_mask_0, x = q_49_cast_fp16)[name = string("op_6277_cast_fp16")];
            int32 var_6279 = const()[name = string("op_6279"), val = int32(-1)];
            bool var_6280_interleave_0 = const()[name = string("op_6280_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_6280_cast_fp16 = concat(axis = var_6279, interleave = var_6280_interleave_0, values = (var_6272_cast_fp16, var_6277_cast_fp16))[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6281_cast_fp16 = mul(x = var_6280_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6281_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_51_cast_fp16 = add(x = var_6266_cast_fp16, y = var_6281_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6284_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6284_cast_fp16")];
            tensor<int32, [4]> var_6289_begin_0 = const()[name = string("op_6289_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6289_end_0 = const()[name = string("op_6289_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_6289_end_mask_0 = const()[name = string("op_6289_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_6289_cast_fp16 = slice_by_index(begin = var_6289_begin_0, end = var_6289_end_0, end_mask = var_6289_end_mask_0, x = k_73_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_6290_cast_fp16 = mul(x = var_6289_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_6290_cast_fp16")];
            tensor<int32, [4]> var_6295_begin_0 = const()[name = string("op_6295_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6295_end_0 = const()[name = string("op_6295_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_6295_end_mask_0 = const()[name = string("op_6295_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_6295_cast_fp16 = slice_by_index(begin = var_6295_begin_0, end = var_6295_end_0, end_mask = var_6295_end_mask_0, x = k_73_cast_fp16)[name = string("op_6295_cast_fp16")];
            int32 var_6297 = const()[name = string("op_6297"), val = int32(-1)];
            bool var_6298_interleave_0 = const()[name = string("op_6298_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_6298_cast_fp16 = concat(axis = var_6297, interleave = var_6298_interleave_0, values = (var_6290_cast_fp16, var_6295_cast_fp16))[name = string("op_6298_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6299_cast_fp16 = mul(x = var_6298_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6299_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_75_cast_fp16 = add(x = var_6284_cast_fp16, y = var_6299_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_24_reps_0 = const()[name = string("tile_24_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = k_75_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [4, 1, 256, 256]> tile_24_cast_fp16 = tile(reps = tile_24_reps_0, x = transpose_48_cast_fp16)[name = string("tile_24_cast_fp16")];
            tensor<int32, [5]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_48_cast_fp16 = reshape(shape = concat_48, x = tile_24_cast_fp16)[name = string("reshape_48_cast_fp16")];
            tensor<int32, [5]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = reshape_48_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [4, 1, 256, 256]> reshape_49_cast_fp16 = reshape(shape = concat_49, x = transpose_49_cast_fp16)[name = string("reshape_49_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_25_reps_0 = const()[name = string("tile_25_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = v_49_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [4, 1, 256, 256]> tile_25_cast_fp16 = tile(reps = tile_25_reps_0, x = transpose_50_cast_fp16)[name = string("tile_25_cast_fp16")];
            tensor<int32, [5]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_50_cast_fp16 = reshape(shape = concat_50, x = tile_25_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [5]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [4, 1, 256, 256]> reshape_51_cast_fp16 = reshape(shape = concat_51, x = transpose_51_cast_fp16)[name = string("reshape_51_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6313_transpose_x_0 = const()[name = string("op_6313_transpose_x_0"), val = bool(false)];
            bool var_6313_transpose_y_0 = const()[name = string("op_6313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = reshape_49_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 4, 256, 256]> var_6313_cast_fp16 = matmul(transpose_x = var_6313_transpose_x_0, transpose_y = var_6313_transpose_y_0, x = q_51_cast_fp16, y = transpose_108_cast_fp16)[name = string("op_6313_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6315_cast_fp16 = add(x = var_6313_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6315_cast_fp16")];
            int32 var_6316 = const()[name = string("op_6316"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_201_cast_fp16 = softmax(axis = var_6316, x = var_6315_cast_fp16)[name = string("w_201_cast_fp16")];
            bool var_6319_transpose_x_0 = const()[name = string("op_6319_transpose_x_0"), val = bool(false)];
            bool var_6319_transpose_y_0 = const()[name = string("op_6319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = reshape_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 4, 256, 256]> var_6319_cast_fp16 = matmul(transpose_x = var_6319_transpose_x_0, transpose_y = var_6319_transpose_y_0, x = w_201_cast_fp16, y = v_51_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<int32, [4]> var_6322_perm_0 = const()[name = string("op_6322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6326 = const()[name = string("op_6326"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_6322_cast_fp16 = transpose(perm = var_6322_perm_0, x = var_6319_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 256, 1024]> input_147_cast_fp16 = reshape(shape = var_6326, x = var_6322_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72710208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73234560))))[name = string("tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6331_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6331_cast_fp16")];
            bool var_6334 = const()[name = string("op_6334"), val = bool(true)];
            tensor<int32, [1]> var_6335_axes_0 = const()[name = string("op_6335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6335_cast_fp16 = reduce_max(axes = var_6335_axes_0, keep_dims = var_6334, x = var_6331_cast_fp16)[name = string("op_6335_cast_fp16")];
            fp16 var_6336_to_fp16 = const()[name = string("op_6336_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_203_cast_fp16 = maximum(x = var_6335_cast_fp16, y = var_6336_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_203_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_203_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6339_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6339_cast_fp16")];
            tensor<int32, [1]> var_6344_axes_0 = const()[name = string("op_6344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6344_keep_dims_0 = const()[name = string("op_6344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6344_cast_fp16 = reduce_mean(axes = var_6344_axes_0, keep_dims = var_6344_keep_dims_0, x = var_6339_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6345_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6345_cast_fp16")];
            fp32 var_6346_epsilon_0 = const()[name = string("op_6346_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6346_cast_fp16 = inverse(epsilon = var_6346_epsilon_0, x = var_6345_cast_fp16)[name = string("op_6346_cast_fp16")];
            fp16 var_6347_to_fp16 = const()[name = string("op_6347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6348_cast_fp16 = mul(x = var_6346_cast_fp16, y = var_6347_to_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6350_cast_fp16 = add(x = var_6344_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            fp32 var_6351_epsilon_0 = const()[name = string("op_6351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6351_cast_fp16 = rsqrt(epsilon = var_6351_epsilon_0, x = var_6350_cast_fp16)[name = string("op_6351_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_177_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6351_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73235648)))];
            tensor<fp16, [1, 256, 512]> var_6353_cast_fp16 = mul(x = out_177_cast_fp16, y = tm_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = var_6353_cast_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6356_cast_fp16 = abs(x = x_227_cast_fp16)[name = string("op_6356_cast_fp16")];
            bool var_6359 = const()[name = string("op_6359"), val = bool(true)];
            tensor<int32, [1]> var_6360_axes_0 = const()[name = string("op_6360_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6360_cast_fp16 = reduce_max(axes = var_6360_axes_0, keep_dims = var_6359, x = var_6356_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_205_cast_fp16 = maximum(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_205_cast_fp16 = real_div(x = x_227_cast_fp16, y = s_205_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6364_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<int32, [1]> var_6369_axes_0 = const()[name = string("op_6369_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6369_keep_dims_0 = const()[name = string("op_6369_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6369_cast_fp16 = reduce_mean(axes = var_6369_axes_0, keep_dims = var_6369_keep_dims_0, x = var_6364_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6370_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6370_cast_fp16")];
            fp32 var_6371_epsilon_0 = const()[name = string("op_6371_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6371_cast_fp16 = inverse(epsilon = var_6371_epsilon_0, x = var_6370_cast_fp16)[name = string("op_6371_cast_fp16")];
            fp16 var_6372_to_fp16 = const()[name = string("op_6372_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6373_cast_fp16 = mul(x = var_6371_cast_fp16, y = var_6372_to_fp16)[name = string("op_6373_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6375_cast_fp16 = add(x = var_6369_cast_fp16, y = var_6373_cast_fp16)[name = string("op_6375_cast_fp16")];
            fp32 var_6376_epsilon_0 = const()[name = string("op_6376_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6376_cast_fp16 = rsqrt(epsilon = var_6376_epsilon_0, x = var_6375_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_179_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6376_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73236736)))];
            tensor<fp16, [1, 256, 512]> input_149_cast_fp16 = mul(x = out_179_cast_fp16, y = tm_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73237824))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74286464))))[name = string("tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_113_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_113_cast_fp16")];
            string var_6383_mode_0 = const()[name = string("op_6383_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_6383_cast_fp16 = gelu(mode = var_6383_mode_0, x = linear_113_cast_fp16)[name = string("op_6383_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74290624))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339264))))[name = string("tm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_114_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_151_cast_fp16 = mul(x = var_6383_cast_fp16, y = linear_114_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75343424))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76392064))))[name = string("tm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6391_cast_fp16 = abs(x = linear_115_cast_fp16)[name = string("op_6391_cast_fp16")];
            bool var_6394 = const()[name = string("op_6394"), val = bool(true)];
            tensor<int32, [1]> var_6395_axes_0 = const()[name = string("op_6395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6395_cast_fp16 = reduce_max(axes = var_6395_axes_0, keep_dims = var_6394, x = var_6391_cast_fp16)[name = string("op_6395_cast_fp16")];
            fp16 var_6396_to_fp16 = const()[name = string("op_6396_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_207_cast_fp16 = maximum(x = var_6395_cast_fp16, y = var_6396_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_207_cast_fp16 = real_div(x = linear_115_cast_fp16, y = s_207_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6399_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6399_cast_fp16")];
            tensor<int32, [1]> var_6404_axes_0 = const()[name = string("op_6404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6404_keep_dims_0 = const()[name = string("op_6404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6404_cast_fp16 = reduce_mean(axes = var_6404_axes_0, keep_dims = var_6404_keep_dims_0, x = var_6399_cast_fp16)[name = string("op_6404_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6405_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6405_cast_fp16")];
            fp32 var_6406_epsilon_0 = const()[name = string("op_6406_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6406_cast_fp16 = inverse(epsilon = var_6406_epsilon_0, x = var_6405_cast_fp16)[name = string("op_6406_cast_fp16")];
            fp16 var_6407_to_fp16 = const()[name = string("op_6407_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6408_cast_fp16 = mul(x = var_6406_cast_fp16, y = var_6407_to_fp16)[name = string("op_6408_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6410_cast_fp16 = add(x = var_6404_cast_fp16, y = var_6408_cast_fp16)[name = string("op_6410_cast_fp16")];
            fp32 var_6411_epsilon_0 = const()[name = string("op_6411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6411_cast_fp16 = rsqrt(epsilon = var_6411_epsilon_0, x = var_6410_cast_fp16)[name = string("op_6411_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_181_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6411_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76393152)))];
            tensor<fp16, [1, 256, 512]> var_6413_cast_fp16 = mul(x = out_181_cast_fp16, y = tm_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6413_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_153_cast_fp16 = add(x = x_227_cast_fp16, y = var_6413_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76394240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76656448))))[name = string("tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_116_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_116_cast_fp16")];
            string x_231_mode_0 = const()[name = string("x_231_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_231_cast_fp16 = gelu(mode = x_231_mode_0, x = linear_116_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_6433_begin_0 = const()[name = string("op_6433_begin_0"), val = tensor<int32, [4]>([0, 0, 12, 0])];
            tensor<int32, [4]> var_6433_end_0 = const()[name = string("op_6433_end_0"), val = tensor<int32, [4]>([1, 256, 13, 512])];
            tensor<bool, [4]> var_6433_end_mask_0 = const()[name = string("op_6433_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6433_squeeze_mask_0 = const()[name = string("op_6433_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_6433_cast_fp16 = slice_by_index(begin = var_6433_begin_0, end = var_6433_end_0, end_mask = var_6433_end_mask_0, squeeze_mask = var_6433_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6433_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_155_cast_fp16 = mul(x = x_231_cast_fp16, y = var_6433_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76657536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76919744))))[name = string("tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6443_cast_fp16 = abs(x = linear_117_cast_fp16)[name = string("op_6443_cast_fp16")];
            bool var_6446 = const()[name = string("op_6446"), val = bool(true)];
            tensor<int32, [1]> var_6447_axes_0 = const()[name = string("op_6447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6447_cast_fp16 = reduce_max(axes = var_6447_axes_0, keep_dims = var_6446, x = var_6443_cast_fp16)[name = string("op_6447_cast_fp16")];
            fp16 var_6448_to_fp16 = const()[name = string("op_6448_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_209_cast_fp16 = maximum(x = var_6447_cast_fp16, y = var_6448_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_209_cast_fp16 = real_div(x = linear_117_cast_fp16, y = s_209_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6451_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6451_cast_fp16")];
            tensor<int32, [1]> var_6456_axes_0 = const()[name = string("op_6456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6456_keep_dims_0 = const()[name = string("op_6456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6456_cast_fp16 = reduce_mean(axes = var_6456_axes_0, keep_dims = var_6456_keep_dims_0, x = var_6451_cast_fp16)[name = string("op_6456_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6457_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6457_cast_fp16")];
            fp32 var_6458_epsilon_0 = const()[name = string("op_6458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6458_cast_fp16 = inverse(epsilon = var_6458_epsilon_0, x = var_6457_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6459_to_fp16 = const()[name = string("op_6459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6460_cast_fp16 = mul(x = var_6458_cast_fp16, y = var_6459_to_fp16)[name = string("op_6460_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6462_cast_fp16 = add(x = var_6456_cast_fp16, y = var_6460_cast_fp16)[name = string("op_6462_cast_fp16")];
            fp32 var_6463_epsilon_0 = const()[name = string("op_6463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6463_cast_fp16 = rsqrt(epsilon = var_6463_epsilon_0, x = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_183_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6463_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76920832)))];
            tensor<fp16, [1, 256, 512]> var_6465_cast_fp16 = mul(x = out_183_cast_fp16, y = tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6465_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_25_cast_fp16 = add(x = input_153_cast_fp16, y = var_6465_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<fp16, [1]> tm_layers_12_layer_scalar_to_fp16 = const()[name = string("tm_layers_12_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.8cp-1])];
            tensor<fp16, [1, 256, 512]> x_235_cast_fp16 = mul(x = h_25_cast_fp16, y = tm_layers_12_layer_scalar_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6469_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_6469_cast_fp16")];
            bool var_6472 = const()[name = string("op_6472"), val = bool(true)];
            tensor<int32, [1]> var_6473_axes_0 = const()[name = string("op_6473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6473_cast_fp16 = reduce_max(axes = var_6473_axes_0, keep_dims = var_6472, x = var_6469_cast_fp16)[name = string("op_6473_cast_fp16")];
            fp16 var_6474_to_fp16 = const()[name = string("op_6474_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_211_cast_fp16 = maximum(x = var_6473_cast_fp16, y = var_6474_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_211_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_211_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6477_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6477_cast_fp16")];
            tensor<int32, [1]> var_6482_axes_0 = const()[name = string("op_6482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6482_keep_dims_0 = const()[name = string("op_6482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6482_cast_fp16 = reduce_mean(axes = var_6482_axes_0, keep_dims = var_6482_keep_dims_0, x = var_6477_cast_fp16)[name = string("op_6482_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6483_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6483_cast_fp16")];
            fp32 var_6484_epsilon_0 = const()[name = string("op_6484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6484_cast_fp16 = inverse(epsilon = var_6484_epsilon_0, x = var_6483_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp16 var_6485_to_fp16 = const()[name = string("op_6485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6486_cast_fp16 = mul(x = var_6484_cast_fp16, y = var_6485_to_fp16)[name = string("op_6486_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6488_cast_fp16 = add(x = var_6482_cast_fp16, y = var_6486_cast_fp16)[name = string("op_6488_cast_fp16")];
            fp32 var_6489_epsilon_0 = const()[name = string("op_6489_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6489_cast_fp16 = rsqrt(epsilon = var_6489_epsilon_0, x = var_6488_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_185_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6489_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76921920)))];
            tensor<fp16, [1, 256, 512]> input_157_cast_fp16 = mul(x = out_185_cast_fp16, y = tm_layers_13_input_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76923008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77447360))))[name = string("tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<int32, [4]> var_6499 = const()[name = string("op_6499"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_6500_cast_fp16 = reshape(shape = var_6499, x = linear_118_cast_fp16)[name = string("op_6500_cast_fp16")];
            tensor<int32, [4]> x_237_perm_0 = const()[name = string("x_237_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77449472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77711680))))[name = string("tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6511 = const()[name = string("op_6511"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6512_cast_fp16 = reshape(shape = var_6511, x = linear_119_cast_fp16)[name = string("op_6512_cast_fp16")];
            tensor<int32, [4]> x_239_perm_0 = const()[name = string("x_239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77712768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77974976))))[name = string("tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6523 = const()[name = string("op_6523"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6524_cast_fp16 = reshape(shape = var_6523, x = linear_120_cast_fp16)[name = string("op_6524_cast_fp16")];
            tensor<int32, [4]> x_241_perm_0 = const()[name = string("x_241_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_237_cast_fp16 = transpose(perm = x_237_perm_0, x = var_6500_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 4, 256, 256]> var_6528_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_6528_cast_fp16")];
            bool var_6531 = const()[name = string("op_6531"), val = bool(true)];
            tensor<int32, [1]> var_6532_axes_0 = const()[name = string("op_6532_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_6532_cast_fp16 = reduce_max(axes = var_6532_axes_0, keep_dims = var_6531, x = var_6528_cast_fp16)[name = string("op_6532_cast_fp16")];
            fp16 var_6533_to_fp16 = const()[name = string("op_6533_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_213_cast_fp16 = maximum(x = var_6532_cast_fp16, y = var_6533_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_213_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_213_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6536_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6536_cast_fp16")];
            tensor<int32, [1]> var_6541_axes_0 = const()[name = string("op_6541_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6541_keep_dims_0 = const()[name = string("op_6541_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_6541_cast_fp16 = reduce_mean(axes = var_6541_axes_0, keep_dims = var_6541_keep_dims_0, x = var_6536_cast_fp16)[name = string("op_6541_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6542_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp32 var_6543_epsilon_0 = const()[name = string("op_6543_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_6543_cast_fp16 = inverse(epsilon = var_6543_epsilon_0, x = var_6542_cast_fp16)[name = string("op_6543_cast_fp16")];
            fp16 var_6544_to_fp16 = const()[name = string("op_6544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_6545_cast_fp16 = mul(x = var_6543_cast_fp16, y = var_6544_to_fp16)[name = string("op_6545_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6547_cast_fp16 = add(x = var_6541_cast_fp16, y = var_6545_cast_fp16)[name = string("op_6547_cast_fp16")];
            fp32 var_6548_epsilon_0 = const()[name = string("op_6548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_6548_cast_fp16 = rsqrt(epsilon = var_6548_epsilon_0, x = var_6547_cast_fp16)[name = string("op_6548_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_187_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6548_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976064)))];
            tensor<fp16, [1, 4, 256, 256]> q_53_cast_fp16 = mul(x = out_187_cast_fp16, y = tm_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("q_53_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_239_cast_fp16 = transpose(perm = x_239_perm_0, x = var_6512_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2, 256, 256]> var_6551_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_6551_cast_fp16")];
            bool var_6554 = const()[name = string("op_6554"), val = bool(true)];
            tensor<int32, [1]> var_6555_axes_0 = const()[name = string("op_6555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6555_cast_fp16 = reduce_max(axes = var_6555_axes_0, keep_dims = var_6554, x = var_6551_cast_fp16)[name = string("op_6555_cast_fp16")];
            fp16 var_6556_to_fp16 = const()[name = string("op_6556_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_215_cast_fp16 = maximum(x = var_6555_cast_fp16, y = var_6556_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_215_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_215_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6559_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6559_cast_fp16")];
            tensor<int32, [1]> var_6564_axes_0 = const()[name = string("op_6564_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6564_keep_dims_0 = const()[name = string("op_6564_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6564_cast_fp16 = reduce_mean(axes = var_6564_axes_0, keep_dims = var_6564_keep_dims_0, x = var_6559_cast_fp16)[name = string("op_6564_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6565_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6565_cast_fp16")];
            fp32 var_6566_epsilon_0 = const()[name = string("op_6566_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6566_cast_fp16 = inverse(epsilon = var_6566_epsilon_0, x = var_6565_cast_fp16)[name = string("op_6566_cast_fp16")];
            fp16 var_6567_to_fp16 = const()[name = string("op_6567_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6568_cast_fp16 = mul(x = var_6566_cast_fp16, y = var_6567_to_fp16)[name = string("op_6568_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6570_cast_fp16 = add(x = var_6564_cast_fp16, y = var_6568_cast_fp16)[name = string("op_6570_cast_fp16")];
            fp32 var_6571_epsilon_0 = const()[name = string("op_6571_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6571_cast_fp16 = rsqrt(epsilon = var_6571_epsilon_0, x = var_6570_cast_fp16)[name = string("op_6571_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_189_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6571_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976640)))];
            tensor<fp16, [1, 2, 256, 256]> k_79_cast_fp16 = mul(x = out_189_cast_fp16, y = tm_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("k_79_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_241_cast_fp16 = transpose(perm = x_241_perm_0, x = var_6524_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 2, 256, 256]> var_6574_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_6574_cast_fp16")];
            bool var_6577 = const()[name = string("op_6577"), val = bool(true)];
            tensor<int32, [1]> var_6578_axes_0 = const()[name = string("op_6578_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6578_cast_fp16 = reduce_max(axes = var_6578_axes_0, keep_dims = var_6577, x = var_6574_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_217_cast_fp16 = maximum(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_217_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_217_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6582_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6587_axes_0 = const()[name = string("op_6587_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6587_keep_dims_0 = const()[name = string("op_6587_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6587_cast_fp16 = reduce_mean(axes = var_6587_axes_0, keep_dims = var_6587_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6588_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6588_cast_fp16")];
            fp32 var_6589_epsilon_0 = const()[name = string("op_6589_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6589_cast_fp16 = inverse(epsilon = var_6589_epsilon_0, x = var_6588_cast_fp16)[name = string("op_6589_cast_fp16")];
            fp16 var_6590_to_fp16 = const()[name = string("op_6590_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6591_cast_fp16 = mul(x = var_6589_cast_fp16, y = var_6590_to_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6593_cast_fp16 = add(x = var_6587_cast_fp16, y = var_6591_cast_fp16)[name = string("op_6593_cast_fp16")];
            fp32 var_6594_epsilon_0 = const()[name = string("op_6594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6594_cast_fp16 = rsqrt(epsilon = var_6594_epsilon_0, x = var_6593_cast_fp16)[name = string("op_6594_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_53_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6594_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6596_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6596_cast_fp16")];
            tensor<int32, [4]> var_6601_begin_0 = const()[name = string("op_6601_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6601_end_0 = const()[name = string("op_6601_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_6601_end_mask_0 = const()[name = string("op_6601_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_6601_cast_fp16 = slice_by_index(begin = var_6601_begin_0, end = var_6601_end_0, end_mask = var_6601_end_mask_0, x = q_53_cast_fp16)[name = string("op_6601_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_6602_cast_fp16 = mul(x = var_6601_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_6602_cast_fp16")];
            tensor<int32, [4]> var_6607_begin_0 = const()[name = string("op_6607_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6607_end_0 = const()[name = string("op_6607_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_6607_end_mask_0 = const()[name = string("op_6607_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_6607_cast_fp16 = slice_by_index(begin = var_6607_begin_0, end = var_6607_end_0, end_mask = var_6607_end_mask_0, x = q_53_cast_fp16)[name = string("op_6607_cast_fp16")];
            int32 var_6609 = const()[name = string("op_6609"), val = int32(-1)];
            bool var_6610_interleave_0 = const()[name = string("op_6610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_6610_cast_fp16 = concat(axis = var_6609, interleave = var_6610_interleave_0, values = (var_6602_cast_fp16, var_6607_cast_fp16))[name = string("op_6610_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6611_cast_fp16 = mul(x = var_6610_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_55_cast_fp16 = add(x = var_6596_cast_fp16, y = var_6611_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6614_cast_fp16 = mul(x = k_79_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = k_79_cast_fp16)[name = string("op_6619_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_6620_cast_fp16 = mul(x = var_6619_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_6620_cast_fp16")];
            tensor<int32, [4]> var_6625_begin_0 = const()[name = string("op_6625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6625_end_0 = const()[name = string("op_6625_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_6625_end_mask_0 = const()[name = string("op_6625_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_6625_cast_fp16 = slice_by_index(begin = var_6625_begin_0, end = var_6625_end_0, end_mask = var_6625_end_mask_0, x = k_79_cast_fp16)[name = string("op_6625_cast_fp16")];
            int32 var_6627 = const()[name = string("op_6627"), val = int32(-1)];
            bool var_6628_interleave_0 = const()[name = string("op_6628_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_6628_cast_fp16 = concat(axis = var_6627, interleave = var_6628_interleave_0, values = (var_6620_cast_fp16, var_6625_cast_fp16))[name = string("op_6628_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6629_cast_fp16 = mul(x = var_6628_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6629_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_81_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6629_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_26_reps_0 = const()[name = string("tile_26_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = k_81_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [4, 1, 256, 256]> tile_26_cast_fp16 = tile(reps = tile_26_reps_0, x = transpose_52_cast_fp16)[name = string("tile_26_cast_fp16")];
            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_52_cast_fp16 = reshape(shape = concat_52, x = tile_26_cast_fp16)[name = string("reshape_52_cast_fp16")];
            tensor<int32, [5]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = reshape_52_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [4, 1, 256, 256]> reshape_53_cast_fp16 = reshape(shape = concat_53, x = transpose_53_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_27_reps_0 = const()[name = string("tile_27_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = v_53_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [4, 1, 256, 256]> tile_27_cast_fp16 = tile(reps = tile_27_reps_0, x = transpose_54_cast_fp16)[name = string("tile_27_cast_fp16")];
            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_54_cast_fp16 = reshape(shape = concat_54, x = tile_27_cast_fp16)[name = string("reshape_54_cast_fp16")];
            tensor<int32, [5]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = reshape_54_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [4, 1, 256, 256]> reshape_55_cast_fp16 = reshape(shape = concat_55, x = transpose_55_cast_fp16)[name = string("reshape_55_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6643_transpose_x_0 = const()[name = string("op_6643_transpose_x_0"), val = bool(false)];
            bool var_6643_transpose_y_0 = const()[name = string("op_6643_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 4, 256, 256]> var_6643_cast_fp16 = matmul(transpose_x = var_6643_transpose_x_0, transpose_y = var_6643_transpose_y_0, x = q_55_cast_fp16, y = transpose_109_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6645_cast_fp16 = add(x = var_6643_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6646 = const()[name = string("op_6646"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_217_cast_fp16 = softmax(axis = var_6646, x = var_6645_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6649_transpose_x_0 = const()[name = string("op_6649_transpose_x_0"), val = bool(false)];
            bool var_6649_transpose_y_0 = const()[name = string("op_6649_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = reshape_55_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 4, 256, 256]> var_6649_cast_fp16 = matmul(transpose_x = var_6649_transpose_x_0, transpose_y = var_6649_transpose_y_0, x = w_217_cast_fp16, y = v_55_cast_fp16)[name = string("op_6649_cast_fp16")];
            tensor<int32, [4]> var_6652_perm_0 = const()[name = string("op_6652_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6656 = const()[name = string("op_6656"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_6652_cast_fp16 = transpose(perm = var_6652_perm_0, x = var_6649_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 256, 1024]> input_159_cast_fp16 = reshape(shape = var_6656, x = var_6652_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77977216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78501568))))[name = string("tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6661_cast_fp16 = abs(x = linear_121_cast_fp16)[name = string("op_6661_cast_fp16")];
            bool var_6664 = const()[name = string("op_6664"), val = bool(true)];
            tensor<int32, [1]> var_6665_axes_0 = const()[name = string("op_6665_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6665_cast_fp16 = reduce_max(axes = var_6665_axes_0, keep_dims = var_6664, x = var_6661_cast_fp16)[name = string("op_6665_cast_fp16")];
            fp16 var_6666_to_fp16 = const()[name = string("op_6666_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_219_cast_fp16 = maximum(x = var_6665_cast_fp16, y = var_6666_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_219_cast_fp16 = real_div(x = linear_121_cast_fp16, y = s_219_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6669_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6669_cast_fp16")];
            tensor<int32, [1]> var_6674_axes_0 = const()[name = string("op_6674_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6674_keep_dims_0 = const()[name = string("op_6674_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6674_cast_fp16 = reduce_mean(axes = var_6674_axes_0, keep_dims = var_6674_keep_dims_0, x = var_6669_cast_fp16)[name = string("op_6674_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6675_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6675_cast_fp16")];
            fp32 var_6676_epsilon_0 = const()[name = string("op_6676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6676_cast_fp16 = inverse(epsilon = var_6676_epsilon_0, x = var_6675_cast_fp16)[name = string("op_6676_cast_fp16")];
            fp16 var_6677_to_fp16 = const()[name = string("op_6677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6678_cast_fp16 = mul(x = var_6676_cast_fp16, y = var_6677_to_fp16)[name = string("op_6678_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6680_cast_fp16 = add(x = var_6674_cast_fp16, y = var_6678_cast_fp16)[name = string("op_6680_cast_fp16")];
            fp32 var_6681_epsilon_0 = const()[name = string("op_6681_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6681_cast_fp16 = rsqrt(epsilon = var_6681_epsilon_0, x = var_6680_cast_fp16)[name = string("op_6681_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_191_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6681_cast_fp16)[name = string("out_191_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78502656)))];
            tensor<fp16, [1, 256, 512]> var_6683_cast_fp16 = mul(x = out_191_cast_fp16, y = tm_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6683_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_245_cast_fp16 = add(x = x_235_cast_fp16, y = var_6683_cast_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6686_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_6686_cast_fp16")];
            bool var_6689 = const()[name = string("op_6689"), val = bool(true)];
            tensor<int32, [1]> var_6690_axes_0 = const()[name = string("op_6690_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6690_cast_fp16 = reduce_max(axes = var_6690_axes_0, keep_dims = var_6689, x = var_6686_cast_fp16)[name = string("op_6690_cast_fp16")];
            fp16 var_6691_to_fp16 = const()[name = string("op_6691_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_221_cast_fp16 = maximum(x = var_6690_cast_fp16, y = var_6691_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_221_cast_fp16 = real_div(x = x_245_cast_fp16, y = s_221_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6694_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6694_cast_fp16")];
            tensor<int32, [1]> var_6699_axes_0 = const()[name = string("op_6699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6699_keep_dims_0 = const()[name = string("op_6699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6699_cast_fp16 = reduce_mean(axes = var_6699_axes_0, keep_dims = var_6699_keep_dims_0, x = var_6694_cast_fp16)[name = string("op_6699_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6700_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6700_cast_fp16")];
            fp32 var_6701_epsilon_0 = const()[name = string("op_6701_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6701_cast_fp16 = inverse(epsilon = var_6701_epsilon_0, x = var_6700_cast_fp16)[name = string("op_6701_cast_fp16")];
            fp16 var_6702_to_fp16 = const()[name = string("op_6702_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6703_cast_fp16 = mul(x = var_6701_cast_fp16, y = var_6702_to_fp16)[name = string("op_6703_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6705_cast_fp16 = add(x = var_6699_cast_fp16, y = var_6703_cast_fp16)[name = string("op_6705_cast_fp16")];
            fp32 var_6706_epsilon_0 = const()[name = string("op_6706_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6706_cast_fp16 = rsqrt(epsilon = var_6706_epsilon_0, x = var_6705_cast_fp16)[name = string("op_6706_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_193_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6706_cast_fp16)[name = string("out_193_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78503744)))];
            tensor<fp16, [1, 256, 512]> input_161_cast_fp16 = mul(x = out_193_cast_fp16, y = tm_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78504832))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79553472))))[name = string("tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_122_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_122_cast_fp16")];
            string var_6713_mode_0 = const()[name = string("op_6713_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_6713_cast_fp16 = gelu(mode = var_6713_mode_0, x = linear_122_cast_fp16)[name = string("op_6713_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79557632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80606272))))[name = string("tm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_123_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_163_cast_fp16 = mul(x = var_6713_cast_fp16, y = linear_123_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80610432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81659072))))[name = string("tm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6721_cast_fp16 = abs(x = linear_124_cast_fp16)[name = string("op_6721_cast_fp16")];
            bool var_6724 = const()[name = string("op_6724"), val = bool(true)];
            tensor<int32, [1]> var_6725_axes_0 = const()[name = string("op_6725_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6725_cast_fp16 = reduce_max(axes = var_6725_axes_0, keep_dims = var_6724, x = var_6721_cast_fp16)[name = string("op_6725_cast_fp16")];
            fp16 var_6726_to_fp16 = const()[name = string("op_6726_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_223_cast_fp16 = maximum(x = var_6725_cast_fp16, y = var_6726_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_223_cast_fp16 = real_div(x = linear_124_cast_fp16, y = s_223_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6729_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [1]> var_6734_axes_0 = const()[name = string("op_6734_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6734_keep_dims_0 = const()[name = string("op_6734_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6734_cast_fp16 = reduce_mean(axes = var_6734_axes_0, keep_dims = var_6734_keep_dims_0, x = var_6729_cast_fp16)[name = string("op_6734_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6735_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6735_cast_fp16")];
            fp32 var_6736_epsilon_0 = const()[name = string("op_6736_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6736_cast_fp16 = inverse(epsilon = var_6736_epsilon_0, x = var_6735_cast_fp16)[name = string("op_6736_cast_fp16")];
            fp16 var_6737_to_fp16 = const()[name = string("op_6737_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6738_cast_fp16 = mul(x = var_6736_cast_fp16, y = var_6737_to_fp16)[name = string("op_6738_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6740_cast_fp16 = add(x = var_6734_cast_fp16, y = var_6738_cast_fp16)[name = string("op_6740_cast_fp16")];
            fp32 var_6741_epsilon_0 = const()[name = string("op_6741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6741_cast_fp16 = rsqrt(epsilon = var_6741_epsilon_0, x = var_6740_cast_fp16)[name = string("op_6741_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_195_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6741_cast_fp16)[name = string("out_195_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81660160)))];
            tensor<fp16, [1, 256, 512]> var_6743_cast_fp16 = mul(x = out_195_cast_fp16, y = tm_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6743_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_165_cast_fp16 = add(x = x_245_cast_fp16, y = var_6743_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81661248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81923456))))[name = string("tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_125_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_125_cast_fp16")];
            string x_249_mode_0 = const()[name = string("x_249_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_249_cast_fp16 = gelu(mode = x_249_mode_0, x = linear_125_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<int32, [4]> var_6763_begin_0 = const()[name = string("op_6763_begin_0"), val = tensor<int32, [4]>([0, 0, 13, 0])];
            tensor<int32, [4]> var_6763_end_0 = const()[name = string("op_6763_end_0"), val = tensor<int32, [4]>([1, 256, 14, 512])];
            tensor<bool, [4]> var_6763_end_mask_0 = const()[name = string("op_6763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6763_squeeze_mask_0 = const()[name = string("op_6763_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_6763_cast_fp16 = slice_by_index(begin = var_6763_begin_0, end = var_6763_end_0, end_mask = var_6763_end_mask_0, squeeze_mask = var_6763_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_167_cast_fp16 = mul(x = x_249_cast_fp16, y = var_6763_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81924544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82186752))))[name = string("tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6773_cast_fp16 = abs(x = linear_126_cast_fp16)[name = string("op_6773_cast_fp16")];
            bool var_6776 = const()[name = string("op_6776"), val = bool(true)];
            tensor<int32, [1]> var_6777_axes_0 = const()[name = string("op_6777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6777_cast_fp16 = reduce_max(axes = var_6777_axes_0, keep_dims = var_6776, x = var_6773_cast_fp16)[name = string("op_6777_cast_fp16")];
            fp16 var_6778_to_fp16 = const()[name = string("op_6778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_225_cast_fp16 = maximum(x = var_6777_cast_fp16, y = var_6778_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_225_cast_fp16 = real_div(x = linear_126_cast_fp16, y = s_225_cast_fp16)[name = string("y_225_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6781_cast_fp16 = mul(x = y_225_cast_fp16, y = y_225_cast_fp16)[name = string("op_6781_cast_fp16")];
            tensor<int32, [1]> var_6786_axes_0 = const()[name = string("op_6786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6786_keep_dims_0 = const()[name = string("op_6786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6786_cast_fp16 = reduce_mean(axes = var_6786_axes_0, keep_dims = var_6786_keep_dims_0, x = var_6781_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6787_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6787_cast_fp16")];
            fp32 var_6788_epsilon_0 = const()[name = string("op_6788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6788_cast_fp16 = inverse(epsilon = var_6788_epsilon_0, x = var_6787_cast_fp16)[name = string("op_6788_cast_fp16")];
            fp16 var_6789_to_fp16 = const()[name = string("op_6789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6790_cast_fp16 = mul(x = var_6788_cast_fp16, y = var_6789_to_fp16)[name = string("op_6790_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6792_cast_fp16 = add(x = var_6786_cast_fp16, y = var_6790_cast_fp16)[name = string("op_6792_cast_fp16")];
            fp32 var_6793_epsilon_0 = const()[name = string("op_6793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6793_cast_fp16 = rsqrt(epsilon = var_6793_epsilon_0, x = var_6792_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_197_cast_fp16 = mul(x = y_225_cast_fp16, y = var_6793_cast_fp16)[name = string("out_197_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82187840)))];
            tensor<fp16, [1, 256, 512]> var_6795_cast_fp16 = mul(x = out_197_cast_fp16, y = tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6795_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_27_cast_fp16 = add(x = input_165_cast_fp16, y = var_6795_cast_fp16)[name = string("h_27_cast_fp16")];
            tensor<fp16, [1]> tm_layers_13_layer_scalar_to_fp16 = const()[name = string("tm_layers_13_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9cp-1])];
            tensor<fp16, [1, 256, 512]> x_253_cast_fp16 = mul(x = h_27_cast_fp16, y = tm_layers_13_layer_scalar_to_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6799_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6799_cast_fp16")];
            bool var_6802 = const()[name = string("op_6802"), val = bool(true)];
            tensor<int32, [1]> var_6803_axes_0 = const()[name = string("op_6803_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6803_cast_fp16 = reduce_max(axes = var_6803_axes_0, keep_dims = var_6802, x = var_6799_cast_fp16)[name = string("op_6803_cast_fp16")];
            fp16 var_6804_to_fp16 = const()[name = string("op_6804_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_227_cast_fp16 = maximum(x = var_6803_cast_fp16, y = var_6804_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_227_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_227_cast_fp16)[name = string("y_227_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6807_cast_fp16 = mul(x = y_227_cast_fp16, y = y_227_cast_fp16)[name = string("op_6807_cast_fp16")];
            tensor<int32, [1]> var_6812_axes_0 = const()[name = string("op_6812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6812_keep_dims_0 = const()[name = string("op_6812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6812_cast_fp16 = reduce_mean(axes = var_6812_axes_0, keep_dims = var_6812_keep_dims_0, x = var_6807_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6813_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6813_cast_fp16")];
            fp32 var_6814_epsilon_0 = const()[name = string("op_6814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6814_cast_fp16 = inverse(epsilon = var_6814_epsilon_0, x = var_6813_cast_fp16)[name = string("op_6814_cast_fp16")];
            fp16 var_6815_to_fp16 = const()[name = string("op_6815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6816_cast_fp16 = mul(x = var_6814_cast_fp16, y = var_6815_to_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6818_cast_fp16 = add(x = var_6812_cast_fp16, y = var_6816_cast_fp16)[name = string("op_6818_cast_fp16")];
            fp32 var_6819_epsilon_0 = const()[name = string("op_6819_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6819_cast_fp16 = rsqrt(epsilon = var_6819_epsilon_0, x = var_6818_cast_fp16)[name = string("op_6819_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_199_cast_fp16 = mul(x = y_227_cast_fp16, y = var_6819_cast_fp16)[name = string("out_199_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82188928)))];
            tensor<fp16, [1, 256, 512]> input_169_cast_fp16 = mul(x = out_199_cast_fp16, y = tm_layers_14_input_layernorm_weight_to_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82190016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82714368))))[name = string("tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6829 = const()[name = string("op_6829"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_6830_cast_fp16 = reshape(shape = var_6829, x = linear_127_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82716480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82978688))))[name = string("tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6841 = const()[name = string("op_6841"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6842_cast_fp16 = reshape(shape = var_6841, x = linear_128_cast_fp16)[name = string("op_6842_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82979776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83241984))))[name = string("tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_6853 = const()[name = string("op_6853"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6854_cast_fp16 = reshape(shape = var_6853, x = linear_129_cast_fp16)[name = string("op_6854_cast_fp16")];
            tensor<int32, [4]> x_259_perm_0 = const()[name = string("x_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_6830_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 4, 256, 256]> var_6858_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6858_cast_fp16")];
            bool var_6861 = const()[name = string("op_6861"), val = bool(true)];
            tensor<int32, [1]> var_6862_axes_0 = const()[name = string("op_6862_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_6862_cast_fp16 = reduce_max(axes = var_6862_axes_0, keep_dims = var_6861, x = var_6858_cast_fp16)[name = string("op_6862_cast_fp16")];
            fp16 var_6863_to_fp16 = const()[name = string("op_6863_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_229_cast_fp16 = maximum(x = var_6862_cast_fp16, y = var_6863_to_fp16)[name = string("s_229_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_229_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_229_cast_fp16)[name = string("y_229_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6866_cast_fp16 = mul(x = y_229_cast_fp16, y = y_229_cast_fp16)[name = string("op_6866_cast_fp16")];
            tensor<int32, [1]> var_6871_axes_0 = const()[name = string("op_6871_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6871_keep_dims_0 = const()[name = string("op_6871_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_6871_cast_fp16 = reduce_mean(axes = var_6871_axes_0, keep_dims = var_6871_keep_dims_0, x = var_6866_cast_fp16)[name = string("op_6871_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6872_cast_fp16 = mul(x = s_229_cast_fp16, y = s_229_cast_fp16)[name = string("op_6872_cast_fp16")];
            fp32 var_6873_epsilon_0 = const()[name = string("op_6873_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_6873_cast_fp16 = inverse(epsilon = var_6873_epsilon_0, x = var_6872_cast_fp16)[name = string("op_6873_cast_fp16")];
            fp16 var_6874_to_fp16 = const()[name = string("op_6874_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_6875_cast_fp16 = mul(x = var_6873_cast_fp16, y = var_6874_to_fp16)[name = string("op_6875_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6877_cast_fp16 = add(x = var_6871_cast_fp16, y = var_6875_cast_fp16)[name = string("op_6877_cast_fp16")];
            fp32 var_6878_epsilon_0 = const()[name = string("op_6878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_6878_cast_fp16 = rsqrt(epsilon = var_6878_epsilon_0, x = var_6877_cast_fp16)[name = string("op_6878_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_201_cast_fp16 = mul(x = y_229_cast_fp16, y = var_6878_cast_fp16)[name = string("out_201_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243072)))];
            tensor<fp16, [1, 4, 256, 256]> q_57_cast_fp16 = mul(x = out_201_cast_fp16, y = tm_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("q_57_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_6842_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, 256, 256]> var_6881_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6884 = const()[name = string("op_6884"), val = bool(true)];
            tensor<int32, [1]> var_6885_axes_0 = const()[name = string("op_6885_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6885_cast_fp16 = reduce_max(axes = var_6885_axes_0, keep_dims = var_6884, x = var_6881_cast_fp16)[name = string("op_6885_cast_fp16")];
            fp16 var_6886_to_fp16 = const()[name = string("op_6886_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_231_cast_fp16 = maximum(x = var_6885_cast_fp16, y = var_6886_to_fp16)[name = string("s_231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_231_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_231_cast_fp16)[name = string("y_231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6889_cast_fp16 = mul(x = y_231_cast_fp16, y = y_231_cast_fp16)[name = string("op_6889_cast_fp16")];
            tensor<int32, [1]> var_6894_axes_0 = const()[name = string("op_6894_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6894_keep_dims_0 = const()[name = string("op_6894_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6894_cast_fp16 = reduce_mean(axes = var_6894_axes_0, keep_dims = var_6894_keep_dims_0, x = var_6889_cast_fp16)[name = string("op_6894_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6895_cast_fp16 = mul(x = s_231_cast_fp16, y = s_231_cast_fp16)[name = string("op_6895_cast_fp16")];
            fp32 var_6896_epsilon_0 = const()[name = string("op_6896_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6896_cast_fp16 = inverse(epsilon = var_6896_epsilon_0, x = var_6895_cast_fp16)[name = string("op_6896_cast_fp16")];
            fp16 var_6897_to_fp16 = const()[name = string("op_6897_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6898_cast_fp16 = mul(x = var_6896_cast_fp16, y = var_6897_to_fp16)[name = string("op_6898_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6900_cast_fp16 = add(x = var_6894_cast_fp16, y = var_6898_cast_fp16)[name = string("op_6900_cast_fp16")];
            fp32 var_6901_epsilon_0 = const()[name = string("op_6901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6901_cast_fp16 = rsqrt(epsilon = var_6901_epsilon_0, x = var_6900_cast_fp16)[name = string("op_6901_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_203_cast_fp16 = mul(x = y_231_cast_fp16, y = var_6901_cast_fp16)[name = string("out_203_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243648)))];
            tensor<fp16, [1, 2, 256, 256]> k_85_cast_fp16 = mul(x = out_203_cast_fp16, y = tm_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("k_85_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_259_cast_fp16 = transpose(perm = x_259_perm_0, x = var_6854_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 2, 256, 256]> var_6904_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6904_cast_fp16")];
            bool var_6907 = const()[name = string("op_6907"), val = bool(true)];
            tensor<int32, [1]> var_6908_axes_0 = const()[name = string("op_6908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6908_cast_fp16 = reduce_max(axes = var_6908_axes_0, keep_dims = var_6907, x = var_6904_cast_fp16)[name = string("op_6908_cast_fp16")];
            fp16 var_6909_to_fp16 = const()[name = string("op_6909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_233_cast_fp16 = maximum(x = var_6908_cast_fp16, y = var_6909_to_fp16)[name = string("s_233_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_233_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_233_cast_fp16)[name = string("y_233_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6912_cast_fp16 = mul(x = y_233_cast_fp16, y = y_233_cast_fp16)[name = string("op_6912_cast_fp16")];
            tensor<int32, [1]> var_6917_axes_0 = const()[name = string("op_6917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6917_keep_dims_0 = const()[name = string("op_6917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6917_cast_fp16 = reduce_mean(axes = var_6917_axes_0, keep_dims = var_6917_keep_dims_0, x = var_6912_cast_fp16)[name = string("op_6917_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6918_cast_fp16 = mul(x = s_233_cast_fp16, y = s_233_cast_fp16)[name = string("op_6918_cast_fp16")];
            fp32 var_6919_epsilon_0 = const()[name = string("op_6919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6919_cast_fp16 = inverse(epsilon = var_6919_epsilon_0, x = var_6918_cast_fp16)[name = string("op_6919_cast_fp16")];
            fp16 var_6920_to_fp16 = const()[name = string("op_6920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6921_cast_fp16 = mul(x = var_6919_cast_fp16, y = var_6920_to_fp16)[name = string("op_6921_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6923_cast_fp16 = add(x = var_6917_cast_fp16, y = var_6921_cast_fp16)[name = string("op_6923_cast_fp16")];
            fp32 var_6924_epsilon_0 = const()[name = string("op_6924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6924_cast_fp16 = rsqrt(epsilon = var_6924_epsilon_0, x = var_6923_cast_fp16)[name = string("op_6924_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_57_cast_fp16 = mul(x = y_233_cast_fp16, y = var_6924_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6926_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6926_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = q_57_cast_fp16)[name = string("op_6931_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_6932_cast_fp16 = mul(x = var_6931_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_6932_cast_fp16")];
            tensor<int32, [4]> var_6937_begin_0 = const()[name = string("op_6937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6937_end_0 = const()[name = string("op_6937_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_6937_end_mask_0 = const()[name = string("op_6937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_6937_cast_fp16 = slice_by_index(begin = var_6937_begin_0, end = var_6937_end_0, end_mask = var_6937_end_mask_0, x = q_57_cast_fp16)[name = string("op_6937_cast_fp16")];
            int32 var_6939 = const()[name = string("op_6939"), val = int32(-1)];
            bool var_6940_interleave_0 = const()[name = string("op_6940_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_6940_cast_fp16 = concat(axis = var_6939, interleave = var_6940_interleave_0, values = (var_6932_cast_fp16, var_6937_cast_fp16))[name = string("op_6940_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6941_cast_fp16 = mul(x = var_6940_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6941_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_59_cast_fp16 = add(x = var_6926_cast_fp16, y = var_6941_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6944_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6944_cast_fp16")];
            tensor<int32, [4]> var_6949_begin_0 = const()[name = string("op_6949_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6949_end_0 = const()[name = string("op_6949_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_6949_end_mask_0 = const()[name = string("op_6949_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_6949_cast_fp16 = slice_by_index(begin = var_6949_begin_0, end = var_6949_end_0, end_mask = var_6949_end_mask_0, x = k_85_cast_fp16)[name = string("op_6949_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_6950_cast_fp16 = mul(x = var_6949_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_6950_cast_fp16")];
            tensor<int32, [4]> var_6955_begin_0 = const()[name = string("op_6955_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6955_end_0 = const()[name = string("op_6955_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_6955_end_mask_0 = const()[name = string("op_6955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_6955_cast_fp16 = slice_by_index(begin = var_6955_begin_0, end = var_6955_end_0, end_mask = var_6955_end_mask_0, x = k_85_cast_fp16)[name = string("op_6955_cast_fp16")];
            int32 var_6957 = const()[name = string("op_6957"), val = int32(-1)];
            bool var_6958_interleave_0 = const()[name = string("op_6958_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_6958_cast_fp16 = concat(axis = var_6957, interleave = var_6958_interleave_0, values = (var_6950_cast_fp16, var_6955_cast_fp16))[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6959_cast_fp16 = mul(x = var_6958_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_87_cast_fp16 = add(x = var_6944_cast_fp16, y = var_6959_cast_fp16)[name = string("k_87_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_28_reps_0 = const()[name = string("tile_28_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = k_87_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 1, 256, 256]> tile_28_cast_fp16 = tile(reps = tile_28_reps_0, x = transpose_56_cast_fp16)[name = string("tile_28_cast_fp16")];
            tensor<int32, [5]> concat_56 = const()[name = string("concat_56"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_56_cast_fp16 = reshape(shape = concat_56, x = tile_28_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [5]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_57 = const()[name = string("concat_57"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 1, 256, 256]> reshape_57_cast_fp16 = reshape(shape = concat_57, x = transpose_57_cast_fp16)[name = string("reshape_57_cast_fp16")];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_29_reps_0 = const()[name = string("tile_29_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = v_57_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 1, 256, 256]> tile_29_cast_fp16 = tile(reps = tile_29_reps_0, x = transpose_58_cast_fp16)[name = string("tile_29_cast_fp16")];
            tensor<int32, [5]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_58_cast_fp16 = reshape(shape = concat_58, x = tile_29_cast_fp16)[name = string("reshape_58_cast_fp16")];
            tensor<int32, [5]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = reshape_58_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 1, 256, 256]> reshape_59_cast_fp16 = reshape(shape = concat_59, x = transpose_59_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6973_transpose_x_0 = const()[name = string("op_6973_transpose_x_0"), val = bool(false)];
            bool var_6973_transpose_y_0 = const()[name = string("op_6973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_110_cast_fp16 = transpose(perm = transpose_110_perm_0, x = reshape_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 4, 256, 256]> var_6973_cast_fp16 = matmul(transpose_x = var_6973_transpose_x_0, transpose_y = var_6973_transpose_y_0, x = q_59_cast_fp16, y = transpose_110_cast_fp16)[name = string("op_6973_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6975_cast_fp16 = add(x = var_6973_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6976 = const()[name = string("op_6976"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_233_cast_fp16 = softmax(axis = var_6976, x = var_6975_cast_fp16)[name = string("w_233_cast_fp16")];
            bool var_6979_transpose_x_0 = const()[name = string("op_6979_transpose_x_0"), val = bool(false)];
            bool var_6979_transpose_y_0 = const()[name = string("op_6979_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 4, 256, 256]> var_6979_cast_fp16 = matmul(transpose_x = var_6979_transpose_x_0, transpose_y = var_6979_transpose_y_0, x = w_233_cast_fp16, y = v_59_cast_fp16)[name = string("op_6979_cast_fp16")];
            tensor<int32, [4]> var_6982_perm_0 = const()[name = string("op_6982_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6986 = const()[name = string("op_6986"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_6982_cast_fp16 = transpose(perm = var_6982_perm_0, x = var_6979_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 256, 1024]> input_171_cast_fp16 = reshape(shape = var_6986, x = var_6982_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83244224))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83768576))))[name = string("tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6991_cast_fp16 = abs(x = linear_130_cast_fp16)[name = string("op_6991_cast_fp16")];
            bool var_6994 = const()[name = string("op_6994"), val = bool(true)];
            tensor<int32, [1]> var_6995_axes_0 = const()[name = string("op_6995_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6995_cast_fp16 = reduce_max(axes = var_6995_axes_0, keep_dims = var_6994, x = var_6991_cast_fp16)[name = string("op_6995_cast_fp16")];
            fp16 var_6996_to_fp16 = const()[name = string("op_6996_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_235_cast_fp16 = maximum(x = var_6995_cast_fp16, y = var_6996_to_fp16)[name = string("s_235_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_235_cast_fp16 = real_div(x = linear_130_cast_fp16, y = s_235_cast_fp16)[name = string("y_235_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6999_cast_fp16 = mul(x = y_235_cast_fp16, y = y_235_cast_fp16)[name = string("op_6999_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_6999_cast_fp16)[name = string("op_7004_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7005_cast_fp16 = mul(x = s_235_cast_fp16, y = s_235_cast_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7010_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7010_cast_fp16")];
            fp32 var_7011_epsilon_0 = const()[name = string("op_7011_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7011_cast_fp16 = rsqrt(epsilon = var_7011_epsilon_0, x = var_7010_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_205_cast_fp16 = mul(x = y_235_cast_fp16, y = var_7011_cast_fp16)[name = string("out_205_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83769664)))];
            tensor<fp16, [1, 256, 512]> var_7013_cast_fp16 = mul(x = out_205_cast_fp16, y = tm_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_7013_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = var_7013_cast_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7016_cast_fp16 = abs(x = x_263_cast_fp16)[name = string("op_7016_cast_fp16")];
            bool var_7019 = const()[name = string("op_7019"), val = bool(true)];
            tensor<int32, [1]> var_7020_axes_0 = const()[name = string("op_7020_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7020_cast_fp16 = reduce_max(axes = var_7020_axes_0, keep_dims = var_7019, x = var_7016_cast_fp16)[name = string("op_7020_cast_fp16")];
            fp16 var_7021_to_fp16 = const()[name = string("op_7021_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_237_cast_fp16 = maximum(x = var_7020_cast_fp16, y = var_7021_to_fp16)[name = string("s_237_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_237_cast_fp16 = real_div(x = x_263_cast_fp16, y = s_237_cast_fp16)[name = string("y_237_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7024_cast_fp16 = mul(x = y_237_cast_fp16, y = y_237_cast_fp16)[name = string("op_7024_cast_fp16")];
            tensor<int32, [1]> var_7029_axes_0 = const()[name = string("op_7029_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7029_keep_dims_0 = const()[name = string("op_7029_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7029_cast_fp16 = reduce_mean(axes = var_7029_axes_0, keep_dims = var_7029_keep_dims_0, x = var_7024_cast_fp16)[name = string("op_7029_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7030_cast_fp16 = mul(x = s_237_cast_fp16, y = s_237_cast_fp16)[name = string("op_7030_cast_fp16")];
            fp32 var_7031_epsilon_0 = const()[name = string("op_7031_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7031_cast_fp16 = inverse(epsilon = var_7031_epsilon_0, x = var_7030_cast_fp16)[name = string("op_7031_cast_fp16")];
            fp16 var_7032_to_fp16 = const()[name = string("op_7032_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7033_cast_fp16 = mul(x = var_7031_cast_fp16, y = var_7032_to_fp16)[name = string("op_7033_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7035_cast_fp16 = add(x = var_7029_cast_fp16, y = var_7033_cast_fp16)[name = string("op_7035_cast_fp16")];
            fp32 var_7036_epsilon_0 = const()[name = string("op_7036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7036_cast_fp16 = rsqrt(epsilon = var_7036_epsilon_0, x = var_7035_cast_fp16)[name = string("op_7036_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_207_cast_fp16 = mul(x = y_237_cast_fp16, y = var_7036_cast_fp16)[name = string("out_207_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83770752)))];
            tensor<fp16, [1, 256, 512]> input_173_cast_fp16 = mul(x = out_207_cast_fp16, y = tm_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83771840))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84820480))))[name = string("tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_131_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_131_cast_fp16")];
            string var_7043_mode_0 = const()[name = string("op_7043_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_7043_cast_fp16 = gelu(mode = var_7043_mode_0, x = linear_131_cast_fp16)[name = string("op_7043_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84824640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85873280))))[name = string("tm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_132_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_175_cast_fp16 = mul(x = var_7043_cast_fp16, y = linear_132_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85877440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86926080))))[name = string("tm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7051_cast_fp16 = abs(x = linear_133_cast_fp16)[name = string("op_7051_cast_fp16")];
            bool var_7054 = const()[name = string("op_7054"), val = bool(true)];
            tensor<int32, [1]> var_7055_axes_0 = const()[name = string("op_7055_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7055_cast_fp16 = reduce_max(axes = var_7055_axes_0, keep_dims = var_7054, x = var_7051_cast_fp16)[name = string("op_7055_cast_fp16")];
            fp16 var_7056_to_fp16 = const()[name = string("op_7056_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_239_cast_fp16 = maximum(x = var_7055_cast_fp16, y = var_7056_to_fp16)[name = string("s_239_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_239_cast_fp16 = real_div(x = linear_133_cast_fp16, y = s_239_cast_fp16)[name = string("y_239_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7059_cast_fp16 = mul(x = y_239_cast_fp16, y = y_239_cast_fp16)[name = string("op_7059_cast_fp16")];
            tensor<int32, [1]> var_7064_axes_0 = const()[name = string("op_7064_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7064_keep_dims_0 = const()[name = string("op_7064_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7064_cast_fp16 = reduce_mean(axes = var_7064_axes_0, keep_dims = var_7064_keep_dims_0, x = var_7059_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7065_cast_fp16 = mul(x = s_239_cast_fp16, y = s_239_cast_fp16)[name = string("op_7065_cast_fp16")];
            fp32 var_7066_epsilon_0 = const()[name = string("op_7066_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7066_cast_fp16 = inverse(epsilon = var_7066_epsilon_0, x = var_7065_cast_fp16)[name = string("op_7066_cast_fp16")];
            fp16 var_7067_to_fp16 = const()[name = string("op_7067_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7068_cast_fp16 = mul(x = var_7066_cast_fp16, y = var_7067_to_fp16)[name = string("op_7068_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7070_cast_fp16 = add(x = var_7064_cast_fp16, y = var_7068_cast_fp16)[name = string("op_7070_cast_fp16")];
            fp32 var_7071_epsilon_0 = const()[name = string("op_7071_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7071_cast_fp16 = rsqrt(epsilon = var_7071_epsilon_0, x = var_7070_cast_fp16)[name = string("op_7071_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_209_cast_fp16 = mul(x = y_239_cast_fp16, y = var_7071_cast_fp16)[name = string("out_209_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86927168)))];
            tensor<fp16, [1, 256, 512]> var_7073_cast_fp16 = mul(x = out_209_cast_fp16, y = tm_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7073_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_177_cast_fp16 = add(x = x_263_cast_fp16, y = var_7073_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86928256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87190464))))[name = string("tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_134_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_134_cast_fp16")];
            string x_267_mode_0 = const()[name = string("x_267_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_267_cast_fp16 = gelu(mode = x_267_mode_0, x = linear_134_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<int32, [4]> var_7093_begin_0 = const()[name = string("op_7093_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
            tensor<int32, [4]> var_7093_end_0 = const()[name = string("op_7093_end_0"), val = tensor<int32, [4]>([1, 256, 15, 512])];
            tensor<bool, [4]> var_7093_end_mask_0 = const()[name = string("op_7093_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7093_squeeze_mask_0 = const()[name = string("op_7093_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_7093_cast_fp16 = slice_by_index(begin = var_7093_begin_0, end = var_7093_end_0, end_mask = var_7093_end_mask_0, squeeze_mask = var_7093_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7093_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_179_cast_fp16 = mul(x = x_267_cast_fp16, y = var_7093_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87191552))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87453760))))[name = string("tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7103_cast_fp16 = abs(x = linear_135_cast_fp16)[name = string("op_7103_cast_fp16")];
            bool var_7106 = const()[name = string("op_7106"), val = bool(true)];
            tensor<int32, [1]> var_7107_axes_0 = const()[name = string("op_7107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7107_cast_fp16 = reduce_max(axes = var_7107_axes_0, keep_dims = var_7106, x = var_7103_cast_fp16)[name = string("op_7107_cast_fp16")];
            fp16 var_7108_to_fp16 = const()[name = string("op_7108_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_241_cast_fp16 = maximum(x = var_7107_cast_fp16, y = var_7108_to_fp16)[name = string("s_241_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_241_cast_fp16 = real_div(x = linear_135_cast_fp16, y = s_241_cast_fp16)[name = string("y_241_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7111_cast_fp16 = mul(x = y_241_cast_fp16, y = y_241_cast_fp16)[name = string("op_7111_cast_fp16")];
            tensor<int32, [1]> var_7116_axes_0 = const()[name = string("op_7116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7116_keep_dims_0 = const()[name = string("op_7116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7116_cast_fp16 = reduce_mean(axes = var_7116_axes_0, keep_dims = var_7116_keep_dims_0, x = var_7111_cast_fp16)[name = string("op_7116_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7117_cast_fp16 = mul(x = s_241_cast_fp16, y = s_241_cast_fp16)[name = string("op_7117_cast_fp16")];
            fp32 var_7118_epsilon_0 = const()[name = string("op_7118_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7118_cast_fp16 = inverse(epsilon = var_7118_epsilon_0, x = var_7117_cast_fp16)[name = string("op_7118_cast_fp16")];
            fp16 var_7119_to_fp16 = const()[name = string("op_7119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7120_cast_fp16 = mul(x = var_7118_cast_fp16, y = var_7119_to_fp16)[name = string("op_7120_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7122_cast_fp16 = add(x = var_7116_cast_fp16, y = var_7120_cast_fp16)[name = string("op_7122_cast_fp16")];
            fp32 var_7123_epsilon_0 = const()[name = string("op_7123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7123_cast_fp16 = rsqrt(epsilon = var_7123_epsilon_0, x = var_7122_cast_fp16)[name = string("op_7123_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_211_cast_fp16 = mul(x = y_241_cast_fp16, y = var_7123_cast_fp16)[name = string("out_211_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87454848)))];
            tensor<fp16, [1, 256, 512]> var_7125_cast_fp16 = mul(x = out_211_cast_fp16, y = tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_29_cast_fp16 = add(x = input_177_cast_fp16, y = var_7125_cast_fp16)[name = string("h_29_cast_fp16")];
            tensor<fp16, [1]> tm_layers_14_layer_scalar_to_fp16 = const()[name = string("tm_layers_14_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1p-1])];
            tensor<fp16, [1, 256, 512]> x_271_cast_fp16 = mul(x = h_29_cast_fp16, y = tm_layers_14_layer_scalar_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7129_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_7129_cast_fp16")];
            bool var_7132 = const()[name = string("op_7132"), val = bool(true)];
            tensor<int32, [1]> var_7133_axes_0 = const()[name = string("op_7133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7133_cast_fp16 = reduce_max(axes = var_7133_axes_0, keep_dims = var_7132, x = var_7129_cast_fp16)[name = string("op_7133_cast_fp16")];
            fp16 var_7134_to_fp16 = const()[name = string("op_7134_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_243_cast_fp16 = maximum(x = var_7133_cast_fp16, y = var_7134_to_fp16)[name = string("s_243_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_243_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_243_cast_fp16)[name = string("y_243_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7137_cast_fp16 = mul(x = y_243_cast_fp16, y = y_243_cast_fp16)[name = string("op_7137_cast_fp16")];
            tensor<int32, [1]> var_7142_axes_0 = const()[name = string("op_7142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7142_keep_dims_0 = const()[name = string("op_7142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7142_cast_fp16 = reduce_mean(axes = var_7142_axes_0, keep_dims = var_7142_keep_dims_0, x = var_7137_cast_fp16)[name = string("op_7142_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7143_cast_fp16 = mul(x = s_243_cast_fp16, y = s_243_cast_fp16)[name = string("op_7143_cast_fp16")];
            fp32 var_7144_epsilon_0 = const()[name = string("op_7144_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7144_cast_fp16 = inverse(epsilon = var_7144_epsilon_0, x = var_7143_cast_fp16)[name = string("op_7144_cast_fp16")];
            fp16 var_7145_to_fp16 = const()[name = string("op_7145_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7146_cast_fp16 = mul(x = var_7144_cast_fp16, y = var_7145_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7148_cast_fp16 = add(x = var_7142_cast_fp16, y = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp32 var_7149_epsilon_0 = const()[name = string("op_7149_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7149_cast_fp16 = rsqrt(epsilon = var_7149_epsilon_0, x = var_7148_cast_fp16)[name = string("op_7149_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_213_cast_fp16 = mul(x = y_243_cast_fp16, y = var_7149_cast_fp16)[name = string("out_213_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87455936)))];
            tensor<fp16, [1, 256, 512]> input_181_cast_fp16 = mul(x = out_213_cast_fp16, y = tm_layers_15_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87457024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87981376))))[name = string("tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<int32, [4]> var_7159 = const()[name = string("op_7159"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_7160_cast_fp16 = reshape(shape = var_7159, x = linear_136_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<int32, [4]> x_273_perm_0 = const()[name = string("x_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87983488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88245696))))[name = string("tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<int32, [4]> var_7171 = const()[name = string("op_7171"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_7172_cast_fp16 = reshape(shape = var_7171, x = linear_137_cast_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [4]> x_275_perm_0 = const()[name = string("x_275_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88246784))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88508992))))[name = string("tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_7183 = const()[name = string("op_7183"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_7184_cast_fp16 = reshape(shape = var_7183, x = linear_138_cast_fp16)[name = string("op_7184_cast_fp16")];
            tensor<int32, [4]> x_277_perm_0 = const()[name = string("x_277_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_273_cast_fp16 = transpose(perm = x_273_perm_0, x = var_7160_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 4, 256, 256]> var_7188_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_7188_cast_fp16")];
            bool var_7191 = const()[name = string("op_7191"), val = bool(true)];
            tensor<int32, [1]> var_7192_axes_0 = const()[name = string("op_7192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_7192_cast_fp16 = reduce_max(axes = var_7192_axes_0, keep_dims = var_7191, x = var_7188_cast_fp16)[name = string("op_7192_cast_fp16")];
            fp16 var_7193_to_fp16 = const()[name = string("op_7193_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_245_cast_fp16 = maximum(x = var_7192_cast_fp16, y = var_7193_to_fp16)[name = string("s_245_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_245_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_245_cast_fp16)[name = string("y_245_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7196_cast_fp16 = mul(x = y_245_cast_fp16, y = y_245_cast_fp16)[name = string("op_7196_cast_fp16")];
            tensor<int32, [1]> var_7201_axes_0 = const()[name = string("op_7201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7201_keep_dims_0 = const()[name = string("op_7201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_7201_cast_fp16 = reduce_mean(axes = var_7201_axes_0, keep_dims = var_7201_keep_dims_0, x = var_7196_cast_fp16)[name = string("op_7201_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7202_cast_fp16 = mul(x = s_245_cast_fp16, y = s_245_cast_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_7203_cast_fp16 = inverse(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            fp16 var_7204_to_fp16 = const()[name = string("op_7204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_7205_cast_fp16 = mul(x = var_7203_cast_fp16, y = var_7204_to_fp16)[name = string("op_7205_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7207_cast_fp16 = add(x = var_7201_cast_fp16, y = var_7205_cast_fp16)[name = string("op_7207_cast_fp16")];
            fp32 var_7208_epsilon_0 = const()[name = string("op_7208_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_7208_cast_fp16 = rsqrt(epsilon = var_7208_epsilon_0, x = var_7207_cast_fp16)[name = string("op_7208_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_215_cast_fp16 = mul(x = y_245_cast_fp16, y = var_7208_cast_fp16)[name = string("out_215_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_61_cast_fp16 = mul(x = out_215_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_61_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_275_cast_fp16 = transpose(perm = x_275_perm_0, x = var_7172_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 2, 256, 256]> var_7211_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_7211_cast_fp16")];
            bool var_7214 = const()[name = string("op_7214"), val = bool(true)];
            tensor<int32, [1]> var_7215_axes_0 = const()[name = string("op_7215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_7215_cast_fp16 = reduce_max(axes = var_7215_axes_0, keep_dims = var_7214, x = var_7211_cast_fp16)[name = string("op_7215_cast_fp16")];
            fp16 var_7216_to_fp16 = const()[name = string("op_7216_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_247_cast_fp16 = maximum(x = var_7215_cast_fp16, y = var_7216_to_fp16)[name = string("s_247_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_247_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_247_cast_fp16)[name = string("y_247_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7219_cast_fp16 = mul(x = y_247_cast_fp16, y = y_247_cast_fp16)[name = string("op_7219_cast_fp16")];
            tensor<int32, [1]> var_7224_axes_0 = const()[name = string("op_7224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7224_keep_dims_0 = const()[name = string("op_7224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_7224_cast_fp16 = reduce_mean(axes = var_7224_axes_0, keep_dims = var_7224_keep_dims_0, x = var_7219_cast_fp16)[name = string("op_7224_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7225_cast_fp16 = mul(x = s_247_cast_fp16, y = s_247_cast_fp16)[name = string("op_7225_cast_fp16")];
            fp32 var_7226_epsilon_0 = const()[name = string("op_7226_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_7226_cast_fp16 = inverse(epsilon = var_7226_epsilon_0, x = var_7225_cast_fp16)[name = string("op_7226_cast_fp16")];
            fp16 var_7227_to_fp16 = const()[name = string("op_7227_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_7228_cast_fp16 = mul(x = var_7226_cast_fp16, y = var_7227_to_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7230_cast_fp16 = add(x = var_7224_cast_fp16, y = var_7228_cast_fp16)[name = string("op_7230_cast_fp16")];
            fp32 var_7231_epsilon_0 = const()[name = string("op_7231_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_7231_cast_fp16 = rsqrt(epsilon = var_7231_epsilon_0, x = var_7230_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_217_cast_fp16 = mul(x = y_247_cast_fp16, y = var_7231_cast_fp16)[name = string("out_217_cast_fp16")];
            tensor<fp16, [256]> tm_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510080)))];
            tensor<fp16, [1, 2, 256, 256]> k_91_cast_fp16 = mul(x = out_217_cast_fp16, y = tm_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_277_cast_fp16 = transpose(perm = x_277_perm_0, x = var_7184_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2, 256, 256]> var_7234_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_7234_cast_fp16")];
            bool var_7237 = const()[name = string("op_7237"), val = bool(true)];
            tensor<int32, [1]> var_7238_axes_0 = const()[name = string("op_7238_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_7238_cast_fp16 = reduce_max(axes = var_7238_axes_0, keep_dims = var_7237, x = var_7234_cast_fp16)[name = string("op_7238_cast_fp16")];
            fp16 var_7239_to_fp16 = const()[name = string("op_7239_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_249_cast_fp16 = maximum(x = var_7238_cast_fp16, y = var_7239_to_fp16)[name = string("s_249_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_249_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_249_cast_fp16)[name = string("y_249_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7242_cast_fp16 = mul(x = y_249_cast_fp16, y = y_249_cast_fp16)[name = string("op_7242_cast_fp16")];
            tensor<int32, [1]> var_7247_axes_0 = const()[name = string("op_7247_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7247_keep_dims_0 = const()[name = string("op_7247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_7247_cast_fp16 = reduce_mean(axes = var_7247_axes_0, keep_dims = var_7247_keep_dims_0, x = var_7242_cast_fp16)[name = string("op_7247_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7248_cast_fp16 = mul(x = s_249_cast_fp16, y = s_249_cast_fp16)[name = string("op_7248_cast_fp16")];
            fp32 var_7249_epsilon_0 = const()[name = string("op_7249_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_7249_cast_fp16 = inverse(epsilon = var_7249_epsilon_0, x = var_7248_cast_fp16)[name = string("op_7249_cast_fp16")];
            fp16 var_7250_to_fp16 = const()[name = string("op_7250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_7251_cast_fp16 = mul(x = var_7249_cast_fp16, y = var_7250_to_fp16)[name = string("op_7251_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7253_cast_fp16 = add(x = var_7247_cast_fp16, y = var_7251_cast_fp16)[name = string("op_7253_cast_fp16")];
            fp32 var_7254_epsilon_0 = const()[name = string("op_7254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_7254_cast_fp16 = rsqrt(epsilon = var_7254_epsilon_0, x = var_7253_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_61_cast_fp16 = mul(x = y_249_cast_fp16, y = var_7254_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7256_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7256_cast_fp16")];
            tensor<int32, [4]> var_7261_begin_0 = const()[name = string("op_7261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7261_end_0 = const()[name = string("op_7261_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_7261_end_mask_0 = const()[name = string("op_7261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_7261_cast_fp16 = slice_by_index(begin = var_7261_begin_0, end = var_7261_end_0, end_mask = var_7261_end_mask_0, x = q_61_cast_fp16)[name = string("op_7261_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_7262_cast_fp16 = mul(x = var_7261_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_7262_cast_fp16")];
            tensor<int32, [4]> var_7267_begin_0 = const()[name = string("op_7267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7267_end_0 = const()[name = string("op_7267_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_7267_end_mask_0 = const()[name = string("op_7267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_7267_cast_fp16 = slice_by_index(begin = var_7267_begin_0, end = var_7267_end_0, end_mask = var_7267_end_mask_0, x = q_61_cast_fp16)[name = string("op_7267_cast_fp16")];
            int32 var_7269 = const()[name = string("op_7269"), val = int32(-1)];
            bool var_7270_interleave_0 = const()[name = string("op_7270_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_7270_cast_fp16 = concat(axis = var_7269, interleave = var_7270_interleave_0, values = (var_7262_cast_fp16, var_7267_cast_fp16))[name = string("op_7270_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7271_cast_fp16 = mul(x = var_7270_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7271_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_63_cast_fp16 = add(x = var_7256_cast_fp16, y = var_7271_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7274_cast_fp16 = mul(x = k_91_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7274_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, x = k_91_cast_fp16)[name = string("op_7279_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_7280_cast_fp16 = mul(x = var_7279_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_7280_cast_fp16")];
            tensor<int32, [4]> var_7285_begin_0 = const()[name = string("op_7285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7285_end_0 = const()[name = string("op_7285_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_7285_end_mask_0 = const()[name = string("op_7285_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, x = k_91_cast_fp16)[name = string("op_7285_cast_fp16")];
            int32 var_7287 = const()[name = string("op_7287"), val = int32(-1)];
            bool var_7288_interleave_0 = const()[name = string("op_7288_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_7288_cast_fp16 = concat(axis = var_7287, interleave = var_7288_interleave_0, values = (var_7280_cast_fp16, var_7285_cast_fp16))[name = string("op_7288_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7289_cast_fp16 = mul(x = var_7288_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7289_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_93_cast_fp16 = add(x = var_7274_cast_fp16, y = var_7289_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_30_reps_0 = const()[name = string("tile_30_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = k_93_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 1, 256, 256]> tile_30_cast_fp16 = tile(reps = tile_30_reps_0, x = transpose_60_cast_fp16)[name = string("tile_30_cast_fp16")];
            tensor<int32, [5]> concat_60 = const()[name = string("concat_60"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_60_cast_fp16 = reshape(shape = concat_60, x = tile_30_cast_fp16)[name = string("reshape_60_cast_fp16")];
            tensor<int32, [5]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = reshape_60_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 1, 256, 256]> reshape_61_cast_fp16 = reshape(shape = concat_61, x = transpose_61_cast_fp16)[name = string("reshape_61_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_31_reps_0 = const()[name = string("tile_31_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = v_61_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 1, 256, 256]> tile_31_cast_fp16 = tile(reps = tile_31_reps_0, x = transpose_62_cast_fp16)[name = string("tile_31_cast_fp16")];
            tensor<int32, [5]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_62_cast_fp16 = reshape(shape = concat_62, x = tile_31_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [5]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 1, 256, 256]> reshape_63_cast_fp16 = reshape(shape = concat_63, x = transpose_63_cast_fp16)[name = string("reshape_63_cast_fp16")];
            tensor<int32, [4]> v_63_perm_0 = const()[name = string("v_63_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7303_transpose_x_0 = const()[name = string("op_7303_transpose_x_0"), val = bool(false)];
            bool var_7303_transpose_y_0 = const()[name = string("op_7303_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = reshape_61_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 4, 256, 256]> var_7303_cast_fp16 = matmul(transpose_x = var_7303_transpose_x_0, transpose_y = var_7303_transpose_y_0, x = q_63_cast_fp16, y = transpose_111_cast_fp16)[name = string("op_7303_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7305_cast_fp16 = add(x = var_7303_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7305_cast_fp16")];
            int32 var_7306 = const()[name = string("op_7306"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_249_cast_fp16 = softmax(axis = var_7306, x = var_7305_cast_fp16)[name = string("w_249_cast_fp16")];
            bool var_7309_transpose_x_0 = const()[name = string("op_7309_transpose_x_0"), val = bool(false)];
            bool var_7309_transpose_y_0 = const()[name = string("op_7309_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = reshape_63_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 4, 256, 256]> var_7309_cast_fp16 = matmul(transpose_x = var_7309_transpose_x_0, transpose_y = var_7309_transpose_y_0, x = w_249_cast_fp16, y = v_63_cast_fp16)[name = string("op_7309_cast_fp16")];
            tensor<int32, [4]> var_7312_perm_0 = const()[name = string("op_7312_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7316 = const()[name = string("op_7316"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_7312_cast_fp16 = transpose(perm = var_7312_perm_0, x = var_7309_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 256, 1024]> input_183_cast_fp16 = reshape(shape = var_7316, x = var_7312_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510656))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89035008))))[name = string("tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7321_cast_fp16 = abs(x = linear_139_cast_fp16)[name = string("op_7321_cast_fp16")];
            bool var_7324 = const()[name = string("op_7324"), val = bool(true)];
            tensor<int32, [1]> var_7325_axes_0 = const()[name = string("op_7325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7325_cast_fp16 = reduce_max(axes = var_7325_axes_0, keep_dims = var_7324, x = var_7321_cast_fp16)[name = string("op_7325_cast_fp16")];
            fp16 var_7326_to_fp16 = const()[name = string("op_7326_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_251_cast_fp16 = maximum(x = var_7325_cast_fp16, y = var_7326_to_fp16)[name = string("s_251_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_251_cast_fp16 = real_div(x = linear_139_cast_fp16, y = s_251_cast_fp16)[name = string("y_251_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7329_cast_fp16 = mul(x = y_251_cast_fp16, y = y_251_cast_fp16)[name = string("op_7329_cast_fp16")];
            tensor<int32, [1]> var_7334_axes_0 = const()[name = string("op_7334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7334_keep_dims_0 = const()[name = string("op_7334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7334_cast_fp16 = reduce_mean(axes = var_7334_axes_0, keep_dims = var_7334_keep_dims_0, x = var_7329_cast_fp16)[name = string("op_7334_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7335_cast_fp16 = mul(x = s_251_cast_fp16, y = s_251_cast_fp16)[name = string("op_7335_cast_fp16")];
            fp32 var_7336_epsilon_0 = const()[name = string("op_7336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7336_cast_fp16 = inverse(epsilon = var_7336_epsilon_0, x = var_7335_cast_fp16)[name = string("op_7336_cast_fp16")];
            fp16 var_7337_to_fp16 = const()[name = string("op_7337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7338_cast_fp16 = mul(x = var_7336_cast_fp16, y = var_7337_to_fp16)[name = string("op_7338_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7340_cast_fp16 = add(x = var_7334_cast_fp16, y = var_7338_cast_fp16)[name = string("op_7340_cast_fp16")];
            fp32 var_7341_epsilon_0 = const()[name = string("op_7341_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7341_cast_fp16 = rsqrt(epsilon = var_7341_epsilon_0, x = var_7340_cast_fp16)[name = string("op_7341_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_219_cast_fp16 = mul(x = y_251_cast_fp16, y = var_7341_cast_fp16)[name = string("out_219_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89036096)))];
            tensor<fp16, [1, 256, 512]> var_7343_cast_fp16 = mul(x = out_219_cast_fp16, y = tm_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_7343_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_281_cast_fp16 = add(x = x_271_cast_fp16, y = var_7343_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7346_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_7346_cast_fp16")];
            bool var_7349 = const()[name = string("op_7349"), val = bool(true)];
            tensor<int32, [1]> var_7350_axes_0 = const()[name = string("op_7350_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7350_cast_fp16 = reduce_max(axes = var_7350_axes_0, keep_dims = var_7349, x = var_7346_cast_fp16)[name = string("op_7350_cast_fp16")];
            fp16 var_7351_to_fp16 = const()[name = string("op_7351_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_253_cast_fp16 = maximum(x = var_7350_cast_fp16, y = var_7351_to_fp16)[name = string("s_253_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_253_cast_fp16 = real_div(x = x_281_cast_fp16, y = s_253_cast_fp16)[name = string("y_253_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7354_cast_fp16 = mul(x = y_253_cast_fp16, y = y_253_cast_fp16)[name = string("op_7354_cast_fp16")];
            tensor<int32, [1]> var_7359_axes_0 = const()[name = string("op_7359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7359_keep_dims_0 = const()[name = string("op_7359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7359_cast_fp16 = reduce_mean(axes = var_7359_axes_0, keep_dims = var_7359_keep_dims_0, x = var_7354_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7360_cast_fp16 = mul(x = s_253_cast_fp16, y = s_253_cast_fp16)[name = string("op_7360_cast_fp16")];
            fp32 var_7361_epsilon_0 = const()[name = string("op_7361_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7361_cast_fp16 = inverse(epsilon = var_7361_epsilon_0, x = var_7360_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7362_to_fp16 = const()[name = string("op_7362_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7363_cast_fp16 = mul(x = var_7361_cast_fp16, y = var_7362_to_fp16)[name = string("op_7363_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7365_cast_fp16 = add(x = var_7359_cast_fp16, y = var_7363_cast_fp16)[name = string("op_7365_cast_fp16")];
            fp32 var_7366_epsilon_0 = const()[name = string("op_7366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7366_cast_fp16 = rsqrt(epsilon = var_7366_epsilon_0, x = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_221_cast_fp16 = mul(x = y_253_cast_fp16, y = var_7366_cast_fp16)[name = string("out_221_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89037184)))];
            tensor<fp16, [1, 256, 512]> input_185_cast_fp16 = mul(x = out_221_cast_fp16, y = tm_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89038272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90086912))))[name = string("tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_140_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_140_cast_fp16")];
            string var_7373_mode_0 = const()[name = string("op_7373_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_7373_cast_fp16 = gelu(mode = var_7373_mode_0, x = linear_140_cast_fp16)[name = string("op_7373_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90091072))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91139712))))[name = string("tm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_141_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_187_cast_fp16 = mul(x = var_7373_cast_fp16, y = linear_141_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91143872))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92192512))))[name = string("tm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7381_cast_fp16 = abs(x = linear_142_cast_fp16)[name = string("op_7381_cast_fp16")];
            bool var_7384 = const()[name = string("op_7384"), val = bool(true)];
            tensor<int32, [1]> var_7385_axes_0 = const()[name = string("op_7385_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7385_cast_fp16 = reduce_max(axes = var_7385_axes_0, keep_dims = var_7384, x = var_7381_cast_fp16)[name = string("op_7385_cast_fp16")];
            fp16 var_7386_to_fp16 = const()[name = string("op_7386_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_255_cast_fp16 = maximum(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = string("s_255_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_255_cast_fp16 = real_div(x = linear_142_cast_fp16, y = s_255_cast_fp16)[name = string("y_255_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7389_cast_fp16 = mul(x = y_255_cast_fp16, y = y_255_cast_fp16)[name = string("op_7389_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7389_cast_fp16)[name = string("op_7394_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7395_cast_fp16 = mul(x = s_255_cast_fp16, y = s_255_cast_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7400_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7400_cast_fp16")];
            fp32 var_7401_epsilon_0 = const()[name = string("op_7401_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7401_cast_fp16 = rsqrt(epsilon = var_7401_epsilon_0, x = var_7400_cast_fp16)[name = string("op_7401_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_223_cast_fp16 = mul(x = y_255_cast_fp16, y = var_7401_cast_fp16)[name = string("out_223_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92193600)))];
            tensor<fp16, [1, 256, 512]> var_7403_cast_fp16 = mul(x = out_223_cast_fp16, y = tm_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7403_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_189_cast_fp16 = add(x = x_281_cast_fp16, y = var_7403_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92194688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92456896))))[name = string("tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_143_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_143_cast_fp16")];
            string x_285_mode_0 = const()[name = string("x_285_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_285_cast_fp16 = gelu(mode = x_285_mode_0, x = linear_143_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [4]> var_7423_begin_0 = const()[name = string("op_7423_begin_0"), val = tensor<int32, [4]>([0, 0, 15, 0])];
            tensor<int32, [4]> var_7423_end_0 = const()[name = string("op_7423_end_0"), val = tensor<int32, [4]>([1, 256, 16, 512])];
            tensor<bool, [4]> var_7423_end_mask_0 = const()[name = string("op_7423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7423_squeeze_mask_0 = const()[name = string("op_7423_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_7423_cast_fp16 = slice_by_index(begin = var_7423_begin_0, end = var_7423_end_0, end_mask = var_7423_end_mask_0, squeeze_mask = var_7423_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7423_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_191_cast_fp16 = mul(x = x_285_cast_fp16, y = var_7423_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92457984))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92720192))))[name = string("tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7433_cast_fp16 = abs(x = linear_144_cast_fp16)[name = string("op_7433_cast_fp16")];
            bool var_7436 = const()[name = string("op_7436"), val = bool(true)];
            tensor<int32, [1]> var_7437_axes_0 = const()[name = string("op_7437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7437_cast_fp16 = reduce_max(axes = var_7437_axes_0, keep_dims = var_7436, x = var_7433_cast_fp16)[name = string("op_7437_cast_fp16")];
            fp16 var_7438_to_fp16 = const()[name = string("op_7438_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_257_cast_fp16 = maximum(x = var_7437_cast_fp16, y = var_7438_to_fp16)[name = string("s_257_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_257_cast_fp16 = real_div(x = linear_144_cast_fp16, y = s_257_cast_fp16)[name = string("y_257_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7441_cast_fp16 = mul(x = y_257_cast_fp16, y = y_257_cast_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7446_axes_0 = const()[name = string("op_7446_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7446_keep_dims_0 = const()[name = string("op_7446_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7446_cast_fp16 = reduce_mean(axes = var_7446_axes_0, keep_dims = var_7446_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7447_cast_fp16 = mul(x = s_257_cast_fp16, y = s_257_cast_fp16)[name = string("op_7447_cast_fp16")];
            fp32 var_7448_epsilon_0 = const()[name = string("op_7448_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7448_cast_fp16 = inverse(epsilon = var_7448_epsilon_0, x = var_7447_cast_fp16)[name = string("op_7448_cast_fp16")];
            fp16 var_7449_to_fp16 = const()[name = string("op_7449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7450_cast_fp16 = mul(x = var_7448_cast_fp16, y = var_7449_to_fp16)[name = string("op_7450_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7452_cast_fp16 = add(x = var_7446_cast_fp16, y = var_7450_cast_fp16)[name = string("op_7452_cast_fp16")];
            fp32 var_7453_epsilon_0 = const()[name = string("op_7453_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7453_cast_fp16 = rsqrt(epsilon = var_7453_epsilon_0, x = var_7452_cast_fp16)[name = string("op_7453_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_225_cast_fp16 = mul(x = y_257_cast_fp16, y = var_7453_cast_fp16)[name = string("out_225_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92721280)))];
            tensor<fp16, [1, 256, 512]> var_7455_cast_fp16 = mul(x = out_225_cast_fp16, y = tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7455_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_31_cast_fp16 = add(x = input_189_cast_fp16, y = var_7455_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<fp16, [1]> tm_layers_15_layer_scalar_to_fp16 = const()[name = string("tm_layers_15_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7ap-1])];
            tensor<fp16, [1, 256, 512]> x_289_cast_fp16 = mul(x = h_31_cast_fp16, y = tm_layers_15_layer_scalar_to_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7459_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_7459_cast_fp16")];
            bool var_7462 = const()[name = string("op_7462"), val = bool(true)];
            tensor<int32, [1]> var_7463_axes_0 = const()[name = string("op_7463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7463_cast_fp16 = reduce_max(axes = var_7463_axes_0, keep_dims = var_7462, x = var_7459_cast_fp16)[name = string("op_7463_cast_fp16")];
            fp16 var_7464_to_fp16 = const()[name = string("op_7464_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_259_cast_fp16 = maximum(x = var_7463_cast_fp16, y = var_7464_to_fp16)[name = string("s_259_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_259_cast_fp16 = real_div(x = x_289_cast_fp16, y = s_259_cast_fp16)[name = string("y_259_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7467_cast_fp16 = mul(x = y_259_cast_fp16, y = y_259_cast_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7472_axes_0 = const()[name = string("op_7472_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7472_keep_dims_0 = const()[name = string("op_7472_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7472_cast_fp16 = reduce_mean(axes = var_7472_axes_0, keep_dims = var_7472_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7473_cast_fp16 = mul(x = s_259_cast_fp16, y = s_259_cast_fp16)[name = string("op_7473_cast_fp16")];
            fp32 var_7474_epsilon_0 = const()[name = string("op_7474_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7474_cast_fp16 = inverse(epsilon = var_7474_epsilon_0, x = var_7473_cast_fp16)[name = string("op_7474_cast_fp16")];
            fp16 var_7475_to_fp16 = const()[name = string("op_7475_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7476_cast_fp16 = mul(x = var_7474_cast_fp16, y = var_7475_to_fp16)[name = string("op_7476_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7478_cast_fp16 = add(x = var_7472_cast_fp16, y = var_7476_cast_fp16)[name = string("op_7478_cast_fp16")];
            fp32 var_7479_epsilon_0 = const()[name = string("op_7479_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7479_cast_fp16 = rsqrt(epsilon = var_7479_epsilon_0, x = var_7478_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_227_cast_fp16 = mul(x = y_259_cast_fp16, y = var_7479_cast_fp16)[name = string("out_227_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92722368)))];
            tensor<fp16, [1, 256, 512]> input_193_cast_fp16 = mul(x = out_227_cast_fp16, y = tm_layers_16_input_layernorm_weight_to_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92723456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93247808))))[name = string("tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<int32, [4]> var_7489 = const()[name = string("op_7489"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_7490_cast_fp16 = reshape(shape = var_7489, x = linear_145_cast_fp16)[name = string("op_7490_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93249920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93512128))))[name = string("tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<int32, [4]> var_7501 = const()[name = string("op_7501"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_7502_cast_fp16 = reshape(shape = var_7501, x = linear_146_cast_fp16)[name = string("op_7502_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93513216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93775424))))[name = string("tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7513 = const()[name = string("op_7513"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_7514_cast_fp16 = reshape(shape = var_7513, x = linear_147_cast_fp16)[name = string("op_7514_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_7490_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 4, 256, 256]> var_7518_cast_fp16 = abs(x = x_291_cast_fp16)[name = string("op_7518_cast_fp16")];
            bool var_7521 = const()[name = string("op_7521"), val = bool(true)];
            tensor<int32, [1]> var_7522_axes_0 = const()[name = string("op_7522_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_7522_cast_fp16 = reduce_max(axes = var_7522_axes_0, keep_dims = var_7521, x = var_7518_cast_fp16)[name = string("op_7522_cast_fp16")];
            fp16 var_7523_to_fp16 = const()[name = string("op_7523_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_261_cast_fp16 = maximum(x = var_7522_cast_fp16, y = var_7523_to_fp16)[name = string("s_261_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_261_cast_fp16 = real_div(x = x_291_cast_fp16, y = s_261_cast_fp16)[name = string("y_261_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7526_cast_fp16 = mul(x = y_261_cast_fp16, y = y_261_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<int32, [1]> var_7531_axes_0 = const()[name = string("op_7531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7531_keep_dims_0 = const()[name = string("op_7531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_7531_cast_fp16 = reduce_mean(axes = var_7531_axes_0, keep_dims = var_7531_keep_dims_0, x = var_7526_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7532_cast_fp16 = mul(x = s_261_cast_fp16, y = s_261_cast_fp16)[name = string("op_7532_cast_fp16")];
            fp32 var_7533_epsilon_0 = const()[name = string("op_7533_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_7533_cast_fp16 = inverse(epsilon = var_7533_epsilon_0, x = var_7532_cast_fp16)[name = string("op_7533_cast_fp16")];
            fp16 var_7534_to_fp16 = const()[name = string("op_7534_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_7535_cast_fp16 = mul(x = var_7533_cast_fp16, y = var_7534_to_fp16)[name = string("op_7535_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7537_cast_fp16 = add(x = var_7531_cast_fp16, y = var_7535_cast_fp16)[name = string("op_7537_cast_fp16")];
            fp32 var_7538_epsilon_0 = const()[name = string("op_7538_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_7538_cast_fp16 = rsqrt(epsilon = var_7538_epsilon_0, x = var_7537_cast_fp16)[name = string("op_7538_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_229_cast_fp16 = mul(x = y_261_cast_fp16, y = var_7538_cast_fp16)[name = string("out_229_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_65_cast_fp16 = mul(x = out_229_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_65_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_7502_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 256, 256]> var_7541_cast_fp16 = abs(x = x_293_cast_fp16)[name = string("op_7541_cast_fp16")];
            bool var_7544 = const()[name = string("op_7544"), val = bool(true)];
            tensor<int32, [1]> var_7545_axes_0 = const()[name = string("op_7545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_7545_cast_fp16 = reduce_max(axes = var_7545_axes_0, keep_dims = var_7544, x = var_7541_cast_fp16)[name = string("op_7545_cast_fp16")];
            fp16 var_7546_to_fp16 = const()[name = string("op_7546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_263_cast_fp16 = maximum(x = var_7545_cast_fp16, y = var_7546_to_fp16)[name = string("s_263_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_263_cast_fp16 = real_div(x = x_293_cast_fp16, y = s_263_cast_fp16)[name = string("y_263_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7549_cast_fp16 = mul(x = y_263_cast_fp16, y = y_263_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<int32, [1]> var_7554_axes_0 = const()[name = string("op_7554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7554_keep_dims_0 = const()[name = string("op_7554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_7554_cast_fp16 = reduce_mean(axes = var_7554_axes_0, keep_dims = var_7554_keep_dims_0, x = var_7549_cast_fp16)[name = string("op_7554_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7555_cast_fp16 = mul(x = s_263_cast_fp16, y = s_263_cast_fp16)[name = string("op_7555_cast_fp16")];
            fp32 var_7556_epsilon_0 = const()[name = string("op_7556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_7556_cast_fp16 = inverse(epsilon = var_7556_epsilon_0, x = var_7555_cast_fp16)[name = string("op_7556_cast_fp16")];
            fp16 var_7557_to_fp16 = const()[name = string("op_7557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_7558_cast_fp16 = mul(x = var_7556_cast_fp16, y = var_7557_to_fp16)[name = string("op_7558_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7560_cast_fp16 = add(x = var_7554_cast_fp16, y = var_7558_cast_fp16)[name = string("op_7560_cast_fp16")];
            fp32 var_7561_epsilon_0 = const()[name = string("op_7561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_7561_cast_fp16 = rsqrt(epsilon = var_7561_epsilon_0, x = var_7560_cast_fp16)[name = string("op_7561_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_231_cast_fp16 = mul(x = y_263_cast_fp16, y = var_7561_cast_fp16)[name = string("out_231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_97_cast_fp16 = mul(x = out_231_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_97_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_7514_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 256, 256]> var_7564_cast_fp16 = abs(x = x_295_cast_fp16)[name = string("op_7564_cast_fp16")];
            bool var_7567 = const()[name = string("op_7567"), val = bool(true)];
            tensor<int32, [1]> var_7568_axes_0 = const()[name = string("op_7568_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_7568_cast_fp16 = reduce_max(axes = var_7568_axes_0, keep_dims = var_7567, x = var_7564_cast_fp16)[name = string("op_7568_cast_fp16")];
            fp16 var_7569_to_fp16 = const()[name = string("op_7569_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_265_cast_fp16 = maximum(x = var_7568_cast_fp16, y = var_7569_to_fp16)[name = string("s_265_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_265_cast_fp16 = real_div(x = x_295_cast_fp16, y = s_265_cast_fp16)[name = string("y_265_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7572_cast_fp16 = mul(x = y_265_cast_fp16, y = y_265_cast_fp16)[name = string("op_7572_cast_fp16")];
            tensor<int32, [1]> var_7577_axes_0 = const()[name = string("op_7577_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7577_keep_dims_0 = const()[name = string("op_7577_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_7577_cast_fp16 = reduce_mean(axes = var_7577_axes_0, keep_dims = var_7577_keep_dims_0, x = var_7572_cast_fp16)[name = string("op_7577_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7578_cast_fp16 = mul(x = s_265_cast_fp16, y = s_265_cast_fp16)[name = string("op_7578_cast_fp16")];
            fp32 var_7579_epsilon_0 = const()[name = string("op_7579_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_7579_cast_fp16 = inverse(epsilon = var_7579_epsilon_0, x = var_7578_cast_fp16)[name = string("op_7579_cast_fp16")];
            fp16 var_7580_to_fp16 = const()[name = string("op_7580_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_7581_cast_fp16 = mul(x = var_7579_cast_fp16, y = var_7580_to_fp16)[name = string("op_7581_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7583_cast_fp16 = add(x = var_7577_cast_fp16, y = var_7581_cast_fp16)[name = string("op_7583_cast_fp16")];
            fp32 var_7584_epsilon_0 = const()[name = string("op_7584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_7584_cast_fp16 = rsqrt(epsilon = var_7584_epsilon_0, x = var_7583_cast_fp16)[name = string("op_7584_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_65_cast_fp16 = mul(x = y_265_cast_fp16, y = var_7584_cast_fp16)[name = string("v_65_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7586_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7586_cast_fp16")];
            tensor<int32, [4]> var_7591_begin_0 = const()[name = string("op_7591_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7591_end_0 = const()[name = string("op_7591_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_7591_end_mask_0 = const()[name = string("op_7591_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_7591_cast_fp16 = slice_by_index(begin = var_7591_begin_0, end = var_7591_end_0, end_mask = var_7591_end_mask_0, x = q_65_cast_fp16)[name = string("op_7591_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_7592_cast_fp16 = mul(x = var_7591_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_7592_cast_fp16")];
            tensor<int32, [4]> var_7597_begin_0 = const()[name = string("op_7597_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7597_end_0 = const()[name = string("op_7597_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_7597_end_mask_0 = const()[name = string("op_7597_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, x = q_65_cast_fp16)[name = string("op_7597_cast_fp16")];
            int32 var_7599 = const()[name = string("op_7599"), val = int32(-1)];
            bool var_7600_interleave_0 = const()[name = string("op_7600_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_7600_cast_fp16 = concat(axis = var_7599, interleave = var_7600_interleave_0, values = (var_7592_cast_fp16, var_7597_cast_fp16))[name = string("op_7600_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7601_cast_fp16 = mul(x = var_7600_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7601_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_67_cast_fp16 = add(x = var_7586_cast_fp16, y = var_7601_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7604_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7604_cast_fp16")];
            tensor<int32, [4]> var_7609_begin_0 = const()[name = string("op_7609_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7609_end_0 = const()[name = string("op_7609_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_7609_end_mask_0 = const()[name = string("op_7609_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_7609_cast_fp16 = slice_by_index(begin = var_7609_begin_0, end = var_7609_end_0, end_mask = var_7609_end_mask_0, x = k_97_cast_fp16)[name = string("op_7609_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_7610_cast_fp16 = mul(x = var_7609_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_7610_cast_fp16")];
            tensor<int32, [4]> var_7615_begin_0 = const()[name = string("op_7615_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7615_end_0 = const()[name = string("op_7615_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_7615_end_mask_0 = const()[name = string("op_7615_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_7615_cast_fp16 = slice_by_index(begin = var_7615_begin_0, end = var_7615_end_0, end_mask = var_7615_end_mask_0, x = k_97_cast_fp16)[name = string("op_7615_cast_fp16")];
            int32 var_7617 = const()[name = string("op_7617"), val = int32(-1)];
            bool var_7618_interleave_0 = const()[name = string("op_7618_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_7618_cast_fp16 = concat(axis = var_7617, interleave = var_7618_interleave_0, values = (var_7610_cast_fp16, var_7615_cast_fp16))[name = string("op_7618_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7619_cast_fp16 = mul(x = var_7618_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7619_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_99_cast_fp16 = add(x = var_7604_cast_fp16, y = var_7619_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_32_reps_0 = const()[name = string("tile_32_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = k_99_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 1, 256, 256]> tile_32_cast_fp16 = tile(reps = tile_32_reps_0, x = transpose_64_cast_fp16)[name = string("tile_32_cast_fp16")];
            tensor<int32, [5]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_64_cast_fp16 = reshape(shape = concat_64, x = tile_32_cast_fp16)[name = string("reshape_64_cast_fp16")];
            tensor<int32, [5]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = reshape_64_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 1, 256, 256]> reshape_65_cast_fp16 = reshape(shape = concat_65, x = transpose_65_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_33_reps_0 = const()[name = string("tile_33_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = v_65_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 1, 256, 256]> tile_33_cast_fp16 = tile(reps = tile_33_reps_0, x = transpose_66_cast_fp16)[name = string("tile_33_cast_fp16")];
            tensor<int32, [5]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_66_cast_fp16 = reshape(shape = concat_66, x = tile_33_cast_fp16)[name = string("reshape_66_cast_fp16")];
            tensor<int32, [5]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = reshape_66_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 1, 256, 256]> reshape_67_cast_fp16 = reshape(shape = concat_67, x = transpose_67_cast_fp16)[name = string("reshape_67_cast_fp16")];
            tensor<int32, [4]> v_67_perm_0 = const()[name = string("v_67_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7633_transpose_x_0 = const()[name = string("op_7633_transpose_x_0"), val = bool(false)];
            bool var_7633_transpose_y_0 = const()[name = string("op_7633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_112_cast_fp16 = transpose(perm = transpose_112_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 4, 256, 256]> var_7633_cast_fp16 = matmul(transpose_x = var_7633_transpose_x_0, transpose_y = var_7633_transpose_y_0, x = q_67_cast_fp16, y = transpose_112_cast_fp16)[name = string("op_7633_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7635_cast_fp16 = add(x = var_7633_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7635_cast_fp16")];
            int32 var_7636 = const()[name = string("op_7636"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_265_cast_fp16 = softmax(axis = var_7636, x = var_7635_cast_fp16)[name = string("w_265_cast_fp16")];
            bool var_7639_transpose_x_0 = const()[name = string("op_7639_transpose_x_0"), val = bool(false)];
            bool var_7639_transpose_y_0 = const()[name = string("op_7639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = reshape_67_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 4, 256, 256]> var_7639_cast_fp16 = matmul(transpose_x = var_7639_transpose_x_0, transpose_y = var_7639_transpose_y_0, x = w_265_cast_fp16, y = v_67_cast_fp16)[name = string("op_7639_cast_fp16")];
            tensor<int32, [4]> var_7642_perm_0 = const()[name = string("op_7642_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7646 = const()[name = string("op_7646"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_7642_cast_fp16 = transpose(perm = var_7642_perm_0, x = var_7639_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 256, 1024]> input_195_cast_fp16 = reshape(shape = var_7646, x = var_7642_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93776512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94300864))))[name = string("tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7651_cast_fp16 = abs(x = linear_148_cast_fp16)[name = string("op_7651_cast_fp16")];
            bool var_7654 = const()[name = string("op_7654"), val = bool(true)];
            tensor<int32, [1]> var_7655_axes_0 = const()[name = string("op_7655_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7655_cast_fp16 = reduce_max(axes = var_7655_axes_0, keep_dims = var_7654, x = var_7651_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp16 var_7656_to_fp16 = const()[name = string("op_7656_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_267_cast_fp16 = maximum(x = var_7655_cast_fp16, y = var_7656_to_fp16)[name = string("s_267_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_267_cast_fp16 = real_div(x = linear_148_cast_fp16, y = s_267_cast_fp16)[name = string("y_267_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7659_cast_fp16 = mul(x = y_267_cast_fp16, y = y_267_cast_fp16)[name = string("op_7659_cast_fp16")];
            tensor<int32, [1]> var_7664_axes_0 = const()[name = string("op_7664_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7664_keep_dims_0 = const()[name = string("op_7664_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7664_cast_fp16 = reduce_mean(axes = var_7664_axes_0, keep_dims = var_7664_keep_dims_0, x = var_7659_cast_fp16)[name = string("op_7664_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7665_cast_fp16 = mul(x = s_267_cast_fp16, y = s_267_cast_fp16)[name = string("op_7665_cast_fp16")];
            fp32 var_7666_epsilon_0 = const()[name = string("op_7666_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7666_cast_fp16 = inverse(epsilon = var_7666_epsilon_0, x = var_7665_cast_fp16)[name = string("op_7666_cast_fp16")];
            fp16 var_7667_to_fp16 = const()[name = string("op_7667_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7668_cast_fp16 = mul(x = var_7666_cast_fp16, y = var_7667_to_fp16)[name = string("op_7668_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7670_cast_fp16 = add(x = var_7664_cast_fp16, y = var_7668_cast_fp16)[name = string("op_7670_cast_fp16")];
            fp32 var_7671_epsilon_0 = const()[name = string("op_7671_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7671_cast_fp16 = rsqrt(epsilon = var_7671_epsilon_0, x = var_7670_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_233_cast_fp16 = mul(x = y_267_cast_fp16, y = var_7671_cast_fp16)[name = string("out_233_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94301952)))];
            tensor<fp16, [1, 256, 512]> var_7673_cast_fp16 = mul(x = out_233_cast_fp16, y = tm_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("op_7673_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = var_7673_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7676_cast_fp16 = abs(x = x_299_cast_fp16)[name = string("op_7676_cast_fp16")];
            bool var_7679 = const()[name = string("op_7679"), val = bool(true)];
            tensor<int32, [1]> var_7680_axes_0 = const()[name = string("op_7680_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7680_cast_fp16 = reduce_max(axes = var_7680_axes_0, keep_dims = var_7679, x = var_7676_cast_fp16)[name = string("op_7680_cast_fp16")];
            fp16 var_7681_to_fp16 = const()[name = string("op_7681_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_269_cast_fp16 = maximum(x = var_7680_cast_fp16, y = var_7681_to_fp16)[name = string("s_269_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_269_cast_fp16 = real_div(x = x_299_cast_fp16, y = s_269_cast_fp16)[name = string("y_269_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7684_cast_fp16 = mul(x = y_269_cast_fp16, y = y_269_cast_fp16)[name = string("op_7684_cast_fp16")];
            tensor<int32, [1]> var_7689_axes_0 = const()[name = string("op_7689_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7689_keep_dims_0 = const()[name = string("op_7689_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7689_cast_fp16 = reduce_mean(axes = var_7689_axes_0, keep_dims = var_7689_keep_dims_0, x = var_7684_cast_fp16)[name = string("op_7689_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7690_cast_fp16 = mul(x = s_269_cast_fp16, y = s_269_cast_fp16)[name = string("op_7690_cast_fp16")];
            fp32 var_7691_epsilon_0 = const()[name = string("op_7691_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7691_cast_fp16 = inverse(epsilon = var_7691_epsilon_0, x = var_7690_cast_fp16)[name = string("op_7691_cast_fp16")];
            fp16 var_7692_to_fp16 = const()[name = string("op_7692_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7693_cast_fp16 = mul(x = var_7691_cast_fp16, y = var_7692_to_fp16)[name = string("op_7693_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7695_cast_fp16 = add(x = var_7689_cast_fp16, y = var_7693_cast_fp16)[name = string("op_7695_cast_fp16")];
            fp32 var_7696_epsilon_0 = const()[name = string("op_7696_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7696_cast_fp16 = rsqrt(epsilon = var_7696_epsilon_0, x = var_7695_cast_fp16)[name = string("op_7696_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_235_cast_fp16 = mul(x = y_269_cast_fp16, y = var_7696_cast_fp16)[name = string("out_235_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94303040)))];
            tensor<fp16, [1, 256, 512]> input_197_cast_fp16 = mul(x = out_235_cast_fp16, y = tm_layers_16_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94304128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95352768))))[name = string("tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_149_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_149_cast_fp16")];
            string var_7703_mode_0 = const()[name = string("op_7703_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_7703_cast_fp16 = gelu(mode = var_7703_mode_0, x = linear_149_cast_fp16)[name = string("op_7703_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95356928))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96405568))))[name = string("tm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_150_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_199_cast_fp16 = mul(x = var_7703_cast_fp16, y = linear_150_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96409728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97458368))))[name = string("tm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7711_cast_fp16 = abs(x = linear_151_cast_fp16)[name = string("op_7711_cast_fp16")];
            bool var_7714 = const()[name = string("op_7714"), val = bool(true)];
            tensor<int32, [1]> var_7715_axes_0 = const()[name = string("op_7715_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7715_cast_fp16 = reduce_max(axes = var_7715_axes_0, keep_dims = var_7714, x = var_7711_cast_fp16)[name = string("op_7715_cast_fp16")];
            fp16 var_7716_to_fp16 = const()[name = string("op_7716_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_271_cast_fp16 = maximum(x = var_7715_cast_fp16, y = var_7716_to_fp16)[name = string("s_271_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_271_cast_fp16 = real_div(x = linear_151_cast_fp16, y = s_271_cast_fp16)[name = string("y_271_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7719_cast_fp16 = mul(x = y_271_cast_fp16, y = y_271_cast_fp16)[name = string("op_7719_cast_fp16")];
            tensor<int32, [1]> var_7724_axes_0 = const()[name = string("op_7724_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7724_keep_dims_0 = const()[name = string("op_7724_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7724_cast_fp16 = reduce_mean(axes = var_7724_axes_0, keep_dims = var_7724_keep_dims_0, x = var_7719_cast_fp16)[name = string("op_7724_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7725_cast_fp16 = mul(x = s_271_cast_fp16, y = s_271_cast_fp16)[name = string("op_7725_cast_fp16")];
            fp32 var_7726_epsilon_0 = const()[name = string("op_7726_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7726_cast_fp16 = inverse(epsilon = var_7726_epsilon_0, x = var_7725_cast_fp16)[name = string("op_7726_cast_fp16")];
            fp16 var_7727_to_fp16 = const()[name = string("op_7727_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7728_cast_fp16 = mul(x = var_7726_cast_fp16, y = var_7727_to_fp16)[name = string("op_7728_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7730_cast_fp16 = add(x = var_7724_cast_fp16, y = var_7728_cast_fp16)[name = string("op_7730_cast_fp16")];
            fp32 var_7731_epsilon_0 = const()[name = string("op_7731_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7731_cast_fp16 = rsqrt(epsilon = var_7731_epsilon_0, x = var_7730_cast_fp16)[name = string("op_7731_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_237_cast_fp16 = mul(x = y_271_cast_fp16, y = var_7731_cast_fp16)[name = string("out_237_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97459456)))];
            tensor<fp16, [1, 256, 512]> var_7733_cast_fp16 = mul(x = out_237_cast_fp16, y = tm_layers_16_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7733_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_201_cast_fp16 = add(x = x_299_cast_fp16, y = var_7733_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97460544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97722752))))[name = string("tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_152_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_152_cast_fp16")];
            string x_303_mode_0 = const()[name = string("x_303_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_303_cast_fp16 = gelu(mode = x_303_mode_0, x = linear_152_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [4]> var_7753_begin_0 = const()[name = string("op_7753_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7753_end_0 = const()[name = string("op_7753_end_0"), val = tensor<int32, [4]>([1, 256, 17, 512])];
            tensor<bool, [4]> var_7753_end_mask_0 = const()[name = string("op_7753_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7753_squeeze_mask_0 = const()[name = string("op_7753_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_7753_cast_fp16 = slice_by_index(begin = var_7753_begin_0, end = var_7753_end_0, end_mask = var_7753_end_mask_0, squeeze_mask = var_7753_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7753_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_203_cast_fp16 = mul(x = x_303_cast_fp16, y = var_7753_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97723840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97986048))))[name = string("tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7763_cast_fp16 = abs(x = linear_153_cast_fp16)[name = string("op_7763_cast_fp16")];
            bool var_7766 = const()[name = string("op_7766"), val = bool(true)];
            tensor<int32, [1]> var_7767_axes_0 = const()[name = string("op_7767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7767_cast_fp16 = reduce_max(axes = var_7767_axes_0, keep_dims = var_7766, x = var_7763_cast_fp16)[name = string("op_7767_cast_fp16")];
            fp16 var_7768_to_fp16 = const()[name = string("op_7768_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_273_cast_fp16 = maximum(x = var_7767_cast_fp16, y = var_7768_to_fp16)[name = string("s_273_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_273_cast_fp16 = real_div(x = linear_153_cast_fp16, y = s_273_cast_fp16)[name = string("y_273_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7771_cast_fp16 = mul(x = y_273_cast_fp16, y = y_273_cast_fp16)[name = string("op_7771_cast_fp16")];
            tensor<int32, [1]> var_7776_axes_0 = const()[name = string("op_7776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7776_keep_dims_0 = const()[name = string("op_7776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7776_cast_fp16 = reduce_mean(axes = var_7776_axes_0, keep_dims = var_7776_keep_dims_0, x = var_7771_cast_fp16)[name = string("op_7776_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7777_cast_fp16 = mul(x = s_273_cast_fp16, y = s_273_cast_fp16)[name = string("op_7777_cast_fp16")];
            fp32 var_7778_epsilon_0 = const()[name = string("op_7778_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7778_cast_fp16 = inverse(epsilon = var_7778_epsilon_0, x = var_7777_cast_fp16)[name = string("op_7778_cast_fp16")];
            fp16 var_7779_to_fp16 = const()[name = string("op_7779_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7780_cast_fp16 = mul(x = var_7778_cast_fp16, y = var_7779_to_fp16)[name = string("op_7780_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7782_cast_fp16 = add(x = var_7776_cast_fp16, y = var_7780_cast_fp16)[name = string("op_7782_cast_fp16")];
            fp32 var_7783_epsilon_0 = const()[name = string("op_7783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7783_cast_fp16 = rsqrt(epsilon = var_7783_epsilon_0, x = var_7782_cast_fp16)[name = string("op_7783_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_239_cast_fp16 = mul(x = y_273_cast_fp16, y = var_7783_cast_fp16)[name = string("out_239_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97987136)))];
            tensor<fp16, [1, 256, 512]> var_7785_cast_fp16 = mul(x = out_239_cast_fp16, y = tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_33_cast_fp16 = add(x = input_201_cast_fp16, y = var_7785_cast_fp16)[name = string("h_33_cast_fp16")];
            tensor<fp16, [1]> tm_layers_16_layer_scalar_to_fp16 = const()[name = string("tm_layers_16_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7p-2])];
            tensor<fp16, [1, 256, 512]> x_307_cast_fp16 = mul(x = h_33_cast_fp16, y = tm_layers_16_layer_scalar_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7789_cast_fp16 = abs(x = x_307_cast_fp16)[name = string("op_7789_cast_fp16")];
            bool var_7792 = const()[name = string("op_7792"), val = bool(true)];
            tensor<int32, [1]> var_7793_axes_0 = const()[name = string("op_7793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7793_cast_fp16 = reduce_max(axes = var_7793_axes_0, keep_dims = var_7792, x = var_7789_cast_fp16)[name = string("op_7793_cast_fp16")];
            fp16 var_7794_to_fp16 = const()[name = string("op_7794_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_275_cast_fp16 = maximum(x = var_7793_cast_fp16, y = var_7794_to_fp16)[name = string("s_275_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_275_cast_fp16 = real_div(x = x_307_cast_fp16, y = s_275_cast_fp16)[name = string("y_275_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7797_cast_fp16 = mul(x = y_275_cast_fp16, y = y_275_cast_fp16)[name = string("op_7797_cast_fp16")];
            tensor<int32, [1]> var_7802_axes_0 = const()[name = string("op_7802_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7802_keep_dims_0 = const()[name = string("op_7802_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7802_cast_fp16 = reduce_mean(axes = var_7802_axes_0, keep_dims = var_7802_keep_dims_0, x = var_7797_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7803_cast_fp16 = mul(x = s_275_cast_fp16, y = s_275_cast_fp16)[name = string("op_7803_cast_fp16")];
            fp32 var_7804_epsilon_0 = const()[name = string("op_7804_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7804_cast_fp16 = inverse(epsilon = var_7804_epsilon_0, x = var_7803_cast_fp16)[name = string("op_7804_cast_fp16")];
            fp16 var_7805_to_fp16 = const()[name = string("op_7805_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7806_cast_fp16 = mul(x = var_7804_cast_fp16, y = var_7805_to_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7808_cast_fp16 = add(x = var_7802_cast_fp16, y = var_7806_cast_fp16)[name = string("op_7808_cast_fp16")];
            fp32 var_7809_epsilon_0 = const()[name = string("op_7809_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7809_cast_fp16 = rsqrt(epsilon = var_7809_epsilon_0, x = var_7808_cast_fp16)[name = string("op_7809_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_241_cast_fp16 = mul(x = y_275_cast_fp16, y = var_7809_cast_fp16)[name = string("out_241_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97988224)))];
            tensor<fp16, [1, 256, 512]> input_205_cast_fp16 = mul(x = out_241_cast_fp16, y = tm_layers_17_input_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97989312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99037952))))[name = string("tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_154_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_7819 = const()[name = string("op_7819"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 256, 4, 512]> var_7820_cast_fp16 = reshape(shape = var_7819, x = linear_154_cast_fp16)[name = string("op_7820_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99042112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99304320))))[name = string("tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_7831 = const()[name = string("op_7831"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_7832_cast_fp16 = reshape(shape = var_7831, x = linear_155_cast_fp16)[name = string("op_7832_cast_fp16")];
            tensor<int32, [4]> x_311_perm_0 = const()[name = string("x_311_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99305408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99567616))))[name = string("tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_7843 = const()[name = string("op_7843"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_7844_cast_fp16 = reshape(shape = var_7843, x = linear_156_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<int32, [4]> x_313_perm_0 = const()[name = string("x_313_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 512]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_7820_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 4, 256, 512]> var_7848_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_7848_cast_fp16")];
            bool var_7851 = const()[name = string("op_7851"), val = bool(true)];
            tensor<int32, [1]> var_7852_axes_0 = const()[name = string("op_7852_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_7852_cast_fp16 = reduce_max(axes = var_7852_axes_0, keep_dims = var_7851, x = var_7848_cast_fp16)[name = string("op_7852_cast_fp16")];
            fp16 var_7853_to_fp16 = const()[name = string("op_7853_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_277_cast_fp16 = maximum(x = var_7852_cast_fp16, y = var_7853_to_fp16)[name = string("s_277_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> y_277_cast_fp16 = real_div(x = x_309_cast_fp16, y = s_277_cast_fp16)[name = string("y_277_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_7856_cast_fp16 = mul(x = y_277_cast_fp16, y = y_277_cast_fp16)[name = string("op_7856_cast_fp16")];
            tensor<int32, [1]> var_7861_axes_0 = const()[name = string("op_7861_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7861_keep_dims_0 = const()[name = string("op_7861_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_7861_cast_fp16 = reduce_mean(axes = var_7861_axes_0, keep_dims = var_7861_keep_dims_0, x = var_7856_cast_fp16)[name = string("op_7861_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7862_cast_fp16 = mul(x = s_277_cast_fp16, y = s_277_cast_fp16)[name = string("op_7862_cast_fp16")];
            fp32 var_7863_epsilon_0 = const()[name = string("op_7863_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_7863_cast_fp16 = inverse(epsilon = var_7863_epsilon_0, x = var_7862_cast_fp16)[name = string("op_7863_cast_fp16")];
            fp16 var_7864_to_fp16 = const()[name = string("op_7864_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_7865_cast_fp16 = mul(x = var_7863_cast_fp16, y = var_7864_to_fp16)[name = string("op_7865_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7867_cast_fp16 = add(x = var_7861_cast_fp16, y = var_7865_cast_fp16)[name = string("op_7867_cast_fp16")];
            fp32 var_7868_epsilon_0 = const()[name = string("op_7868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_7868_cast_fp16 = rsqrt(epsilon = var_7868_epsilon_0, x = var_7867_cast_fp16)[name = string("op_7868_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> out_243_cast_fp16 = mul(x = y_277_cast_fp16, y = var_7868_cast_fp16)[name = string("out_243_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99568704)))];
            tensor<fp16, [1, 4, 256, 512]> q_69_cast_fp16 = mul(x = out_243_cast_fp16, y = tm_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("q_69_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_311_cast_fp16 = transpose(perm = x_311_perm_0, x = var_7832_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 1, 256, 512]> var_7871_cast_fp16 = abs(x = x_311_cast_fp16)[name = string("op_7871_cast_fp16")];
            bool var_7874 = const()[name = string("op_7874"), val = bool(true)];
            tensor<int32, [1]> var_7875_axes_0 = const()[name = string("op_7875_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_7875_cast_fp16 = reduce_max(axes = var_7875_axes_0, keep_dims = var_7874, x = var_7871_cast_fp16)[name = string("op_7875_cast_fp16")];
            fp16 var_7876_to_fp16 = const()[name = string("op_7876_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_279_cast_fp16 = maximum(x = var_7875_cast_fp16, y = var_7876_to_fp16)[name = string("s_279_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_279_cast_fp16 = real_div(x = x_311_cast_fp16, y = s_279_cast_fp16)[name = string("y_279_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_7879_cast_fp16 = mul(x = y_279_cast_fp16, y = y_279_cast_fp16)[name = string("op_7879_cast_fp16")];
            tensor<int32, [1]> var_7884_axes_0 = const()[name = string("op_7884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7884_keep_dims_0 = const()[name = string("op_7884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_7884_cast_fp16 = reduce_mean(axes = var_7884_axes_0, keep_dims = var_7884_keep_dims_0, x = var_7879_cast_fp16)[name = string("op_7884_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_7885_cast_fp16 = mul(x = s_279_cast_fp16, y = s_279_cast_fp16)[name = string("op_7885_cast_fp16")];
            fp32 var_7886_epsilon_0 = const()[name = string("op_7886_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_7886_cast_fp16 = inverse(epsilon = var_7886_epsilon_0, x = var_7885_cast_fp16)[name = string("op_7886_cast_fp16")];
            fp16 var_7887_to_fp16 = const()[name = string("op_7887_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_7888_cast_fp16 = mul(x = var_7886_cast_fp16, y = var_7887_to_fp16)[name = string("op_7888_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_7890_cast_fp16 = add(x = var_7884_cast_fp16, y = var_7888_cast_fp16)[name = string("op_7890_cast_fp16")];
            fp32 var_7891_epsilon_0 = const()[name = string("op_7891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_7891_cast_fp16 = rsqrt(epsilon = var_7891_epsilon_0, x = var_7890_cast_fp16)[name = string("op_7891_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> out_245_cast_fp16 = mul(x = y_279_cast_fp16, y = var_7891_cast_fp16)[name = string("out_245_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99569792)))];
            tensor<fp16, [1, 1, 256, 512]> k_103_cast_fp16 = mul(x = out_245_cast_fp16, y = tm_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("k_103_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_313_cast_fp16 = transpose(perm = x_313_perm_0, x = var_7844_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1, 256, 512]> var_7894_cast_fp16 = abs(x = x_313_cast_fp16)[name = string("op_7894_cast_fp16")];
            bool var_7897 = const()[name = string("op_7897"), val = bool(true)];
            tensor<int32, [1]> var_7898_axes_0 = const()[name = string("op_7898_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_7898_cast_fp16 = reduce_max(axes = var_7898_axes_0, keep_dims = var_7897, x = var_7894_cast_fp16)[name = string("op_7898_cast_fp16")];
            fp16 var_7899_to_fp16 = const()[name = string("op_7899_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_281_cast_fp16 = maximum(x = var_7898_cast_fp16, y = var_7899_to_fp16)[name = string("s_281_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_281_cast_fp16 = real_div(x = x_313_cast_fp16, y = s_281_cast_fp16)[name = string("y_281_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_7902_cast_fp16 = mul(x = y_281_cast_fp16, y = y_281_cast_fp16)[name = string("op_7902_cast_fp16")];
            tensor<int32, [1]> var_7907_axes_0 = const()[name = string("op_7907_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7907_keep_dims_0 = const()[name = string("op_7907_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_7907_cast_fp16 = reduce_mean(axes = var_7907_axes_0, keep_dims = var_7907_keep_dims_0, x = var_7902_cast_fp16)[name = string("op_7907_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_7908_cast_fp16 = mul(x = s_281_cast_fp16, y = s_281_cast_fp16)[name = string("op_7908_cast_fp16")];
            fp32 var_7909_epsilon_0 = const()[name = string("op_7909_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_7909_cast_fp16 = inverse(epsilon = var_7909_epsilon_0, x = var_7908_cast_fp16)[name = string("op_7909_cast_fp16")];
            fp16 var_7910_to_fp16 = const()[name = string("op_7910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_7911_cast_fp16 = mul(x = var_7909_cast_fp16, y = var_7910_to_fp16)[name = string("op_7911_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_7913_cast_fp16 = add(x = var_7907_cast_fp16, y = var_7911_cast_fp16)[name = string("op_7913_cast_fp16")];
            fp32 var_7914_epsilon_0 = const()[name = string("op_7914_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_7914_cast_fp16 = rsqrt(epsilon = var_7914_epsilon_0, x = var_7913_cast_fp16)[name = string("op_7914_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> v_69_cast_fp16 = mul(x = y_281_cast_fp16, y = var_7914_cast_fp16)[name = string("v_69_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_7916_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7916_cast_fp16")];
            tensor<int32, [4]> var_7921_begin_0 = const()[name = string("op_7921_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7921_end_0 = const()[name = string("op_7921_end_0"), val = tensor<int32, [4]>([1, 4, 256, 512])];
            tensor<bool, [4]> var_7921_end_mask_0 = const()[name = string("op_7921_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 256]> var_7921_cast_fp16 = slice_by_index(begin = var_7921_begin_0, end = var_7921_end_0, end_mask = var_7921_end_mask_0, x = q_69_cast_fp16)[name = string("op_7921_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 256]> var_7922_cast_fp16 = mul(x = var_7921_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 256]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = q_69_cast_fp16)[name = string("op_7927_cast_fp16")];
            int32 var_7929 = const()[name = string("op_7929"), val = int32(-1)];
            bool var_7930_interleave_0 = const()[name = string("op_7930_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> var_7930_cast_fp16 = concat(axis = var_7929, interleave = var_7930_interleave_0, values = (var_7922_cast_fp16, var_7927_cast_fp16))[name = string("op_7930_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_7931_cast_fp16 = mul(x = var_7930_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7931_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> q_71_cast_fp16 = add(x = var_7916_cast_fp16, y = var_7931_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_7934_cast_fp16 = mul(x = k_103_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7934_cast_fp16")];
            tensor<int32, [4]> var_7939_begin_0 = const()[name = string("op_7939_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7939_end_0 = const()[name = string("op_7939_end_0"), val = tensor<int32, [4]>([1, 1, 256, 512])];
            tensor<bool, [4]> var_7939_end_mask_0 = const()[name = string("op_7939_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 256, 256]> var_7939_cast_fp16 = slice_by_index(begin = var_7939_begin_0, end = var_7939_end_0, end_mask = var_7939_end_mask_0, x = k_103_cast_fp16)[name = string("op_7939_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 256, 256]> var_7940_cast_fp16 = mul(x = var_7939_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_7940_cast_fp16")];
            tensor<int32, [4]> var_7945_begin_0 = const()[name = string("op_7945_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7945_end_0 = const()[name = string("op_7945_end_0"), val = tensor<int32, [4]>([1, 1, 256, 256])];
            tensor<bool, [4]> var_7945_end_mask_0 = const()[name = string("op_7945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 256, 256]> var_7945_cast_fp16 = slice_by_index(begin = var_7945_begin_0, end = var_7945_end_0, end_mask = var_7945_end_mask_0, x = k_103_cast_fp16)[name = string("op_7945_cast_fp16")];
            int32 var_7947 = const()[name = string("op_7947"), val = int32(-1)];
            bool var_7948_interleave_0 = const()[name = string("op_7948_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 256, 512]> var_7948_cast_fp16 = concat(axis = var_7947, interleave = var_7948_interleave_0, values = (var_7940_cast_fp16, var_7945_cast_fp16))[name = string("op_7948_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_7949_cast_fp16 = mul(x = var_7948_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> k_105_cast_fp16 = add(x = var_7934_cast_fp16, y = var_7949_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_34_reps_0 = const()[name = string("tile_34_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = k_105_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 1, 256, 512]> tile_34_cast_fp16 = tile(reps = tile_34_reps_0, x = transpose_68_cast_fp16)[name = string("tile_34_cast_fp16")];
            tensor<int32, [5]> concat_68 = const()[name = string("concat_68"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_68_cast_fp16 = reshape(shape = concat_68, x = tile_34_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [5]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 1, 256, 512]> reshape_69_cast_fp16 = reshape(shape = concat_69, x = transpose_69_cast_fp16)[name = string("reshape_69_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_35_reps_0 = const()[name = string("tile_35_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = v_69_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 1, 256, 512]> tile_35_cast_fp16 = tile(reps = tile_35_reps_0, x = transpose_70_cast_fp16)[name = string("tile_35_cast_fp16")];
            tensor<int32, [5]> concat_70 = const()[name = string("concat_70"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_70_cast_fp16 = reshape(shape = concat_70, x = tile_35_cast_fp16)[name = string("reshape_70_cast_fp16")];
            tensor<int32, [5]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = reshape_70_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 1, 256, 512]> reshape_71_cast_fp16 = reshape(shape = concat_71, x = transpose_71_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7963_transpose_x_0 = const()[name = string("op_7963_transpose_x_0"), val = bool(false)];
            bool var_7963_transpose_y_0 = const()[name = string("op_7963_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = reshape_69_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 4, 256, 256]> var_7963_cast_fp16 = matmul(transpose_x = var_7963_transpose_x_0, transpose_y = var_7963_transpose_y_0, x = q_71_cast_fp16, y = transpose_113_cast_fp16)[name = string("op_7963_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7965_cast_fp16 = add(x = var_7963_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7965_cast_fp16")];
            int32 var_7966 = const()[name = string("op_7966"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_281_cast_fp16 = softmax(axis = var_7966, x = var_7965_cast_fp16)[name = string("w_281_cast_fp16")];
            bool var_7969_transpose_x_0 = const()[name = string("op_7969_transpose_x_0"), val = bool(false)];
            bool var_7969_transpose_y_0 = const()[name = string("op_7969_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> v_71_cast_fp16 = transpose(perm = v_71_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 4, 256, 512]> var_7969_cast_fp16 = matmul(transpose_x = var_7969_transpose_x_0, transpose_y = var_7969_transpose_y_0, x = w_281_cast_fp16, y = v_71_cast_fp16)[name = string("op_7969_cast_fp16")];
            tensor<int32, [4]> var_7972_perm_0 = const()[name = string("op_7972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7976 = const()[name = string("op_7976"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 256, 4, 512]> var_7972_cast_fp16 = transpose(perm = var_7972_perm_0, x = var_7969_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 256, 2048]> input_207_cast_fp16 = reshape(shape = var_7976, x = var_7972_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99570880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619520))))[name = string("tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7981_cast_fp16 = abs(x = linear_157_cast_fp16)[name = string("op_7981_cast_fp16")];
            bool var_7984 = const()[name = string("op_7984"), val = bool(true)];
            tensor<int32, [1]> var_7985_axes_0 = const()[name = string("op_7985_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7985_cast_fp16 = reduce_max(axes = var_7985_axes_0, keep_dims = var_7984, x = var_7981_cast_fp16)[name = string("op_7985_cast_fp16")];
            fp16 var_7986_to_fp16 = const()[name = string("op_7986_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_283_cast_fp16 = maximum(x = var_7985_cast_fp16, y = var_7986_to_fp16)[name = string("s_283_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_283_cast_fp16 = real_div(x = linear_157_cast_fp16, y = s_283_cast_fp16)[name = string("y_283_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7989_cast_fp16 = mul(x = y_283_cast_fp16, y = y_283_cast_fp16)[name = string("op_7989_cast_fp16")];
            tensor<int32, [1]> var_7994_axes_0 = const()[name = string("op_7994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7994_keep_dims_0 = const()[name = string("op_7994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7994_cast_fp16 = reduce_mean(axes = var_7994_axes_0, keep_dims = var_7994_keep_dims_0, x = var_7989_cast_fp16)[name = string("op_7994_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7995_cast_fp16 = mul(x = s_283_cast_fp16, y = s_283_cast_fp16)[name = string("op_7995_cast_fp16")];
            fp32 var_7996_epsilon_0 = const()[name = string("op_7996_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7996_cast_fp16 = inverse(epsilon = var_7996_epsilon_0, x = var_7995_cast_fp16)[name = string("op_7996_cast_fp16")];
            fp16 var_7997_to_fp16 = const()[name = string("op_7997_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7998_cast_fp16 = mul(x = var_7996_cast_fp16, y = var_7997_to_fp16)[name = string("op_7998_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8000_cast_fp16 = add(x = var_7994_cast_fp16, y = var_7998_cast_fp16)[name = string("op_8000_cast_fp16")];
            fp32 var_8001_epsilon_0 = const()[name = string("op_8001_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8001_cast_fp16 = rsqrt(epsilon = var_8001_epsilon_0, x = var_8000_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_247_cast_fp16 = mul(x = y_283_cast_fp16, y = var_8001_cast_fp16)[name = string("out_247_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100620608)))];
            tensor<fp16, [1, 256, 512]> var_8003_cast_fp16 = mul(x = out_247_cast_fp16, y = tm_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("op_8003_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_317_cast_fp16 = add(x = x_307_cast_fp16, y = var_8003_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8006_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_8006_cast_fp16")];
            bool var_8009 = const()[name = string("op_8009"), val = bool(true)];
            tensor<int32, [1]> var_8010_axes_0 = const()[name = string("op_8010_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8010_cast_fp16 = reduce_max(axes = var_8010_axes_0, keep_dims = var_8009, x = var_8006_cast_fp16)[name = string("op_8010_cast_fp16")];
            fp16 var_8011_to_fp16 = const()[name = string("op_8011_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_285_cast_fp16 = maximum(x = var_8010_cast_fp16, y = var_8011_to_fp16)[name = string("s_285_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_285_cast_fp16 = real_div(x = x_317_cast_fp16, y = s_285_cast_fp16)[name = string("y_285_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8014_cast_fp16 = mul(x = y_285_cast_fp16, y = y_285_cast_fp16)[name = string("op_8014_cast_fp16")];
            tensor<int32, [1]> var_8019_axes_0 = const()[name = string("op_8019_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8019_keep_dims_0 = const()[name = string("op_8019_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8019_cast_fp16 = reduce_mean(axes = var_8019_axes_0, keep_dims = var_8019_keep_dims_0, x = var_8014_cast_fp16)[name = string("op_8019_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8020_cast_fp16 = mul(x = s_285_cast_fp16, y = s_285_cast_fp16)[name = string("op_8020_cast_fp16")];
            fp32 var_8021_epsilon_0 = const()[name = string("op_8021_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8021_cast_fp16 = inverse(epsilon = var_8021_epsilon_0, x = var_8020_cast_fp16)[name = string("op_8021_cast_fp16")];
            fp16 var_8022_to_fp16 = const()[name = string("op_8022_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8023_cast_fp16 = mul(x = var_8021_cast_fp16, y = var_8022_to_fp16)[name = string("op_8023_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8025_cast_fp16 = add(x = var_8019_cast_fp16, y = var_8023_cast_fp16)[name = string("op_8025_cast_fp16")];
            fp32 var_8026_epsilon_0 = const()[name = string("op_8026_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8026_cast_fp16 = rsqrt(epsilon = var_8026_epsilon_0, x = var_8025_cast_fp16)[name = string("op_8026_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_249_cast_fp16 = mul(x = y_285_cast_fp16, y = var_8026_cast_fp16)[name = string("out_249_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621696)))];
            tensor<fp16, [1, 256, 512]> input_209_cast_fp16 = mul(x = out_249_cast_fp16, y = tm_layers_17_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100622784))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101671424))))[name = string("tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_158_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_158_cast_fp16")];
            string var_8033_mode_0 = const()[name = string("op_8033_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_8033_cast_fp16 = gelu(mode = var_8033_mode_0, x = linear_158_cast_fp16)[name = string("op_8033_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101675584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102724224))))[name = string("tm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_159_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_211_cast_fp16 = mul(x = var_8033_cast_fp16, y = linear_159_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102728384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103777024))))[name = string("tm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8041_cast_fp16 = abs(x = linear_160_cast_fp16)[name = string("op_8041_cast_fp16")];
            bool var_8044 = const()[name = string("op_8044"), val = bool(true)];
            tensor<int32, [1]> var_8045_axes_0 = const()[name = string("op_8045_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8045_cast_fp16 = reduce_max(axes = var_8045_axes_0, keep_dims = var_8044, x = var_8041_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp16 var_8046_to_fp16 = const()[name = string("op_8046_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_287_cast_fp16 = maximum(x = var_8045_cast_fp16, y = var_8046_to_fp16)[name = string("s_287_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_287_cast_fp16 = real_div(x = linear_160_cast_fp16, y = s_287_cast_fp16)[name = string("y_287_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8049_cast_fp16 = mul(x = y_287_cast_fp16, y = y_287_cast_fp16)[name = string("op_8049_cast_fp16")];
            tensor<int32, [1]> var_8054_axes_0 = const()[name = string("op_8054_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8054_keep_dims_0 = const()[name = string("op_8054_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8054_cast_fp16 = reduce_mean(axes = var_8054_axes_0, keep_dims = var_8054_keep_dims_0, x = var_8049_cast_fp16)[name = string("op_8054_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8055_cast_fp16 = mul(x = s_287_cast_fp16, y = s_287_cast_fp16)[name = string("op_8055_cast_fp16")];
            fp32 var_8056_epsilon_0 = const()[name = string("op_8056_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8056_cast_fp16 = inverse(epsilon = var_8056_epsilon_0, x = var_8055_cast_fp16)[name = string("op_8056_cast_fp16")];
            fp16 var_8057_to_fp16 = const()[name = string("op_8057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8058_cast_fp16 = mul(x = var_8056_cast_fp16, y = var_8057_to_fp16)[name = string("op_8058_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8060_cast_fp16 = add(x = var_8054_cast_fp16, y = var_8058_cast_fp16)[name = string("op_8060_cast_fp16")];
            fp32 var_8061_epsilon_0 = const()[name = string("op_8061_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8061_cast_fp16 = rsqrt(epsilon = var_8061_epsilon_0, x = var_8060_cast_fp16)[name = string("op_8061_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_251_cast_fp16 = mul(x = y_287_cast_fp16, y = var_8061_cast_fp16)[name = string("out_251_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103778112)))];
            tensor<fp16, [1, 256, 512]> var_8063_cast_fp16 = mul(x = out_251_cast_fp16, y = tm_layers_17_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_213_cast_fp16 = add(x = x_317_cast_fp16, y = var_8063_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103779200))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041408))))[name = string("tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_161_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_161_cast_fp16")];
            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_321_cast_fp16 = gelu(mode = x_321_mode_0, x = linear_161_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [4]> var_8083_begin_0 = const()[name = string("op_8083_begin_0"), val = tensor<int32, [4]>([0, 0, 17, 0])];
            tensor<int32, [4]> var_8083_end_0 = const()[name = string("op_8083_end_0"), val = tensor<int32, [4]>([1, 256, 18, 512])];
            tensor<bool, [4]> var_8083_end_mask_0 = const()[name = string("op_8083_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8083_squeeze_mask_0 = const()[name = string("op_8083_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_8083_cast_fp16 = slice_by_index(begin = var_8083_begin_0, end = var_8083_end_0, end_mask = var_8083_end_mask_0, squeeze_mask = var_8083_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8083_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_215_cast_fp16 = mul(x = x_321_cast_fp16, y = var_8083_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104042496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104304704))))[name = string("tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8093_cast_fp16 = abs(x = linear_162_cast_fp16)[name = string("op_8093_cast_fp16")];
            bool var_8096 = const()[name = string("op_8096"), val = bool(true)];
            tensor<int32, [1]> var_8097_axes_0 = const()[name = string("op_8097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8097_cast_fp16 = reduce_max(axes = var_8097_axes_0, keep_dims = var_8096, x = var_8093_cast_fp16)[name = string("op_8097_cast_fp16")];
            fp16 var_8098_to_fp16 = const()[name = string("op_8098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_289_cast_fp16 = maximum(x = var_8097_cast_fp16, y = var_8098_to_fp16)[name = string("s_289_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_289_cast_fp16 = real_div(x = linear_162_cast_fp16, y = s_289_cast_fp16)[name = string("y_289_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8101_cast_fp16 = mul(x = y_289_cast_fp16, y = y_289_cast_fp16)[name = string("op_8101_cast_fp16")];
            tensor<int32, [1]> var_8106_axes_0 = const()[name = string("op_8106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8106_keep_dims_0 = const()[name = string("op_8106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8106_cast_fp16 = reduce_mean(axes = var_8106_axes_0, keep_dims = var_8106_keep_dims_0, x = var_8101_cast_fp16)[name = string("op_8106_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8107_cast_fp16 = mul(x = s_289_cast_fp16, y = s_289_cast_fp16)[name = string("op_8107_cast_fp16")];
            fp32 var_8108_epsilon_0 = const()[name = string("op_8108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8108_cast_fp16 = inverse(epsilon = var_8108_epsilon_0, x = var_8107_cast_fp16)[name = string("op_8108_cast_fp16")];
            fp16 var_8109_to_fp16 = const()[name = string("op_8109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8110_cast_fp16 = mul(x = var_8108_cast_fp16, y = var_8109_to_fp16)[name = string("op_8110_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8112_cast_fp16 = add(x = var_8106_cast_fp16, y = var_8110_cast_fp16)[name = string("op_8112_cast_fp16")];
            fp32 var_8113_epsilon_0 = const()[name = string("op_8113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8113_cast_fp16 = rsqrt(epsilon = var_8113_epsilon_0, x = var_8112_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_253_cast_fp16 = mul(x = y_289_cast_fp16, y = var_8113_cast_fp16)[name = string("out_253_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104305792)))];
            tensor<fp16, [1, 256, 512]> var_8115_cast_fp16 = mul(x = out_253_cast_fp16, y = tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8115_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_35_cast_fp16 = add(x = input_213_cast_fp16, y = var_8115_cast_fp16)[name = string("h_35_cast_fp16")];
            tensor<fp16, [1]> tm_layers_17_layer_scalar_to_fp16 = const()[name = string("tm_layers_17_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e6p-1])];
            tensor<fp16, [1, 256, 512]> x_325_cast_fp16 = mul(x = h_35_cast_fp16, y = tm_layers_17_layer_scalar_to_fp16)[name = string("x_325_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8119_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_8119_cast_fp16")];
            bool var_8122 = const()[name = string("op_8122"), val = bool(true)];
            tensor<int32, [1]> var_8123_axes_0 = const()[name = string("op_8123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8123_cast_fp16 = reduce_max(axes = var_8123_axes_0, keep_dims = var_8122, x = var_8119_cast_fp16)[name = string("op_8123_cast_fp16")];
            fp16 var_8124_to_fp16 = const()[name = string("op_8124_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_291_cast_fp16 = maximum(x = var_8123_cast_fp16, y = var_8124_to_fp16)[name = string("s_291_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_291_cast_fp16 = real_div(x = x_325_cast_fp16, y = s_291_cast_fp16)[name = string("y_291_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8127_cast_fp16 = mul(x = y_291_cast_fp16, y = y_291_cast_fp16)[name = string("op_8127_cast_fp16")];
            tensor<int32, [1]> var_8132_axes_0 = const()[name = string("op_8132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8132_keep_dims_0 = const()[name = string("op_8132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8132_cast_fp16 = reduce_mean(axes = var_8132_axes_0, keep_dims = var_8132_keep_dims_0, x = var_8127_cast_fp16)[name = string("op_8132_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8133_cast_fp16 = mul(x = s_291_cast_fp16, y = s_291_cast_fp16)[name = string("op_8133_cast_fp16")];
            fp32 var_8134_epsilon_0 = const()[name = string("op_8134_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8134_cast_fp16 = inverse(epsilon = var_8134_epsilon_0, x = var_8133_cast_fp16)[name = string("op_8134_cast_fp16")];
            fp16 var_8135_to_fp16 = const()[name = string("op_8135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = string("op_8136_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8138_cast_fp16 = add(x = var_8132_cast_fp16, y = var_8136_cast_fp16)[name = string("op_8138_cast_fp16")];
            fp32 var_8139_epsilon_0 = const()[name = string("op_8139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8139_cast_fp16 = rsqrt(epsilon = var_8139_epsilon_0, x = var_8138_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_255_cast_fp16 = mul(x = y_291_cast_fp16, y = var_8139_cast_fp16)[name = string("out_255_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104306880)))];
            tensor<fp16, [1, 256, 512]> input_217_cast_fp16 = mul(x = out_255_cast_fp16, y = tm_layers_18_input_layernorm_weight_to_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104307968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104832320))))[name = string("tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8149 = const()[name = string("op_8149"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_8150_cast_fp16 = reshape(shape = var_8149, x = linear_163_cast_fp16)[name = string("op_8150_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104834432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105096640))))[name = string("tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [4]> var_8161 = const()[name = string("op_8161"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8162_cast_fp16 = reshape(shape = var_8161, x = linear_164_cast_fp16)[name = string("op_8162_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105097728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105359936))))[name = string("tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<int32, [4]> var_8173 = const()[name = string("op_8173"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8174_cast_fp16 = reshape(shape = var_8173, x = linear_165_cast_fp16)[name = string("op_8174_cast_fp16")];
            tensor<int32, [4]> x_331_perm_0 = const()[name = string("x_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_8150_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 4, 256, 256]> var_8178_cast_fp16 = abs(x = x_327_cast_fp16)[name = string("op_8178_cast_fp16")];
            bool var_8181 = const()[name = string("op_8181"), val = bool(true)];
            tensor<int32, [1]> var_8182_axes_0 = const()[name = string("op_8182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_8182_cast_fp16 = reduce_max(axes = var_8182_axes_0, keep_dims = var_8181, x = var_8178_cast_fp16)[name = string("op_8182_cast_fp16")];
            fp16 var_8183_to_fp16 = const()[name = string("op_8183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_293_cast_fp16 = maximum(x = var_8182_cast_fp16, y = var_8183_to_fp16)[name = string("s_293_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_293_cast_fp16 = real_div(x = x_327_cast_fp16, y = s_293_cast_fp16)[name = string("y_293_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8186_cast_fp16 = mul(x = y_293_cast_fp16, y = y_293_cast_fp16)[name = string("op_8186_cast_fp16")];
            tensor<int32, [1]> var_8191_axes_0 = const()[name = string("op_8191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8191_keep_dims_0 = const()[name = string("op_8191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_8191_cast_fp16 = reduce_mean(axes = var_8191_axes_0, keep_dims = var_8191_keep_dims_0, x = var_8186_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8192_cast_fp16 = mul(x = s_293_cast_fp16, y = s_293_cast_fp16)[name = string("op_8192_cast_fp16")];
            fp32 var_8193_epsilon_0 = const()[name = string("op_8193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_8193_cast_fp16 = inverse(epsilon = var_8193_epsilon_0, x = var_8192_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8197_cast_fp16 = add(x = var_8191_cast_fp16, y = var_8195_cast_fp16)[name = string("op_8197_cast_fp16")];
            fp32 var_8198_epsilon_0 = const()[name = string("op_8198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_8198_cast_fp16 = rsqrt(epsilon = var_8198_epsilon_0, x = var_8197_cast_fp16)[name = string("op_8198_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_257_cast_fp16 = mul(x = y_293_cast_fp16, y = var_8198_cast_fp16)[name = string("out_257_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_73_cast_fp16 = mul(x = out_257_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_73_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_8162_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 256, 256]> var_8201_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_8201_cast_fp16")];
            bool var_8204 = const()[name = string("op_8204"), val = bool(true)];
            tensor<int32, [1]> var_8205_axes_0 = const()[name = string("op_8205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8205_cast_fp16 = reduce_max(axes = var_8205_axes_0, keep_dims = var_8204, x = var_8201_cast_fp16)[name = string("op_8205_cast_fp16")];
            fp16 var_8206_to_fp16 = const()[name = string("op_8206_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_295_cast_fp16 = maximum(x = var_8205_cast_fp16, y = var_8206_to_fp16)[name = string("s_295_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_295_cast_fp16 = real_div(x = x_329_cast_fp16, y = s_295_cast_fp16)[name = string("y_295_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8209_cast_fp16 = mul(x = y_295_cast_fp16, y = y_295_cast_fp16)[name = string("op_8209_cast_fp16")];
            tensor<int32, [1]> var_8214_axes_0 = const()[name = string("op_8214_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8214_keep_dims_0 = const()[name = string("op_8214_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8214_cast_fp16 = reduce_mean(axes = var_8214_axes_0, keep_dims = var_8214_keep_dims_0, x = var_8209_cast_fp16)[name = string("op_8214_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8215_cast_fp16 = mul(x = s_295_cast_fp16, y = s_295_cast_fp16)[name = string("op_8215_cast_fp16")];
            fp32 var_8216_epsilon_0 = const()[name = string("op_8216_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8216_cast_fp16 = inverse(epsilon = var_8216_epsilon_0, x = var_8215_cast_fp16)[name = string("op_8216_cast_fp16")];
            fp16 var_8217_to_fp16 = const()[name = string("op_8217_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8218_cast_fp16 = mul(x = var_8216_cast_fp16, y = var_8217_to_fp16)[name = string("op_8218_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8220_cast_fp16 = add(x = var_8214_cast_fp16, y = var_8218_cast_fp16)[name = string("op_8220_cast_fp16")];
            fp32 var_8221_epsilon_0 = const()[name = string("op_8221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8221_cast_fp16 = rsqrt(epsilon = var_8221_epsilon_0, x = var_8220_cast_fp16)[name = string("op_8221_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_259_cast_fp16 = mul(x = y_295_cast_fp16, y = var_8221_cast_fp16)[name = string("out_259_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_109_cast_fp16 = mul(x = out_259_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_109_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_331_cast_fp16 = transpose(perm = x_331_perm_0, x = var_8174_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 256, 256]> var_8224_cast_fp16 = abs(x = x_331_cast_fp16)[name = string("op_8224_cast_fp16")];
            bool var_8227 = const()[name = string("op_8227"), val = bool(true)];
            tensor<int32, [1]> var_8228_axes_0 = const()[name = string("op_8228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8228_cast_fp16 = reduce_max(axes = var_8228_axes_0, keep_dims = var_8227, x = var_8224_cast_fp16)[name = string("op_8228_cast_fp16")];
            fp16 var_8229_to_fp16 = const()[name = string("op_8229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_297_cast_fp16 = maximum(x = var_8228_cast_fp16, y = var_8229_to_fp16)[name = string("s_297_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_297_cast_fp16 = real_div(x = x_331_cast_fp16, y = s_297_cast_fp16)[name = string("y_297_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8232_cast_fp16 = mul(x = y_297_cast_fp16, y = y_297_cast_fp16)[name = string("op_8232_cast_fp16")];
            tensor<int32, [1]> var_8237_axes_0 = const()[name = string("op_8237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8237_keep_dims_0 = const()[name = string("op_8237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8237_cast_fp16 = reduce_mean(axes = var_8237_axes_0, keep_dims = var_8237_keep_dims_0, x = var_8232_cast_fp16)[name = string("op_8237_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8238_cast_fp16 = mul(x = s_297_cast_fp16, y = s_297_cast_fp16)[name = string("op_8238_cast_fp16")];
            fp32 var_8239_epsilon_0 = const()[name = string("op_8239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8239_cast_fp16 = inverse(epsilon = var_8239_epsilon_0, x = var_8238_cast_fp16)[name = string("op_8239_cast_fp16")];
            fp16 var_8240_to_fp16 = const()[name = string("op_8240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8241_cast_fp16 = mul(x = var_8239_cast_fp16, y = var_8240_to_fp16)[name = string("op_8241_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8243_cast_fp16 = add(x = var_8237_cast_fp16, y = var_8241_cast_fp16)[name = string("op_8243_cast_fp16")];
            fp32 var_8244_epsilon_0 = const()[name = string("op_8244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8244_cast_fp16 = rsqrt(epsilon = var_8244_epsilon_0, x = var_8243_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_73_cast_fp16 = mul(x = y_297_cast_fp16, y = var_8244_cast_fp16)[name = string("v_73_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8246_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8246_cast_fp16")];
            tensor<int32, [4]> var_8251_begin_0 = const()[name = string("op_8251_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8251_end_0 = const()[name = string("op_8251_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_8251_end_mask_0 = const()[name = string("op_8251_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_8251_cast_fp16 = slice_by_index(begin = var_8251_begin_0, end = var_8251_end_0, end_mask = var_8251_end_mask_0, x = q_73_cast_fp16)[name = string("op_8251_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_8252_cast_fp16 = mul(x = var_8251_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8257_begin_0 = const()[name = string("op_8257_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8257_end_0 = const()[name = string("op_8257_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_8257_end_mask_0 = const()[name = string("op_8257_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_8257_cast_fp16 = slice_by_index(begin = var_8257_begin_0, end = var_8257_end_0, end_mask = var_8257_end_mask_0, x = q_73_cast_fp16)[name = string("op_8257_cast_fp16")];
            int32 var_8259 = const()[name = string("op_8259"), val = int32(-1)];
            bool var_8260_interleave_0 = const()[name = string("op_8260_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_8260_cast_fp16 = concat(axis = var_8259, interleave = var_8260_interleave_0, values = (var_8252_cast_fp16, var_8257_cast_fp16))[name = string("op_8260_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8261_cast_fp16 = mul(x = var_8260_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8261_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_75_cast_fp16 = add(x = var_8246_cast_fp16, y = var_8261_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8264_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8264_cast_fp16")];
            tensor<int32, [4]> var_8269_begin_0 = const()[name = string("op_8269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8269_end_0 = const()[name = string("op_8269_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_8269_end_mask_0 = const()[name = string("op_8269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_8269_cast_fp16 = slice_by_index(begin = var_8269_begin_0, end = var_8269_end_0, end_mask = var_8269_end_mask_0, x = k_109_cast_fp16)[name = string("op_8269_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_8270_cast_fp16 = mul(x = var_8269_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_8270_cast_fp16")];
            tensor<int32, [4]> var_8275_begin_0 = const()[name = string("op_8275_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8275_end_0 = const()[name = string("op_8275_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_8275_end_mask_0 = const()[name = string("op_8275_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_8275_cast_fp16 = slice_by_index(begin = var_8275_begin_0, end = var_8275_end_0, end_mask = var_8275_end_mask_0, x = k_109_cast_fp16)[name = string("op_8275_cast_fp16")];
            int32 var_8277 = const()[name = string("op_8277"), val = int32(-1)];
            bool var_8278_interleave_0 = const()[name = string("op_8278_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_8278_cast_fp16 = concat(axis = var_8277, interleave = var_8278_interleave_0, values = (var_8270_cast_fp16, var_8275_cast_fp16))[name = string("op_8278_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8279_cast_fp16 = mul(x = var_8278_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8279_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_111_cast_fp16 = add(x = var_8264_cast_fp16, y = var_8279_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_36_reps_0 = const()[name = string("tile_36_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_72_cast_fp16 = transpose(perm = transpose_72_perm_0, x = k_111_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 1, 256, 256]> tile_36_cast_fp16 = tile(reps = tile_36_reps_0, x = transpose_72_cast_fp16)[name = string("tile_36_cast_fp16")];
            tensor<int32, [5]> concat_72 = const()[name = string("concat_72"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_72_cast_fp16 = reshape(shape = concat_72, x = tile_36_cast_fp16)[name = string("reshape_72_cast_fp16")];
            tensor<int32, [5]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_73 = const()[name = string("concat_73"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = reshape_72_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 1, 256, 256]> reshape_73_cast_fp16 = reshape(shape = concat_73, x = transpose_73_cast_fp16)[name = string("reshape_73_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_37_reps_0 = const()[name = string("tile_37_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = v_73_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 1, 256, 256]> tile_37_cast_fp16 = tile(reps = tile_37_reps_0, x = transpose_74_cast_fp16)[name = string("tile_37_cast_fp16")];
            tensor<int32, [5]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_74_cast_fp16 = reshape(shape = concat_74, x = tile_37_cast_fp16)[name = string("reshape_74_cast_fp16")];
            tensor<int32, [5]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = reshape_74_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 1, 256, 256]> reshape_75_cast_fp16 = reshape(shape = concat_75, x = transpose_75_cast_fp16)[name = string("reshape_75_cast_fp16")];
            tensor<int32, [4]> v_75_perm_0 = const()[name = string("v_75_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8293_transpose_x_0 = const()[name = string("op_8293_transpose_x_0"), val = bool(false)];
            bool var_8293_transpose_y_0 = const()[name = string("op_8293_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = reshape_73_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 4, 256, 256]> var_8293_cast_fp16 = matmul(transpose_x = var_8293_transpose_x_0, transpose_y = var_8293_transpose_y_0, x = q_75_cast_fp16, y = transpose_114_cast_fp16)[name = string("op_8293_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8295_cast_fp16 = add(x = var_8293_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8295_cast_fp16")];
            int32 var_8296 = const()[name = string("op_8296"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_297_cast_fp16 = softmax(axis = var_8296, x = var_8295_cast_fp16)[name = string("w_297_cast_fp16")];
            bool var_8299_transpose_x_0 = const()[name = string("op_8299_transpose_x_0"), val = bool(false)];
            bool var_8299_transpose_y_0 = const()[name = string("op_8299_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_75_cast_fp16 = transpose(perm = v_75_perm_0, x = reshape_75_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 4, 256, 256]> var_8299_cast_fp16 = matmul(transpose_x = var_8299_transpose_x_0, transpose_y = var_8299_transpose_y_0, x = w_297_cast_fp16, y = v_75_cast_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [4]> var_8302_perm_0 = const()[name = string("op_8302_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8306 = const()[name = string("op_8306"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_8302_cast_fp16 = transpose(perm = var_8302_perm_0, x = var_8299_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 256, 1024]> input_219_cast_fp16 = reshape(shape = var_8306, x = var_8302_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105361024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105885376))))[name = string("tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8311_cast_fp16 = abs(x = linear_166_cast_fp16)[name = string("op_8311_cast_fp16")];
            bool var_8314 = const()[name = string("op_8314"), val = bool(true)];
            tensor<int32, [1]> var_8315_axes_0 = const()[name = string("op_8315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8315_cast_fp16 = reduce_max(axes = var_8315_axes_0, keep_dims = var_8314, x = var_8311_cast_fp16)[name = string("op_8315_cast_fp16")];
            fp16 var_8316_to_fp16 = const()[name = string("op_8316_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_299_cast_fp16 = maximum(x = var_8315_cast_fp16, y = var_8316_to_fp16)[name = string("s_299_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_299_cast_fp16 = real_div(x = linear_166_cast_fp16, y = s_299_cast_fp16)[name = string("y_299_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8319_cast_fp16 = mul(x = y_299_cast_fp16, y = y_299_cast_fp16)[name = string("op_8319_cast_fp16")];
            tensor<int32, [1]> var_8324_axes_0 = const()[name = string("op_8324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8324_keep_dims_0 = const()[name = string("op_8324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8324_cast_fp16 = reduce_mean(axes = var_8324_axes_0, keep_dims = var_8324_keep_dims_0, x = var_8319_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8325_cast_fp16 = mul(x = s_299_cast_fp16, y = s_299_cast_fp16)[name = string("op_8325_cast_fp16")];
            fp32 var_8326_epsilon_0 = const()[name = string("op_8326_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8326_cast_fp16 = inverse(epsilon = var_8326_epsilon_0, x = var_8325_cast_fp16)[name = string("op_8326_cast_fp16")];
            fp16 var_8327_to_fp16 = const()[name = string("op_8327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8328_cast_fp16 = mul(x = var_8326_cast_fp16, y = var_8327_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8330_cast_fp16 = add(x = var_8324_cast_fp16, y = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_261_cast_fp16 = mul(x = y_299_cast_fp16, y = var_8331_cast_fp16)[name = string("out_261_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105886464)))];
            tensor<fp16, [1, 256, 512]> var_8333_cast_fp16 = mul(x = out_261_cast_fp16, y = tm_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("op_8333_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = var_8333_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8336_cast_fp16 = abs(x = x_335_cast_fp16)[name = string("op_8336_cast_fp16")];
            bool var_8339 = const()[name = string("op_8339"), val = bool(true)];
            tensor<int32, [1]> var_8340_axes_0 = const()[name = string("op_8340_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8340_cast_fp16 = reduce_max(axes = var_8340_axes_0, keep_dims = var_8339, x = var_8336_cast_fp16)[name = string("op_8340_cast_fp16")];
            fp16 var_8341_to_fp16 = const()[name = string("op_8341_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_301_cast_fp16 = maximum(x = var_8340_cast_fp16, y = var_8341_to_fp16)[name = string("s_301_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_301_cast_fp16 = real_div(x = x_335_cast_fp16, y = s_301_cast_fp16)[name = string("y_301_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8344_cast_fp16 = mul(x = y_301_cast_fp16, y = y_301_cast_fp16)[name = string("op_8344_cast_fp16")];
            tensor<int32, [1]> var_8349_axes_0 = const()[name = string("op_8349_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8349_keep_dims_0 = const()[name = string("op_8349_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8349_cast_fp16 = reduce_mean(axes = var_8349_axes_0, keep_dims = var_8349_keep_dims_0, x = var_8344_cast_fp16)[name = string("op_8349_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8350_cast_fp16 = mul(x = s_301_cast_fp16, y = s_301_cast_fp16)[name = string("op_8350_cast_fp16")];
            fp32 var_8351_epsilon_0 = const()[name = string("op_8351_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8351_cast_fp16 = inverse(epsilon = var_8351_epsilon_0, x = var_8350_cast_fp16)[name = string("op_8351_cast_fp16")];
            fp16 var_8352_to_fp16 = const()[name = string("op_8352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8353_cast_fp16 = mul(x = var_8351_cast_fp16, y = var_8352_to_fp16)[name = string("op_8353_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8355_cast_fp16 = add(x = var_8349_cast_fp16, y = var_8353_cast_fp16)[name = string("op_8355_cast_fp16")];
            fp32 var_8356_epsilon_0 = const()[name = string("op_8356_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8356_cast_fp16 = rsqrt(epsilon = var_8356_epsilon_0, x = var_8355_cast_fp16)[name = string("op_8356_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_263_cast_fp16 = mul(x = y_301_cast_fp16, y = var_8356_cast_fp16)[name = string("out_263_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105887552)))];
            tensor<fp16, [1, 256, 512]> input_221_cast_fp16 = mul(x = out_263_cast_fp16, y = tm_layers_18_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105888640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106937280))))[name = string("tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_167_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_167_cast_fp16")];
            string var_8363_mode_0 = const()[name = string("op_8363_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_8363_cast_fp16 = gelu(mode = var_8363_mode_0, x = linear_167_cast_fp16)[name = string("op_8363_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106941440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107990080))))[name = string("tm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_168_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_223_cast_fp16 = mul(x = var_8363_cast_fp16, y = linear_168_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107994240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109042880))))[name = string("tm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8371_cast_fp16 = abs(x = linear_169_cast_fp16)[name = string("op_8371_cast_fp16")];
            bool var_8374 = const()[name = string("op_8374"), val = bool(true)];
            tensor<int32, [1]> var_8375_axes_0 = const()[name = string("op_8375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8375_cast_fp16 = reduce_max(axes = var_8375_axes_0, keep_dims = var_8374, x = var_8371_cast_fp16)[name = string("op_8375_cast_fp16")];
            fp16 var_8376_to_fp16 = const()[name = string("op_8376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_303_cast_fp16 = maximum(x = var_8375_cast_fp16, y = var_8376_to_fp16)[name = string("s_303_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_303_cast_fp16 = real_div(x = linear_169_cast_fp16, y = s_303_cast_fp16)[name = string("y_303_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8379_cast_fp16 = mul(x = y_303_cast_fp16, y = y_303_cast_fp16)[name = string("op_8379_cast_fp16")];
            tensor<int32, [1]> var_8384_axes_0 = const()[name = string("op_8384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8384_keep_dims_0 = const()[name = string("op_8384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8384_cast_fp16 = reduce_mean(axes = var_8384_axes_0, keep_dims = var_8384_keep_dims_0, x = var_8379_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8385_cast_fp16 = mul(x = s_303_cast_fp16, y = s_303_cast_fp16)[name = string("op_8385_cast_fp16")];
            fp32 var_8386_epsilon_0 = const()[name = string("op_8386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8386_cast_fp16 = inverse(epsilon = var_8386_epsilon_0, x = var_8385_cast_fp16)[name = string("op_8386_cast_fp16")];
            fp16 var_8387_to_fp16 = const()[name = string("op_8387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8388_cast_fp16 = mul(x = var_8386_cast_fp16, y = var_8387_to_fp16)[name = string("op_8388_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8390_cast_fp16 = add(x = var_8384_cast_fp16, y = var_8388_cast_fp16)[name = string("op_8390_cast_fp16")];
            fp32 var_8391_epsilon_0 = const()[name = string("op_8391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8391_cast_fp16 = rsqrt(epsilon = var_8391_epsilon_0, x = var_8390_cast_fp16)[name = string("op_8391_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_265_cast_fp16 = mul(x = y_303_cast_fp16, y = var_8391_cast_fp16)[name = string("out_265_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109043968)))];
            tensor<fp16, [1, 256, 512]> var_8393_cast_fp16 = mul(x = out_265_cast_fp16, y = tm_layers_18_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8393_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_225_cast_fp16 = add(x = x_335_cast_fp16, y = var_8393_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109045056))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109307264))))[name = string("tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_170_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_170_cast_fp16")];
            string x_339_mode_0 = const()[name = string("x_339_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_339_cast_fp16 = gelu(mode = x_339_mode_0, x = linear_170_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [4]> var_8413_begin_0 = const()[name = string("op_8413_begin_0"), val = tensor<int32, [4]>([0, 0, 18, 0])];
            tensor<int32, [4]> var_8413_end_0 = const()[name = string("op_8413_end_0"), val = tensor<int32, [4]>([1, 256, 19, 512])];
            tensor<bool, [4]> var_8413_end_mask_0 = const()[name = string("op_8413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8413_squeeze_mask_0 = const()[name = string("op_8413_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_8413_cast_fp16 = slice_by_index(begin = var_8413_begin_0, end = var_8413_end_0, end_mask = var_8413_end_mask_0, squeeze_mask = var_8413_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_227_cast_fp16 = mul(x = x_339_cast_fp16, y = var_8413_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109308352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109570560))))[name = string("tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8423_cast_fp16 = abs(x = linear_171_cast_fp16)[name = string("op_8423_cast_fp16")];
            bool var_8426 = const()[name = string("op_8426"), val = bool(true)];
            tensor<int32, [1]> var_8427_axes_0 = const()[name = string("op_8427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8427_cast_fp16 = reduce_max(axes = var_8427_axes_0, keep_dims = var_8426, x = var_8423_cast_fp16)[name = string("op_8427_cast_fp16")];
            fp16 var_8428_to_fp16 = const()[name = string("op_8428_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_305_cast_fp16 = maximum(x = var_8427_cast_fp16, y = var_8428_to_fp16)[name = string("s_305_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_305_cast_fp16 = real_div(x = linear_171_cast_fp16, y = s_305_cast_fp16)[name = string("y_305_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8431_cast_fp16 = mul(x = y_305_cast_fp16, y = y_305_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<int32, [1]> var_8436_axes_0 = const()[name = string("op_8436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8436_keep_dims_0 = const()[name = string("op_8436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8436_cast_fp16 = reduce_mean(axes = var_8436_axes_0, keep_dims = var_8436_keep_dims_0, x = var_8431_cast_fp16)[name = string("op_8436_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8437_cast_fp16 = mul(x = s_305_cast_fp16, y = s_305_cast_fp16)[name = string("op_8437_cast_fp16")];
            fp32 var_8438_epsilon_0 = const()[name = string("op_8438_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8438_cast_fp16 = inverse(epsilon = var_8438_epsilon_0, x = var_8437_cast_fp16)[name = string("op_8438_cast_fp16")];
            fp16 var_8439_to_fp16 = const()[name = string("op_8439_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8440_cast_fp16 = mul(x = var_8438_cast_fp16, y = var_8439_to_fp16)[name = string("op_8440_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8442_cast_fp16 = add(x = var_8436_cast_fp16, y = var_8440_cast_fp16)[name = string("op_8442_cast_fp16")];
            fp32 var_8443_epsilon_0 = const()[name = string("op_8443_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8443_cast_fp16 = rsqrt(epsilon = var_8443_epsilon_0, x = var_8442_cast_fp16)[name = string("op_8443_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_267_cast_fp16 = mul(x = y_305_cast_fp16, y = var_8443_cast_fp16)[name = string("out_267_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109571648)))];
            tensor<fp16, [1, 256, 512]> var_8445_cast_fp16 = mul(x = out_267_cast_fp16, y = tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8445_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_37_cast_fp16 = add(x = input_225_cast_fp16, y = var_8445_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<fp16, [1]> tm_layers_18_layer_scalar_to_fp16 = const()[name = string("tm_layers_18_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.68p-1])];
            tensor<fp16, [1, 256, 512]> x_343_cast_fp16 = mul(x = h_37_cast_fp16, y = tm_layers_18_layer_scalar_to_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8449_cast_fp16 = abs(x = x_343_cast_fp16)[name = string("op_8449_cast_fp16")];
            bool var_8452 = const()[name = string("op_8452"), val = bool(true)];
            tensor<int32, [1]> var_8453_axes_0 = const()[name = string("op_8453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8453_cast_fp16 = reduce_max(axes = var_8453_axes_0, keep_dims = var_8452, x = var_8449_cast_fp16)[name = string("op_8453_cast_fp16")];
            fp16 var_8454_to_fp16 = const()[name = string("op_8454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_307_cast_fp16 = maximum(x = var_8453_cast_fp16, y = var_8454_to_fp16)[name = string("s_307_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_307_cast_fp16 = real_div(x = x_343_cast_fp16, y = s_307_cast_fp16)[name = string("y_307_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8457_cast_fp16 = mul(x = y_307_cast_fp16, y = y_307_cast_fp16)[name = string("op_8457_cast_fp16")];
            tensor<int32, [1]> var_8462_axes_0 = const()[name = string("op_8462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8462_keep_dims_0 = const()[name = string("op_8462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8462_cast_fp16 = reduce_mean(axes = var_8462_axes_0, keep_dims = var_8462_keep_dims_0, x = var_8457_cast_fp16)[name = string("op_8462_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8463_cast_fp16 = mul(x = s_307_cast_fp16, y = s_307_cast_fp16)[name = string("op_8463_cast_fp16")];
            fp32 var_8464_epsilon_0 = const()[name = string("op_8464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8464_cast_fp16 = inverse(epsilon = var_8464_epsilon_0, x = var_8463_cast_fp16)[name = string("op_8464_cast_fp16")];
            fp16 var_8465_to_fp16 = const()[name = string("op_8465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8466_cast_fp16 = mul(x = var_8464_cast_fp16, y = var_8465_to_fp16)[name = string("op_8466_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8468_cast_fp16 = add(x = var_8462_cast_fp16, y = var_8466_cast_fp16)[name = string("op_8468_cast_fp16")];
            fp32 var_8469_epsilon_0 = const()[name = string("op_8469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8469_cast_fp16 = rsqrt(epsilon = var_8469_epsilon_0, x = var_8468_cast_fp16)[name = string("op_8469_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_269_cast_fp16 = mul(x = y_307_cast_fp16, y = var_8469_cast_fp16)[name = string("out_269_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109572736)))];
            tensor<fp16, [1, 256, 512]> input_229_cast_fp16 = mul(x = out_269_cast_fp16, y = tm_layers_19_input_layernorm_weight_to_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109573824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110098176))))[name = string("tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [4]> var_8479 = const()[name = string("op_8479"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_8480_cast_fp16 = reshape(shape = var_8479, x = linear_172_cast_fp16)[name = string("op_8480_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110100288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110362496))))[name = string("tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<int32, [4]> var_8491 = const()[name = string("op_8491"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8492_cast_fp16 = reshape(shape = var_8491, x = linear_173_cast_fp16)[name = string("op_8492_cast_fp16")];
            tensor<int32, [4]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110363584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110625792))))[name = string("tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<int32, [4]> var_8503 = const()[name = string("op_8503"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8504_cast_fp16 = reshape(shape = var_8503, x = linear_174_cast_fp16)[name = string("op_8504_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_8480_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 4, 256, 256]> var_8508_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_8508_cast_fp16")];
            bool var_8511 = const()[name = string("op_8511"), val = bool(true)];
            tensor<int32, [1]> var_8512_axes_0 = const()[name = string("op_8512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_8512_cast_fp16 = reduce_max(axes = var_8512_axes_0, keep_dims = var_8511, x = var_8508_cast_fp16)[name = string("op_8512_cast_fp16")];
            fp16 var_8513_to_fp16 = const()[name = string("op_8513_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_309_cast_fp16 = maximum(x = var_8512_cast_fp16, y = var_8513_to_fp16)[name = string("s_309_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_309_cast_fp16 = real_div(x = x_345_cast_fp16, y = s_309_cast_fp16)[name = string("y_309_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8516_cast_fp16 = mul(x = y_309_cast_fp16, y = y_309_cast_fp16)[name = string("op_8516_cast_fp16")];
            tensor<int32, [1]> var_8521_axes_0 = const()[name = string("op_8521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8521_keep_dims_0 = const()[name = string("op_8521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_8521_cast_fp16 = reduce_mean(axes = var_8521_axes_0, keep_dims = var_8521_keep_dims_0, x = var_8516_cast_fp16)[name = string("op_8521_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8522_cast_fp16 = mul(x = s_309_cast_fp16, y = s_309_cast_fp16)[name = string("op_8522_cast_fp16")];
            fp32 var_8523_epsilon_0 = const()[name = string("op_8523_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_8523_cast_fp16 = inverse(epsilon = var_8523_epsilon_0, x = var_8522_cast_fp16)[name = string("op_8523_cast_fp16")];
            fp16 var_8524_to_fp16 = const()[name = string("op_8524_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_8525_cast_fp16 = mul(x = var_8523_cast_fp16, y = var_8524_to_fp16)[name = string("op_8525_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8527_cast_fp16 = add(x = var_8521_cast_fp16, y = var_8525_cast_fp16)[name = string("op_8527_cast_fp16")];
            fp32 var_8528_epsilon_0 = const()[name = string("op_8528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_8528_cast_fp16 = rsqrt(epsilon = var_8528_epsilon_0, x = var_8527_cast_fp16)[name = string("op_8528_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_271_cast_fp16 = mul(x = y_309_cast_fp16, y = var_8528_cast_fp16)[name = string("out_271_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_77_cast_fp16 = mul(x = out_271_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_77_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = var_8492_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 256, 256]> var_8531_cast_fp16 = abs(x = x_347_cast_fp16)[name = string("op_8531_cast_fp16")];
            bool var_8534 = const()[name = string("op_8534"), val = bool(true)];
            tensor<int32, [1]> var_8535_axes_0 = const()[name = string("op_8535_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8535_cast_fp16 = reduce_max(axes = var_8535_axes_0, keep_dims = var_8534, x = var_8531_cast_fp16)[name = string("op_8535_cast_fp16")];
            fp16 var_8536_to_fp16 = const()[name = string("op_8536_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_311_cast_fp16 = maximum(x = var_8535_cast_fp16, y = var_8536_to_fp16)[name = string("s_311_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_311_cast_fp16 = real_div(x = x_347_cast_fp16, y = s_311_cast_fp16)[name = string("y_311_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8539_cast_fp16 = mul(x = y_311_cast_fp16, y = y_311_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<int32, [1]> var_8544_axes_0 = const()[name = string("op_8544_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8544_keep_dims_0 = const()[name = string("op_8544_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8544_cast_fp16 = reduce_mean(axes = var_8544_axes_0, keep_dims = var_8544_keep_dims_0, x = var_8539_cast_fp16)[name = string("op_8544_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8545_cast_fp16 = mul(x = s_311_cast_fp16, y = s_311_cast_fp16)[name = string("op_8545_cast_fp16")];
            fp32 var_8546_epsilon_0 = const()[name = string("op_8546_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8546_cast_fp16 = inverse(epsilon = var_8546_epsilon_0, x = var_8545_cast_fp16)[name = string("op_8546_cast_fp16")];
            fp16 var_8547_to_fp16 = const()[name = string("op_8547_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8548_cast_fp16 = mul(x = var_8546_cast_fp16, y = var_8547_to_fp16)[name = string("op_8548_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8550_cast_fp16 = add(x = var_8544_cast_fp16, y = var_8548_cast_fp16)[name = string("op_8550_cast_fp16")];
            fp32 var_8551_epsilon_0 = const()[name = string("op_8551_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8551_cast_fp16 = rsqrt(epsilon = var_8551_epsilon_0, x = var_8550_cast_fp16)[name = string("op_8551_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_273_cast_fp16 = mul(x = y_311_cast_fp16, y = var_8551_cast_fp16)[name = string("out_273_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_115_cast_fp16 = mul(x = out_273_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_115_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_8504_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 2, 256, 256]> var_8554_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_8554_cast_fp16")];
            bool var_8557 = const()[name = string("op_8557"), val = bool(true)];
            tensor<int32, [1]> var_8558_axes_0 = const()[name = string("op_8558_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8558_cast_fp16 = reduce_max(axes = var_8558_axes_0, keep_dims = var_8557, x = var_8554_cast_fp16)[name = string("op_8558_cast_fp16")];
            fp16 var_8559_to_fp16 = const()[name = string("op_8559_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_313_cast_fp16 = maximum(x = var_8558_cast_fp16, y = var_8559_to_fp16)[name = string("s_313_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_313_cast_fp16 = real_div(x = x_349_cast_fp16, y = s_313_cast_fp16)[name = string("y_313_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8562_cast_fp16 = mul(x = y_313_cast_fp16, y = y_313_cast_fp16)[name = string("op_8562_cast_fp16")];
            tensor<int32, [1]> var_8567_axes_0 = const()[name = string("op_8567_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8567_keep_dims_0 = const()[name = string("op_8567_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8567_cast_fp16 = reduce_mean(axes = var_8567_axes_0, keep_dims = var_8567_keep_dims_0, x = var_8562_cast_fp16)[name = string("op_8567_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8568_cast_fp16 = mul(x = s_313_cast_fp16, y = s_313_cast_fp16)[name = string("op_8568_cast_fp16")];
            fp32 var_8569_epsilon_0 = const()[name = string("op_8569_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8569_cast_fp16 = inverse(epsilon = var_8569_epsilon_0, x = var_8568_cast_fp16)[name = string("op_8569_cast_fp16")];
            fp16 var_8570_to_fp16 = const()[name = string("op_8570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8571_cast_fp16 = mul(x = var_8569_cast_fp16, y = var_8570_to_fp16)[name = string("op_8571_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8573_cast_fp16 = add(x = var_8567_cast_fp16, y = var_8571_cast_fp16)[name = string("op_8573_cast_fp16")];
            fp32 var_8574_epsilon_0 = const()[name = string("op_8574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8574_cast_fp16 = rsqrt(epsilon = var_8574_epsilon_0, x = var_8573_cast_fp16)[name = string("op_8574_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_77_cast_fp16 = mul(x = y_313_cast_fp16, y = var_8574_cast_fp16)[name = string("v_77_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8576_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8576_cast_fp16")];
            tensor<int32, [4]> var_8581_begin_0 = const()[name = string("op_8581_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8581_end_0 = const()[name = string("op_8581_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_8581_end_mask_0 = const()[name = string("op_8581_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_8581_cast_fp16 = slice_by_index(begin = var_8581_begin_0, end = var_8581_end_0, end_mask = var_8581_end_mask_0, x = q_77_cast_fp16)[name = string("op_8581_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_8582_cast_fp16 = mul(x = var_8581_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_8582_cast_fp16")];
            tensor<int32, [4]> var_8587_begin_0 = const()[name = string("op_8587_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8587_end_0 = const()[name = string("op_8587_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_8587_end_mask_0 = const()[name = string("op_8587_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_8587_cast_fp16 = slice_by_index(begin = var_8587_begin_0, end = var_8587_end_0, end_mask = var_8587_end_mask_0, x = q_77_cast_fp16)[name = string("op_8587_cast_fp16")];
            int32 var_8589 = const()[name = string("op_8589"), val = int32(-1)];
            bool var_8590_interleave_0 = const()[name = string("op_8590_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_8590_cast_fp16 = concat(axis = var_8589, interleave = var_8590_interleave_0, values = (var_8582_cast_fp16, var_8587_cast_fp16))[name = string("op_8590_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8591_cast_fp16 = mul(x = var_8590_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8591_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_79_cast_fp16 = add(x = var_8576_cast_fp16, y = var_8591_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8594_cast_fp16 = mul(x = k_115_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8594_cast_fp16")];
            tensor<int32, [4]> var_8599_begin_0 = const()[name = string("op_8599_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8599_end_0 = const()[name = string("op_8599_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_8599_end_mask_0 = const()[name = string("op_8599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_8599_cast_fp16 = slice_by_index(begin = var_8599_begin_0, end = var_8599_end_0, end_mask = var_8599_end_mask_0, x = k_115_cast_fp16)[name = string("op_8599_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_8600_cast_fp16 = mul(x = var_8599_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_8600_cast_fp16")];
            tensor<int32, [4]> var_8605_begin_0 = const()[name = string("op_8605_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8605_end_0 = const()[name = string("op_8605_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_8605_end_mask_0 = const()[name = string("op_8605_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_8605_cast_fp16 = slice_by_index(begin = var_8605_begin_0, end = var_8605_end_0, end_mask = var_8605_end_mask_0, x = k_115_cast_fp16)[name = string("op_8605_cast_fp16")];
            int32 var_8607 = const()[name = string("op_8607"), val = int32(-1)];
            bool var_8608_interleave_0 = const()[name = string("op_8608_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_8608_cast_fp16 = concat(axis = var_8607, interleave = var_8608_interleave_0, values = (var_8600_cast_fp16, var_8605_cast_fp16))[name = string("op_8608_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8609_cast_fp16 = mul(x = var_8608_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8609_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_117_cast_fp16 = add(x = var_8594_cast_fp16, y = var_8609_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_38_reps_0 = const()[name = string("tile_38_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = k_117_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 1, 256, 256]> tile_38_cast_fp16 = tile(reps = tile_38_reps_0, x = transpose_76_cast_fp16)[name = string("tile_38_cast_fp16")];
            tensor<int32, [5]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_76_cast_fp16 = reshape(shape = concat_76, x = tile_38_cast_fp16)[name = string("reshape_76_cast_fp16")];
            tensor<int32, [5]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_77 = const()[name = string("concat_77"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_77_cast_fp16 = transpose(perm = transpose_77_perm_0, x = reshape_76_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 1, 256, 256]> reshape_77_cast_fp16 = reshape(shape = concat_77, x = transpose_77_cast_fp16)[name = string("reshape_77_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_39_reps_0 = const()[name = string("tile_39_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_78_cast_fp16 = transpose(perm = transpose_78_perm_0, x = v_77_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 1, 256, 256]> tile_39_cast_fp16 = tile(reps = tile_39_reps_0, x = transpose_78_cast_fp16)[name = string("tile_39_cast_fp16")];
            tensor<int32, [5]> concat_78 = const()[name = string("concat_78"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_78_cast_fp16 = reshape(shape = concat_78, x = tile_39_cast_fp16)[name = string("reshape_78_cast_fp16")];
            tensor<int32, [5]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = reshape_78_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 1, 256, 256]> reshape_79_cast_fp16 = reshape(shape = concat_79, x = transpose_79_cast_fp16)[name = string("reshape_79_cast_fp16")];
            tensor<int32, [4]> v_79_perm_0 = const()[name = string("v_79_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8623_transpose_x_0 = const()[name = string("op_8623_transpose_x_0"), val = bool(false)];
            bool var_8623_transpose_y_0 = const()[name = string("op_8623_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = reshape_77_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4, 256, 256]> var_8623_cast_fp16 = matmul(transpose_x = var_8623_transpose_x_0, transpose_y = var_8623_transpose_y_0, x = q_79_cast_fp16, y = transpose_115_cast_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8625_cast_fp16")];
            int32 var_8626 = const()[name = string("op_8626"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_313_cast_fp16 = softmax(axis = var_8626, x = var_8625_cast_fp16)[name = string("w_313_cast_fp16")];
            bool var_8629_transpose_x_0 = const()[name = string("op_8629_transpose_x_0"), val = bool(false)];
            bool var_8629_transpose_y_0 = const()[name = string("op_8629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = reshape_79_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 4, 256, 256]> var_8629_cast_fp16 = matmul(transpose_x = var_8629_transpose_x_0, transpose_y = var_8629_transpose_y_0, x = w_313_cast_fp16, y = v_79_cast_fp16)[name = string("op_8629_cast_fp16")];
            tensor<int32, [4]> var_8632_perm_0 = const()[name = string("op_8632_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_8632_cast_fp16 = transpose(perm = var_8632_perm_0, x = var_8629_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 256, 1024]> input_231_cast_fp16 = reshape(shape = var_8636, x = var_8632_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110626880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111151232))))[name = string("tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8641_cast_fp16 = abs(x = linear_175_cast_fp16)[name = string("op_8641_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<int32, [1]> var_8645_axes_0 = const()[name = string("op_8645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8645_cast_fp16 = reduce_max(axes = var_8645_axes_0, keep_dims = var_8644, x = var_8641_cast_fp16)[name = string("op_8645_cast_fp16")];
            fp16 var_8646_to_fp16 = const()[name = string("op_8646_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_315_cast_fp16 = maximum(x = var_8645_cast_fp16, y = var_8646_to_fp16)[name = string("s_315_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_315_cast_fp16 = real_div(x = linear_175_cast_fp16, y = s_315_cast_fp16)[name = string("y_315_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8649_cast_fp16 = mul(x = y_315_cast_fp16, y = y_315_cast_fp16)[name = string("op_8649_cast_fp16")];
            tensor<int32, [1]> var_8654_axes_0 = const()[name = string("op_8654_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8654_keep_dims_0 = const()[name = string("op_8654_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8654_cast_fp16 = reduce_mean(axes = var_8654_axes_0, keep_dims = var_8654_keep_dims_0, x = var_8649_cast_fp16)[name = string("op_8654_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8655_cast_fp16 = mul(x = s_315_cast_fp16, y = s_315_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp32 var_8656_epsilon_0 = const()[name = string("op_8656_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8656_cast_fp16 = inverse(epsilon = var_8656_epsilon_0, x = var_8655_cast_fp16)[name = string("op_8656_cast_fp16")];
            fp16 var_8657_to_fp16 = const()[name = string("op_8657_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8658_cast_fp16 = mul(x = var_8656_cast_fp16, y = var_8657_to_fp16)[name = string("op_8658_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8660_cast_fp16 = add(x = var_8654_cast_fp16, y = var_8658_cast_fp16)[name = string("op_8660_cast_fp16")];
            fp32 var_8661_epsilon_0 = const()[name = string("op_8661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8661_cast_fp16 = rsqrt(epsilon = var_8661_epsilon_0, x = var_8660_cast_fp16)[name = string("op_8661_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_275_cast_fp16 = mul(x = y_315_cast_fp16, y = var_8661_cast_fp16)[name = string("out_275_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111152320)))];
            tensor<fp16, [1, 256, 512]> var_8663_cast_fp16 = mul(x = out_275_cast_fp16, y = tm_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("op_8663_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_353_cast_fp16 = add(x = x_343_cast_fp16, y = var_8663_cast_fp16)[name = string("x_353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8666_cast_fp16 = abs(x = x_353_cast_fp16)[name = string("op_8666_cast_fp16")];
            bool var_8669 = const()[name = string("op_8669"), val = bool(true)];
            tensor<int32, [1]> var_8670_axes_0 = const()[name = string("op_8670_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8670_cast_fp16 = reduce_max(axes = var_8670_axes_0, keep_dims = var_8669, x = var_8666_cast_fp16)[name = string("op_8670_cast_fp16")];
            fp16 var_8671_to_fp16 = const()[name = string("op_8671_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_317_cast_fp16 = maximum(x = var_8670_cast_fp16, y = var_8671_to_fp16)[name = string("s_317_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_317_cast_fp16 = real_div(x = x_353_cast_fp16, y = s_317_cast_fp16)[name = string("y_317_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8674_cast_fp16 = mul(x = y_317_cast_fp16, y = y_317_cast_fp16)[name = string("op_8674_cast_fp16")];
            tensor<int32, [1]> var_8679_axes_0 = const()[name = string("op_8679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8679_keep_dims_0 = const()[name = string("op_8679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8679_cast_fp16 = reduce_mean(axes = var_8679_axes_0, keep_dims = var_8679_keep_dims_0, x = var_8674_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8680_cast_fp16 = mul(x = s_317_cast_fp16, y = s_317_cast_fp16)[name = string("op_8680_cast_fp16")];
            fp32 var_8681_epsilon_0 = const()[name = string("op_8681_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8681_cast_fp16 = inverse(epsilon = var_8681_epsilon_0, x = var_8680_cast_fp16)[name = string("op_8681_cast_fp16")];
            fp16 var_8682_to_fp16 = const()[name = string("op_8682_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8683_cast_fp16 = mul(x = var_8681_cast_fp16, y = var_8682_to_fp16)[name = string("op_8683_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8685_cast_fp16 = add(x = var_8679_cast_fp16, y = var_8683_cast_fp16)[name = string("op_8685_cast_fp16")];
            fp32 var_8686_epsilon_0 = const()[name = string("op_8686_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8686_cast_fp16 = rsqrt(epsilon = var_8686_epsilon_0, x = var_8685_cast_fp16)[name = string("op_8686_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_277_cast_fp16 = mul(x = y_317_cast_fp16, y = var_8686_cast_fp16)[name = string("out_277_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111153408)))];
            tensor<fp16, [1, 256, 512]> input_233_cast_fp16 = mul(x = out_277_cast_fp16, y = tm_layers_19_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111154496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112203136))))[name = string("tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_176_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_176_cast_fp16")];
            string var_8693_mode_0 = const()[name = string("op_8693_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_8693_cast_fp16 = gelu(mode = var_8693_mode_0, x = linear_176_cast_fp16)[name = string("op_8693_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112207296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113255936))))[name = string("tm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_177_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_235_cast_fp16 = mul(x = var_8693_cast_fp16, y = linear_177_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113260096))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114308736))))[name = string("tm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8701_cast_fp16 = abs(x = linear_178_cast_fp16)[name = string("op_8701_cast_fp16")];
            bool var_8704 = const()[name = string("op_8704"), val = bool(true)];
            tensor<int32, [1]> var_8705_axes_0 = const()[name = string("op_8705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8705_cast_fp16 = reduce_max(axes = var_8705_axes_0, keep_dims = var_8704, x = var_8701_cast_fp16)[name = string("op_8705_cast_fp16")];
            fp16 var_8706_to_fp16 = const()[name = string("op_8706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_319_cast_fp16 = maximum(x = var_8705_cast_fp16, y = var_8706_to_fp16)[name = string("s_319_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_319_cast_fp16 = real_div(x = linear_178_cast_fp16, y = s_319_cast_fp16)[name = string("y_319_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8709_cast_fp16 = mul(x = y_319_cast_fp16, y = y_319_cast_fp16)[name = string("op_8709_cast_fp16")];
            tensor<int32, [1]> var_8714_axes_0 = const()[name = string("op_8714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8714_keep_dims_0 = const()[name = string("op_8714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8714_cast_fp16 = reduce_mean(axes = var_8714_axes_0, keep_dims = var_8714_keep_dims_0, x = var_8709_cast_fp16)[name = string("op_8714_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8715_cast_fp16 = mul(x = s_319_cast_fp16, y = s_319_cast_fp16)[name = string("op_8715_cast_fp16")];
            fp32 var_8716_epsilon_0 = const()[name = string("op_8716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8716_cast_fp16 = inverse(epsilon = var_8716_epsilon_0, x = var_8715_cast_fp16)[name = string("op_8716_cast_fp16")];
            fp16 var_8717_to_fp16 = const()[name = string("op_8717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8718_cast_fp16 = mul(x = var_8716_cast_fp16, y = var_8717_to_fp16)[name = string("op_8718_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8720_cast_fp16 = add(x = var_8714_cast_fp16, y = var_8718_cast_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8721_cast_fp16 = rsqrt(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_279_cast_fp16 = mul(x = y_319_cast_fp16, y = var_8721_cast_fp16)[name = string("out_279_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114309824)))];
            tensor<fp16, [1, 256, 512]> var_8723_cast_fp16 = mul(x = out_279_cast_fp16, y = tm_layers_19_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_237_cast_fp16 = add(x = x_353_cast_fp16, y = var_8723_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114310912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114573120))))[name = string("tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_179_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_237_cast_fp16)[name = string("linear_179_cast_fp16")];
            string x_357_mode_0 = const()[name = string("x_357_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_357_cast_fp16 = gelu(mode = x_357_mode_0, x = linear_179_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [4]> var_8743_begin_0 = const()[name = string("op_8743_begin_0"), val = tensor<int32, [4]>([0, 0, 19, 0])];
            tensor<int32, [4]> var_8743_end_0 = const()[name = string("op_8743_end_0"), val = tensor<int32, [4]>([1, 256, 20, 512])];
            tensor<bool, [4]> var_8743_end_mask_0 = const()[name = string("op_8743_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8743_squeeze_mask_0 = const()[name = string("op_8743_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_8743_cast_fp16 = slice_by_index(begin = var_8743_begin_0, end = var_8743_end_0, end_mask = var_8743_end_mask_0, squeeze_mask = var_8743_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8743_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_239_cast_fp16 = mul(x = x_357_cast_fp16, y = var_8743_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114574208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114836416))))[name = string("tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8753_cast_fp16 = abs(x = linear_180_cast_fp16)[name = string("op_8753_cast_fp16")];
            bool var_8756 = const()[name = string("op_8756"), val = bool(true)];
            tensor<int32, [1]> var_8757_axes_0 = const()[name = string("op_8757_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8757_cast_fp16 = reduce_max(axes = var_8757_axes_0, keep_dims = var_8756, x = var_8753_cast_fp16)[name = string("op_8757_cast_fp16")];
            fp16 var_8758_to_fp16 = const()[name = string("op_8758_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_321_cast_fp16 = maximum(x = var_8757_cast_fp16, y = var_8758_to_fp16)[name = string("s_321_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_321_cast_fp16 = real_div(x = linear_180_cast_fp16, y = s_321_cast_fp16)[name = string("y_321_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8761_cast_fp16 = mul(x = y_321_cast_fp16, y = y_321_cast_fp16)[name = string("op_8761_cast_fp16")];
            tensor<int32, [1]> var_8766_axes_0 = const()[name = string("op_8766_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8766_keep_dims_0 = const()[name = string("op_8766_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8766_cast_fp16 = reduce_mean(axes = var_8766_axes_0, keep_dims = var_8766_keep_dims_0, x = var_8761_cast_fp16)[name = string("op_8766_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8767_cast_fp16 = mul(x = s_321_cast_fp16, y = s_321_cast_fp16)[name = string("op_8767_cast_fp16")];
            fp32 var_8768_epsilon_0 = const()[name = string("op_8768_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8768_cast_fp16 = inverse(epsilon = var_8768_epsilon_0, x = var_8767_cast_fp16)[name = string("op_8768_cast_fp16")];
            fp16 var_8769_to_fp16 = const()[name = string("op_8769_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8770_cast_fp16 = mul(x = var_8768_cast_fp16, y = var_8769_to_fp16)[name = string("op_8770_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8772_cast_fp16 = add(x = var_8766_cast_fp16, y = var_8770_cast_fp16)[name = string("op_8772_cast_fp16")];
            fp32 var_8773_epsilon_0 = const()[name = string("op_8773_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8773_cast_fp16 = rsqrt(epsilon = var_8773_epsilon_0, x = var_8772_cast_fp16)[name = string("op_8773_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_281_cast_fp16 = mul(x = y_321_cast_fp16, y = var_8773_cast_fp16)[name = string("out_281_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114837504)))];
            tensor<fp16, [1, 256, 512]> var_8775_cast_fp16 = mul(x = out_281_cast_fp16, y = tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_39_cast_fp16 = add(x = input_237_cast_fp16, y = var_8775_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<fp16, [1]> tm_layers_19_layer_scalar_to_fp16 = const()[name = string("tm_layers_19_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fp-1])];
            tensor<fp16, [1, 256, 512]> x_361_cast_fp16 = mul(x = h_39_cast_fp16, y = tm_layers_19_layer_scalar_to_fp16)[name = string("x_361_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8779_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_8779_cast_fp16")];
            bool var_8782 = const()[name = string("op_8782"), val = bool(true)];
            tensor<int32, [1]> var_8783_axes_0 = const()[name = string("op_8783_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8783_cast_fp16 = reduce_max(axes = var_8783_axes_0, keep_dims = var_8782, x = var_8779_cast_fp16)[name = string("op_8783_cast_fp16")];
            fp16 var_8784_to_fp16 = const()[name = string("op_8784_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_323_cast_fp16 = maximum(x = var_8783_cast_fp16, y = var_8784_to_fp16)[name = string("s_323_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_323_cast_fp16 = real_div(x = x_361_cast_fp16, y = s_323_cast_fp16)[name = string("y_323_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8787_cast_fp16 = mul(x = y_323_cast_fp16, y = y_323_cast_fp16)[name = string("op_8787_cast_fp16")];
            tensor<int32, [1]> var_8792_axes_0 = const()[name = string("op_8792_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8792_keep_dims_0 = const()[name = string("op_8792_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8792_cast_fp16 = reduce_mean(axes = var_8792_axes_0, keep_dims = var_8792_keep_dims_0, x = var_8787_cast_fp16)[name = string("op_8792_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8793_cast_fp16 = mul(x = s_323_cast_fp16, y = s_323_cast_fp16)[name = string("op_8793_cast_fp16")];
            fp32 var_8794_epsilon_0 = const()[name = string("op_8794_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8794_cast_fp16 = inverse(epsilon = var_8794_epsilon_0, x = var_8793_cast_fp16)[name = string("op_8794_cast_fp16")];
            fp16 var_8795_to_fp16 = const()[name = string("op_8795_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8796_cast_fp16 = mul(x = var_8794_cast_fp16, y = var_8795_to_fp16)[name = string("op_8796_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8798_cast_fp16 = add(x = var_8792_cast_fp16, y = var_8796_cast_fp16)[name = string("op_8798_cast_fp16")];
            fp32 var_8799_epsilon_0 = const()[name = string("op_8799_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8799_cast_fp16 = rsqrt(epsilon = var_8799_epsilon_0, x = var_8798_cast_fp16)[name = string("op_8799_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_283_cast_fp16 = mul(x = y_323_cast_fp16, y = var_8799_cast_fp16)[name = string("out_283_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114838592)))];
            tensor<fp16, [1, 256, 512]> input_241_cast_fp16 = mul(x = out_283_cast_fp16, y = tm_layers_20_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114839680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115364032))))[name = string("tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<int32, [4]> var_8809 = const()[name = string("op_8809"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_8810_cast_fp16 = reshape(shape = var_8809, x = linear_181_cast_fp16)[name = string("op_8810_cast_fp16")];
            tensor<int32, [4]> x_363_perm_0 = const()[name = string("x_363_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115366144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115628352))))[name = string("tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_8821 = const()[name = string("op_8821"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8822_cast_fp16 = reshape(shape = var_8821, x = linear_182_cast_fp16)[name = string("op_8822_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115629440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115891648))))[name = string("tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_8833 = const()[name = string("op_8833"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8834_cast_fp16 = reshape(shape = var_8833, x = linear_183_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<int32, [4]> x_367_perm_0 = const()[name = string("x_367_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_363_cast_fp16 = transpose(perm = x_363_perm_0, x = var_8810_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 4, 256, 256]> var_8838_cast_fp16 = abs(x = x_363_cast_fp16)[name = string("op_8838_cast_fp16")];
            bool var_8841 = const()[name = string("op_8841"), val = bool(true)];
            tensor<int32, [1]> var_8842_axes_0 = const()[name = string("op_8842_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_8842_cast_fp16 = reduce_max(axes = var_8842_axes_0, keep_dims = var_8841, x = var_8838_cast_fp16)[name = string("op_8842_cast_fp16")];
            fp16 var_8843_to_fp16 = const()[name = string("op_8843_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_325_cast_fp16 = maximum(x = var_8842_cast_fp16, y = var_8843_to_fp16)[name = string("s_325_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_325_cast_fp16 = real_div(x = x_363_cast_fp16, y = s_325_cast_fp16)[name = string("y_325_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8846_cast_fp16 = mul(x = y_325_cast_fp16, y = y_325_cast_fp16)[name = string("op_8846_cast_fp16")];
            tensor<int32, [1]> var_8851_axes_0 = const()[name = string("op_8851_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8851_keep_dims_0 = const()[name = string("op_8851_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_8851_cast_fp16 = reduce_mean(axes = var_8851_axes_0, keep_dims = var_8851_keep_dims_0, x = var_8846_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8852_cast_fp16 = mul(x = s_325_cast_fp16, y = s_325_cast_fp16)[name = string("op_8852_cast_fp16")];
            fp32 var_8853_epsilon_0 = const()[name = string("op_8853_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_8853_cast_fp16 = inverse(epsilon = var_8853_epsilon_0, x = var_8852_cast_fp16)[name = string("op_8853_cast_fp16")];
            fp16 var_8854_to_fp16 = const()[name = string("op_8854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_8855_cast_fp16 = mul(x = var_8853_cast_fp16, y = var_8854_to_fp16)[name = string("op_8855_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8857_cast_fp16 = add(x = var_8851_cast_fp16, y = var_8855_cast_fp16)[name = string("op_8857_cast_fp16")];
            fp32 var_8858_epsilon_0 = const()[name = string("op_8858_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_8858_cast_fp16 = rsqrt(epsilon = var_8858_epsilon_0, x = var_8857_cast_fp16)[name = string("op_8858_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_285_cast_fp16 = mul(x = y_325_cast_fp16, y = var_8858_cast_fp16)[name = string("out_285_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_81_cast_fp16 = mul(x = out_285_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_81_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_8822_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 256, 256]> var_8861_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_8861_cast_fp16")];
            bool var_8864 = const()[name = string("op_8864"), val = bool(true)];
            tensor<int32, [1]> var_8865_axes_0 = const()[name = string("op_8865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8865_cast_fp16 = reduce_max(axes = var_8865_axes_0, keep_dims = var_8864, x = var_8861_cast_fp16)[name = string("op_8865_cast_fp16")];
            fp16 var_8866_to_fp16 = const()[name = string("op_8866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_327_cast_fp16 = maximum(x = var_8865_cast_fp16, y = var_8866_to_fp16)[name = string("s_327_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_327_cast_fp16 = real_div(x = x_365_cast_fp16, y = s_327_cast_fp16)[name = string("y_327_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8869_cast_fp16 = mul(x = y_327_cast_fp16, y = y_327_cast_fp16)[name = string("op_8869_cast_fp16")];
            tensor<int32, [1]> var_8874_axes_0 = const()[name = string("op_8874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8874_keep_dims_0 = const()[name = string("op_8874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8874_cast_fp16 = reduce_mean(axes = var_8874_axes_0, keep_dims = var_8874_keep_dims_0, x = var_8869_cast_fp16)[name = string("op_8874_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8875_cast_fp16 = mul(x = s_327_cast_fp16, y = s_327_cast_fp16)[name = string("op_8875_cast_fp16")];
            fp32 var_8876_epsilon_0 = const()[name = string("op_8876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8876_cast_fp16 = inverse(epsilon = var_8876_epsilon_0, x = var_8875_cast_fp16)[name = string("op_8876_cast_fp16")];
            fp16 var_8877_to_fp16 = const()[name = string("op_8877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8878_cast_fp16 = mul(x = var_8876_cast_fp16, y = var_8877_to_fp16)[name = string("op_8878_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8880_cast_fp16 = add(x = var_8874_cast_fp16, y = var_8878_cast_fp16)[name = string("op_8880_cast_fp16")];
            fp32 var_8881_epsilon_0 = const()[name = string("op_8881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8881_cast_fp16 = rsqrt(epsilon = var_8881_epsilon_0, x = var_8880_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_287_cast_fp16 = mul(x = y_327_cast_fp16, y = var_8881_cast_fp16)[name = string("out_287_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_121_cast_fp16 = mul(x = out_287_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_121_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_367_cast_fp16 = transpose(perm = x_367_perm_0, x = var_8834_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 256, 256]> var_8884_cast_fp16 = abs(x = x_367_cast_fp16)[name = string("op_8884_cast_fp16")];
            bool var_8887 = const()[name = string("op_8887"), val = bool(true)];
            tensor<int32, [1]> var_8888_axes_0 = const()[name = string("op_8888_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8888_cast_fp16 = reduce_max(axes = var_8888_axes_0, keep_dims = var_8887, x = var_8884_cast_fp16)[name = string("op_8888_cast_fp16")];
            fp16 var_8889_to_fp16 = const()[name = string("op_8889_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_329_cast_fp16 = maximum(x = var_8888_cast_fp16, y = var_8889_to_fp16)[name = string("s_329_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_329_cast_fp16 = real_div(x = x_367_cast_fp16, y = s_329_cast_fp16)[name = string("y_329_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8892_cast_fp16 = mul(x = y_329_cast_fp16, y = y_329_cast_fp16)[name = string("op_8892_cast_fp16")];
            tensor<int32, [1]> var_8897_axes_0 = const()[name = string("op_8897_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8897_keep_dims_0 = const()[name = string("op_8897_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8897_cast_fp16 = reduce_mean(axes = var_8897_axes_0, keep_dims = var_8897_keep_dims_0, x = var_8892_cast_fp16)[name = string("op_8897_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8898_cast_fp16 = mul(x = s_329_cast_fp16, y = s_329_cast_fp16)[name = string("op_8898_cast_fp16")];
            fp32 var_8899_epsilon_0 = const()[name = string("op_8899_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8899_cast_fp16 = inverse(epsilon = var_8899_epsilon_0, x = var_8898_cast_fp16)[name = string("op_8899_cast_fp16")];
            fp16 var_8900_to_fp16 = const()[name = string("op_8900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8901_cast_fp16 = mul(x = var_8899_cast_fp16, y = var_8900_to_fp16)[name = string("op_8901_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8903_cast_fp16 = add(x = var_8897_cast_fp16, y = var_8901_cast_fp16)[name = string("op_8903_cast_fp16")];
            fp32 var_8904_epsilon_0 = const()[name = string("op_8904_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8904_cast_fp16 = rsqrt(epsilon = var_8904_epsilon_0, x = var_8903_cast_fp16)[name = string("op_8904_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_81_cast_fp16 = mul(x = y_329_cast_fp16, y = var_8904_cast_fp16)[name = string("v_81_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8906_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8906_cast_fp16")];
            tensor<int32, [4]> var_8911_begin_0 = const()[name = string("op_8911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8911_end_0 = const()[name = string("op_8911_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_8911_end_mask_0 = const()[name = string("op_8911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_8911_cast_fp16 = slice_by_index(begin = var_8911_begin_0, end = var_8911_end_0, end_mask = var_8911_end_mask_0, x = q_81_cast_fp16)[name = string("op_8911_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_8912_cast_fp16 = mul(x = var_8911_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_8912_cast_fp16")];
            tensor<int32, [4]> var_8917_begin_0 = const()[name = string("op_8917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8917_end_0 = const()[name = string("op_8917_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_8917_end_mask_0 = const()[name = string("op_8917_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_8917_cast_fp16 = slice_by_index(begin = var_8917_begin_0, end = var_8917_end_0, end_mask = var_8917_end_mask_0, x = q_81_cast_fp16)[name = string("op_8917_cast_fp16")];
            int32 var_8919 = const()[name = string("op_8919"), val = int32(-1)];
            bool var_8920_interleave_0 = const()[name = string("op_8920_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_8920_cast_fp16 = concat(axis = var_8919, interleave = var_8920_interleave_0, values = (var_8912_cast_fp16, var_8917_cast_fp16))[name = string("op_8920_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8921_cast_fp16 = mul(x = var_8920_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_83_cast_fp16 = add(x = var_8906_cast_fp16, y = var_8921_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8924_cast_fp16 = mul(x = k_121_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8924_cast_fp16")];
            tensor<int32, [4]> var_8929_begin_0 = const()[name = string("op_8929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8929_end_0 = const()[name = string("op_8929_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_8929_end_mask_0 = const()[name = string("op_8929_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, x = k_121_cast_fp16)[name = string("op_8929_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_8930_cast_fp16 = mul(x = var_8929_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_8930_cast_fp16")];
            tensor<int32, [4]> var_8935_begin_0 = const()[name = string("op_8935_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8935_end_0 = const()[name = string("op_8935_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_8935_end_mask_0 = const()[name = string("op_8935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, x = k_121_cast_fp16)[name = string("op_8935_cast_fp16")];
            int32 var_8937 = const()[name = string("op_8937"), val = int32(-1)];
            bool var_8938_interleave_0 = const()[name = string("op_8938_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_8938_cast_fp16 = concat(axis = var_8937, interleave = var_8938_interleave_0, values = (var_8930_cast_fp16, var_8935_cast_fp16))[name = string("op_8938_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8939_cast_fp16 = mul(x = var_8938_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8939_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_123_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8939_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_40_reps_0 = const()[name = string("tile_40_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = k_123_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 1, 256, 256]> tile_40_cast_fp16 = tile(reps = tile_40_reps_0, x = transpose_80_cast_fp16)[name = string("tile_40_cast_fp16")];
            tensor<int32, [5]> concat_80 = const()[name = string("concat_80"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_80_cast_fp16 = reshape(shape = concat_80, x = tile_40_cast_fp16)[name = string("reshape_80_cast_fp16")];
            tensor<int32, [5]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_81 = const()[name = string("concat_81"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = reshape_80_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 1, 256, 256]> reshape_81_cast_fp16 = reshape(shape = concat_81, x = transpose_81_cast_fp16)[name = string("reshape_81_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_41_reps_0 = const()[name = string("tile_41_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_82_cast_fp16 = transpose(perm = transpose_82_perm_0, x = v_81_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 1, 256, 256]> tile_41_cast_fp16 = tile(reps = tile_41_reps_0, x = transpose_82_cast_fp16)[name = string("tile_41_cast_fp16")];
            tensor<int32, [5]> concat_82 = const()[name = string("concat_82"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_82_cast_fp16 = reshape(shape = concat_82, x = tile_41_cast_fp16)[name = string("reshape_82_cast_fp16")];
            tensor<int32, [5]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_83 = const()[name = string("concat_83"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = reshape_82_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 1, 256, 256]> reshape_83_cast_fp16 = reshape(shape = concat_83, x = transpose_83_cast_fp16)[name = string("reshape_83_cast_fp16")];
            tensor<int32, [4]> v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8953_transpose_x_0 = const()[name = string("op_8953_transpose_x_0"), val = bool(false)];
            bool var_8953_transpose_y_0 = const()[name = string("op_8953_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = reshape_81_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 4, 256, 256]> var_8953_cast_fp16 = matmul(transpose_x = var_8953_transpose_x_0, transpose_y = var_8953_transpose_y_0, x = q_83_cast_fp16, y = transpose_116_cast_fp16)[name = string("op_8953_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8955_cast_fp16 = add(x = var_8953_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8955_cast_fp16")];
            int32 var_8956 = const()[name = string("op_8956"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_329_cast_fp16 = softmax(axis = var_8956, x = var_8955_cast_fp16)[name = string("w_329_cast_fp16")];
            bool var_8959_transpose_x_0 = const()[name = string("op_8959_transpose_x_0"), val = bool(false)];
            bool var_8959_transpose_y_0 = const()[name = string("op_8959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_83_cast_fp16 = transpose(perm = v_83_perm_0, x = reshape_83_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 4, 256, 256]> var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = w_329_cast_fp16, y = v_83_cast_fp16)[name = string("op_8959_cast_fp16")];
            tensor<int32, [4]> var_8962_perm_0 = const()[name = string("op_8962_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8966 = const()[name = string("op_8966"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_8962_cast_fp16 = transpose(perm = var_8962_perm_0, x = var_8959_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 256, 1024]> input_243_cast_fp16 = reshape(shape = var_8966, x = var_8962_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115892736))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116417088))))[name = string("tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8971_cast_fp16 = abs(x = linear_184_cast_fp16)[name = string("op_8971_cast_fp16")];
            bool var_8974 = const()[name = string("op_8974"), val = bool(true)];
            tensor<int32, [1]> var_8975_axes_0 = const()[name = string("op_8975_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8975_cast_fp16 = reduce_max(axes = var_8975_axes_0, keep_dims = var_8974, x = var_8971_cast_fp16)[name = string("op_8975_cast_fp16")];
            fp16 var_8976_to_fp16 = const()[name = string("op_8976_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_331_cast_fp16 = maximum(x = var_8975_cast_fp16, y = var_8976_to_fp16)[name = string("s_331_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_331_cast_fp16 = real_div(x = linear_184_cast_fp16, y = s_331_cast_fp16)[name = string("y_331_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8979_cast_fp16 = mul(x = y_331_cast_fp16, y = y_331_cast_fp16)[name = string("op_8979_cast_fp16")];
            tensor<int32, [1]> var_8984_axes_0 = const()[name = string("op_8984_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8984_keep_dims_0 = const()[name = string("op_8984_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8984_cast_fp16 = reduce_mean(axes = var_8984_axes_0, keep_dims = var_8984_keep_dims_0, x = var_8979_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8985_cast_fp16 = mul(x = s_331_cast_fp16, y = s_331_cast_fp16)[name = string("op_8985_cast_fp16")];
            fp32 var_8986_epsilon_0 = const()[name = string("op_8986_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8986_cast_fp16 = inverse(epsilon = var_8986_epsilon_0, x = var_8985_cast_fp16)[name = string("op_8986_cast_fp16")];
            fp16 var_8987_to_fp16 = const()[name = string("op_8987_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8988_cast_fp16 = mul(x = var_8986_cast_fp16, y = var_8987_to_fp16)[name = string("op_8988_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8990_cast_fp16 = add(x = var_8984_cast_fp16, y = var_8988_cast_fp16)[name = string("op_8990_cast_fp16")];
            fp32 var_8991_epsilon_0 = const()[name = string("op_8991_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8991_cast_fp16 = rsqrt(epsilon = var_8991_epsilon_0, x = var_8990_cast_fp16)[name = string("op_8991_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_289_cast_fp16 = mul(x = y_331_cast_fp16, y = var_8991_cast_fp16)[name = string("out_289_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116418176)))];
            tensor<fp16, [1, 256, 512]> var_8993_cast_fp16 = mul(x = out_289_cast_fp16, y = tm_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("op_8993_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_371_cast_fp16 = add(x = x_361_cast_fp16, y = var_8993_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8996_cast_fp16 = abs(x = x_371_cast_fp16)[name = string("op_8996_cast_fp16")];
            bool var_8999 = const()[name = string("op_8999"), val = bool(true)];
            tensor<int32, [1]> var_9000_axes_0 = const()[name = string("op_9000_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9000_cast_fp16 = reduce_max(axes = var_9000_axes_0, keep_dims = var_8999, x = var_8996_cast_fp16)[name = string("op_9000_cast_fp16")];
            fp16 var_9001_to_fp16 = const()[name = string("op_9001_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_333_cast_fp16 = maximum(x = var_9000_cast_fp16, y = var_9001_to_fp16)[name = string("s_333_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_333_cast_fp16 = real_div(x = x_371_cast_fp16, y = s_333_cast_fp16)[name = string("y_333_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9004_cast_fp16 = mul(x = y_333_cast_fp16, y = y_333_cast_fp16)[name = string("op_9004_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9004_cast_fp16)[name = string("op_9009_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9010_cast_fp16 = mul(x = s_333_cast_fp16, y = s_333_cast_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9015_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9015_cast_fp16")];
            fp32 var_9016_epsilon_0 = const()[name = string("op_9016_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9016_cast_fp16 = rsqrt(epsilon = var_9016_epsilon_0, x = var_9015_cast_fp16)[name = string("op_9016_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_291_cast_fp16 = mul(x = y_333_cast_fp16, y = var_9016_cast_fp16)[name = string("out_291_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116419264)))];
            tensor<fp16, [1, 256, 512]> input_245_cast_fp16 = mul(x = out_291_cast_fp16, y = tm_layers_20_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116420352))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117468992))))[name = string("tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_185_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_185_cast_fp16")];
            string var_9023_mode_0 = const()[name = string("op_9023_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_9023_cast_fp16 = gelu(mode = var_9023_mode_0, x = linear_185_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117473152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118521792))))[name = string("tm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_186_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_247_cast_fp16 = mul(x = var_9023_cast_fp16, y = linear_186_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118525952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119574592))))[name = string("tm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9031_cast_fp16 = abs(x = linear_187_cast_fp16)[name = string("op_9031_cast_fp16")];
            bool var_9034 = const()[name = string("op_9034"), val = bool(true)];
            tensor<int32, [1]> var_9035_axes_0 = const()[name = string("op_9035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9035_cast_fp16 = reduce_max(axes = var_9035_axes_0, keep_dims = var_9034, x = var_9031_cast_fp16)[name = string("op_9035_cast_fp16")];
            fp16 var_9036_to_fp16 = const()[name = string("op_9036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_335_cast_fp16 = maximum(x = var_9035_cast_fp16, y = var_9036_to_fp16)[name = string("s_335_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_335_cast_fp16 = real_div(x = linear_187_cast_fp16, y = s_335_cast_fp16)[name = string("y_335_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9039_cast_fp16 = mul(x = y_335_cast_fp16, y = y_335_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [1]> var_9044_axes_0 = const()[name = string("op_9044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9044_keep_dims_0 = const()[name = string("op_9044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9044_cast_fp16 = reduce_mean(axes = var_9044_axes_0, keep_dims = var_9044_keep_dims_0, x = var_9039_cast_fp16)[name = string("op_9044_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9045_cast_fp16 = mul(x = s_335_cast_fp16, y = s_335_cast_fp16)[name = string("op_9045_cast_fp16")];
            fp32 var_9046_epsilon_0 = const()[name = string("op_9046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9046_cast_fp16 = inverse(epsilon = var_9046_epsilon_0, x = var_9045_cast_fp16)[name = string("op_9046_cast_fp16")];
            fp16 var_9047_to_fp16 = const()[name = string("op_9047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9048_cast_fp16 = mul(x = var_9046_cast_fp16, y = var_9047_to_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9050_cast_fp16 = add(x = var_9044_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9050_cast_fp16")];
            fp32 var_9051_epsilon_0 = const()[name = string("op_9051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9051_cast_fp16 = rsqrt(epsilon = var_9051_epsilon_0, x = var_9050_cast_fp16)[name = string("op_9051_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_293_cast_fp16 = mul(x = y_335_cast_fp16, y = var_9051_cast_fp16)[name = string("out_293_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119575680)))];
            tensor<fp16, [1, 256, 512]> var_9053_cast_fp16 = mul(x = out_293_cast_fp16, y = tm_layers_20_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9053_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_249_cast_fp16 = add(x = x_371_cast_fp16, y = var_9053_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119576768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119838976))))[name = string("tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_188_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_188_cast_fp16")];
            string x_375_mode_0 = const()[name = string("x_375_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_375_cast_fp16 = gelu(mode = x_375_mode_0, x = linear_188_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [4]> var_9073_begin_0 = const()[name = string("op_9073_begin_0"), val = tensor<int32, [4]>([0, 0, 20, 0])];
            tensor<int32, [4]> var_9073_end_0 = const()[name = string("op_9073_end_0"), val = tensor<int32, [4]>([1, 256, 21, 512])];
            tensor<bool, [4]> var_9073_end_mask_0 = const()[name = string("op_9073_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9073_squeeze_mask_0 = const()[name = string("op_9073_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_9073_cast_fp16 = slice_by_index(begin = var_9073_begin_0, end = var_9073_end_0, end_mask = var_9073_end_mask_0, squeeze_mask = var_9073_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_251_cast_fp16 = mul(x = x_375_cast_fp16, y = var_9073_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119840064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120102272))))[name = string("tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9083_cast_fp16 = abs(x = linear_189_cast_fp16)[name = string("op_9083_cast_fp16")];
            bool var_9086 = const()[name = string("op_9086"), val = bool(true)];
            tensor<int32, [1]> var_9087_axes_0 = const()[name = string("op_9087_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9087_cast_fp16 = reduce_max(axes = var_9087_axes_0, keep_dims = var_9086, x = var_9083_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_337_cast_fp16 = maximum(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("s_337_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_337_cast_fp16 = real_div(x = linear_189_cast_fp16, y = s_337_cast_fp16)[name = string("y_337_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9091_cast_fp16 = mul(x = y_337_cast_fp16, y = y_337_cast_fp16)[name = string("op_9091_cast_fp16")];
            tensor<int32, [1]> var_9096_axes_0 = const()[name = string("op_9096_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9096_keep_dims_0 = const()[name = string("op_9096_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9096_cast_fp16 = reduce_mean(axes = var_9096_axes_0, keep_dims = var_9096_keep_dims_0, x = var_9091_cast_fp16)[name = string("op_9096_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9097_cast_fp16 = mul(x = s_337_cast_fp16, y = s_337_cast_fp16)[name = string("op_9097_cast_fp16")];
            fp32 var_9098_epsilon_0 = const()[name = string("op_9098_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9098_cast_fp16 = inverse(epsilon = var_9098_epsilon_0, x = var_9097_cast_fp16)[name = string("op_9098_cast_fp16")];
            fp16 var_9099_to_fp16 = const()[name = string("op_9099_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9100_cast_fp16 = mul(x = var_9098_cast_fp16, y = var_9099_to_fp16)[name = string("op_9100_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9102_cast_fp16 = add(x = var_9096_cast_fp16, y = var_9100_cast_fp16)[name = string("op_9102_cast_fp16")];
            fp32 var_9103_epsilon_0 = const()[name = string("op_9103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9103_cast_fp16 = rsqrt(epsilon = var_9103_epsilon_0, x = var_9102_cast_fp16)[name = string("op_9103_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_295_cast_fp16 = mul(x = y_337_cast_fp16, y = var_9103_cast_fp16)[name = string("out_295_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120103360)))];
            tensor<fp16, [1, 256, 512]> var_9105_cast_fp16 = mul(x = out_295_cast_fp16, y = tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9105_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_41_cast_fp16 = add(x = input_249_cast_fp16, y = var_9105_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<fp16, [1]> tm_layers_20_layer_scalar_to_fp16 = const()[name = string("tm_layers_20_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ecp-1])];
            tensor<fp16, [1, 256, 512]> x_379_cast_fp16 = mul(x = h_41_cast_fp16, y = tm_layers_20_layer_scalar_to_fp16)[name = string("x_379_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9109_cast_fp16 = abs(x = x_379_cast_fp16)[name = string("op_9109_cast_fp16")];
            bool var_9112 = const()[name = string("op_9112"), val = bool(true)];
            tensor<int32, [1]> var_9113_axes_0 = const()[name = string("op_9113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9113_cast_fp16 = reduce_max(axes = var_9113_axes_0, keep_dims = var_9112, x = var_9109_cast_fp16)[name = string("op_9113_cast_fp16")];
            fp16 var_9114_to_fp16 = const()[name = string("op_9114_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_339_cast_fp16 = maximum(x = var_9113_cast_fp16, y = var_9114_to_fp16)[name = string("s_339_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_339_cast_fp16 = real_div(x = x_379_cast_fp16, y = s_339_cast_fp16)[name = string("y_339_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9117_cast_fp16 = mul(x = y_339_cast_fp16, y = y_339_cast_fp16)[name = string("op_9117_cast_fp16")];
            tensor<int32, [1]> var_9122_axes_0 = const()[name = string("op_9122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9122_keep_dims_0 = const()[name = string("op_9122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9122_cast_fp16 = reduce_mean(axes = var_9122_axes_0, keep_dims = var_9122_keep_dims_0, x = var_9117_cast_fp16)[name = string("op_9122_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9123_cast_fp16 = mul(x = s_339_cast_fp16, y = s_339_cast_fp16)[name = string("op_9123_cast_fp16")];
            fp32 var_9124_epsilon_0 = const()[name = string("op_9124_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9124_cast_fp16 = inverse(epsilon = var_9124_epsilon_0, x = var_9123_cast_fp16)[name = string("op_9124_cast_fp16")];
            fp16 var_9125_to_fp16 = const()[name = string("op_9125_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9126_cast_fp16 = mul(x = var_9124_cast_fp16, y = var_9125_to_fp16)[name = string("op_9126_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9128_cast_fp16 = add(x = var_9122_cast_fp16, y = var_9126_cast_fp16)[name = string("op_9128_cast_fp16")];
            fp32 var_9129_epsilon_0 = const()[name = string("op_9129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9129_cast_fp16 = rsqrt(epsilon = var_9129_epsilon_0, x = var_9128_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_297_cast_fp16 = mul(x = y_339_cast_fp16, y = var_9129_cast_fp16)[name = string("out_297_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120104448)))];
            tensor<fp16, [1, 256, 512]> input_253_cast_fp16 = mul(x = out_297_cast_fp16, y = tm_layers_21_input_layernorm_weight_to_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120105536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120629888))))[name = string("tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9139 = const()[name = string("op_9139"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_9140_cast_fp16 = reshape(shape = var_9139, x = linear_190_cast_fp16)[name = string("op_9140_cast_fp16")];
            tensor<int32, [4]> x_381_perm_0 = const()[name = string("x_381_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120632000))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120894208))))[name = string("tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9151 = const()[name = string("op_9151"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_9152_cast_fp16 = reshape(shape = var_9151, x = linear_191_cast_fp16)[name = string("op_9152_cast_fp16")];
            tensor<int32, [4]> x_383_perm_0 = const()[name = string("x_383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120895296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121157504))))[name = string("tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_9163 = const()[name = string("op_9163"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_9164_cast_fp16 = reshape(shape = var_9163, x = linear_192_cast_fp16)[name = string("op_9164_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_381_cast_fp16 = transpose(perm = x_381_perm_0, x = var_9140_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 4, 256, 256]> var_9168_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_9168_cast_fp16")];
            bool var_9171 = const()[name = string("op_9171"), val = bool(true)];
            tensor<int32, [1]> var_9172_axes_0 = const()[name = string("op_9172_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_9172_cast_fp16 = reduce_max(axes = var_9172_axes_0, keep_dims = var_9171, x = var_9168_cast_fp16)[name = string("op_9172_cast_fp16")];
            fp16 var_9173_to_fp16 = const()[name = string("op_9173_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_341_cast_fp16 = maximum(x = var_9172_cast_fp16, y = var_9173_to_fp16)[name = string("s_341_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_341_cast_fp16 = real_div(x = x_381_cast_fp16, y = s_341_cast_fp16)[name = string("y_341_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9176_cast_fp16 = mul(x = y_341_cast_fp16, y = y_341_cast_fp16)[name = string("op_9176_cast_fp16")];
            tensor<int32, [1]> var_9181_axes_0 = const()[name = string("op_9181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9181_keep_dims_0 = const()[name = string("op_9181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_9181_cast_fp16 = reduce_mean(axes = var_9181_axes_0, keep_dims = var_9181_keep_dims_0, x = var_9176_cast_fp16)[name = string("op_9181_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9182_cast_fp16 = mul(x = s_341_cast_fp16, y = s_341_cast_fp16)[name = string("op_9182_cast_fp16")];
            fp32 var_9183_epsilon_0 = const()[name = string("op_9183_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_9183_cast_fp16 = inverse(epsilon = var_9183_epsilon_0, x = var_9182_cast_fp16)[name = string("op_9183_cast_fp16")];
            fp16 var_9184_to_fp16 = const()[name = string("op_9184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_9185_cast_fp16 = mul(x = var_9183_cast_fp16, y = var_9184_to_fp16)[name = string("op_9185_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9187_cast_fp16 = add(x = var_9181_cast_fp16, y = var_9185_cast_fp16)[name = string("op_9187_cast_fp16")];
            fp32 var_9188_epsilon_0 = const()[name = string("op_9188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_9188_cast_fp16 = rsqrt(epsilon = var_9188_epsilon_0, x = var_9187_cast_fp16)[name = string("op_9188_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_299_cast_fp16 = mul(x = y_341_cast_fp16, y = var_9188_cast_fp16)[name = string("out_299_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_85_cast_fp16 = mul(x = out_299_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_85_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_383_cast_fp16 = transpose(perm = x_383_perm_0, x = var_9152_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2, 256, 256]> var_9191_cast_fp16 = abs(x = x_383_cast_fp16)[name = string("op_9191_cast_fp16")];
            bool var_9194 = const()[name = string("op_9194"), val = bool(true)];
            tensor<int32, [1]> var_9195_axes_0 = const()[name = string("op_9195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_9195_cast_fp16 = reduce_max(axes = var_9195_axes_0, keep_dims = var_9194, x = var_9191_cast_fp16)[name = string("op_9195_cast_fp16")];
            fp16 var_9196_to_fp16 = const()[name = string("op_9196_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_343_cast_fp16 = maximum(x = var_9195_cast_fp16, y = var_9196_to_fp16)[name = string("s_343_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_343_cast_fp16 = real_div(x = x_383_cast_fp16, y = s_343_cast_fp16)[name = string("y_343_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9199_cast_fp16 = mul(x = y_343_cast_fp16, y = y_343_cast_fp16)[name = string("op_9199_cast_fp16")];
            tensor<int32, [1]> var_9204_axes_0 = const()[name = string("op_9204_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9204_keep_dims_0 = const()[name = string("op_9204_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_9204_cast_fp16 = reduce_mean(axes = var_9204_axes_0, keep_dims = var_9204_keep_dims_0, x = var_9199_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9205_cast_fp16 = mul(x = s_343_cast_fp16, y = s_343_cast_fp16)[name = string("op_9205_cast_fp16")];
            fp32 var_9206_epsilon_0 = const()[name = string("op_9206_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_9206_cast_fp16 = inverse(epsilon = var_9206_epsilon_0, x = var_9205_cast_fp16)[name = string("op_9206_cast_fp16")];
            fp16 var_9207_to_fp16 = const()[name = string("op_9207_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_9208_cast_fp16 = mul(x = var_9206_cast_fp16, y = var_9207_to_fp16)[name = string("op_9208_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9210_cast_fp16 = add(x = var_9204_cast_fp16, y = var_9208_cast_fp16)[name = string("op_9210_cast_fp16")];
            fp32 var_9211_epsilon_0 = const()[name = string("op_9211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_9211_cast_fp16 = rsqrt(epsilon = var_9211_epsilon_0, x = var_9210_cast_fp16)[name = string("op_9211_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_301_cast_fp16 = mul(x = y_343_cast_fp16, y = var_9211_cast_fp16)[name = string("out_301_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_127_cast_fp16 = mul(x = out_301_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_127_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_9164_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2, 256, 256]> var_9214_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_9214_cast_fp16")];
            bool var_9217 = const()[name = string("op_9217"), val = bool(true)];
            tensor<int32, [1]> var_9218_axes_0 = const()[name = string("op_9218_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_9218_cast_fp16 = reduce_max(axes = var_9218_axes_0, keep_dims = var_9217, x = var_9214_cast_fp16)[name = string("op_9218_cast_fp16")];
            fp16 var_9219_to_fp16 = const()[name = string("op_9219_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_345_cast_fp16 = maximum(x = var_9218_cast_fp16, y = var_9219_to_fp16)[name = string("s_345_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_345_cast_fp16 = real_div(x = x_385_cast_fp16, y = s_345_cast_fp16)[name = string("y_345_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9222_cast_fp16 = mul(x = y_345_cast_fp16, y = y_345_cast_fp16)[name = string("op_9222_cast_fp16")];
            tensor<int32, [1]> var_9227_axes_0 = const()[name = string("op_9227_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9227_keep_dims_0 = const()[name = string("op_9227_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_9227_cast_fp16 = reduce_mean(axes = var_9227_axes_0, keep_dims = var_9227_keep_dims_0, x = var_9222_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9228_cast_fp16 = mul(x = s_345_cast_fp16, y = s_345_cast_fp16)[name = string("op_9228_cast_fp16")];
            fp32 var_9229_epsilon_0 = const()[name = string("op_9229_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_9229_cast_fp16 = inverse(epsilon = var_9229_epsilon_0, x = var_9228_cast_fp16)[name = string("op_9229_cast_fp16")];
            fp16 var_9230_to_fp16 = const()[name = string("op_9230_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_9231_cast_fp16 = mul(x = var_9229_cast_fp16, y = var_9230_to_fp16)[name = string("op_9231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9233_cast_fp16 = add(x = var_9227_cast_fp16, y = var_9231_cast_fp16)[name = string("op_9233_cast_fp16")];
            fp32 var_9234_epsilon_0 = const()[name = string("op_9234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_9234_cast_fp16 = rsqrt(epsilon = var_9234_epsilon_0, x = var_9233_cast_fp16)[name = string("op_9234_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_85_cast_fp16 = mul(x = y_345_cast_fp16, y = var_9234_cast_fp16)[name = string("v_85_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9236_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9236_cast_fp16")];
            tensor<int32, [4]> var_9241_begin_0 = const()[name = string("op_9241_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9241_end_0 = const()[name = string("op_9241_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_9241_end_mask_0 = const()[name = string("op_9241_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_9241_cast_fp16 = slice_by_index(begin = var_9241_begin_0, end = var_9241_end_0, end_mask = var_9241_end_mask_0, x = q_85_cast_fp16)[name = string("op_9241_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_9242_cast_fp16 = mul(x = var_9241_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_9242_cast_fp16")];
            tensor<int32, [4]> var_9247_begin_0 = const()[name = string("op_9247_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9247_end_0 = const()[name = string("op_9247_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_9247_end_mask_0 = const()[name = string("op_9247_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_9247_cast_fp16 = slice_by_index(begin = var_9247_begin_0, end = var_9247_end_0, end_mask = var_9247_end_mask_0, x = q_85_cast_fp16)[name = string("op_9247_cast_fp16")];
            int32 var_9249 = const()[name = string("op_9249"), val = int32(-1)];
            bool var_9250_interleave_0 = const()[name = string("op_9250_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_9250_cast_fp16 = concat(axis = var_9249, interleave = var_9250_interleave_0, values = (var_9242_cast_fp16, var_9247_cast_fp16))[name = string("op_9250_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9251_cast_fp16 = mul(x = var_9250_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9251_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_87_cast_fp16 = add(x = var_9236_cast_fp16, y = var_9251_cast_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9254_cast_fp16 = mul(x = k_127_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9254_cast_fp16")];
            tensor<int32, [4]> var_9259_begin_0 = const()[name = string("op_9259_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9259_end_0 = const()[name = string("op_9259_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_9259_end_mask_0 = const()[name = string("op_9259_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_9259_cast_fp16 = slice_by_index(begin = var_9259_begin_0, end = var_9259_end_0, end_mask = var_9259_end_mask_0, x = k_127_cast_fp16)[name = string("op_9259_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_9260_cast_fp16 = mul(x = var_9259_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_9260_cast_fp16")];
            tensor<int32, [4]> var_9265_begin_0 = const()[name = string("op_9265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9265_end_0 = const()[name = string("op_9265_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_9265_end_mask_0 = const()[name = string("op_9265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_9265_cast_fp16 = slice_by_index(begin = var_9265_begin_0, end = var_9265_end_0, end_mask = var_9265_end_mask_0, x = k_127_cast_fp16)[name = string("op_9265_cast_fp16")];
            int32 var_9267 = const()[name = string("op_9267"), val = int32(-1)];
            bool var_9268_interleave_0 = const()[name = string("op_9268_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_9268_cast_fp16 = concat(axis = var_9267, interleave = var_9268_interleave_0, values = (var_9260_cast_fp16, var_9265_cast_fp16))[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9269_cast_fp16 = mul(x = var_9268_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_129_cast_fp16 = add(x = var_9254_cast_fp16, y = var_9269_cast_fp16)[name = string("k_129_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_42_reps_0 = const()[name = string("tile_42_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = k_129_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 1, 256, 256]> tile_42_cast_fp16 = tile(reps = tile_42_reps_0, x = transpose_84_cast_fp16)[name = string("tile_42_cast_fp16")];
            tensor<int32, [5]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_84_cast_fp16 = reshape(shape = concat_84, x = tile_42_cast_fp16)[name = string("reshape_84_cast_fp16")];
            tensor<int32, [5]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = reshape_84_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 1, 256, 256]> reshape_85_cast_fp16 = reshape(shape = concat_85, x = transpose_85_cast_fp16)[name = string("reshape_85_cast_fp16")];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_43_reps_0 = const()[name = string("tile_43_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_86_cast_fp16 = transpose(perm = transpose_86_perm_0, x = v_85_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 1, 256, 256]> tile_43_cast_fp16 = tile(reps = tile_43_reps_0, x = transpose_86_cast_fp16)[name = string("tile_43_cast_fp16")];
            tensor<int32, [5]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_86_cast_fp16 = reshape(shape = concat_86, x = tile_43_cast_fp16)[name = string("reshape_86_cast_fp16")];
            tensor<int32, [5]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_87_cast_fp16 = transpose(perm = transpose_87_perm_0, x = reshape_86_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 1, 256, 256]> reshape_87_cast_fp16 = reshape(shape = concat_87, x = transpose_87_cast_fp16)[name = string("reshape_87_cast_fp16")];
            tensor<int32, [4]> v_87_perm_0 = const()[name = string("v_87_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9283_transpose_x_0 = const()[name = string("op_9283_transpose_x_0"), val = bool(false)];
            bool var_9283_transpose_y_0 = const()[name = string("op_9283_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_117_cast_fp16 = transpose(perm = transpose_117_perm_0, x = reshape_85_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 4, 256, 256]> var_9283_cast_fp16 = matmul(transpose_x = var_9283_transpose_x_0, transpose_y = var_9283_transpose_y_0, x = q_87_cast_fp16, y = transpose_117_cast_fp16)[name = string("op_9283_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9285_cast_fp16 = add(x = var_9283_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9285_cast_fp16")];
            int32 var_9286 = const()[name = string("op_9286"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_345_cast_fp16 = softmax(axis = var_9286, x = var_9285_cast_fp16)[name = string("w_345_cast_fp16")];
            bool var_9289_transpose_x_0 = const()[name = string("op_9289_transpose_x_0"), val = bool(false)];
            bool var_9289_transpose_y_0 = const()[name = string("op_9289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_87_cast_fp16 = transpose(perm = v_87_perm_0, x = reshape_87_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 4, 256, 256]> var_9289_cast_fp16 = matmul(transpose_x = var_9289_transpose_x_0, transpose_y = var_9289_transpose_y_0, x = w_345_cast_fp16, y = v_87_cast_fp16)[name = string("op_9289_cast_fp16")];
            tensor<int32, [4]> var_9292_perm_0 = const()[name = string("op_9292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9296 = const()[name = string("op_9296"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_9292_cast_fp16 = transpose(perm = var_9292_perm_0, x = var_9289_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 256, 1024]> input_255_cast_fp16 = reshape(shape = var_9296, x = var_9292_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121158592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121682944))))[name = string("tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9301_cast_fp16 = abs(x = linear_193_cast_fp16)[name = string("op_9301_cast_fp16")];
            bool var_9304 = const()[name = string("op_9304"), val = bool(true)];
            tensor<int32, [1]> var_9305_axes_0 = const()[name = string("op_9305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9305_cast_fp16 = reduce_max(axes = var_9305_axes_0, keep_dims = var_9304, x = var_9301_cast_fp16)[name = string("op_9305_cast_fp16")];
            fp16 var_9306_to_fp16 = const()[name = string("op_9306_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_347_cast_fp16 = maximum(x = var_9305_cast_fp16, y = var_9306_to_fp16)[name = string("s_347_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_347_cast_fp16 = real_div(x = linear_193_cast_fp16, y = s_347_cast_fp16)[name = string("y_347_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9309_cast_fp16 = mul(x = y_347_cast_fp16, y = y_347_cast_fp16)[name = string("op_9309_cast_fp16")];
            tensor<int32, [1]> var_9314_axes_0 = const()[name = string("op_9314_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9314_keep_dims_0 = const()[name = string("op_9314_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9314_cast_fp16 = reduce_mean(axes = var_9314_axes_0, keep_dims = var_9314_keep_dims_0, x = var_9309_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9315_cast_fp16 = mul(x = s_347_cast_fp16, y = s_347_cast_fp16)[name = string("op_9315_cast_fp16")];
            fp32 var_9316_epsilon_0 = const()[name = string("op_9316_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9316_cast_fp16 = inverse(epsilon = var_9316_epsilon_0, x = var_9315_cast_fp16)[name = string("op_9316_cast_fp16")];
            fp16 var_9317_to_fp16 = const()[name = string("op_9317_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9318_cast_fp16 = mul(x = var_9316_cast_fp16, y = var_9317_to_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9320_cast_fp16 = add(x = var_9314_cast_fp16, y = var_9318_cast_fp16)[name = string("op_9320_cast_fp16")];
            fp32 var_9321_epsilon_0 = const()[name = string("op_9321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9321_cast_fp16 = rsqrt(epsilon = var_9321_epsilon_0, x = var_9320_cast_fp16)[name = string("op_9321_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_303_cast_fp16 = mul(x = y_347_cast_fp16, y = var_9321_cast_fp16)[name = string("out_303_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121684032)))];
            tensor<fp16, [1, 256, 512]> var_9323_cast_fp16 = mul(x = out_303_cast_fp16, y = tm_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("op_9323_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_389_cast_fp16 = add(x = x_379_cast_fp16, y = var_9323_cast_fp16)[name = string("x_389_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9326_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_9326_cast_fp16")];
            bool var_9329 = const()[name = string("op_9329"), val = bool(true)];
            tensor<int32, [1]> var_9330_axes_0 = const()[name = string("op_9330_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9330_cast_fp16 = reduce_max(axes = var_9330_axes_0, keep_dims = var_9329, x = var_9326_cast_fp16)[name = string("op_9330_cast_fp16")];
            fp16 var_9331_to_fp16 = const()[name = string("op_9331_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_349_cast_fp16 = maximum(x = var_9330_cast_fp16, y = var_9331_to_fp16)[name = string("s_349_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_349_cast_fp16 = real_div(x = x_389_cast_fp16, y = s_349_cast_fp16)[name = string("y_349_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9334_cast_fp16 = mul(x = y_349_cast_fp16, y = y_349_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [1]> var_9339_axes_0 = const()[name = string("op_9339_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9339_keep_dims_0 = const()[name = string("op_9339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9339_cast_fp16 = reduce_mean(axes = var_9339_axes_0, keep_dims = var_9339_keep_dims_0, x = var_9334_cast_fp16)[name = string("op_9339_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9340_cast_fp16 = mul(x = s_349_cast_fp16, y = s_349_cast_fp16)[name = string("op_9340_cast_fp16")];
            fp32 var_9341_epsilon_0 = const()[name = string("op_9341_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9341_cast_fp16 = inverse(epsilon = var_9341_epsilon_0, x = var_9340_cast_fp16)[name = string("op_9341_cast_fp16")];
            fp16 var_9342_to_fp16 = const()[name = string("op_9342_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9343_cast_fp16 = mul(x = var_9341_cast_fp16, y = var_9342_to_fp16)[name = string("op_9343_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9345_cast_fp16 = add(x = var_9339_cast_fp16, y = var_9343_cast_fp16)[name = string("op_9345_cast_fp16")];
            fp32 var_9346_epsilon_0 = const()[name = string("op_9346_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9346_cast_fp16 = rsqrt(epsilon = var_9346_epsilon_0, x = var_9345_cast_fp16)[name = string("op_9346_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_305_cast_fp16 = mul(x = y_349_cast_fp16, y = var_9346_cast_fp16)[name = string("out_305_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121685120)))];
            tensor<fp16, [1, 256, 512]> input_257_cast_fp16 = mul(x = out_305_cast_fp16, y = tm_layers_21_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121686208))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122734848))))[name = string("tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_194_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_194_cast_fp16")];
            string var_9353_mode_0 = const()[name = string("op_9353_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_9353_cast_fp16 = gelu(mode = var_9353_mode_0, x = linear_194_cast_fp16)[name = string("op_9353_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122739008))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123787648))))[name = string("tm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_195_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_259_cast_fp16 = mul(x = var_9353_cast_fp16, y = linear_195_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123791808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124840448))))[name = string("tm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9361_cast_fp16 = abs(x = linear_196_cast_fp16)[name = string("op_9361_cast_fp16")];
            bool var_9364 = const()[name = string("op_9364"), val = bool(true)];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9365_cast_fp16 = reduce_max(axes = var_9365_axes_0, keep_dims = var_9364, x = var_9361_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9366_to_fp16 = const()[name = string("op_9366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_351_cast_fp16 = maximum(x = var_9365_cast_fp16, y = var_9366_to_fp16)[name = string("s_351_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_351_cast_fp16 = real_div(x = linear_196_cast_fp16, y = s_351_cast_fp16)[name = string("y_351_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9369_cast_fp16 = mul(x = y_351_cast_fp16, y = y_351_cast_fp16)[name = string("op_9369_cast_fp16")];
            tensor<int32, [1]> var_9374_axes_0 = const()[name = string("op_9374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9374_keep_dims_0 = const()[name = string("op_9374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9374_cast_fp16 = reduce_mean(axes = var_9374_axes_0, keep_dims = var_9374_keep_dims_0, x = var_9369_cast_fp16)[name = string("op_9374_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9375_cast_fp16 = mul(x = s_351_cast_fp16, y = s_351_cast_fp16)[name = string("op_9375_cast_fp16")];
            fp32 var_9376_epsilon_0 = const()[name = string("op_9376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9376_cast_fp16 = inverse(epsilon = var_9376_epsilon_0, x = var_9375_cast_fp16)[name = string("op_9376_cast_fp16")];
            fp16 var_9377_to_fp16 = const()[name = string("op_9377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9378_cast_fp16 = mul(x = var_9376_cast_fp16, y = var_9377_to_fp16)[name = string("op_9378_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9380_cast_fp16 = add(x = var_9374_cast_fp16, y = var_9378_cast_fp16)[name = string("op_9380_cast_fp16")];
            fp32 var_9381_epsilon_0 = const()[name = string("op_9381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9381_cast_fp16 = rsqrt(epsilon = var_9381_epsilon_0, x = var_9380_cast_fp16)[name = string("op_9381_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_307_cast_fp16 = mul(x = y_351_cast_fp16, y = var_9381_cast_fp16)[name = string("out_307_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124841536)))];
            tensor<fp16, [1, 256, 512]> var_9383_cast_fp16 = mul(x = out_307_cast_fp16, y = tm_layers_21_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9383_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_261_cast_fp16 = add(x = x_389_cast_fp16, y = var_9383_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124842624))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125104832))))[name = string("tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_197_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_197_cast_fp16")];
            string x_393_mode_0 = const()[name = string("x_393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_393_cast_fp16 = gelu(mode = x_393_mode_0, x = linear_197_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<int32, [4]> var_9403_begin_0 = const()[name = string("op_9403_begin_0"), val = tensor<int32, [4]>([0, 0, 21, 0])];
            tensor<int32, [4]> var_9403_end_0 = const()[name = string("op_9403_end_0"), val = tensor<int32, [4]>([1, 256, 22, 512])];
            tensor<bool, [4]> var_9403_end_mask_0 = const()[name = string("op_9403_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9403_squeeze_mask_0 = const()[name = string("op_9403_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_9403_cast_fp16 = slice_by_index(begin = var_9403_begin_0, end = var_9403_end_0, end_mask = var_9403_end_mask_0, squeeze_mask = var_9403_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9403_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_263_cast_fp16 = mul(x = x_393_cast_fp16, y = var_9403_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125105920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125368128))))[name = string("tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9413_cast_fp16 = abs(x = linear_198_cast_fp16)[name = string("op_9413_cast_fp16")];
            bool var_9416 = const()[name = string("op_9416"), val = bool(true)];
            tensor<int32, [1]> var_9417_axes_0 = const()[name = string("op_9417_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9417_cast_fp16 = reduce_max(axes = var_9417_axes_0, keep_dims = var_9416, x = var_9413_cast_fp16)[name = string("op_9417_cast_fp16")];
            fp16 var_9418_to_fp16 = const()[name = string("op_9418_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_353_cast_fp16 = maximum(x = var_9417_cast_fp16, y = var_9418_to_fp16)[name = string("s_353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_353_cast_fp16 = real_div(x = linear_198_cast_fp16, y = s_353_cast_fp16)[name = string("y_353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9421_cast_fp16 = mul(x = y_353_cast_fp16, y = y_353_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<int32, [1]> var_9426_axes_0 = const()[name = string("op_9426_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9426_keep_dims_0 = const()[name = string("op_9426_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9426_cast_fp16 = reduce_mean(axes = var_9426_axes_0, keep_dims = var_9426_keep_dims_0, x = var_9421_cast_fp16)[name = string("op_9426_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9427_cast_fp16 = mul(x = s_353_cast_fp16, y = s_353_cast_fp16)[name = string("op_9427_cast_fp16")];
            fp32 var_9428_epsilon_0 = const()[name = string("op_9428_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9428_cast_fp16 = inverse(epsilon = var_9428_epsilon_0, x = var_9427_cast_fp16)[name = string("op_9428_cast_fp16")];
            fp16 var_9429_to_fp16 = const()[name = string("op_9429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9430_cast_fp16 = mul(x = var_9428_cast_fp16, y = var_9429_to_fp16)[name = string("op_9430_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9432_cast_fp16 = add(x = var_9426_cast_fp16, y = var_9430_cast_fp16)[name = string("op_9432_cast_fp16")];
            fp32 var_9433_epsilon_0 = const()[name = string("op_9433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9433_cast_fp16 = rsqrt(epsilon = var_9433_epsilon_0, x = var_9432_cast_fp16)[name = string("op_9433_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_309_cast_fp16 = mul(x = y_353_cast_fp16, y = var_9433_cast_fp16)[name = string("out_309_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125369216)))];
            tensor<fp16, [1, 256, 512]> var_9435_cast_fp16 = mul(x = out_309_cast_fp16, y = tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9435_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_43_cast_fp16 = add(x = input_261_cast_fp16, y = var_9435_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<fp16, [1]> tm_layers_21_layer_scalar_to_fp16 = const()[name = string("tm_layers_21_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e8p-1])];
            tensor<fp16, [1, 256, 512]> x_397_cast_fp16 = mul(x = h_43_cast_fp16, y = tm_layers_21_layer_scalar_to_fp16)[name = string("x_397_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9439_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_9439_cast_fp16")];
            bool var_9442 = const()[name = string("op_9442"), val = bool(true)];
            tensor<int32, [1]> var_9443_axes_0 = const()[name = string("op_9443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9443_cast_fp16 = reduce_max(axes = var_9443_axes_0, keep_dims = var_9442, x = var_9439_cast_fp16)[name = string("op_9443_cast_fp16")];
            fp16 var_9444_to_fp16 = const()[name = string("op_9444_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_355_cast_fp16 = maximum(x = var_9443_cast_fp16, y = var_9444_to_fp16)[name = string("s_355_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_355_cast_fp16 = real_div(x = x_397_cast_fp16, y = s_355_cast_fp16)[name = string("y_355_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9447_cast_fp16 = mul(x = y_355_cast_fp16, y = y_355_cast_fp16)[name = string("op_9447_cast_fp16")];
            tensor<int32, [1]> var_9452_axes_0 = const()[name = string("op_9452_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9452_keep_dims_0 = const()[name = string("op_9452_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9452_cast_fp16 = reduce_mean(axes = var_9452_axes_0, keep_dims = var_9452_keep_dims_0, x = var_9447_cast_fp16)[name = string("op_9452_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9453_cast_fp16 = mul(x = s_355_cast_fp16, y = s_355_cast_fp16)[name = string("op_9453_cast_fp16")];
            fp32 var_9454_epsilon_0 = const()[name = string("op_9454_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9454_cast_fp16 = inverse(epsilon = var_9454_epsilon_0, x = var_9453_cast_fp16)[name = string("op_9454_cast_fp16")];
            fp16 var_9455_to_fp16 = const()[name = string("op_9455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9456_cast_fp16 = mul(x = var_9454_cast_fp16, y = var_9455_to_fp16)[name = string("op_9456_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9458_cast_fp16 = add(x = var_9452_cast_fp16, y = var_9456_cast_fp16)[name = string("op_9458_cast_fp16")];
            fp32 var_9459_epsilon_0 = const()[name = string("op_9459_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9459_cast_fp16 = rsqrt(epsilon = var_9459_epsilon_0, x = var_9458_cast_fp16)[name = string("op_9459_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_311_cast_fp16 = mul(x = y_355_cast_fp16, y = var_9459_cast_fp16)[name = string("out_311_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125370304)))];
            tensor<fp16, [1, 256, 512]> input_265_cast_fp16 = mul(x = out_311_cast_fp16, y = tm_layers_22_input_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125371392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125895744))))[name = string("tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<int32, [4]> var_9469 = const()[name = string("op_9469"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_9470_cast_fp16 = reshape(shape = var_9469, x = linear_199_cast_fp16)[name = string("op_9470_cast_fp16")];
            tensor<int32, [4]> x_399_perm_0 = const()[name = string("x_399_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125897856))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126160064))))[name = string("tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<int32, [4]> var_9481 = const()[name = string("op_9481"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_9482_cast_fp16 = reshape(shape = var_9481, x = linear_200_cast_fp16)[name = string("op_9482_cast_fp16")];
            tensor<int32, [4]> x_401_perm_0 = const()[name = string("x_401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126161152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126423360))))[name = string("tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<int32, [4]> var_9493 = const()[name = string("op_9493"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_9494_cast_fp16 = reshape(shape = var_9493, x = linear_201_cast_fp16)[name = string("op_9494_cast_fp16")];
            tensor<int32, [4]> x_403_perm_0 = const()[name = string("x_403_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_399_cast_fp16 = transpose(perm = x_399_perm_0, x = var_9470_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 4, 256, 256]> var_9498_cast_fp16 = abs(x = x_399_cast_fp16)[name = string("op_9498_cast_fp16")];
            bool var_9501 = const()[name = string("op_9501"), val = bool(true)];
            tensor<int32, [1]> var_9502_axes_0 = const()[name = string("op_9502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_9502_cast_fp16 = reduce_max(axes = var_9502_axes_0, keep_dims = var_9501, x = var_9498_cast_fp16)[name = string("op_9502_cast_fp16")];
            fp16 var_9503_to_fp16 = const()[name = string("op_9503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_357_cast_fp16 = maximum(x = var_9502_cast_fp16, y = var_9503_to_fp16)[name = string("s_357_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_357_cast_fp16 = real_div(x = x_399_cast_fp16, y = s_357_cast_fp16)[name = string("y_357_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9506_cast_fp16 = mul(x = y_357_cast_fp16, y = y_357_cast_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9511_axes_0 = const()[name = string("op_9511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9511_keep_dims_0 = const()[name = string("op_9511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_9511_cast_fp16 = reduce_mean(axes = var_9511_axes_0, keep_dims = var_9511_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9512_cast_fp16 = mul(x = s_357_cast_fp16, y = s_357_cast_fp16)[name = string("op_9512_cast_fp16")];
            fp32 var_9513_epsilon_0 = const()[name = string("op_9513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_9513_cast_fp16 = inverse(epsilon = var_9513_epsilon_0, x = var_9512_cast_fp16)[name = string("op_9513_cast_fp16")];
            fp16 var_9514_to_fp16 = const()[name = string("op_9514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_9515_cast_fp16 = mul(x = var_9513_cast_fp16, y = var_9514_to_fp16)[name = string("op_9515_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9517_cast_fp16 = add(x = var_9511_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9517_cast_fp16")];
            fp32 var_9518_epsilon_0 = const()[name = string("op_9518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_9518_cast_fp16 = rsqrt(epsilon = var_9518_epsilon_0, x = var_9517_cast_fp16)[name = string("op_9518_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_313_cast_fp16 = mul(x = y_357_cast_fp16, y = var_9518_cast_fp16)[name = string("out_313_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_89_cast_fp16 = mul(x = out_313_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_89_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_401_cast_fp16 = transpose(perm = x_401_perm_0, x = var_9482_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 256, 256]> var_9521_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_9521_cast_fp16")];
            bool var_9524 = const()[name = string("op_9524"), val = bool(true)];
            tensor<int32, [1]> var_9525_axes_0 = const()[name = string("op_9525_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_9525_cast_fp16 = reduce_max(axes = var_9525_axes_0, keep_dims = var_9524, x = var_9521_cast_fp16)[name = string("op_9525_cast_fp16")];
            fp16 var_9526_to_fp16 = const()[name = string("op_9526_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_359_cast_fp16 = maximum(x = var_9525_cast_fp16, y = var_9526_to_fp16)[name = string("s_359_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_359_cast_fp16 = real_div(x = x_401_cast_fp16, y = s_359_cast_fp16)[name = string("y_359_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9529_cast_fp16 = mul(x = y_359_cast_fp16, y = y_359_cast_fp16)[name = string("op_9529_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9529_cast_fp16)[name = string("op_9534_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9535_cast_fp16 = mul(x = s_359_cast_fp16, y = s_359_cast_fp16)[name = string("op_9535_cast_fp16")];
            fp32 var_9536_epsilon_0 = const()[name = string("op_9536_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_9536_cast_fp16 = inverse(epsilon = var_9536_epsilon_0, x = var_9535_cast_fp16)[name = string("op_9536_cast_fp16")];
            fp16 var_9537_to_fp16 = const()[name = string("op_9537_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_9538_cast_fp16 = mul(x = var_9536_cast_fp16, y = var_9537_to_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9540_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9538_cast_fp16)[name = string("op_9540_cast_fp16")];
            fp32 var_9541_epsilon_0 = const()[name = string("op_9541_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_9541_cast_fp16 = rsqrt(epsilon = var_9541_epsilon_0, x = var_9540_cast_fp16)[name = string("op_9541_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_315_cast_fp16 = mul(x = y_359_cast_fp16, y = var_9541_cast_fp16)[name = string("out_315_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_133_cast_fp16 = mul(x = out_315_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_133_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_403_cast_fp16 = transpose(perm = x_403_perm_0, x = var_9494_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 256, 256]> var_9544_cast_fp16 = abs(x = x_403_cast_fp16)[name = string("op_9544_cast_fp16")];
            bool var_9547 = const()[name = string("op_9547"), val = bool(true)];
            tensor<int32, [1]> var_9548_axes_0 = const()[name = string("op_9548_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_9548_cast_fp16 = reduce_max(axes = var_9548_axes_0, keep_dims = var_9547, x = var_9544_cast_fp16)[name = string("op_9548_cast_fp16")];
            fp16 var_9549_to_fp16 = const()[name = string("op_9549_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_361_cast_fp16 = maximum(x = var_9548_cast_fp16, y = var_9549_to_fp16)[name = string("s_361_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_361_cast_fp16 = real_div(x = x_403_cast_fp16, y = s_361_cast_fp16)[name = string("y_361_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9552_cast_fp16 = mul(x = y_361_cast_fp16, y = y_361_cast_fp16)[name = string("op_9552_cast_fp16")];
            tensor<int32, [1]> var_9557_axes_0 = const()[name = string("op_9557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9557_keep_dims_0 = const()[name = string("op_9557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_9557_cast_fp16 = reduce_mean(axes = var_9557_axes_0, keep_dims = var_9557_keep_dims_0, x = var_9552_cast_fp16)[name = string("op_9557_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9558_cast_fp16 = mul(x = s_361_cast_fp16, y = s_361_cast_fp16)[name = string("op_9558_cast_fp16")];
            fp32 var_9559_epsilon_0 = const()[name = string("op_9559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_9559_cast_fp16 = inverse(epsilon = var_9559_epsilon_0, x = var_9558_cast_fp16)[name = string("op_9559_cast_fp16")];
            fp16 var_9560_to_fp16 = const()[name = string("op_9560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_9561_cast_fp16 = mul(x = var_9559_cast_fp16, y = var_9560_to_fp16)[name = string("op_9561_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9563_cast_fp16 = add(x = var_9557_cast_fp16, y = var_9561_cast_fp16)[name = string("op_9563_cast_fp16")];
            fp32 var_9564_epsilon_0 = const()[name = string("op_9564_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_9564_cast_fp16 = rsqrt(epsilon = var_9564_epsilon_0, x = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_89_cast_fp16 = mul(x = y_361_cast_fp16, y = var_9564_cast_fp16)[name = string("v_89_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9566_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9566_cast_fp16")];
            tensor<int32, [4]> var_9571_begin_0 = const()[name = string("op_9571_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9571_end_0 = const()[name = string("op_9571_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_9571_end_mask_0 = const()[name = string("op_9571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_9571_cast_fp16 = slice_by_index(begin = var_9571_begin_0, end = var_9571_end_0, end_mask = var_9571_end_mask_0, x = q_89_cast_fp16)[name = string("op_9571_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_9572_cast_fp16 = mul(x = var_9571_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_9572_cast_fp16")];
            tensor<int32, [4]> var_9577_begin_0 = const()[name = string("op_9577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9577_end_0 = const()[name = string("op_9577_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_9577_end_mask_0 = const()[name = string("op_9577_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_9577_cast_fp16 = slice_by_index(begin = var_9577_begin_0, end = var_9577_end_0, end_mask = var_9577_end_mask_0, x = q_89_cast_fp16)[name = string("op_9577_cast_fp16")];
            int32 var_9579 = const()[name = string("op_9579"), val = int32(-1)];
            bool var_9580_interleave_0 = const()[name = string("op_9580_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_9580_cast_fp16 = concat(axis = var_9579, interleave = var_9580_interleave_0, values = (var_9572_cast_fp16, var_9577_cast_fp16))[name = string("op_9580_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9581_cast_fp16 = mul(x = var_9580_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9581_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_91_cast_fp16 = add(x = var_9566_cast_fp16, y = var_9581_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9584_cast_fp16 = mul(x = k_133_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9584_cast_fp16")];
            tensor<int32, [4]> var_9589_begin_0 = const()[name = string("op_9589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9589_end_0 = const()[name = string("op_9589_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_9589_end_mask_0 = const()[name = string("op_9589_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_9589_cast_fp16 = slice_by_index(begin = var_9589_begin_0, end = var_9589_end_0, end_mask = var_9589_end_mask_0, x = k_133_cast_fp16)[name = string("op_9589_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_9590_cast_fp16 = mul(x = var_9589_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_9590_cast_fp16")];
            tensor<int32, [4]> var_9595_begin_0 = const()[name = string("op_9595_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9595_end_0 = const()[name = string("op_9595_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_9595_end_mask_0 = const()[name = string("op_9595_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_9595_cast_fp16 = slice_by_index(begin = var_9595_begin_0, end = var_9595_end_0, end_mask = var_9595_end_mask_0, x = k_133_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool var_9598_interleave_0 = const()[name = string("op_9598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_9598_cast_fp16 = concat(axis = var_9597, interleave = var_9598_interleave_0, values = (var_9590_cast_fp16, var_9595_cast_fp16))[name = string("op_9598_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9599_cast_fp16 = mul(x = var_9598_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9599_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_135_cast_fp16 = add(x = var_9584_cast_fp16, y = var_9599_cast_fp16)[name = string("k_135_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_44_reps_0 = const()[name = string("tile_44_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = k_135_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 1, 256, 256]> tile_44_cast_fp16 = tile(reps = tile_44_reps_0, x = transpose_88_cast_fp16)[name = string("tile_44_cast_fp16")];
            tensor<int32, [5]> concat_88 = const()[name = string("concat_88"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_88_cast_fp16 = reshape(shape = concat_88, x = tile_44_cast_fp16)[name = string("reshape_88_cast_fp16")];
            tensor<int32, [5]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = reshape_88_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 1, 256, 256]> reshape_89_cast_fp16 = reshape(shape = concat_89, x = transpose_89_cast_fp16)[name = string("reshape_89_cast_fp16")];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_45_reps_0 = const()[name = string("tile_45_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = v_89_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 1, 256, 256]> tile_45_cast_fp16 = tile(reps = tile_45_reps_0, x = transpose_90_cast_fp16)[name = string("tile_45_cast_fp16")];
            tensor<int32, [5]> concat_90 = const()[name = string("concat_90"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_90_cast_fp16 = reshape(shape = concat_90, x = tile_45_cast_fp16)[name = string("reshape_90_cast_fp16")];
            tensor<int32, [5]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_91 = const()[name = string("concat_91"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = reshape_90_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 1, 256, 256]> reshape_91_cast_fp16 = reshape(shape = concat_91, x = transpose_91_cast_fp16)[name = string("reshape_91_cast_fp16")];
            tensor<int32, [4]> v_91_perm_0 = const()[name = string("v_91_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9613_transpose_x_0 = const()[name = string("op_9613_transpose_x_0"), val = bool(false)];
            bool var_9613_transpose_y_0 = const()[name = string("op_9613_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_118_cast_fp16 = transpose(perm = transpose_118_perm_0, x = reshape_89_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 4, 256, 256]> var_9613_cast_fp16 = matmul(transpose_x = var_9613_transpose_x_0, transpose_y = var_9613_transpose_y_0, x = q_91_cast_fp16, y = transpose_118_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9615_cast_fp16 = add(x = var_9613_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9615_cast_fp16")];
            int32 var_9616 = const()[name = string("op_9616"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_361_cast_fp16 = softmax(axis = var_9616, x = var_9615_cast_fp16)[name = string("w_361_cast_fp16")];
            bool var_9619_transpose_x_0 = const()[name = string("op_9619_transpose_x_0"), val = bool(false)];
            bool var_9619_transpose_y_0 = const()[name = string("op_9619_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = reshape_91_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 4, 256, 256]> var_9619_cast_fp16 = matmul(transpose_x = var_9619_transpose_x_0, transpose_y = var_9619_transpose_y_0, x = w_361_cast_fp16, y = v_91_cast_fp16)[name = string("op_9619_cast_fp16")];
            tensor<int32, [4]> var_9622_perm_0 = const()[name = string("op_9622_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9626 = const()[name = string("op_9626"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_9622_cast_fp16 = transpose(perm = var_9622_perm_0, x = var_9619_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 256, 1024]> input_267_cast_fp16 = reshape(shape = var_9626, x = var_9622_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126424448))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126948800))))[name = string("tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9631_cast_fp16 = abs(x = linear_202_cast_fp16)[name = string("op_9631_cast_fp16")];
            bool var_9634 = const()[name = string("op_9634"), val = bool(true)];
            tensor<int32, [1]> var_9635_axes_0 = const()[name = string("op_9635_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9635_cast_fp16 = reduce_max(axes = var_9635_axes_0, keep_dims = var_9634, x = var_9631_cast_fp16)[name = string("op_9635_cast_fp16")];
            fp16 var_9636_to_fp16 = const()[name = string("op_9636_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_363_cast_fp16 = maximum(x = var_9635_cast_fp16, y = var_9636_to_fp16)[name = string("s_363_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_363_cast_fp16 = real_div(x = linear_202_cast_fp16, y = s_363_cast_fp16)[name = string("y_363_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9639_cast_fp16 = mul(x = y_363_cast_fp16, y = y_363_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [1]> var_9644_axes_0 = const()[name = string("op_9644_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9644_keep_dims_0 = const()[name = string("op_9644_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9644_cast_fp16 = reduce_mean(axes = var_9644_axes_0, keep_dims = var_9644_keep_dims_0, x = var_9639_cast_fp16)[name = string("op_9644_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9645_cast_fp16 = mul(x = s_363_cast_fp16, y = s_363_cast_fp16)[name = string("op_9645_cast_fp16")];
            fp32 var_9646_epsilon_0 = const()[name = string("op_9646_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9646_cast_fp16 = inverse(epsilon = var_9646_epsilon_0, x = var_9645_cast_fp16)[name = string("op_9646_cast_fp16")];
            fp16 var_9647_to_fp16 = const()[name = string("op_9647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9648_cast_fp16 = mul(x = var_9646_cast_fp16, y = var_9647_to_fp16)[name = string("op_9648_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9650_cast_fp16 = add(x = var_9644_cast_fp16, y = var_9648_cast_fp16)[name = string("op_9650_cast_fp16")];
            fp32 var_9651_epsilon_0 = const()[name = string("op_9651_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9651_cast_fp16 = rsqrt(epsilon = var_9651_epsilon_0, x = var_9650_cast_fp16)[name = string("op_9651_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_317_cast_fp16 = mul(x = y_363_cast_fp16, y = var_9651_cast_fp16)[name = string("out_317_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126949888)))];
            tensor<fp16, [1, 256, 512]> var_9653_cast_fp16 = mul(x = out_317_cast_fp16, y = tm_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_407_cast_fp16 = add(x = x_397_cast_fp16, y = var_9653_cast_fp16)[name = string("x_407_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9656_cast_fp16 = abs(x = x_407_cast_fp16)[name = string("op_9656_cast_fp16")];
            bool var_9659 = const()[name = string("op_9659"), val = bool(true)];
            tensor<int32, [1]> var_9660_axes_0 = const()[name = string("op_9660_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9660_cast_fp16 = reduce_max(axes = var_9660_axes_0, keep_dims = var_9659, x = var_9656_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp16 var_9661_to_fp16 = const()[name = string("op_9661_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_365_cast_fp16 = maximum(x = var_9660_cast_fp16, y = var_9661_to_fp16)[name = string("s_365_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_365_cast_fp16 = real_div(x = x_407_cast_fp16, y = s_365_cast_fp16)[name = string("y_365_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9664_cast_fp16 = mul(x = y_365_cast_fp16, y = y_365_cast_fp16)[name = string("op_9664_cast_fp16")];
            tensor<int32, [1]> var_9669_axes_0 = const()[name = string("op_9669_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9669_keep_dims_0 = const()[name = string("op_9669_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9669_cast_fp16 = reduce_mean(axes = var_9669_axes_0, keep_dims = var_9669_keep_dims_0, x = var_9664_cast_fp16)[name = string("op_9669_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9670_cast_fp16 = mul(x = s_365_cast_fp16, y = s_365_cast_fp16)[name = string("op_9670_cast_fp16")];
            fp32 var_9671_epsilon_0 = const()[name = string("op_9671_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9671_cast_fp16 = inverse(epsilon = var_9671_epsilon_0, x = var_9670_cast_fp16)[name = string("op_9671_cast_fp16")];
            fp16 var_9672_to_fp16 = const()[name = string("op_9672_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9673_cast_fp16 = mul(x = var_9671_cast_fp16, y = var_9672_to_fp16)[name = string("op_9673_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9675_cast_fp16 = add(x = var_9669_cast_fp16, y = var_9673_cast_fp16)[name = string("op_9675_cast_fp16")];
            fp32 var_9676_epsilon_0 = const()[name = string("op_9676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9676_cast_fp16 = rsqrt(epsilon = var_9676_epsilon_0, x = var_9675_cast_fp16)[name = string("op_9676_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_319_cast_fp16 = mul(x = y_365_cast_fp16, y = var_9676_cast_fp16)[name = string("out_319_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126950976)))];
            tensor<fp16, [1, 256, 512]> input_269_cast_fp16 = mul(x = out_319_cast_fp16, y = tm_layers_22_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126952064))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128000704))))[name = string("tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_203_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_203_cast_fp16")];
            string var_9683_mode_0 = const()[name = string("op_9683_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_9683_cast_fp16 = gelu(mode = var_9683_mode_0, x = linear_203_cast_fp16)[name = string("op_9683_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128004864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129053504))))[name = string("tm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_204_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_271_cast_fp16 = mul(x = var_9683_cast_fp16, y = linear_204_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129057664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130106304))))[name = string("tm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9691_cast_fp16 = abs(x = linear_205_cast_fp16)[name = string("op_9691_cast_fp16")];
            bool var_9694 = const()[name = string("op_9694"), val = bool(true)];
            tensor<int32, [1]> var_9695_axes_0 = const()[name = string("op_9695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9695_cast_fp16 = reduce_max(axes = var_9695_axes_0, keep_dims = var_9694, x = var_9691_cast_fp16)[name = string("op_9695_cast_fp16")];
            fp16 var_9696_to_fp16 = const()[name = string("op_9696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_367_cast_fp16 = maximum(x = var_9695_cast_fp16, y = var_9696_to_fp16)[name = string("s_367_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_367_cast_fp16 = real_div(x = linear_205_cast_fp16, y = s_367_cast_fp16)[name = string("y_367_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9699_cast_fp16 = mul(x = y_367_cast_fp16, y = y_367_cast_fp16)[name = string("op_9699_cast_fp16")];
            tensor<int32, [1]> var_9704_axes_0 = const()[name = string("op_9704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9704_keep_dims_0 = const()[name = string("op_9704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9704_cast_fp16 = reduce_mean(axes = var_9704_axes_0, keep_dims = var_9704_keep_dims_0, x = var_9699_cast_fp16)[name = string("op_9704_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9705_cast_fp16 = mul(x = s_367_cast_fp16, y = s_367_cast_fp16)[name = string("op_9705_cast_fp16")];
            fp32 var_9706_epsilon_0 = const()[name = string("op_9706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9706_cast_fp16 = inverse(epsilon = var_9706_epsilon_0, x = var_9705_cast_fp16)[name = string("op_9706_cast_fp16")];
            fp16 var_9707_to_fp16 = const()[name = string("op_9707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9708_cast_fp16 = mul(x = var_9706_cast_fp16, y = var_9707_to_fp16)[name = string("op_9708_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9710_cast_fp16 = add(x = var_9704_cast_fp16, y = var_9708_cast_fp16)[name = string("op_9710_cast_fp16")];
            fp32 var_9711_epsilon_0 = const()[name = string("op_9711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9711_cast_fp16 = rsqrt(epsilon = var_9711_epsilon_0, x = var_9710_cast_fp16)[name = string("op_9711_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_321_cast_fp16 = mul(x = y_367_cast_fp16, y = var_9711_cast_fp16)[name = string("out_321_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130107392)))];
            tensor<fp16, [1, 256, 512]> var_9713_cast_fp16 = mul(x = out_321_cast_fp16, y = tm_layers_22_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9713_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_273_cast_fp16 = add(x = x_407_cast_fp16, y = var_9713_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130108480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130370688))))[name = string("tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_206_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_206_cast_fp16")];
            string x_411_mode_0 = const()[name = string("x_411_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_411_cast_fp16 = gelu(mode = x_411_mode_0, x = linear_206_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [4]> var_9733_begin_0 = const()[name = string("op_9733_begin_0"), val = tensor<int32, [4]>([0, 0, 22, 0])];
            tensor<int32, [4]> var_9733_end_0 = const()[name = string("op_9733_end_0"), val = tensor<int32, [4]>([1, 256, 23, 512])];
            tensor<bool, [4]> var_9733_end_mask_0 = const()[name = string("op_9733_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9733_squeeze_mask_0 = const()[name = string("op_9733_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_9733_cast_fp16 = slice_by_index(begin = var_9733_begin_0, end = var_9733_end_0, end_mask = var_9733_end_mask_0, squeeze_mask = var_9733_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9733_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_275_cast_fp16 = mul(x = x_411_cast_fp16, y = var_9733_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130371776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130633984))))[name = string("tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9743_cast_fp16 = abs(x = linear_207_cast_fp16)[name = string("op_9743_cast_fp16")];
            bool var_9746 = const()[name = string("op_9746"), val = bool(true)];
            tensor<int32, [1]> var_9747_axes_0 = const()[name = string("op_9747_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9747_cast_fp16 = reduce_max(axes = var_9747_axes_0, keep_dims = var_9746, x = var_9743_cast_fp16)[name = string("op_9747_cast_fp16")];
            fp16 var_9748_to_fp16 = const()[name = string("op_9748_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_369_cast_fp16 = maximum(x = var_9747_cast_fp16, y = var_9748_to_fp16)[name = string("s_369_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_369_cast_fp16 = real_div(x = linear_207_cast_fp16, y = s_369_cast_fp16)[name = string("y_369_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9751_cast_fp16 = mul(x = y_369_cast_fp16, y = y_369_cast_fp16)[name = string("op_9751_cast_fp16")];
            tensor<int32, [1]> var_9756_axes_0 = const()[name = string("op_9756_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9756_keep_dims_0 = const()[name = string("op_9756_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9756_cast_fp16 = reduce_mean(axes = var_9756_axes_0, keep_dims = var_9756_keep_dims_0, x = var_9751_cast_fp16)[name = string("op_9756_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9757_cast_fp16 = mul(x = s_369_cast_fp16, y = s_369_cast_fp16)[name = string("op_9757_cast_fp16")];
            fp32 var_9758_epsilon_0 = const()[name = string("op_9758_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9758_cast_fp16 = inverse(epsilon = var_9758_epsilon_0, x = var_9757_cast_fp16)[name = string("op_9758_cast_fp16")];
            fp16 var_9759_to_fp16 = const()[name = string("op_9759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9760_cast_fp16 = mul(x = var_9758_cast_fp16, y = var_9759_to_fp16)[name = string("op_9760_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9762_cast_fp16 = add(x = var_9756_cast_fp16, y = var_9760_cast_fp16)[name = string("op_9762_cast_fp16")];
            fp32 var_9763_epsilon_0 = const()[name = string("op_9763_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9763_cast_fp16 = rsqrt(epsilon = var_9763_epsilon_0, x = var_9762_cast_fp16)[name = string("op_9763_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_323_cast_fp16 = mul(x = y_369_cast_fp16, y = var_9763_cast_fp16)[name = string("out_323_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130635072)))];
            tensor<fp16, [1, 256, 512]> var_9765_cast_fp16 = mul(x = out_323_cast_fp16, y = tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9765_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_45_cast_fp16 = add(x = input_273_cast_fp16, y = var_9765_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<fp16, [1]> tm_layers_22_layer_scalar_to_fp16 = const()[name = string("tm_layers_22_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 256, 512]> x_415_cast_fp16 = mul(x = h_45_cast_fp16, y = tm_layers_22_layer_scalar_to_fp16)[name = string("x_415_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9769_cast_fp16 = abs(x = x_415_cast_fp16)[name = string("op_9769_cast_fp16")];
            bool var_9772 = const()[name = string("op_9772"), val = bool(true)];
            tensor<int32, [1]> var_9773_axes_0 = const()[name = string("op_9773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9773_cast_fp16 = reduce_max(axes = var_9773_axes_0, keep_dims = var_9772, x = var_9769_cast_fp16)[name = string("op_9773_cast_fp16")];
            fp16 var_9774_to_fp16 = const()[name = string("op_9774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_371_cast_fp16 = maximum(x = var_9773_cast_fp16, y = var_9774_to_fp16)[name = string("s_371_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_371_cast_fp16 = real_div(x = x_415_cast_fp16, y = s_371_cast_fp16)[name = string("y_371_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9777_cast_fp16 = mul(x = y_371_cast_fp16, y = y_371_cast_fp16)[name = string("op_9777_cast_fp16")];
            tensor<int32, [1]> var_9782_axes_0 = const()[name = string("op_9782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9782_keep_dims_0 = const()[name = string("op_9782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9782_cast_fp16 = reduce_mean(axes = var_9782_axes_0, keep_dims = var_9782_keep_dims_0, x = var_9777_cast_fp16)[name = string("op_9782_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9783_cast_fp16 = mul(x = s_371_cast_fp16, y = s_371_cast_fp16)[name = string("op_9783_cast_fp16")];
            fp32 var_9784_epsilon_0 = const()[name = string("op_9784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9784_cast_fp16 = inverse(epsilon = var_9784_epsilon_0, x = var_9783_cast_fp16)[name = string("op_9784_cast_fp16")];
            fp16 var_9785_to_fp16 = const()[name = string("op_9785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9786_cast_fp16 = mul(x = var_9784_cast_fp16, y = var_9785_to_fp16)[name = string("op_9786_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9788_cast_fp16 = add(x = var_9782_cast_fp16, y = var_9786_cast_fp16)[name = string("op_9788_cast_fp16")];
            fp32 var_9789_epsilon_0 = const()[name = string("op_9789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9789_cast_fp16 = rsqrt(epsilon = var_9789_epsilon_0, x = var_9788_cast_fp16)[name = string("op_9789_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_325_cast_fp16 = mul(x = y_371_cast_fp16, y = var_9789_cast_fp16)[name = string("out_325_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130636160)))];
            tensor<fp16, [1, 256, 512]> input_277_cast_fp16 = mul(x = out_325_cast_fp16, y = tm_layers_23_input_layernorm_weight_to_fp16)[name = string("input_277_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130637248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131685888))))[name = string("tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_208_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<int32, [4]> var_9799 = const()[name = string("op_9799"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 256, 4, 512]> var_9800_cast_fp16 = reshape(shape = var_9799, x = linear_208_cast_fp16)[name = string("op_9800_cast_fp16")];
            tensor<int32, [4]> x_417_perm_0 = const()[name = string("x_417_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131690048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131952256))))[name = string("tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<int32, [4]> var_9811 = const()[name = string("op_9811"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_9812_cast_fp16 = reshape(shape = var_9811, x = linear_209_cast_fp16)[name = string("op_9812_cast_fp16")];
            tensor<int32, [4]> x_419_perm_0 = const()[name = string("x_419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131953344))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132215552))))[name = string("tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<int32, [4]> var_9823 = const()[name = string("op_9823"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_9824_cast_fp16 = reshape(shape = var_9823, x = linear_210_cast_fp16)[name = string("op_9824_cast_fp16")];
            tensor<int32, [4]> x_421_perm_0 = const()[name = string("x_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 512]> x_417_cast_fp16 = transpose(perm = x_417_perm_0, x = var_9800_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 4, 256, 512]> var_9828_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_9828_cast_fp16")];
            bool var_9831 = const()[name = string("op_9831"), val = bool(true)];
            tensor<int32, [1]> var_9832_axes_0 = const()[name = string("op_9832_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_9832_cast_fp16 = reduce_max(axes = var_9832_axes_0, keep_dims = var_9831, x = var_9828_cast_fp16)[name = string("op_9832_cast_fp16")];
            fp16 var_9833_to_fp16 = const()[name = string("op_9833_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_373_cast_fp16 = maximum(x = var_9832_cast_fp16, y = var_9833_to_fp16)[name = string("s_373_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> y_373_cast_fp16 = real_div(x = x_417_cast_fp16, y = s_373_cast_fp16)[name = string("y_373_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_9836_cast_fp16 = mul(x = y_373_cast_fp16, y = y_373_cast_fp16)[name = string("op_9836_cast_fp16")];
            tensor<int32, [1]> var_9841_axes_0 = const()[name = string("op_9841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9841_keep_dims_0 = const()[name = string("op_9841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_9841_cast_fp16 = reduce_mean(axes = var_9841_axes_0, keep_dims = var_9841_keep_dims_0, x = var_9836_cast_fp16)[name = string("op_9841_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9842_cast_fp16 = mul(x = s_373_cast_fp16, y = s_373_cast_fp16)[name = string("op_9842_cast_fp16")];
            fp32 var_9843_epsilon_0 = const()[name = string("op_9843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_9843_cast_fp16 = inverse(epsilon = var_9843_epsilon_0, x = var_9842_cast_fp16)[name = string("op_9843_cast_fp16")];
            fp16 var_9844_to_fp16 = const()[name = string("op_9844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_9845_cast_fp16 = mul(x = var_9843_cast_fp16, y = var_9844_to_fp16)[name = string("op_9845_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9847_cast_fp16 = add(x = var_9841_cast_fp16, y = var_9845_cast_fp16)[name = string("op_9847_cast_fp16")];
            fp32 var_9848_epsilon_0 = const()[name = string("op_9848_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_9848_cast_fp16 = rsqrt(epsilon = var_9848_epsilon_0, x = var_9847_cast_fp16)[name = string("op_9848_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> out_327_cast_fp16 = mul(x = y_373_cast_fp16, y = var_9848_cast_fp16)[name = string("out_327_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132216640)))];
            tensor<fp16, [1, 4, 256, 512]> q_93_cast_fp16 = mul(x = out_327_cast_fp16, y = tm_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("q_93_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_419_cast_fp16 = transpose(perm = x_419_perm_0, x = var_9812_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1, 256, 512]> var_9851_cast_fp16 = abs(x = x_419_cast_fp16)[name = string("op_9851_cast_fp16")];
            bool var_9854 = const()[name = string("op_9854"), val = bool(true)];
            tensor<int32, [1]> var_9855_axes_0 = const()[name = string("op_9855_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_9855_cast_fp16 = reduce_max(axes = var_9855_axes_0, keep_dims = var_9854, x = var_9851_cast_fp16)[name = string("op_9855_cast_fp16")];
            fp16 var_9856_to_fp16 = const()[name = string("op_9856_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_375_cast_fp16 = maximum(x = var_9855_cast_fp16, y = var_9856_to_fp16)[name = string("s_375_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_375_cast_fp16 = real_div(x = x_419_cast_fp16, y = s_375_cast_fp16)[name = string("y_375_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_9859_cast_fp16 = mul(x = y_375_cast_fp16, y = y_375_cast_fp16)[name = string("op_9859_cast_fp16")];
            tensor<int32, [1]> var_9864_axes_0 = const()[name = string("op_9864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9864_keep_dims_0 = const()[name = string("op_9864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_9864_cast_fp16 = reduce_mean(axes = var_9864_axes_0, keep_dims = var_9864_keep_dims_0, x = var_9859_cast_fp16)[name = string("op_9864_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_9865_cast_fp16 = mul(x = s_375_cast_fp16, y = s_375_cast_fp16)[name = string("op_9865_cast_fp16")];
            fp32 var_9866_epsilon_0 = const()[name = string("op_9866_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_9866_cast_fp16 = inverse(epsilon = var_9866_epsilon_0, x = var_9865_cast_fp16)[name = string("op_9866_cast_fp16")];
            fp16 var_9867_to_fp16 = const()[name = string("op_9867_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_9868_cast_fp16 = mul(x = var_9866_cast_fp16, y = var_9867_to_fp16)[name = string("op_9868_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_9870_cast_fp16 = add(x = var_9864_cast_fp16, y = var_9868_cast_fp16)[name = string("op_9870_cast_fp16")];
            fp32 var_9871_epsilon_0 = const()[name = string("op_9871_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_9871_cast_fp16 = rsqrt(epsilon = var_9871_epsilon_0, x = var_9870_cast_fp16)[name = string("op_9871_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> out_329_cast_fp16 = mul(x = y_375_cast_fp16, y = var_9871_cast_fp16)[name = string("out_329_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132217728)))];
            tensor<fp16, [1, 1, 256, 512]> k_139_cast_fp16 = mul(x = out_329_cast_fp16, y = tm_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("k_139_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_421_cast_fp16 = transpose(perm = x_421_perm_0, x = var_9824_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 1, 256, 512]> var_9874_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_9874_cast_fp16")];
            bool var_9877 = const()[name = string("op_9877"), val = bool(true)];
            tensor<int32, [1]> var_9878_axes_0 = const()[name = string("op_9878_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_9878_cast_fp16 = reduce_max(axes = var_9878_axes_0, keep_dims = var_9877, x = var_9874_cast_fp16)[name = string("op_9878_cast_fp16")];
            fp16 var_9879_to_fp16 = const()[name = string("op_9879_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_377_cast_fp16 = maximum(x = var_9878_cast_fp16, y = var_9879_to_fp16)[name = string("s_377_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_377_cast_fp16 = real_div(x = x_421_cast_fp16, y = s_377_cast_fp16)[name = string("y_377_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_9882_cast_fp16 = mul(x = y_377_cast_fp16, y = y_377_cast_fp16)[name = string("op_9882_cast_fp16")];
            tensor<int32, [1]> var_9887_axes_0 = const()[name = string("op_9887_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9887_keep_dims_0 = const()[name = string("op_9887_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_9887_cast_fp16 = reduce_mean(axes = var_9887_axes_0, keep_dims = var_9887_keep_dims_0, x = var_9882_cast_fp16)[name = string("op_9887_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_9888_cast_fp16 = mul(x = s_377_cast_fp16, y = s_377_cast_fp16)[name = string("op_9888_cast_fp16")];
            fp32 var_9889_epsilon_0 = const()[name = string("op_9889_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_9889_cast_fp16 = inverse(epsilon = var_9889_epsilon_0, x = var_9888_cast_fp16)[name = string("op_9889_cast_fp16")];
            fp16 var_9890_to_fp16 = const()[name = string("op_9890_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_9891_cast_fp16 = mul(x = var_9889_cast_fp16, y = var_9890_to_fp16)[name = string("op_9891_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_9893_cast_fp16 = add(x = var_9887_cast_fp16, y = var_9891_cast_fp16)[name = string("op_9893_cast_fp16")];
            fp32 var_9894_epsilon_0 = const()[name = string("op_9894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_9894_cast_fp16 = rsqrt(epsilon = var_9894_epsilon_0, x = var_9893_cast_fp16)[name = string("op_9894_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> v_93_cast_fp16 = mul(x = y_377_cast_fp16, y = var_9894_cast_fp16)[name = string("v_93_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_9896_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9896_cast_fp16")];
            tensor<int32, [4]> var_9901_begin_0 = const()[name = string("op_9901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9901_end_0 = const()[name = string("op_9901_end_0"), val = tensor<int32, [4]>([1, 4, 256, 512])];
            tensor<bool, [4]> var_9901_end_mask_0 = const()[name = string("op_9901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 256]> var_9901_cast_fp16 = slice_by_index(begin = var_9901_begin_0, end = var_9901_end_0, end_mask = var_9901_end_mask_0, x = q_93_cast_fp16)[name = string("op_9901_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 256]> var_9902_cast_fp16 = mul(x = var_9901_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_9902_cast_fp16")];
            tensor<int32, [4]> var_9907_begin_0 = const()[name = string("op_9907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9907_end_0 = const()[name = string("op_9907_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_9907_end_mask_0 = const()[name = string("op_9907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 256]> var_9907_cast_fp16 = slice_by_index(begin = var_9907_begin_0, end = var_9907_end_0, end_mask = var_9907_end_mask_0, x = q_93_cast_fp16)[name = string("op_9907_cast_fp16")];
            int32 var_9909 = const()[name = string("op_9909"), val = int32(-1)];
            bool var_9910_interleave_0 = const()[name = string("op_9910_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> var_9910_cast_fp16 = concat(axis = var_9909, interleave = var_9910_interleave_0, values = (var_9902_cast_fp16, var_9907_cast_fp16))[name = string("op_9910_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_9911_cast_fp16 = mul(x = var_9910_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9911_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> q_cast_fp16 = add(x = var_9896_cast_fp16, y = var_9911_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_9914_cast_fp16 = mul(x = k_139_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9914_cast_fp16")];
            tensor<int32, [4]> var_9919_begin_0 = const()[name = string("op_9919_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9919_end_0 = const()[name = string("op_9919_end_0"), val = tensor<int32, [4]>([1, 1, 256, 512])];
            tensor<bool, [4]> var_9919_end_mask_0 = const()[name = string("op_9919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 256, 256]> var_9919_cast_fp16 = slice_by_index(begin = var_9919_begin_0, end = var_9919_end_0, end_mask = var_9919_end_mask_0, x = k_139_cast_fp16)[name = string("op_9919_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 256, 256]> var_9920_cast_fp16 = mul(x = var_9919_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_9920_cast_fp16")];
            tensor<int32, [4]> var_9925_begin_0 = const()[name = string("op_9925_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9925_end_0 = const()[name = string("op_9925_end_0"), val = tensor<int32, [4]>([1, 1, 256, 256])];
            tensor<bool, [4]> var_9925_end_mask_0 = const()[name = string("op_9925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 256, 256]> var_9925_cast_fp16 = slice_by_index(begin = var_9925_begin_0, end = var_9925_end_0, end_mask = var_9925_end_mask_0, x = k_139_cast_fp16)[name = string("op_9925_cast_fp16")];
            int32 var_9927 = const()[name = string("op_9927"), val = int32(-1)];
            bool var_9928_interleave_0 = const()[name = string("op_9928_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 256, 512]> var_9928_cast_fp16 = concat(axis = var_9927, interleave = var_9928_interleave_0, values = (var_9920_cast_fp16, var_9925_cast_fp16))[name = string("op_9928_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_9929_cast_fp16 = mul(x = var_9928_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9929_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> k_141_cast_fp16 = add(x = var_9914_cast_fp16, y = var_9929_cast_fp16)[name = string("k_141_cast_fp16")];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_46_reps_0 = const()[name = string("tile_46_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_92_cast_fp16 = transpose(perm = transpose_92_perm_0, x = k_141_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 1, 256, 512]> tile_46_cast_fp16 = tile(reps = tile_46_reps_0, x = transpose_92_cast_fp16)[name = string("tile_46_cast_fp16")];
            tensor<int32, [5]> concat_92 = const()[name = string("concat_92"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_92_cast_fp16 = reshape(shape = concat_92, x = tile_46_cast_fp16)[name = string("reshape_92_cast_fp16")];
            tensor<int32, [5]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_93 = const()[name = string("concat_93"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = reshape_92_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 1, 256, 512]> reshape_93_cast_fp16 = reshape(shape = concat_93, x = transpose_93_cast_fp16)[name = string("reshape_93_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_47_reps_0 = const()[name = string("tile_47_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_94_cast_fp16 = transpose(perm = transpose_94_perm_0, x = v_93_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 1, 256, 512]> tile_47_cast_fp16 = tile(reps = tile_47_reps_0, x = transpose_94_cast_fp16)[name = string("tile_47_cast_fp16")];
            tensor<int32, [5]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_94_cast_fp16 = reshape(shape = concat_94, x = tile_47_cast_fp16)[name = string("reshape_94_cast_fp16")];
            tensor<int32, [5]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = reshape_94_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 1, 256, 512]> reshape_95_cast_fp16 = reshape(shape = concat_95, x = transpose_95_cast_fp16)[name = string("reshape_95_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9943_transpose_x_0 = const()[name = string("op_9943_transpose_x_0"), val = bool(false)];
            bool var_9943_transpose_y_0 = const()[name = string("op_9943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = reshape_93_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 4, 256, 256]> var_9943_cast_fp16 = matmul(transpose_x = var_9943_transpose_x_0, transpose_y = var_9943_transpose_y_0, x = q_cast_fp16, y = transpose_119_cast_fp16)[name = string("op_9943_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9945_cast_fp16 = add(x = var_9943_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9945_cast_fp16")];
            int32 var_9946 = const()[name = string("op_9946"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> w_377_cast_fp16 = softmax(axis = var_9946, x = var_9945_cast_fp16)[name = string("w_377_cast_fp16")];
            bool var_9949_transpose_x_0 = const()[name = string("op_9949_transpose_x_0"), val = bool(false)];
            bool var_9949_transpose_y_0 = const()[name = string("op_9949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> v_cast_fp16 = transpose(perm = v_perm_0, x = reshape_95_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 4, 256, 512]> var_9949_cast_fp16 = matmul(transpose_x = var_9949_transpose_x_0, transpose_y = var_9949_transpose_y_0, x = w_377_cast_fp16, y = v_cast_fp16)[name = string("op_9949_cast_fp16")];
            tensor<int32, [4]> var_9952_perm_0 = const()[name = string("op_9952_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9956 = const()[name = string("op_9956"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 256, 4, 512]> var_9952_cast_fp16 = transpose(perm = var_9952_perm_0, x = var_9949_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 256, 2048]> input_279_cast_fp16 = reshape(shape = var_9956, x = var_9952_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132218816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133267456))))[name = string("tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9961_cast_fp16 = abs(x = linear_211_cast_fp16)[name = string("op_9961_cast_fp16")];
            bool var_9964 = const()[name = string("op_9964"), val = bool(true)];
            tensor<int32, [1]> var_9965_axes_0 = const()[name = string("op_9965_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9965_cast_fp16 = reduce_max(axes = var_9965_axes_0, keep_dims = var_9964, x = var_9961_cast_fp16)[name = string("op_9965_cast_fp16")];
            fp16 var_9966_to_fp16 = const()[name = string("op_9966_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_379_cast_fp16 = maximum(x = var_9965_cast_fp16, y = var_9966_to_fp16)[name = string("s_379_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_379_cast_fp16 = real_div(x = linear_211_cast_fp16, y = s_379_cast_fp16)[name = string("y_379_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9969_cast_fp16 = mul(x = y_379_cast_fp16, y = y_379_cast_fp16)[name = string("op_9969_cast_fp16")];
            tensor<int32, [1]> var_9974_axes_0 = const()[name = string("op_9974_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9974_keep_dims_0 = const()[name = string("op_9974_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9974_cast_fp16 = reduce_mean(axes = var_9974_axes_0, keep_dims = var_9974_keep_dims_0, x = var_9969_cast_fp16)[name = string("op_9974_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9975_cast_fp16 = mul(x = s_379_cast_fp16, y = s_379_cast_fp16)[name = string("op_9975_cast_fp16")];
            fp32 var_9976_epsilon_0 = const()[name = string("op_9976_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9976_cast_fp16 = inverse(epsilon = var_9976_epsilon_0, x = var_9975_cast_fp16)[name = string("op_9976_cast_fp16")];
            fp16 var_9977_to_fp16 = const()[name = string("op_9977_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9978_cast_fp16 = mul(x = var_9976_cast_fp16, y = var_9977_to_fp16)[name = string("op_9978_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9980_cast_fp16 = add(x = var_9974_cast_fp16, y = var_9978_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp32 var_9981_epsilon_0 = const()[name = string("op_9981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9981_cast_fp16 = rsqrt(epsilon = var_9981_epsilon_0, x = var_9980_cast_fp16)[name = string("op_9981_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_331_cast_fp16 = mul(x = y_379_cast_fp16, y = var_9981_cast_fp16)[name = string("out_331_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133268544)))];
            tensor<fp16, [1, 256, 512]> var_9983_cast_fp16 = mul(x = out_331_cast_fp16, y = tm_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("op_9983_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_425_cast_fp16 = add(x = x_415_cast_fp16, y = var_9983_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9986_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_9986_cast_fp16")];
            bool var_9989 = const()[name = string("op_9989"), val = bool(true)];
            tensor<int32, [1]> var_9990_axes_0 = const()[name = string("op_9990_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9990_cast_fp16 = reduce_max(axes = var_9990_axes_0, keep_dims = var_9989, x = var_9986_cast_fp16)[name = string("op_9990_cast_fp16")];
            fp16 var_9991_to_fp16 = const()[name = string("op_9991_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_381_cast_fp16 = maximum(x = var_9990_cast_fp16, y = var_9991_to_fp16)[name = string("s_381_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_381_cast_fp16 = real_div(x = x_425_cast_fp16, y = s_381_cast_fp16)[name = string("y_381_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9994_cast_fp16 = mul(x = y_381_cast_fp16, y = y_381_cast_fp16)[name = string("op_9994_cast_fp16")];
            tensor<int32, [1]> var_9999_axes_0 = const()[name = string("op_9999_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9999_keep_dims_0 = const()[name = string("op_9999_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9999_cast_fp16 = reduce_mean(axes = var_9999_axes_0, keep_dims = var_9999_keep_dims_0, x = var_9994_cast_fp16)[name = string("op_9999_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10000_cast_fp16 = mul(x = s_381_cast_fp16, y = s_381_cast_fp16)[name = string("op_10000_cast_fp16")];
            fp32 var_10001_epsilon_0 = const()[name = string("op_10001_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_10001_cast_fp16 = inverse(epsilon = var_10001_epsilon_0, x = var_10000_cast_fp16)[name = string("op_10001_cast_fp16")];
            fp16 var_10002_to_fp16 = const()[name = string("op_10002_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_10003_cast_fp16 = mul(x = var_10001_cast_fp16, y = var_10002_to_fp16)[name = string("op_10003_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10005_cast_fp16 = add(x = var_9999_cast_fp16, y = var_10003_cast_fp16)[name = string("op_10005_cast_fp16")];
            fp32 var_10006_epsilon_0 = const()[name = string("op_10006_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_10006_cast_fp16 = rsqrt(epsilon = var_10006_epsilon_0, x = var_10005_cast_fp16)[name = string("op_10006_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_333_cast_fp16 = mul(x = y_381_cast_fp16, y = var_10006_cast_fp16)[name = string("out_333_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133269632)))];
            tensor<fp16, [1, 256, 512]> input_281_cast_fp16 = mul(x = out_333_cast_fp16, y = tm_layers_23_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133270720))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134319360))))[name = string("tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_212_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_212_cast_fp16")];
            string var_10013_mode_0 = const()[name = string("op_10013_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_10013_cast_fp16 = gelu(mode = var_10013_mode_0, x = linear_212_cast_fp16)[name = string("op_10013_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134323520))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135372160))))[name = string("tm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_213_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_213_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_283_cast_fp16 = mul(x = var_10013_cast_fp16, y = linear_213_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135376320))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136424960))))[name = string("tm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_283_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10021_cast_fp16 = abs(x = linear_214_cast_fp16)[name = string("op_10021_cast_fp16")];
            bool var_10024 = const()[name = string("op_10024"), val = bool(true)];
            tensor<int32, [1]> var_10025_axes_0 = const()[name = string("op_10025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_10025_cast_fp16 = reduce_max(axes = var_10025_axes_0, keep_dims = var_10024, x = var_10021_cast_fp16)[name = string("op_10025_cast_fp16")];
            fp16 var_10026_to_fp16 = const()[name = string("op_10026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_383_cast_fp16 = maximum(x = var_10025_cast_fp16, y = var_10026_to_fp16)[name = string("s_383_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_383_cast_fp16 = real_div(x = linear_214_cast_fp16, y = s_383_cast_fp16)[name = string("y_383_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10029_cast_fp16 = mul(x = y_383_cast_fp16, y = y_383_cast_fp16)[name = string("op_10029_cast_fp16")];
            tensor<int32, [1]> var_10034_axes_0 = const()[name = string("op_10034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10034_keep_dims_0 = const()[name = string("op_10034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_10034_cast_fp16 = reduce_mean(axes = var_10034_axes_0, keep_dims = var_10034_keep_dims_0, x = var_10029_cast_fp16)[name = string("op_10034_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10035_cast_fp16 = mul(x = s_383_cast_fp16, y = s_383_cast_fp16)[name = string("op_10035_cast_fp16")];
            fp32 var_10036_epsilon_0 = const()[name = string("op_10036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_10036_cast_fp16 = inverse(epsilon = var_10036_epsilon_0, x = var_10035_cast_fp16)[name = string("op_10036_cast_fp16")];
            fp16 var_10037_to_fp16 = const()[name = string("op_10037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_10038_cast_fp16 = mul(x = var_10036_cast_fp16, y = var_10037_to_fp16)[name = string("op_10038_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10040_cast_fp16 = add(x = var_10034_cast_fp16, y = var_10038_cast_fp16)[name = string("op_10040_cast_fp16")];
            fp32 var_10041_epsilon_0 = const()[name = string("op_10041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_10041_cast_fp16 = rsqrt(epsilon = var_10041_epsilon_0, x = var_10040_cast_fp16)[name = string("op_10041_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_335_cast_fp16 = mul(x = y_383_cast_fp16, y = var_10041_cast_fp16)[name = string("out_335_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136426048)))];
            tensor<fp16, [1, 256, 512]> var_10043_cast_fp16 = mul(x = out_335_cast_fp16, y = tm_layers_23_post_feedforward_layernorm_weight_to_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_285_cast_fp16 = add(x = x_425_cast_fp16, y = var_10043_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136427136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136689344))))[name = string("tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_215_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = string("linear_215_cast_fp16")];
            string x_429_mode_0 = const()[name = string("x_429_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> x_429_cast_fp16 = gelu(mode = x_429_mode_0, x = linear_215_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [4]> var_10063_begin_0 = const()[name = string("op_10063_begin_0"), val = tensor<int32, [4]>([0, 0, 23, 0])];
            tensor<int32, [4]> var_10063_end_0 = const()[name = string("op_10063_end_0"), val = tensor<int32, [4]>([1, 256, 24, 512])];
            tensor<bool, [4]> var_10063_end_mask_0 = const()[name = string("op_10063_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_10063_squeeze_mask_0 = const()[name = string("op_10063_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_10063_cast_fp16 = slice_by_index(begin = var_10063_begin_0, end = var_10063_end_0, end_mask = var_10063_end_mask_0, squeeze_mask = var_10063_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_287_cast_fp16 = mul(x = x_429_cast_fp16, y = var_10063_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136690432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136952640))))[name = string("tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10073_cast_fp16 = abs(x = linear_216_cast_fp16)[name = string("op_10073_cast_fp16")];
            bool var_10076 = const()[name = string("op_10076"), val = bool(true)];
            tensor<int32, [1]> var_10077_axes_0 = const()[name = string("op_10077_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_10077_cast_fp16 = reduce_max(axes = var_10077_axes_0, keep_dims = var_10076, x = var_10073_cast_fp16)[name = string("op_10077_cast_fp16")];
            fp16 var_10078_to_fp16 = const()[name = string("op_10078_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_385_cast_fp16 = maximum(x = var_10077_cast_fp16, y = var_10078_to_fp16)[name = string("s_385_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_385_cast_fp16 = real_div(x = linear_216_cast_fp16, y = s_385_cast_fp16)[name = string("y_385_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10081_cast_fp16 = mul(x = y_385_cast_fp16, y = y_385_cast_fp16)[name = string("op_10081_cast_fp16")];
            tensor<int32, [1]> var_10086_axes_0 = const()[name = string("op_10086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10086_keep_dims_0 = const()[name = string("op_10086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_10086_cast_fp16 = reduce_mean(axes = var_10086_axes_0, keep_dims = var_10086_keep_dims_0, x = var_10081_cast_fp16)[name = string("op_10086_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10087_cast_fp16 = mul(x = s_385_cast_fp16, y = s_385_cast_fp16)[name = string("op_10087_cast_fp16")];
            fp32 var_10088_epsilon_0 = const()[name = string("op_10088_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_10088_cast_fp16 = inverse(epsilon = var_10088_epsilon_0, x = var_10087_cast_fp16)[name = string("op_10088_cast_fp16")];
            fp16 var_10089_to_fp16 = const()[name = string("op_10089_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_10090_cast_fp16 = mul(x = var_10088_cast_fp16, y = var_10089_to_fp16)[name = string("op_10090_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10092_cast_fp16 = add(x = var_10086_cast_fp16, y = var_10090_cast_fp16)[name = string("op_10092_cast_fp16")];
            fp32 var_10093_epsilon_0 = const()[name = string("op_10093_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_10093_cast_fp16 = rsqrt(epsilon = var_10093_epsilon_0, x = var_10092_cast_fp16)[name = string("op_10093_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_337_cast_fp16 = mul(x = y_385_cast_fp16, y = var_10093_cast_fp16)[name = string("out_337_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136953728)))];
            tensor<fp16, [1, 256, 512]> var_10095_cast_fp16 = mul(x = out_337_cast_fp16, y = tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_10095_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_47_cast_fp16 = add(x = input_285_cast_fp16, y = var_10095_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<fp16, [1]> tm_layers_23_layer_scalar_to_fp16 = const()[name = string("tm_layers_23_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d6p-4])];
            tensor<fp16, [1, 256, 512]> x_cast_fp16 = mul(x = h_47_cast_fp16, y = tm_layers_23_layer_scalar_to_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10099_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_10099_cast_fp16")];
            bool var_10102 = const()[name = string("op_10102"), val = bool(true)];
            tensor<int32, [1]> var_10103_axes_0 = const()[name = string("op_10103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_10103_cast_fp16 = reduce_max(axes = var_10103_axes_0, keep_dims = var_10102, x = var_10099_cast_fp16)[name = string("op_10103_cast_fp16")];
            fp16 var_10104_to_fp16 = const()[name = string("op_10104_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_387_cast_fp16 = maximum(x = var_10103_cast_fp16, y = var_10104_to_fp16)[name = string("s_387_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_387_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10107_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_10107_cast_fp16")];
            tensor<int32, [1]> var_10112_axes_0 = const()[name = string("op_10112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10112_keep_dims_0 = const()[name = string("op_10112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_10112_cast_fp16 = reduce_mean(axes = var_10112_axes_0, keep_dims = var_10112_keep_dims_0, x = var_10107_cast_fp16)[name = string("op_10112_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10113_cast_fp16 = mul(x = s_387_cast_fp16, y = s_387_cast_fp16)[name = string("op_10113_cast_fp16")];
            fp32 var_10114_epsilon_0 = const()[name = string("op_10114_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_10114_cast_fp16 = inverse(epsilon = var_10114_epsilon_0, x = var_10113_cast_fp16)[name = string("op_10114_cast_fp16")];
            fp16 var_10115_to_fp16 = const()[name = string("op_10115_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_10116_cast_fp16 = mul(x = var_10114_cast_fp16, y = var_10115_to_fp16)[name = string("op_10116_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10118_cast_fp16 = add(x = var_10112_cast_fp16, y = var_10116_cast_fp16)[name = string("op_10118_cast_fp16")];
            fp32 var_10119_epsilon_0 = const()[name = string("op_10119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_10119_cast_fp16 = rsqrt(epsilon = var_10119_epsilon_0, x = var_10118_cast_fp16)[name = string("op_10119_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_cast_fp16 = mul(x = y_cast_fp16, y = var_10119_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [512]> tm_norm_weight_to_fp16 = const()[name = string("tm_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136954816)))];
            tensor<fp16, [1, 256, 512]> h_cast_fp16 = mul(x = out_cast_fp16, y = tm_norm_weight_to_fp16)[name = string("h_cast_fp16")];
            tensor<int32, [1]> var_10126_axes_0 = const()[name = string("op_10126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10126_keep_dims_0 = const()[name = string("op_10126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10126_cast_fp16 = reduce_sum(axes = var_10126_axes_0, keep_dims = var_10126_keep_dims_0, x = attention_mask)[name = string("op_10126_cast_fp16")];
            tensor<fp16, [1, 256]> var_10127_cast_fp16 = real_div(x = attention_mask, y = var_10126_cast_fp16)[name = string("op_10127_cast_fp16")];
            tensor<int32, [1]> var_10134_axes_0 = const()[name = string("op_10134_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 256]> var_10134_cast_fp16 = expand_dims(axes = var_10134_axes_0, x = var_10127_cast_fp16)[name = string("op_10134_cast_fp16")];
            bool input_transpose_x_0 = const()[name = string("input_transpose_x_0"), val = bool(false)];
            bool input_transpose_y_0 = const()[name = string("input_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512]> input_cast_fp16 = matmul(transpose_x = input_transpose_x_0, transpose_y = input_transpose_y_0, x = var_10134_cast_fp16, y = h_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [768, 512]> tm_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136955904))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137349184))))[name = string("tm_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_217_bias_0_to_fp16 = const()[name = string("linear_217_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137350784)))];
            tensor<fp16, [1, 1, 768]> linear_217_cast_fp16 = linear(bias = linear_217_bias_0_to_fp16, weight = tm_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<int32, [3]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_10151_squeeze_mask_0 = const()[name = string("op_10151_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, squeeze_mask = var_10151_squeeze_mask_0, x = linear_217_cast_fp16)[name = string("op_10151_cast_fp16")];
            tensor<fp16, [1, 768]> var_10157_cast_fp16 = abs(x = var_10151_cast_fp16)[name = string("op_10157_cast_fp16")];
            bool var_10160 = const()[name = string("op_10160"), val = bool(true)];
            tensor<int32, [1]> s_axes_0 = const()[name = string("s_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> s_cast_fp16 = reduce_max(axes = s_axes_0, keep_dims = var_10160, x = var_10157_cast_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [1, 768]> pooled_cast_fp16 = real_div(x = var_10151_cast_fp16, y = s_cast_fp16)[name = string("pooled_cast_fp16")];
            tensor<fp16, [1, 768]> var_10163_cast_fp16 = mul(x = pooled_cast_fp16, y = pooled_cast_fp16)[name = string("op_10163_cast_fp16")];
            tensor<int32, [1]> var_10168_axes_0 = const()[name = string("op_10168_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10168_keep_dims_0 = const()[name = string("op_10168_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10168_cast_fp16 = reduce_sum(axes = var_10168_axes_0, keep_dims = var_10168_keep_dims_0, x = var_10163_cast_fp16)[name = string("op_10168_cast_fp16")];
            fp32 var_10169_epsilon_0 = const()[name = string("op_10169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_10169_cast_fp16 = rsqrt(epsilon = var_10169_epsilon_0, x = var_10168_cast_fp16)[name = string("op_10169_cast_fp16")];
            tensor<fp16, [1, 768]> var_10170_cast_fp16 = mul(x = pooled_cast_fp16, y = var_10169_cast_fp16)[name = string("op_10170_cast_fp16")];
            string var_10170_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_10170_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 768]> embedding = cast(dtype = var_10170_cast_fp16_to_fp32_dtype_0, x = var_10170_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func embed_32<ios18>(tensor<fp16, [1, 32]> attention_mask, tensor<fp16, [1, 32, 512]> inputs_embeds) {
            tensor<fp16, [12288, 512]> tm_ple_per_layer_model_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [12288, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1728))), scale = tensor<fp16, [12288, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6293248))))[name = string("tm_ple_per_layer_model_projection_weight_to_fp16_quantized")];
            tensor<fp16, [12288]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [12288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6317888)))];
            tensor<fp16, [1, 32, 12288]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = tm_ple_per_layer_model_projection_weight_to_fp16_quantized, x = inputs_embeds)[name = string("linear_0_cast_fp16")];
            fp16 var_2129_to_fp16 = const()[name = string("op_2129_to_fp16"), val = fp16(0x1.6ap-5)];
            tensor<fp16, [1, 32, 12288]> ple_1_cast_fp16 = mul(x = linear_0_cast_fp16, y = var_2129_to_fp16)[name = string("ple_1_cast_fp16")];
            tensor<int32, [4]> var_2135 = const()[name = string("op_2135"), val = tensor<int32, [4]>([1, -1, 24, 512])];
            tensor<fp16, [1, 32, 24, 512]> x_1_cast_fp16 = reshape(shape = var_2135, x = ple_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 32, 24, 512]> var_2137_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_2137_cast_fp16")];
            bool var_2140 = const()[name = string("op_2140"), val = bool(true)];
            tensor<int32, [1]> var_2141_axes_0 = const()[name = string("op_2141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 24, 1]> var_2141_cast_fp16 = reduce_max(axes = var_2141_axes_0, keep_dims = var_2140, x = var_2137_cast_fp16)[name = string("op_2141_cast_fp16")];
            fp16 var_2142_to_fp16 = const()[name = string("op_2142_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 24, 1]> s_1_cast_fp16 = maximum(x = var_2141_cast_fp16, y = var_2142_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 32, 24, 512]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 32, 24, 512]> var_2145_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [1]> var_2150_axes_0 = const()[name = string("op_2150_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2150_keep_dims_0 = const()[name = string("op_2150_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 24, 1]> var_2150_cast_fp16 = reduce_mean(axes = var_2150_axes_0, keep_dims = var_2150_keep_dims_0, x = var_2145_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 32, 24, 1]> var_2151_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 24, 1]> var_2152_cast_fp16 = inverse(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            fp16 var_2153_to_fp16 = const()[name = string("op_2153_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 24, 1]> var_2154_cast_fp16 = mul(x = var_2152_cast_fp16, y = var_2153_to_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 32, 24, 1]> var_2156_cast_fp16 = add(x = var_2150_cast_fp16, y = var_2154_cast_fp16)[name = string("op_2156_cast_fp16")];
            fp32 var_2157_epsilon_0 = const()[name = string("op_2157_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 24, 1]> var_2157_cast_fp16 = rsqrt(epsilon = var_2157_epsilon_0, x = var_2156_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<fp16, [1, 32, 24, 512]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_2157_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [512]> tm_ple_per_layer_projection_norm_weight_to_fp16 = const()[name = string("tm_ple_per_layer_projection_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6342528)))];
            tensor<fp16, [1, 32, 24, 512]> ple_3_cast_fp16 = mul(x = out_1_cast_fp16, y = tm_ple_per_layer_projection_norm_weight_to_fp16)[name = string("ple_3_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 32]> var_2162_cast_fp16 = sub(x = var_2160_to_fp16, y = attention_mask)[name = string("op_2162_cast_fp16")];
            fp16 var_2163_to_fp16 = const()[name = string("op_2163_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 32]> var_2164_cast_fp16 = mul(x = var_2162_cast_fp16, y = var_2163_to_fp16)[name = string("op_2164_cast_fp16")];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 32]> var_2171_cast_fp16 = expand_dims(axes = var_2171_axes_0, x = var_2164_cast_fp16)[name = string("op_2171_cast_fp16")];
            tensor<int32, [1]> var_2173_axes_0 = const()[name = string("op_2173_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 32]> var_2173_cast_fp16 = expand_dims(axes = var_2173_axes_0, x = var_2171_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2179_cast_fp16 = abs(x = inputs_embeds)[name = string("op_2179_cast_fp16")];
            bool var_2182 = const()[name = string("op_2182"), val = bool(true)];
            tensor<int32, [1]> var_2183_axes_0 = const()[name = string("op_2183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2183_cast_fp16 = reduce_max(axes = var_2183_axes_0, keep_dims = var_2182, x = var_2179_cast_fp16)[name = string("op_2183_cast_fp16")];
            fp16 var_2184_to_fp16 = const()[name = string("op_2184_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_3_cast_fp16 = maximum(x = var_2183_cast_fp16, y = var_2184_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_3_cast_fp16 = real_div(x = inputs_embeds, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2187_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> var_2192_axes_0 = const()[name = string("op_2192_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2192_keep_dims_0 = const()[name = string("op_2192_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2192_cast_fp16 = reduce_mean(axes = var_2192_axes_0, keep_dims = var_2192_keep_dims_0, x = var_2187_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2193_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp32 var_2194_epsilon_0 = const()[name = string("op_2194_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2194_cast_fp16 = inverse(epsilon = var_2194_epsilon_0, x = var_2193_cast_fp16)[name = string("op_2194_cast_fp16")];
            fp16 var_2195_to_fp16 = const()[name = string("op_2195_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2196_cast_fp16 = mul(x = var_2194_cast_fp16, y = var_2195_to_fp16)[name = string("op_2196_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2198_cast_fp16 = add(x = var_2192_cast_fp16, y = var_2196_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_2199_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6343616)))];
            tensor<fp16, [1, 32, 512]> input_1_cast_fp16 = mul(x = out_3_cast_fp16, y = tm_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6344704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6869056))))[name = string("tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6871168)))];
            tensor<fp16, [1, 32, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_1_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6873280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7135488))))[name = string("tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7136576)))];
            tensor<fp16, [1, 32, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_2221 = const()[name = string("op_2221"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_2222_cast_fp16 = reshape(shape = var_2221, x = linear_2_cast_fp16)[name = string("op_2222_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7137664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7399872))))[name = string("tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_2234_cast_fp16 = reshape(shape = var_2233, x = linear_3_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<int32, [4]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_2210_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 4, 32, 256]> var_2238_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_2238_cast_fp16")];
            bool var_2241 = const()[name = string("op_2241"), val = bool(true)];
            tensor<int32, [1]> var_2242_axes_0 = const()[name = string("op_2242_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_2242_cast_fp16 = reduce_max(axes = var_2242_axes_0, keep_dims = var_2241, x = var_2238_cast_fp16)[name = string("op_2242_cast_fp16")];
            fp16 var_2243_to_fp16 = const()[name = string("op_2243_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_5_cast_fp16 = maximum(x = var_2242_cast_fp16, y = var_2243_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_2246_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<int32, [1]> var_2251_axes_0 = const()[name = string("op_2251_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2251_keep_dims_0 = const()[name = string("op_2251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_2251_cast_fp16 = reduce_mean(axes = var_2251_axes_0, keep_dims = var_2251_keep_dims_0, x = var_2246_cast_fp16)[name = string("op_2251_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_2252_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_2252_cast_fp16")];
            fp32 var_2253_epsilon_0 = const()[name = string("op_2253_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_2253_cast_fp16 = inverse(epsilon = var_2253_epsilon_0, x = var_2252_cast_fp16)[name = string("op_2253_cast_fp16")];
            fp16 var_2254_to_fp16 = const()[name = string("op_2254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_2255_cast_fp16 = mul(x = var_2253_cast_fp16, y = var_2254_to_fp16)[name = string("op_2255_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_2257_cast_fp16 = add(x = var_2251_cast_fp16, y = var_2255_cast_fp16)[name = string("op_2257_cast_fp16")];
            fp32 var_2258_epsilon_0 = const()[name = string("op_2258_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_2258_cast_fp16 = rsqrt(epsilon = var_2258_epsilon_0, x = var_2257_cast_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_2258_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7400960)))];
            tensor<fp16, [1, 4, 32, 256]> q_1_cast_fp16 = mul(x = out_5_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_2222_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 2, 32, 256]> var_2261_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_2261_cast_fp16")];
            bool var_2264 = const()[name = string("op_2264"), val = bool(true)];
            tensor<int32, [1]> var_2265_axes_0 = const()[name = string("op_2265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_2265_cast_fp16 = reduce_max(axes = var_2265_axes_0, keep_dims = var_2264, x = var_2261_cast_fp16)[name = string("op_2265_cast_fp16")];
            fp16 var_2266_to_fp16 = const()[name = string("op_2266_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_7_cast_fp16 = maximum(x = var_2265_cast_fp16, y = var_2266_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2269_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> var_2274_axes_0 = const()[name = string("op_2274_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2274_keep_dims_0 = const()[name = string("op_2274_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_2274_cast_fp16 = reduce_mean(axes = var_2274_axes_0, keep_dims = var_2274_keep_dims_0, x = var_2269_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2275_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_2276_cast_fp16 = inverse(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            fp16 var_2277_to_fp16 = const()[name = string("op_2277_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_2278_cast_fp16 = mul(x = var_2276_cast_fp16, y = var_2277_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2280_cast_fp16 = add(x = var_2274_cast_fp16, y = var_2278_cast_fp16)[name = string("op_2280_cast_fp16")];
            fp32 var_2281_epsilon_0 = const()[name = string("op_2281_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_2281_cast_fp16 = rsqrt(epsilon = var_2281_epsilon_0, x = var_2280_cast_fp16)[name = string("op_2281_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_7_cast_fp16 = mul(x = y_7_cast_fp16, y = var_2281_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7401536)))];
            tensor<fp16, [1, 2, 32, 256]> k_1_cast_fp16 = mul(x = out_7_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = var_2234_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2, 32, 256]> var_2284_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_2284_cast_fp16")];
            bool var_2287 = const()[name = string("op_2287"), val = bool(true)];
            tensor<int32, [1]> var_2288_axes_0 = const()[name = string("op_2288_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_2288_cast_fp16 = reduce_max(axes = var_2288_axes_0, keep_dims = var_2287, x = var_2284_cast_fp16)[name = string("op_2288_cast_fp16")];
            fp16 var_2289_to_fp16 = const()[name = string("op_2289_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_9_cast_fp16 = maximum(x = var_2288_cast_fp16, y = var_2289_to_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_9_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2292_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_2292_cast_fp16")];
            tensor<int32, [1]> var_2297_axes_0 = const()[name = string("op_2297_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2297_keep_dims_0 = const()[name = string("op_2297_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_2297_cast_fp16 = reduce_mean(axes = var_2297_axes_0, keep_dims = var_2297_keep_dims_0, x = var_2292_cast_fp16)[name = string("op_2297_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2298_cast_fp16 = mul(x = s_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_2298_cast_fp16")];
            fp32 var_2299_epsilon_0 = const()[name = string("op_2299_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_2299_cast_fp16 = inverse(epsilon = var_2299_epsilon_0, x = var_2298_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 var_2300_to_fp16 = const()[name = string("op_2300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_2301_cast_fp16 = mul(x = var_2299_cast_fp16, y = var_2300_to_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2303_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2301_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp32 var_2304_epsilon_0 = const()[name = string("op_2304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_2304_cast_fp16 = rsqrt(epsilon = var_2304_epsilon_0, x = var_2303_cast_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_1_cast_fp16 = mul(x = y_9_cast_fp16, y = var_2304_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<fp16, [32, 256]> cos_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [32, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137426944))), scale = tensor<fp16, [32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137435200))))[name = string("cos_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 32, 256]> var_2306_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2306_cast_fp16")];
            tensor<int32, [4]> var_2311_begin_0 = const()[name = string("op_2311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2311_end_0 = const()[name = string("op_2311_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_2311_end_mask_0 = const()[name = string("op_2311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_2311_cast_fp16 = slice_by_index(begin = var_2311_begin_0, end = var_2311_end_0, end_mask = var_2311_end_mask_0, x = q_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = q_1_cast_fp16)[name = string("op_2317_cast_fp16")];
            int32 var_2319 = const()[name = string("op_2319"), val = int32(-1)];
            bool var_2320_interleave_0 = const()[name = string("op_2320_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_2320_cast_fp16 = concat(axis = var_2319, interleave = var_2320_interleave_0, values = (var_2312_cast_fp16, var_2317_cast_fp16))[name = string("op_2320_cast_fp16")];
            tensor<fp16, [32, 256]> sin_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [32, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137435328))), scale = tensor<fp16, [32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137443584))))[name = string("sin_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 32, 256]> var_2321_cast_fp16 = mul(x = var_2320_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_3_cast_fp16 = add(x = var_2306_cast_fp16, y = var_2321_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2324_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2329_begin_0 = const()[name = string("op_2329_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2329_end_0 = const()[name = string("op_2329_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_2329_end_mask_0 = const()[name = string("op_2329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, x = k_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_2330_cast_fp16 = mul(x = var_2329_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_2330_cast_fp16")];
            tensor<int32, [4]> var_2335_begin_0 = const()[name = string("op_2335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2335_end_0 = const()[name = string("op_2335_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_2335_end_mask_0 = const()[name = string("op_2335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_2335_cast_fp16 = slice_by_index(begin = var_2335_begin_0, end = var_2335_end_0, end_mask = var_2335_end_mask_0, x = k_1_cast_fp16)[name = string("op_2335_cast_fp16")];
            int32 var_2337 = const()[name = string("op_2337"), val = int32(-1)];
            bool var_2338_interleave_0 = const()[name = string("op_2338_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_2338_cast_fp16 = concat(axis = var_2337, interleave = var_2338_interleave_0, values = (var_2330_cast_fp16, var_2335_cast_fp16))[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2339_cast_fp16 = mul(x = var_2338_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2339_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_3_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2339_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = k_3_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [4, 1, 32, 256]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
            tensor<int32, [5]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_0_cast_fp16 = reshape(shape = concat_0, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [4, 1, 32, 256]> reshape_1_cast_fp16 = reshape(shape = concat_1, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = v_1_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [4, 1, 32, 256]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
            tensor<int32, [5]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_2_cast_fp16 = reshape(shape = concat_2, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [4, 1, 32, 256]> reshape_3_cast_fp16 = reshape(shape = concat_3, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2353_transpose_x_0 = const()[name = string("op_2353_transpose_x_0"), val = bool(false)];
            bool var_2353_transpose_y_0 = const()[name = string("op_2353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 4, 32, 32]> var_2353_cast_fp16 = matmul(transpose_x = var_2353_transpose_x_0, transpose_y = var_2353_transpose_y_0, x = q_3_cast_fp16, y = transpose_96_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_2355_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2356 = const()[name = string("op_2356"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_9_cast_fp16 = softmax(axis = var_2356, x = var_2355_cast_fp16)[name = string("w_9_cast_fp16")];
            bool var_2359_transpose_x_0 = const()[name = string("op_2359_transpose_x_0"), val = bool(false)];
            bool var_2359_transpose_y_0 = const()[name = string("op_2359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 4, 32, 256]> var_2359_cast_fp16 = matmul(transpose_x = var_2359_transpose_x_0, transpose_y = var_2359_transpose_y_0, x = w_9_cast_fp16, y = v_3_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<int32, [4]> var_2362_perm_0 = const()[name = string("op_2362_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2366 = const()[name = string("op_2366"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_2362_cast_fp16 = transpose(perm = var_2362_perm_0, x = var_2359_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 32, 1024]> input_3_cast_fp16 = reshape(shape = var_2366, x = var_2362_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7402112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7926464))))[name = string("tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2371_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_2371_cast_fp16")];
            bool var_2374 = const()[name = string("op_2374"), val = bool(true)];
            tensor<int32, [1]> var_2375_axes_0 = const()[name = string("op_2375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2375_cast_fp16 = reduce_max(axes = var_2375_axes_0, keep_dims = var_2374, x = var_2371_cast_fp16)[name = string("op_2375_cast_fp16")];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_11_cast_fp16 = maximum(x = var_2375_cast_fp16, y = var_2376_to_fp16)[name = string("s_11_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_11_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_11_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2379_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_2379_cast_fp16")];
            tensor<int32, [1]> var_2384_axes_0 = const()[name = string("op_2384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2384_keep_dims_0 = const()[name = string("op_2384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2384_cast_fp16 = reduce_mean(axes = var_2384_axes_0, keep_dims = var_2384_keep_dims_0, x = var_2379_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2385_cast_fp16 = mul(x = s_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2385_cast_fp16")];
            fp32 var_2386_epsilon_0 = const()[name = string("op_2386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2386_cast_fp16 = inverse(epsilon = var_2386_epsilon_0, x = var_2385_cast_fp16)[name = string("op_2386_cast_fp16")];
            fp16 var_2387_to_fp16 = const()[name = string("op_2387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2388_cast_fp16 = mul(x = var_2386_cast_fp16, y = var_2387_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2390_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2388_cast_fp16)[name = string("op_2390_cast_fp16")];
            fp32 var_2391_epsilon_0 = const()[name = string("op_2391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2391_cast_fp16 = rsqrt(epsilon = var_2391_epsilon_0, x = var_2390_cast_fp16)[name = string("op_2391_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_2391_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7927552)))];
            tensor<fp16, [1, 32, 512]> var_2393_cast_fp16 = mul(x = out_9_cast_fp16, y = tm_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_11_cast_fp16 = add(x = inputs_embeds, y = var_2393_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2396_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_2396_cast_fp16")];
            bool var_2399 = const()[name = string("op_2399"), val = bool(true)];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_2399, x = var_2396_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 var_2401_to_fp16 = const()[name = string("op_2401_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_13_cast_fp16 = maximum(x = var_2400_cast_fp16, y = var_2401_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = s_13_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2404_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [1]> var_2409_axes_0 = const()[name = string("op_2409_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2409_keep_dims_0 = const()[name = string("op_2409_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2409_cast_fp16 = reduce_mean(axes = var_2409_axes_0, keep_dims = var_2409_keep_dims_0, x = var_2404_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2410_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2411_cast_fp16 = inverse(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2413_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("op_2413_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2415_cast_fp16 = add(x = var_2409_cast_fp16, y = var_2413_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp32 var_2416_epsilon_0 = const()[name = string("op_2416_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2416_cast_fp16 = rsqrt(epsilon = var_2416_epsilon_0, x = var_2415_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_2416_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7928640)))];
            tensor<fp16, [1, 32, 512]> input_5_cast_fp16 = mul(x = out_11_cast_fp16, y = tm_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7929728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8978368))))[name = string("tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8982528)))];
            tensor<fp16, [1, 32, 2048]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_2423_mode_0 = const()[name = string("op_2423_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_2423_cast_fp16 = gelu(mode = var_2423_mode_0, x = linear_5_cast_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8986688))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10035328))))[name = string("tm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_7_cast_fp16 = mul(x = var_2423_cast_fp16, y = linear_6_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10039488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11088128))))[name = string("tm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2431_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_2431_cast_fp16")];
            bool var_2434 = const()[name = string("op_2434"), val = bool(true)];
            tensor<int32, [1]> var_2435_axes_0 = const()[name = string("op_2435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2435_cast_fp16 = reduce_max(axes = var_2435_axes_0, keep_dims = var_2434, x = var_2431_cast_fp16)[name = string("op_2435_cast_fp16")];
            fp16 var_2436_to_fp16 = const()[name = string("op_2436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_15_cast_fp16 = maximum(x = var_2435_cast_fp16, y = var_2436_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_15_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_15_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2439_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_2439_cast_fp16")];
            tensor<int32, [1]> var_2444_axes_0 = const()[name = string("op_2444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2444_keep_dims_0 = const()[name = string("op_2444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2444_cast_fp16 = reduce_mean(axes = var_2444_axes_0, keep_dims = var_2444_keep_dims_0, x = var_2439_cast_fp16)[name = string("op_2444_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2445_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_2445_cast_fp16")];
            fp32 var_2446_epsilon_0 = const()[name = string("op_2446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2446_cast_fp16 = inverse(epsilon = var_2446_epsilon_0, x = var_2445_cast_fp16)[name = string("op_2446_cast_fp16")];
            fp16 var_2447_to_fp16 = const()[name = string("op_2447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2448_cast_fp16 = mul(x = var_2446_cast_fp16, y = var_2447_to_fp16)[name = string("op_2448_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2450_cast_fp16 = add(x = var_2444_cast_fp16, y = var_2448_cast_fp16)[name = string("op_2450_cast_fp16")];
            fp32 var_2451_epsilon_0 = const()[name = string("op_2451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2451_cast_fp16 = rsqrt(epsilon = var_2451_epsilon_0, x = var_2450_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_2451_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11089216)))];
            tensor<fp16, [1, 32, 512]> var_2453_cast_fp16 = mul(x = out_13_cast_fp16, y = tm_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_9_cast_fp16 = add(x = x_11_cast_fp16, y = var_2453_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11090304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11352512))))[name = string("tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_8_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_8_cast_fp16")];
            string x_15_mode_0 = const()[name = string("x_15_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_15_cast_fp16 = gelu(mode = x_15_mode_0, x = linear_8_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([1, 32, 1, 512])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2473_squeeze_mask_0 = const()[name = string("op_2473_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, squeeze_mask = var_2473_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2473_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_11_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2473_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11353600))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11615808))))[name = string("tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2483_cast_fp16 = abs(x = linear_9_cast_fp16)[name = string("op_2483_cast_fp16")];
            bool var_2486 = const()[name = string("op_2486"), val = bool(true)];
            tensor<int32, [1]> var_2487_axes_0 = const()[name = string("op_2487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2487_cast_fp16 = reduce_max(axes = var_2487_axes_0, keep_dims = var_2486, x = var_2483_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp16 var_2488_to_fp16 = const()[name = string("op_2488_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_17_cast_fp16 = maximum(x = var_2487_cast_fp16, y = var_2488_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_17_cast_fp16 = real_div(x = linear_9_cast_fp16, y = s_17_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2491_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_2491_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2496_keep_dims_0 = const()[name = string("op_2496_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2496_cast_fp16 = reduce_mean(axes = var_2496_axes_0, keep_dims = var_2496_keep_dims_0, x = var_2491_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2497_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_2497_cast_fp16")];
            fp32 var_2498_epsilon_0 = const()[name = string("op_2498_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2498_cast_fp16 = inverse(epsilon = var_2498_epsilon_0, x = var_2497_cast_fp16)[name = string("op_2498_cast_fp16")];
            fp16 var_2499_to_fp16 = const()[name = string("op_2499_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2500_cast_fp16 = mul(x = var_2498_cast_fp16, y = var_2499_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2502_cast_fp16 = add(x = var_2496_cast_fp16, y = var_2500_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp32 var_2503_epsilon_0 = const()[name = string("op_2503_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2503_cast_fp16 = rsqrt(epsilon = var_2503_epsilon_0, x = var_2502_cast_fp16)[name = string("op_2503_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_2503_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11616896)))];
            tensor<fp16, [1, 32, 512]> var_2505_cast_fp16 = mul(x = out_15_cast_fp16, y = tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_1_cast_fp16 = add(x = input_9_cast_fp16, y = var_2505_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<fp16, [1]> tm_layers_0_layer_scalar_to_fp16 = const()[name = string("tm_layers_0_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9p-2])];
            tensor<fp16, [1, 32, 512]> x_19_cast_fp16 = mul(x = h_1_cast_fp16, y = tm_layers_0_layer_scalar_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2509_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool var_2512 = const()[name = string("op_2512"), val = bool(true)];
            tensor<int32, [1]> var_2513_axes_0 = const()[name = string("op_2513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2513_cast_fp16 = reduce_max(axes = var_2513_axes_0, keep_dims = var_2512, x = var_2509_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_19_cast_fp16 = maximum(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_19_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2517_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [1]> var_2522_axes_0 = const()[name = string("op_2522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2522_keep_dims_0 = const()[name = string("op_2522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2522_cast_fp16 = reduce_mean(axes = var_2522_axes_0, keep_dims = var_2522_keep_dims_0, x = var_2517_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2523_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_2523_cast_fp16")];
            fp32 var_2524_epsilon_0 = const()[name = string("op_2524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2524_cast_fp16 = inverse(epsilon = var_2524_epsilon_0, x = var_2523_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp16 var_2525_to_fp16 = const()[name = string("op_2525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2526_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_2525_to_fp16)[name = string("op_2526_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2528_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2526_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_2529_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11617984)))];
            tensor<fp16, [1, 32, 512]> input_13_cast_fp16 = mul(x = out_17_cast_fp16, y = tm_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11619072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12143424))))[name = string("tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_2540_cast_fp16 = reshape(shape = var_2539, x = linear_10_cast_fp16)[name = string("op_2540_cast_fp16")];
            tensor<int32, [4]> x_21_perm_0 = const()[name = string("x_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12145536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12407744))))[name = string("tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<int32, [4]> var_2551 = const()[name = string("op_2551"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_2552_cast_fp16 = reshape(shape = var_2551, x = linear_11_cast_fp16)[name = string("op_2552_cast_fp16")];
            tensor<int32, [4]> x_23_perm_0 = const()[name = string("x_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12408832))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671040))))[name = string("tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [4]> var_2563 = const()[name = string("op_2563"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_2564_cast_fp16 = reshape(shape = var_2563, x = linear_12_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = var_2540_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 4, 32, 256]> var_2568_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_2568_cast_fp16")];
            bool var_2571 = const()[name = string("op_2571"), val = bool(true)];
            tensor<int32, [1]> var_2572_axes_0 = const()[name = string("op_2572_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_2572_cast_fp16 = reduce_max(axes = var_2572_axes_0, keep_dims = var_2571, x = var_2568_cast_fp16)[name = string("op_2572_cast_fp16")];
            fp16 var_2573_to_fp16 = const()[name = string("op_2573_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_21_cast_fp16 = maximum(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_21_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_2576_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> var_2581_axes_0 = const()[name = string("op_2581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2581_keep_dims_0 = const()[name = string("op_2581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_2581_cast_fp16 = reduce_mean(axes = var_2581_axes_0, keep_dims = var_2581_keep_dims_0, x = var_2576_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_2582_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp32 var_2583_epsilon_0 = const()[name = string("op_2583_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_2583_cast_fp16 = inverse(epsilon = var_2583_epsilon_0, x = var_2582_cast_fp16)[name = string("op_2583_cast_fp16")];
            fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_2585_cast_fp16 = mul(x = var_2583_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_2587_cast_fp16 = add(x = var_2581_cast_fp16, y = var_2585_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_19_cast_fp16 = mul(x = y_21_cast_fp16, y = var_2588_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672128)))];
            tensor<fp16, [1, 4, 32, 256]> q_5_cast_fp16 = mul(x = out_19_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_23_cast_fp16 = transpose(perm = x_23_perm_0, x = var_2552_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2, 32, 256]> var_2591_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2594 = const()[name = string("op_2594"), val = bool(true)];
            tensor<int32, [1]> var_2595_axes_0 = const()[name = string("op_2595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_2595_cast_fp16 = reduce_max(axes = var_2595_axes_0, keep_dims = var_2594, x = var_2591_cast_fp16)[name = string("op_2595_cast_fp16")];
            fp16 var_2596_to_fp16 = const()[name = string("op_2596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_23_cast_fp16 = maximum(x = var_2595_cast_fp16, y = var_2596_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_23_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2599_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2599_cast_fp16")];
            tensor<int32, [1]> var_2604_axes_0 = const()[name = string("op_2604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2604_keep_dims_0 = const()[name = string("op_2604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_2604_cast_fp16 = reduce_mean(axes = var_2604_axes_0, keep_dims = var_2604_keep_dims_0, x = var_2599_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2605_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_2605_cast_fp16")];
            fp32 var_2606_epsilon_0 = const()[name = string("op_2606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_2606_cast_fp16 = inverse(epsilon = var_2606_epsilon_0, x = var_2605_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp16 var_2607_to_fp16 = const()[name = string("op_2607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_2608_cast_fp16 = mul(x = var_2606_cast_fp16, y = var_2607_to_fp16)[name = string("op_2608_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2610_cast_fp16 = add(x = var_2604_cast_fp16, y = var_2608_cast_fp16)[name = string("op_2610_cast_fp16")];
            fp32 var_2611_epsilon_0 = const()[name = string("op_2611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_2611_cast_fp16 = rsqrt(epsilon = var_2611_epsilon_0, x = var_2610_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_21_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2611_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672704)))];
            tensor<fp16, [1, 2, 32, 256]> k_7_cast_fp16 = mul(x = out_21_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_2564_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 2, 32, 256]> var_2614_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_2614_cast_fp16")];
            bool var_2617 = const()[name = string("op_2617"), val = bool(true)];
            tensor<int32, [1]> var_2618_axes_0 = const()[name = string("op_2618_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_2618_cast_fp16 = reduce_max(axes = var_2618_axes_0, keep_dims = var_2617, x = var_2614_cast_fp16)[name = string("op_2618_cast_fp16")];
            fp16 var_2619_to_fp16 = const()[name = string("op_2619_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_25_cast_fp16 = maximum(x = var_2618_cast_fp16, y = var_2619_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_25_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2622_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2622_cast_fp16")];
            tensor<int32, [1]> var_2627_axes_0 = const()[name = string("op_2627_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2627_keep_dims_0 = const()[name = string("op_2627_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_2627_cast_fp16 = reduce_mean(axes = var_2627_axes_0, keep_dims = var_2627_keep_dims_0, x = var_2622_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2628_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_2628_cast_fp16")];
            fp32 var_2629_epsilon_0 = const()[name = string("op_2629_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_2629_cast_fp16 = inverse(epsilon = var_2629_epsilon_0, x = var_2628_cast_fp16)[name = string("op_2629_cast_fp16")];
            fp16 var_2630_to_fp16 = const()[name = string("op_2630_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_2631_cast_fp16 = mul(x = var_2629_cast_fp16, y = var_2630_to_fp16)[name = string("op_2631_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2633_cast_fp16 = add(x = var_2627_cast_fp16, y = var_2631_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_2634_cast_fp16 = rsqrt(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_5_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2634_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_2636_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2636_cast_fp16")];
            tensor<int32, [4]> var_2641_begin_0 = const()[name = string("op_2641_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2641_end_0 = const()[name = string("op_2641_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_2641_end_mask_0 = const()[name = string("op_2641_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_2641_cast_fp16 = slice_by_index(begin = var_2641_begin_0, end = var_2641_end_0, end_mask = var_2641_end_mask_0, x = q_5_cast_fp16)[name = string("op_2641_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_2642_cast_fp16 = mul(x = var_2641_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_2642_cast_fp16")];
            tensor<int32, [4]> var_2647_begin_0 = const()[name = string("op_2647_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2647_end_0 = const()[name = string("op_2647_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_2647_end_mask_0 = const()[name = string("op_2647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_2647_cast_fp16 = slice_by_index(begin = var_2647_begin_0, end = var_2647_end_0, end_mask = var_2647_end_mask_0, x = q_5_cast_fp16)[name = string("op_2647_cast_fp16")];
            int32 var_2649 = const()[name = string("op_2649"), val = int32(-1)];
            bool var_2650_interleave_0 = const()[name = string("op_2650_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_2650_cast_fp16 = concat(axis = var_2649, interleave = var_2650_interleave_0, values = (var_2642_cast_fp16, var_2647_cast_fp16))[name = string("op_2650_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2651_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_7_cast_fp16 = add(x = var_2636_cast_fp16, y = var_2651_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2654_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = k_7_cast_fp16)[name = string("op_2659_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_2660_cast_fp16 = mul(x = var_2659_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2660_cast_fp16")];
            tensor<int32, [4]> var_2665_begin_0 = const()[name = string("op_2665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2665_end_0 = const()[name = string("op_2665_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_2665_end_mask_0 = const()[name = string("op_2665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_2665_cast_fp16 = slice_by_index(begin = var_2665_begin_0, end = var_2665_end_0, end_mask = var_2665_end_mask_0, x = k_7_cast_fp16)[name = string("op_2665_cast_fp16")];
            int32 var_2667 = const()[name = string("op_2667"), val = int32(-1)];
            bool var_2668_interleave_0 = const()[name = string("op_2668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_2668_cast_fp16 = concat(axis = var_2667, interleave = var_2668_interleave_0, values = (var_2660_cast_fp16, var_2665_cast_fp16))[name = string("op_2668_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2669_cast_fp16 = mul(x = var_2668_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_9_cast_fp16 = add(x = var_2654_cast_fp16, y = var_2669_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = k_9_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [4, 1, 32, 256]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_4_cast_fp16 = reshape(shape = concat_4, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [4, 1, 32, 256]> reshape_5_cast_fp16 = reshape(shape = concat_5, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = v_5_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [4, 1, 32, 256]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_6_cast_fp16 = reshape(shape = concat_6, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [4, 1, 32, 256]> reshape_7_cast_fp16 = reshape(shape = concat_7, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2683_transpose_x_0 = const()[name = string("op_2683_transpose_x_0"), val = bool(false)];
            bool var_2683_transpose_y_0 = const()[name = string("op_2683_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_97_cast_fp16 = transpose(perm = transpose_97_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 4, 32, 32]> var_2683_cast_fp16 = matmul(transpose_x = var_2683_transpose_x_0, transpose_y = var_2683_transpose_y_0, x = q_7_cast_fp16, y = transpose_97_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_2685_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2685_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_25_cast_fp16 = softmax(axis = var_2686, x = var_2685_cast_fp16)[name = string("w_25_cast_fp16")];
            bool var_2689_transpose_x_0 = const()[name = string("op_2689_transpose_x_0"), val = bool(false)];
            bool var_2689_transpose_y_0 = const()[name = string("op_2689_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 4, 32, 256]> var_2689_cast_fp16 = matmul(transpose_x = var_2689_transpose_x_0, transpose_y = var_2689_transpose_y_0, x = w_25_cast_fp16, y = v_7_cast_fp16)[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> var_2692_perm_0 = const()[name = string("op_2692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2696 = const()[name = string("op_2696"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_2692_cast_fp16 = transpose(perm = var_2692_perm_0, x = var_2689_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 32, 1024]> input_15_cast_fp16 = reshape(shape = var_2696, x = var_2692_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13197632))))[name = string("tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2701_cast_fp16 = abs(x = linear_13_cast_fp16)[name = string("op_2701_cast_fp16")];
            bool var_2704 = const()[name = string("op_2704"), val = bool(true)];
            tensor<int32, [1]> var_2705_axes_0 = const()[name = string("op_2705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2705_cast_fp16 = reduce_max(axes = var_2705_axes_0, keep_dims = var_2704, x = var_2701_cast_fp16)[name = string("op_2705_cast_fp16")];
            fp16 var_2706_to_fp16 = const()[name = string("op_2706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_27_cast_fp16 = maximum(x = var_2705_cast_fp16, y = var_2706_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_27_cast_fp16 = real_div(x = linear_13_cast_fp16, y = s_27_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2709_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2709_cast_fp16")];
            tensor<int32, [1]> var_2714_axes_0 = const()[name = string("op_2714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2714_keep_dims_0 = const()[name = string("op_2714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2714_cast_fp16 = reduce_mean(axes = var_2714_axes_0, keep_dims = var_2714_keep_dims_0, x = var_2709_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2715_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2715_cast_fp16")];
            fp32 var_2716_epsilon_0 = const()[name = string("op_2716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2716_cast_fp16 = inverse(epsilon = var_2716_epsilon_0, x = var_2715_cast_fp16)[name = string("op_2716_cast_fp16")];
            fp16 var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2718_cast_fp16 = mul(x = var_2716_cast_fp16, y = var_2717_to_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2720_cast_fp16 = add(x = var_2714_cast_fp16, y = var_2718_cast_fp16)[name = string("op_2720_cast_fp16")];
            fp32 var_2721_epsilon_0 = const()[name = string("op_2721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2721_cast_fp16 = rsqrt(epsilon = var_2721_epsilon_0, x = var_2720_cast_fp16)[name = string("op_2721_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2721_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13198720)))];
            tensor<fp16, [1, 32, 512]> var_2723_cast_fp16 = mul(x = out_23_cast_fp16, y = tm_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2723_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_29_cast_fp16 = add(x = x_19_cast_fp16, y = var_2723_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2726_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_2726_cast_fp16")];
            bool var_2729 = const()[name = string("op_2729"), val = bool(true)];
            tensor<int32, [1]> var_2730_axes_0 = const()[name = string("op_2730_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2730_cast_fp16 = reduce_max(axes = var_2730_axes_0, keep_dims = var_2729, x = var_2726_cast_fp16)[name = string("op_2730_cast_fp16")];
            fp16 var_2731_to_fp16 = const()[name = string("op_2731_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_29_cast_fp16 = maximum(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_29_cast_fp16 = real_div(x = x_29_cast_fp16, y = s_29_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2734_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2734_cast_fp16")];
            tensor<int32, [1]> var_2739_axes_0 = const()[name = string("op_2739_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2739_keep_dims_0 = const()[name = string("op_2739_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2739_cast_fp16 = reduce_mean(axes = var_2739_axes_0, keep_dims = var_2739_keep_dims_0, x = var_2734_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2740_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2740_cast_fp16")];
            fp32 var_2741_epsilon_0 = const()[name = string("op_2741_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2741_cast_fp16 = inverse(epsilon = var_2741_epsilon_0, x = var_2740_cast_fp16)[name = string("op_2741_cast_fp16")];
            fp16 var_2742_to_fp16 = const()[name = string("op_2742_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2743_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_2742_to_fp16)[name = string("op_2743_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2745_cast_fp16 = add(x = var_2739_cast_fp16, y = var_2743_cast_fp16)[name = string("op_2745_cast_fp16")];
            fp32 var_2746_epsilon_0 = const()[name = string("op_2746_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2746_cast_fp16 = rsqrt(epsilon = var_2746_epsilon_0, x = var_2745_cast_fp16)[name = string("op_2746_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2746_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13199808)))];
            tensor<fp16, [1, 32, 512]> input_17_cast_fp16 = mul(x = out_25_cast_fp16, y = tm_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13200896))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14249536))))[name = string("tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_14_cast_fp16")];
            string var_2753_mode_0 = const()[name = string("op_2753_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_2753_cast_fp16 = gelu(mode = var_2753_mode_0, x = linear_14_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14253696))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15302336))))[name = string("tm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_15_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_19_cast_fp16 = mul(x = var_2753_cast_fp16, y = linear_15_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15306496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16355136))))[name = string("tm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2761_cast_fp16 = abs(x = linear_16_cast_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2764 = const()[name = string("op_2764"), val = bool(true)];
            tensor<int32, [1]> var_2765_axes_0 = const()[name = string("op_2765_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2765_cast_fp16 = reduce_max(axes = var_2765_axes_0, keep_dims = var_2764, x = var_2761_cast_fp16)[name = string("op_2765_cast_fp16")];
            fp16 var_2766_to_fp16 = const()[name = string("op_2766_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_31_cast_fp16 = maximum(x = var_2765_cast_fp16, y = var_2766_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_31_cast_fp16 = real_div(x = linear_16_cast_fp16, y = s_31_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2769_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [1]> var_2774_axes_0 = const()[name = string("op_2774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2774_keep_dims_0 = const()[name = string("op_2774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2774_cast_fp16 = reduce_mean(axes = var_2774_axes_0, keep_dims = var_2774_keep_dims_0, x = var_2769_cast_fp16)[name = string("op_2774_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2775_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp32 var_2776_epsilon_0 = const()[name = string("op_2776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2776_cast_fp16 = inverse(epsilon = var_2776_epsilon_0, x = var_2775_cast_fp16)[name = string("op_2776_cast_fp16")];
            fp16 var_2777_to_fp16 = const()[name = string("op_2777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2778_cast_fp16 = mul(x = var_2776_cast_fp16, y = var_2777_to_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2780_cast_fp16 = add(x = var_2774_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2780_cast_fp16")];
            fp32 var_2781_epsilon_0 = const()[name = string("op_2781_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2781_cast_fp16 = rsqrt(epsilon = var_2781_epsilon_0, x = var_2780_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2781_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16356224)))];
            tensor<fp16, [1, 32, 512]> var_2783_cast_fp16 = mul(x = out_27_cast_fp16, y = tm_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2783_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_21_cast_fp16 = add(x = x_29_cast_fp16, y = var_2783_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16357312))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16619520))))[name = string("tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_17_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            string x_33_mode_0 = const()[name = string("x_33_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_33_cast_fp16 = gelu(mode = x_33_mode_0, x = linear_17_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> var_2803_begin_0 = const()[name = string("op_2803_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2803_end_0 = const()[name = string("op_2803_end_0"), val = tensor<int32, [4]>([1, 32, 2, 512])];
            tensor<bool, [4]> var_2803_end_mask_0 = const()[name = string("op_2803_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2803_squeeze_mask_0 = const()[name = string("op_2803_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_2803_cast_fp16 = slice_by_index(begin = var_2803_begin_0, end = var_2803_end_0, end_mask = var_2803_end_mask_0, squeeze_mask = var_2803_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_23_cast_fp16 = mul(x = x_33_cast_fp16, y = var_2803_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16620608))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16882816))))[name = string("tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2813_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2813_cast_fp16")];
            bool var_2816 = const()[name = string("op_2816"), val = bool(true)];
            tensor<int32, [1]> var_2817_axes_0 = const()[name = string("op_2817_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2817_cast_fp16 = reduce_max(axes = var_2817_axes_0, keep_dims = var_2816, x = var_2813_cast_fp16)[name = string("op_2817_cast_fp16")];
            fp16 var_2818_to_fp16 = const()[name = string("op_2818_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_33_cast_fp16 = maximum(x = var_2817_cast_fp16, y = var_2818_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_33_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_33_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2821_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<int32, [1]> var_2826_axes_0 = const()[name = string("op_2826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2826_keep_dims_0 = const()[name = string("op_2826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2826_cast_fp16 = reduce_mean(axes = var_2826_axes_0, keep_dims = var_2826_keep_dims_0, x = var_2821_cast_fp16)[name = string("op_2826_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2827_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp32 var_2828_epsilon_0 = const()[name = string("op_2828_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2828_cast_fp16 = inverse(epsilon = var_2828_epsilon_0, x = var_2827_cast_fp16)[name = string("op_2828_cast_fp16")];
            fp16 var_2829_to_fp16 = const()[name = string("op_2829_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2830_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_2829_to_fp16)[name = string("op_2830_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2832_cast_fp16 = add(x = var_2826_cast_fp16, y = var_2830_cast_fp16)[name = string("op_2832_cast_fp16")];
            fp32 var_2833_epsilon_0 = const()[name = string("op_2833_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2833_cast_fp16 = rsqrt(epsilon = var_2833_epsilon_0, x = var_2832_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2833_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16883904)))];
            tensor<fp16, [1, 32, 512]> var_2835_cast_fp16 = mul(x = out_29_cast_fp16, y = tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_3_cast_fp16 = add(x = input_21_cast_fp16, y = var_2835_cast_fp16)[name = string("h_3_cast_fp16")];
            tensor<fp16, [1]> tm_layers_1_layer_scalar_to_fp16 = const()[name = string("tm_layers_1_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.5ap-3])];
            tensor<fp16, [1, 32, 512]> x_37_cast_fp16 = mul(x = h_3_cast_fp16, y = tm_layers_1_layer_scalar_to_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2839_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2839_cast_fp16")];
            bool var_2842 = const()[name = string("op_2842"), val = bool(true)];
            tensor<int32, [1]> var_2843_axes_0 = const()[name = string("op_2843_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2843_cast_fp16 = reduce_max(axes = var_2843_axes_0, keep_dims = var_2842, x = var_2839_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp16 var_2844_to_fp16 = const()[name = string("op_2844_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_35_cast_fp16 = maximum(x = var_2843_cast_fp16, y = var_2844_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_35_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_35_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_2847_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2847_cast_fp16")];
            tensor<int32, [1]> var_2852_axes_0 = const()[name = string("op_2852_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2852_keep_dims_0 = const()[name = string("op_2852_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_2852_cast_fp16 = reduce_mean(axes = var_2852_axes_0, keep_dims = var_2852_keep_dims_0, x = var_2847_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2853_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp32 var_2854_epsilon_0 = const()[name = string("op_2854_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2854_cast_fp16 = inverse(epsilon = var_2854_epsilon_0, x = var_2853_cast_fp16)[name = string("op_2854_cast_fp16")];
            fp16 var_2855_to_fp16 = const()[name = string("op_2855_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2856_cast_fp16 = mul(x = var_2854_cast_fp16, y = var_2855_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2858_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2856_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp32 var_2859_epsilon_0 = const()[name = string("op_2859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2859_cast_fp16 = rsqrt(epsilon = var_2859_epsilon_0, x = var_2858_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_31_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2859_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16884992)))];
            tensor<fp16, [1, 32, 512]> input_25_cast_fp16 = mul(x = out_31_cast_fp16, y = tm_layers_2_input_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16886080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17410432))))[name = string("tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [4]> var_2869 = const()[name = string("op_2869"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_2870_cast_fp16 = reshape(shape = var_2869, x = linear_19_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17412544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17674752))))[name = string("tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [4]> var_2881 = const()[name = string("op_2881"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_2882_cast_fp16 = reshape(shape = var_2881, x = linear_20_cast_fp16)[name = string("op_2882_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17675840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17938048))))[name = string("tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_2894_cast_fp16 = reshape(shape = var_2893, x = linear_21_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [4]> x_43_perm_0 = const()[name = string("x_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2870_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 4, 32, 256]> var_2898_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2898_cast_fp16")];
            bool var_2901 = const()[name = string("op_2901"), val = bool(true)];
            tensor<int32, [1]> var_2902_axes_0 = const()[name = string("op_2902_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_2902_cast_fp16 = reduce_max(axes = var_2902_axes_0, keep_dims = var_2901, x = var_2898_cast_fp16)[name = string("op_2902_cast_fp16")];
            fp16 var_2903_to_fp16 = const()[name = string("op_2903_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_37_cast_fp16 = maximum(x = var_2902_cast_fp16, y = var_2903_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_37_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_37_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_2906_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<int32, [1]> var_2911_axes_0 = const()[name = string("op_2911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2911_keep_dims_0 = const()[name = string("op_2911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_2911_cast_fp16 = reduce_mean(axes = var_2911_axes_0, keep_dims = var_2911_keep_dims_0, x = var_2906_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_2912_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2912_cast_fp16")];
            fp32 var_2913_epsilon_0 = const()[name = string("op_2913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_2913_cast_fp16 = inverse(epsilon = var_2913_epsilon_0, x = var_2912_cast_fp16)[name = string("op_2913_cast_fp16")];
            fp16 var_2914_to_fp16 = const()[name = string("op_2914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_2915_cast_fp16 = mul(x = var_2913_cast_fp16, y = var_2914_to_fp16)[name = string("op_2915_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_2917_cast_fp16 = add(x = var_2911_cast_fp16, y = var_2915_cast_fp16)[name = string("op_2917_cast_fp16")];
            fp32 var_2918_epsilon_0 = const()[name = string("op_2918_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_2918_cast_fp16 = rsqrt(epsilon = var_2918_epsilon_0, x = var_2917_cast_fp16)[name = string("op_2918_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_33_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2918_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_9_cast_fp16 = mul(x = out_33_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2882_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 2, 32, 256]> var_2921_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2921_cast_fp16")];
            bool var_2924 = const()[name = string("op_2924"), val = bool(true)];
            tensor<int32, [1]> var_2925_axes_0 = const()[name = string("op_2925_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_2925_cast_fp16 = reduce_max(axes = var_2925_axes_0, keep_dims = var_2924, x = var_2921_cast_fp16)[name = string("op_2925_cast_fp16")];
            fp16 var_2926_to_fp16 = const()[name = string("op_2926_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_39_cast_fp16 = maximum(x = var_2925_cast_fp16, y = var_2926_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_39_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_39_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2929_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2929_cast_fp16")];
            tensor<int32, [1]> var_2934_axes_0 = const()[name = string("op_2934_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2934_keep_dims_0 = const()[name = string("op_2934_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_2934_cast_fp16 = reduce_mean(axes = var_2934_axes_0, keep_dims = var_2934_keep_dims_0, x = var_2929_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2935_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2935_cast_fp16")];
            fp32 var_2936_epsilon_0 = const()[name = string("op_2936_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_2936_cast_fp16 = inverse(epsilon = var_2936_epsilon_0, x = var_2935_cast_fp16)[name = string("op_2936_cast_fp16")];
            fp16 var_2937_to_fp16 = const()[name = string("op_2937_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_2938_cast_fp16 = mul(x = var_2936_cast_fp16, y = var_2937_to_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2940_cast_fp16 = add(x = var_2934_cast_fp16, y = var_2938_cast_fp16)[name = string("op_2940_cast_fp16")];
            fp32 var_2941_epsilon_0 = const()[name = string("op_2941_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_2941_cast_fp16 = rsqrt(epsilon = var_2941_epsilon_0, x = var_2940_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_35_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2941_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_13_cast_fp16 = mul(x = out_35_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_43_cast_fp16 = transpose(perm = x_43_perm_0, x = var_2894_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 2, 32, 256]> var_2944_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2944_cast_fp16")];
            bool var_2947 = const()[name = string("op_2947"), val = bool(true)];
            tensor<int32, [1]> var_2948_axes_0 = const()[name = string("op_2948_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_2948_cast_fp16 = reduce_max(axes = var_2948_axes_0, keep_dims = var_2947, x = var_2944_cast_fp16)[name = string("op_2948_cast_fp16")];
            fp16 var_2949_to_fp16 = const()[name = string("op_2949_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_41_cast_fp16 = maximum(x = var_2948_cast_fp16, y = var_2949_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_41_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_41_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2952_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2952_cast_fp16")];
            tensor<int32, [1]> var_2957_axes_0 = const()[name = string("op_2957_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2957_keep_dims_0 = const()[name = string("op_2957_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_2957_cast_fp16 = reduce_mean(axes = var_2957_axes_0, keep_dims = var_2957_keep_dims_0, x = var_2952_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2958_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2958_cast_fp16")];
            fp32 var_2959_epsilon_0 = const()[name = string("op_2959_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_2959_cast_fp16 = inverse(epsilon = var_2959_epsilon_0, x = var_2958_cast_fp16)[name = string("op_2959_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_2961_cast_fp16 = mul(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_2963_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2961_cast_fp16)[name = string("op_2963_cast_fp16")];
            fp32 var_2964_epsilon_0 = const()[name = string("op_2964_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_2964_cast_fp16 = rsqrt(epsilon = var_2964_epsilon_0, x = var_2963_cast_fp16)[name = string("op_2964_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_9_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2964_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_2966_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2966_cast_fp16")];
            tensor<int32, [4]> var_2971_begin_0 = const()[name = string("op_2971_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2971_end_0 = const()[name = string("op_2971_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_2971_end_mask_0 = const()[name = string("op_2971_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_2971_cast_fp16 = slice_by_index(begin = var_2971_begin_0, end = var_2971_end_0, end_mask = var_2971_end_mask_0, x = q_9_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_2972_cast_fp16 = mul(x = var_2971_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2972_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            int32 var_2979 = const()[name = string("op_2979"), val = int32(-1)];
            bool var_2980_interleave_0 = const()[name = string("op_2980_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_2980_cast_fp16 = concat(axis = var_2979, interleave = var_2980_interleave_0, values = (var_2972_cast_fp16, var_2977_cast_fp16))[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_2981_cast_fp16 = mul(x = var_2980_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2981_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_11_cast_fp16 = add(x = var_2966_cast_fp16, y = var_2981_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2984_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2989_begin_0 = const()[name = string("op_2989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2989_end_0 = const()[name = string("op_2989_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_2989_end_mask_0 = const()[name = string("op_2989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_2989_cast_fp16 = slice_by_index(begin = var_2989_begin_0, end = var_2989_end_0, end_mask = var_2989_end_mask_0, x = k_13_cast_fp16)[name = string("op_2989_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_2990_cast_fp16 = mul(x = var_2989_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2990_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = k_13_cast_fp16)[name = string("op_2995_cast_fp16")];
            int32 var_2997 = const()[name = string("op_2997"), val = int32(-1)];
            bool var_2998_interleave_0 = const()[name = string("op_2998_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_2998_cast_fp16 = concat(axis = var_2997, interleave = var_2998_interleave_0, values = (var_2990_cast_fp16, var_2995_cast_fp16))[name = string("op_2998_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_2999_cast_fp16 = mul(x = var_2998_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2999_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_15_cast_fp16 = add(x = var_2984_cast_fp16, y = var_2999_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = k_15_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [4, 1, 32, 256]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
            tensor<int32, [5]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_8_cast_fp16 = reshape(shape = concat_8, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [4, 1, 32, 256]> reshape_9_cast_fp16 = reshape(shape = concat_9, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = v_9_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [4, 1, 32, 256]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
            tensor<int32, [5]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_10_cast_fp16 = reshape(shape = concat_10, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [4, 1, 32, 256]> reshape_11_cast_fp16 = reshape(shape = concat_11, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3013_transpose_x_0 = const()[name = string("op_3013_transpose_x_0"), val = bool(false)];
            bool var_3013_transpose_y_0 = const()[name = string("op_3013_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 4, 32, 32]> var_3013_cast_fp16 = matmul(transpose_x = var_3013_transpose_x_0, transpose_y = var_3013_transpose_y_0, x = q_11_cast_fp16, y = transpose_98_cast_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_3015_cast_fp16 = add(x = var_3013_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3015_cast_fp16")];
            int32 var_3016 = const()[name = string("op_3016"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_41_cast_fp16 = softmax(axis = var_3016, x = var_3015_cast_fp16)[name = string("w_41_cast_fp16")];
            bool var_3019_transpose_x_0 = const()[name = string("op_3019_transpose_x_0"), val = bool(false)];
            bool var_3019_transpose_y_0 = const()[name = string("op_3019_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 4, 32, 256]> var_3019_cast_fp16 = matmul(transpose_x = var_3019_transpose_x_0, transpose_y = var_3019_transpose_y_0, x = w_41_cast_fp16, y = v_11_cast_fp16)[name = string("op_3019_cast_fp16")];
            tensor<int32, [4]> var_3022_perm_0 = const()[name = string("op_3022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3026 = const()[name = string("op_3026"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_3022_cast_fp16 = transpose(perm = var_3022_perm_0, x = var_3019_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 32, 1024]> input_27_cast_fp16 = reshape(shape = var_3026, x = var_3022_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17939136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18463488))))[name = string("tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3031_cast_fp16 = abs(x = linear_22_cast_fp16)[name = string("op_3031_cast_fp16")];
            bool var_3034 = const()[name = string("op_3034"), val = bool(true)];
            tensor<int32, [1]> var_3035_axes_0 = const()[name = string("op_3035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3035_cast_fp16 = reduce_max(axes = var_3035_axes_0, keep_dims = var_3034, x = var_3031_cast_fp16)[name = string("op_3035_cast_fp16")];
            fp16 var_3036_to_fp16 = const()[name = string("op_3036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_43_cast_fp16 = maximum(x = var_3035_cast_fp16, y = var_3036_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_43_cast_fp16 = real_div(x = linear_22_cast_fp16, y = s_43_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3039_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3039_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3045_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_3045_cast_fp16")];
            fp32 var_3046_epsilon_0 = const()[name = string("op_3046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3046_cast_fp16 = inverse(epsilon = var_3046_epsilon_0, x = var_3045_cast_fp16)[name = string("op_3046_cast_fp16")];
            fp16 var_3047_to_fp16 = const()[name = string("op_3047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3048_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3047_to_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3050_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3048_cast_fp16)[name = string("op_3050_cast_fp16")];
            fp32 var_3051_epsilon_0 = const()[name = string("op_3051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3051_cast_fp16 = rsqrt(epsilon = var_3051_epsilon_0, x = var_3050_cast_fp16)[name = string("op_3051_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_3051_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18464576)))];
            tensor<fp16, [1, 32, 512]> var_3053_cast_fp16 = mul(x = out_37_cast_fp16, y = tm_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = var_3053_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3056_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_3056_cast_fp16")];
            bool var_3059 = const()[name = string("op_3059"), val = bool(true)];
            tensor<int32, [1]> var_3060_axes_0 = const()[name = string("op_3060_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3060_cast_fp16 = reduce_max(axes = var_3060_axes_0, keep_dims = var_3059, x = var_3056_cast_fp16)[name = string("op_3060_cast_fp16")];
            fp16 var_3061_to_fp16 = const()[name = string("op_3061_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_45_cast_fp16 = maximum(x = var_3060_cast_fp16, y = var_3061_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_45_cast_fp16 = real_div(x = x_47_cast_fp16, y = s_45_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3064_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<int32, [1]> var_3069_axes_0 = const()[name = string("op_3069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3069_keep_dims_0 = const()[name = string("op_3069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3069_cast_fp16 = reduce_mean(axes = var_3069_axes_0, keep_dims = var_3069_keep_dims_0, x = var_3064_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3070_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp32 var_3071_epsilon_0 = const()[name = string("op_3071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3071_cast_fp16 = inverse(epsilon = var_3071_epsilon_0, x = var_3070_cast_fp16)[name = string("op_3071_cast_fp16")];
            fp16 var_3072_to_fp16 = const()[name = string("op_3072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3073_cast_fp16 = mul(x = var_3071_cast_fp16, y = var_3072_to_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3075_cast_fp16 = add(x = var_3069_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3075_cast_fp16")];
            fp32 var_3076_epsilon_0 = const()[name = string("op_3076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3076_cast_fp16 = rsqrt(epsilon = var_3076_epsilon_0, x = var_3075_cast_fp16)[name = string("op_3076_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_3076_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18465664)))];
            tensor<fp16, [1, 32, 512]> input_29_cast_fp16 = mul(x = out_39_cast_fp16, y = tm_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18466752))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19515392))))[name = string("tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_23_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_23_cast_fp16")];
            string var_3083_mode_0 = const()[name = string("op_3083_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_3083_cast_fp16 = gelu(mode = var_3083_mode_0, x = linear_23_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19519552))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20568192))))[name = string("tm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_24_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_31_cast_fp16 = mul(x = var_3083_cast_fp16, y = linear_24_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20572352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21620992))))[name = string("tm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3091_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_3091_cast_fp16")];
            bool var_3094 = const()[name = string("op_3094"), val = bool(true)];
            tensor<int32, [1]> var_3095_axes_0 = const()[name = string("op_3095_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3095_cast_fp16 = reduce_max(axes = var_3095_axes_0, keep_dims = var_3094, x = var_3091_cast_fp16)[name = string("op_3095_cast_fp16")];
            fp16 var_3096_to_fp16 = const()[name = string("op_3096_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_47_cast_fp16 = maximum(x = var_3095_cast_fp16, y = var_3096_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_47_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_47_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3099_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3104_keep_dims_0 = const()[name = string("op_3104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3104_cast_fp16 = reduce_mean(axes = var_3104_axes_0, keep_dims = var_3104_keep_dims_0, x = var_3099_cast_fp16)[name = string("op_3104_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3105_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_3105_cast_fp16")];
            fp32 var_3106_epsilon_0 = const()[name = string("op_3106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3106_cast_fp16 = inverse(epsilon = var_3106_epsilon_0, x = var_3105_cast_fp16)[name = string("op_3106_cast_fp16")];
            fp16 var_3107_to_fp16 = const()[name = string("op_3107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3108_cast_fp16 = mul(x = var_3106_cast_fp16, y = var_3107_to_fp16)[name = string("op_3108_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3110_cast_fp16 = add(x = var_3104_cast_fp16, y = var_3108_cast_fp16)[name = string("op_3110_cast_fp16")];
            fp32 var_3111_epsilon_0 = const()[name = string("op_3111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3111_cast_fp16 = rsqrt(epsilon = var_3111_epsilon_0, x = var_3110_cast_fp16)[name = string("op_3111_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_3111_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21622080)))];
            tensor<fp16, [1, 32, 512]> var_3113_cast_fp16 = mul(x = out_41_cast_fp16, y = tm_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_33_cast_fp16 = add(x = x_47_cast_fp16, y = var_3113_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21623168))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21885376))))[name = string("tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_26_cast_fp16")];
            string x_51_mode_0 = const()[name = string("x_51_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_51_cast_fp16 = gelu(mode = x_51_mode_0, x = linear_26_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([1, 32, 3, 512])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3133_squeeze_mask_0 = const()[name = string("op_3133_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, squeeze_mask = var_3133_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_35_cast_fp16 = mul(x = x_51_cast_fp16, y = var_3133_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21886464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22148672))))[name = string("tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3143_cast_fp16 = abs(x = linear_27_cast_fp16)[name = string("op_3143_cast_fp16")];
            bool var_3146 = const()[name = string("op_3146"), val = bool(true)];
            tensor<int32, [1]> var_3147_axes_0 = const()[name = string("op_3147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3147_cast_fp16 = reduce_max(axes = var_3147_axes_0, keep_dims = var_3146, x = var_3143_cast_fp16)[name = string("op_3147_cast_fp16")];
            fp16 var_3148_to_fp16 = const()[name = string("op_3148_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_49_cast_fp16 = maximum(x = var_3147_cast_fp16, y = var_3148_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_49_cast_fp16 = real_div(x = linear_27_cast_fp16, y = s_49_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3151_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_3151_cast_fp16")];
            tensor<int32, [1]> var_3156_axes_0 = const()[name = string("op_3156_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3156_keep_dims_0 = const()[name = string("op_3156_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3156_cast_fp16 = reduce_mean(axes = var_3156_axes_0, keep_dims = var_3156_keep_dims_0, x = var_3151_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3157_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_3157_cast_fp16")];
            fp32 var_3158_epsilon_0 = const()[name = string("op_3158_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3158_cast_fp16 = inverse(epsilon = var_3158_epsilon_0, x = var_3157_cast_fp16)[name = string("op_3158_cast_fp16")];
            fp16 var_3159_to_fp16 = const()[name = string("op_3159_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3160_cast_fp16 = mul(x = var_3158_cast_fp16, y = var_3159_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3162_cast_fp16 = add(x = var_3156_cast_fp16, y = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp32 var_3163_epsilon_0 = const()[name = string("op_3163_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3163_cast_fp16 = rsqrt(epsilon = var_3163_epsilon_0, x = var_3162_cast_fp16)[name = string("op_3163_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_43_cast_fp16 = mul(x = y_49_cast_fp16, y = var_3163_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22149760)))];
            tensor<fp16, [1, 32, 512]> var_3165_cast_fp16 = mul(x = out_43_cast_fp16, y = tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_5_cast_fp16 = add(x = input_33_cast_fp16, y = var_3165_cast_fp16)[name = string("h_5_cast_fp16")];
            tensor<fp16, [1]> tm_layers_2_layer_scalar_to_fp16 = const()[name = string("tm_layers_2_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ep-1])];
            tensor<fp16, [1, 32, 512]> x_55_cast_fp16 = mul(x = h_5_cast_fp16, y = tm_layers_2_layer_scalar_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3169_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_3169_cast_fp16")];
            bool var_3172 = const()[name = string("op_3172"), val = bool(true)];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3173_cast_fp16 = reduce_max(axes = var_3173_axes_0, keep_dims = var_3172, x = var_3169_cast_fp16)[name = string("op_3173_cast_fp16")];
            fp16 var_3174_to_fp16 = const()[name = string("op_3174_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_51_cast_fp16 = maximum(x = var_3173_cast_fp16, y = var_3174_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_51_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_51_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3177_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_3177_cast_fp16")];
            tensor<int32, [1]> var_3182_axes_0 = const()[name = string("op_3182_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3182_keep_dims_0 = const()[name = string("op_3182_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3182_cast_fp16 = reduce_mean(axes = var_3182_axes_0, keep_dims = var_3182_keep_dims_0, x = var_3177_cast_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3183_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_3183_cast_fp16")];
            fp32 var_3184_epsilon_0 = const()[name = string("op_3184_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3184_cast_fp16 = inverse(epsilon = var_3184_epsilon_0, x = var_3183_cast_fp16)[name = string("op_3184_cast_fp16")];
            fp16 var_3185_to_fp16 = const()[name = string("op_3185_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3186_cast_fp16 = mul(x = var_3184_cast_fp16, y = var_3185_to_fp16)[name = string("op_3186_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3188_cast_fp16 = add(x = var_3182_cast_fp16, y = var_3186_cast_fp16)[name = string("op_3188_cast_fp16")];
            fp32 var_3189_epsilon_0 = const()[name = string("op_3189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3189_cast_fp16 = rsqrt(epsilon = var_3189_epsilon_0, x = var_3188_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_45_cast_fp16 = mul(x = y_51_cast_fp16, y = var_3189_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22150848)))];
            tensor<fp16, [1, 32, 512]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = tm_layers_3_input_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22151936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22676288))))[name = string("tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_3200_cast_fp16 = reshape(shape = var_3199, x = linear_28_cast_fp16)[name = string("op_3200_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22678400))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22940608))))[name = string("tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_29_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_59_perm_0 = const()[name = string("x_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22941696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23203904))))[name = string("tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_3223 = const()[name = string("op_3223"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_3224_cast_fp16 = reshape(shape = var_3223, x = linear_30_cast_fp16)[name = string("op_3224_cast_fp16")];
            tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_3200_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 4, 32, 256]> var_3228_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3231 = const()[name = string("op_3231"), val = bool(true)];
            tensor<int32, [1]> var_3232_axes_0 = const()[name = string("op_3232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_3232_cast_fp16 = reduce_max(axes = var_3232_axes_0, keep_dims = var_3231, x = var_3228_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_53_cast_fp16 = maximum(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_53_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_53_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_3236_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_3236_cast_fp16")];
            tensor<int32, [1]> var_3241_axes_0 = const()[name = string("op_3241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3241_keep_dims_0 = const()[name = string("op_3241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_3241_cast_fp16 = reduce_mean(axes = var_3241_axes_0, keep_dims = var_3241_keep_dims_0, x = var_3236_cast_fp16)[name = string("op_3241_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_3242_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_3242_cast_fp16")];
            fp32 var_3243_epsilon_0 = const()[name = string("op_3243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_3243_cast_fp16 = inverse(epsilon = var_3243_epsilon_0, x = var_3242_cast_fp16)[name = string("op_3243_cast_fp16")];
            fp16 var_3244_to_fp16 = const()[name = string("op_3244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_3245_cast_fp16 = mul(x = var_3243_cast_fp16, y = var_3244_to_fp16)[name = string("op_3245_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_3247_cast_fp16 = add(x = var_3241_cast_fp16, y = var_3245_cast_fp16)[name = string("op_3247_cast_fp16")];
            fp32 var_3248_epsilon_0 = const()[name = string("op_3248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_3248_cast_fp16 = rsqrt(epsilon = var_3248_epsilon_0, x = var_3247_cast_fp16)[name = string("op_3248_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_47_cast_fp16 = mul(x = y_53_cast_fp16, y = var_3248_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23204992)))];
            tensor<fp16, [1, 4, 32, 256]> q_13_cast_fp16 = mul(x = out_47_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_59_cast_fp16 = transpose(perm = x_59_perm_0, x = var_3212_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 2, 32, 256]> var_3251_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_3251_cast_fp16")];
            bool var_3254 = const()[name = string("op_3254"), val = bool(true)];
            tensor<int32, [1]> var_3255_axes_0 = const()[name = string("op_3255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_3255_cast_fp16 = reduce_max(axes = var_3255_axes_0, keep_dims = var_3254, x = var_3251_cast_fp16)[name = string("op_3255_cast_fp16")];
            fp16 var_3256_to_fp16 = const()[name = string("op_3256_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_55_cast_fp16 = maximum(x = var_3255_cast_fp16, y = var_3256_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_55_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_55_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_3259_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<int32, [1]> var_3264_axes_0 = const()[name = string("op_3264_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3264_keep_dims_0 = const()[name = string("op_3264_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_3264_cast_fp16 = reduce_mean(axes = var_3264_axes_0, keep_dims = var_3264_keep_dims_0, x = var_3259_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_3265_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_3265_cast_fp16")];
            fp32 var_3266_epsilon_0 = const()[name = string("op_3266_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_3266_cast_fp16 = inverse(epsilon = var_3266_epsilon_0, x = var_3265_cast_fp16)[name = string("op_3266_cast_fp16")];
            fp16 var_3267_to_fp16 = const()[name = string("op_3267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_3268_cast_fp16 = mul(x = var_3266_cast_fp16, y = var_3267_to_fp16)[name = string("op_3268_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_3270_cast_fp16 = add(x = var_3264_cast_fp16, y = var_3268_cast_fp16)[name = string("op_3270_cast_fp16")];
            fp32 var_3271_epsilon_0 = const()[name = string("op_3271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_3271_cast_fp16 = rsqrt(epsilon = var_3271_epsilon_0, x = var_3270_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_49_cast_fp16 = mul(x = y_55_cast_fp16, y = var_3271_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23205568)))];
            tensor<fp16, [1, 2, 32, 256]> k_19_cast_fp16 = mul(x = out_49_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_3224_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2, 32, 256]> var_3274_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_3274_cast_fp16")];
            bool var_3277 = const()[name = string("op_3277"), val = bool(true)];
            tensor<int32, [1]> var_3278_axes_0 = const()[name = string("op_3278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_3278_cast_fp16 = reduce_max(axes = var_3278_axes_0, keep_dims = var_3277, x = var_3274_cast_fp16)[name = string("op_3278_cast_fp16")];
            fp16 var_3279_to_fp16 = const()[name = string("op_3279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_57_cast_fp16 = maximum(x = var_3278_cast_fp16, y = var_3279_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_57_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_57_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_3282_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_3282_cast_fp16")];
            tensor<int32, [1]> var_3287_axes_0 = const()[name = string("op_3287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3287_keep_dims_0 = const()[name = string("op_3287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_3287_cast_fp16 = reduce_mean(axes = var_3287_axes_0, keep_dims = var_3287_keep_dims_0, x = var_3282_cast_fp16)[name = string("op_3287_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_3288_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_3289_cast_fp16 = inverse(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            fp16 var_3290_to_fp16 = const()[name = string("op_3290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_3291_cast_fp16 = mul(x = var_3289_cast_fp16, y = var_3290_to_fp16)[name = string("op_3291_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_3293_cast_fp16 = add(x = var_3287_cast_fp16, y = var_3291_cast_fp16)[name = string("op_3293_cast_fp16")];
            fp32 var_3294_epsilon_0 = const()[name = string("op_3294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_3294_cast_fp16 = rsqrt(epsilon = var_3294_epsilon_0, x = var_3293_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_13_cast_fp16 = mul(x = y_57_cast_fp16, y = var_3294_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_3296_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3296_cast_fp16")];
            tensor<int32, [4]> var_3301_begin_0 = const()[name = string("op_3301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3301_end_0 = const()[name = string("op_3301_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_3301_end_mask_0 = const()[name = string("op_3301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_3301_cast_fp16 = slice_by_index(begin = var_3301_begin_0, end = var_3301_end_0, end_mask = var_3301_end_mask_0, x = q_13_cast_fp16)[name = string("op_3301_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_3302_cast_fp16 = mul(x = var_3301_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3307_begin_0 = const()[name = string("op_3307_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3307_end_0 = const()[name = string("op_3307_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_3307_end_mask_0 = const()[name = string("op_3307_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_3307_cast_fp16 = slice_by_index(begin = var_3307_begin_0, end = var_3307_end_0, end_mask = var_3307_end_mask_0, x = q_13_cast_fp16)[name = string("op_3307_cast_fp16")];
            int32 var_3309 = const()[name = string("op_3309"), val = int32(-1)];
            bool var_3310_interleave_0 = const()[name = string("op_3310_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_3310_cast_fp16 = concat(axis = var_3309, interleave = var_3310_interleave_0, values = (var_3302_cast_fp16, var_3307_cast_fp16))[name = string("op_3310_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_3311_cast_fp16 = mul(x = var_3310_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3311_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_15_cast_fp16 = add(x = var_3296_cast_fp16, y = var_3311_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_3314_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3314_cast_fp16")];
            tensor<int32, [4]> var_3319_begin_0 = const()[name = string("op_3319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3319_end_0 = const()[name = string("op_3319_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_3319_end_mask_0 = const()[name = string("op_3319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_3319_cast_fp16 = slice_by_index(begin = var_3319_begin_0, end = var_3319_end_0, end_mask = var_3319_end_mask_0, x = k_19_cast_fp16)[name = string("op_3319_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_3320_cast_fp16 = mul(x = var_3319_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_3320_cast_fp16")];
            tensor<int32, [4]> var_3325_begin_0 = const()[name = string("op_3325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3325_end_0 = const()[name = string("op_3325_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_3325_end_mask_0 = const()[name = string("op_3325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_3325_cast_fp16 = slice_by_index(begin = var_3325_begin_0, end = var_3325_end_0, end_mask = var_3325_end_mask_0, x = k_19_cast_fp16)[name = string("op_3325_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), val = int32(-1)];
            bool var_3328_interleave_0 = const()[name = string("op_3328_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_3328_cast_fp16 = concat(axis = var_3327, interleave = var_3328_interleave_0, values = (var_3320_cast_fp16, var_3325_cast_fp16))[name = string("op_3328_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_3329_cast_fp16 = mul(x = var_3328_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3329_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_21_cast_fp16 = add(x = var_3314_cast_fp16, y = var_3329_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = k_21_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [4, 1, 32, 256]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_12_cast_fp16 = reshape(shape = concat_12, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [4, 1, 32, 256]> reshape_13_cast_fp16 = reshape(shape = concat_13, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = v_13_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [4, 1, 32, 256]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_14_cast_fp16 = reshape(shape = concat_14, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [4, 1, 32, 256]> reshape_15_cast_fp16 = reshape(shape = concat_15, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3343_transpose_x_0 = const()[name = string("op_3343_transpose_x_0"), val = bool(false)];
            bool var_3343_transpose_y_0 = const()[name = string("op_3343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 4, 32, 32]> var_3343_cast_fp16 = matmul(transpose_x = var_3343_transpose_x_0, transpose_y = var_3343_transpose_y_0, x = q_15_cast_fp16, y = transpose_99_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_3345_cast_fp16 = add(x = var_3343_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3345_cast_fp16")];
            int32 var_3346 = const()[name = string("op_3346"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_57_cast_fp16 = softmax(axis = var_3346, x = var_3345_cast_fp16)[name = string("w_57_cast_fp16")];
            bool var_3349_transpose_x_0 = const()[name = string("op_3349_transpose_x_0"), val = bool(false)];
            bool var_3349_transpose_y_0 = const()[name = string("op_3349_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 4, 32, 256]> var_3349_cast_fp16 = matmul(transpose_x = var_3349_transpose_x_0, transpose_y = var_3349_transpose_y_0, x = w_57_cast_fp16, y = v_15_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<int32, [4]> var_3352_perm_0 = const()[name = string("op_3352_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3356 = const()[name = string("op_3356"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_3352_cast_fp16 = transpose(perm = var_3352_perm_0, x = var_3349_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 32, 1024]> input_39_cast_fp16 = reshape(shape = var_3356, x = var_3352_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23206144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23730496))))[name = string("tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3361_cast_fp16 = abs(x = linear_31_cast_fp16)[name = string("op_3361_cast_fp16")];
            bool var_3364 = const()[name = string("op_3364"), val = bool(true)];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3365_cast_fp16 = reduce_max(axes = var_3365_axes_0, keep_dims = var_3364, x = var_3361_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_59_cast_fp16 = maximum(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_59_cast_fp16 = real_div(x = linear_31_cast_fp16, y = s_59_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3369_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<int32, [1]> var_3374_axes_0 = const()[name = string("op_3374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3374_keep_dims_0 = const()[name = string("op_3374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3374_cast_fp16 = reduce_mean(axes = var_3374_axes_0, keep_dims = var_3374_keep_dims_0, x = var_3369_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3375_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_3375_cast_fp16")];
            fp32 var_3376_epsilon_0 = const()[name = string("op_3376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3376_cast_fp16 = inverse(epsilon = var_3376_epsilon_0, x = var_3375_cast_fp16)[name = string("op_3376_cast_fp16")];
            fp16 var_3377_to_fp16 = const()[name = string("op_3377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3378_cast_fp16 = mul(x = var_3376_cast_fp16, y = var_3377_to_fp16)[name = string("op_3378_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3380_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3380_cast_fp16")];
            fp32 var_3381_epsilon_0 = const()[name = string("op_3381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3381_cast_fp16 = rsqrt(epsilon = var_3381_epsilon_0, x = var_3380_cast_fp16)[name = string("op_3381_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_3381_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23731584)))];
            tensor<fp16, [1, 32, 512]> var_3383_cast_fp16 = mul(x = out_51_cast_fp16, y = tm_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_65_cast_fp16 = add(x = x_55_cast_fp16, y = var_3383_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3386_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_3386_cast_fp16")];
            bool var_3389 = const()[name = string("op_3389"), val = bool(true)];
            tensor<int32, [1]> var_3390_axes_0 = const()[name = string("op_3390_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3390_cast_fp16 = reduce_max(axes = var_3390_axes_0, keep_dims = var_3389, x = var_3386_cast_fp16)[name = string("op_3390_cast_fp16")];
            fp16 var_3391_to_fp16 = const()[name = string("op_3391_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_61_cast_fp16 = maximum(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_61_cast_fp16 = real_div(x = x_65_cast_fp16, y = s_61_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3394_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<int32, [1]> var_3399_axes_0 = const()[name = string("op_3399_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3399_keep_dims_0 = const()[name = string("op_3399_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3399_cast_fp16 = reduce_mean(axes = var_3399_axes_0, keep_dims = var_3399_keep_dims_0, x = var_3394_cast_fp16)[name = string("op_3399_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3400_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_3400_cast_fp16")];
            fp32 var_3401_epsilon_0 = const()[name = string("op_3401_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3401_cast_fp16 = inverse(epsilon = var_3401_epsilon_0, x = var_3400_cast_fp16)[name = string("op_3401_cast_fp16")];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3403_cast_fp16 = mul(x = var_3401_cast_fp16, y = var_3402_to_fp16)[name = string("op_3403_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3405_cast_fp16 = add(x = var_3399_cast_fp16, y = var_3403_cast_fp16)[name = string("op_3405_cast_fp16")];
            fp32 var_3406_epsilon_0 = const()[name = string("op_3406_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3406_cast_fp16 = rsqrt(epsilon = var_3406_epsilon_0, x = var_3405_cast_fp16)[name = string("op_3406_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_3406_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23732672)))];
            tensor<fp16, [1, 32, 512]> input_41_cast_fp16 = mul(x = out_53_cast_fp16, y = tm_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23733760))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24782400))))[name = string("tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_32_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_32_cast_fp16")];
            string var_3413_mode_0 = const()[name = string("op_3413_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_3413_cast_fp16 = gelu(mode = var_3413_mode_0, x = linear_32_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24786560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25835200))))[name = string("tm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_43_cast_fp16 = mul(x = var_3413_cast_fp16, y = linear_33_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25839360))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26888000))))[name = string("tm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3421_cast_fp16 = abs(x = linear_34_cast_fp16)[name = string("op_3421_cast_fp16")];
            bool var_3424 = const()[name = string("op_3424"), val = bool(true)];
            tensor<int32, [1]> var_3425_axes_0 = const()[name = string("op_3425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3425_cast_fp16 = reduce_max(axes = var_3425_axes_0, keep_dims = var_3424, x = var_3421_cast_fp16)[name = string("op_3425_cast_fp16")];
            fp16 var_3426_to_fp16 = const()[name = string("op_3426_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_63_cast_fp16 = maximum(x = var_3425_cast_fp16, y = var_3426_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_63_cast_fp16 = real_div(x = linear_34_cast_fp16, y = s_63_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3429_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_3429_cast_fp16")];
            tensor<int32, [1]> var_3434_axes_0 = const()[name = string("op_3434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3434_keep_dims_0 = const()[name = string("op_3434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3434_cast_fp16 = reduce_mean(axes = var_3434_axes_0, keep_dims = var_3434_keep_dims_0, x = var_3429_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3435_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp32 var_3436_epsilon_0 = const()[name = string("op_3436_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3436_cast_fp16 = inverse(epsilon = var_3436_epsilon_0, x = var_3435_cast_fp16)[name = string("op_3436_cast_fp16")];
            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3438_cast_fp16 = mul(x = var_3436_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3440_cast_fp16 = add(x = var_3434_cast_fp16, y = var_3438_cast_fp16)[name = string("op_3440_cast_fp16")];
            fp32 var_3441_epsilon_0 = const()[name = string("op_3441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3441_cast_fp16 = rsqrt(epsilon = var_3441_epsilon_0, x = var_3440_cast_fp16)[name = string("op_3441_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_55_cast_fp16 = mul(x = y_63_cast_fp16, y = var_3441_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26889088)))];
            tensor<fp16, [1, 32, 512]> var_3443_cast_fp16 = mul(x = out_55_cast_fp16, y = tm_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_45_cast_fp16 = add(x = x_65_cast_fp16, y = var_3443_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26890176))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27152384))))[name = string("tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_35_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_35_cast_fp16")];
            string x_69_mode_0 = const()[name = string("x_69_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_69_cast_fp16 = gelu(mode = x_69_mode_0, x = linear_35_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [4]> var_3463_begin_0 = const()[name = string("op_3463_begin_0"), val = tensor<int32, [4]>([0, 0, 3, 0])];
            tensor<int32, [4]> var_3463_end_0 = const()[name = string("op_3463_end_0"), val = tensor<int32, [4]>([1, 32, 4, 512])];
            tensor<bool, [4]> var_3463_end_mask_0 = const()[name = string("op_3463_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3463_squeeze_mask_0 = const()[name = string("op_3463_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_3463_cast_fp16 = slice_by_index(begin = var_3463_begin_0, end = var_3463_end_0, end_mask = var_3463_end_mask_0, squeeze_mask = var_3463_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3463_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_47_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3463_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27153472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27415680))))[name = string("tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3473_cast_fp16 = abs(x = linear_36_cast_fp16)[name = string("op_3473_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<int32, [1]> var_3477_axes_0 = const()[name = string("op_3477_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3477_cast_fp16 = reduce_max(axes = var_3477_axes_0, keep_dims = var_3476, x = var_3473_cast_fp16)[name = string("op_3477_cast_fp16")];
            fp16 var_3478_to_fp16 = const()[name = string("op_3478_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_65_cast_fp16 = maximum(x = var_3477_cast_fp16, y = var_3478_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_65_cast_fp16 = real_div(x = linear_36_cast_fp16, y = s_65_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3481_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [1]> var_3486_axes_0 = const()[name = string("op_3486_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3486_keep_dims_0 = const()[name = string("op_3486_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3486_cast_fp16 = reduce_mean(axes = var_3486_axes_0, keep_dims = var_3486_keep_dims_0, x = var_3481_cast_fp16)[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3487_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp32 var_3488_epsilon_0 = const()[name = string("op_3488_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3488_cast_fp16 = inverse(epsilon = var_3488_epsilon_0, x = var_3487_cast_fp16)[name = string("op_3488_cast_fp16")];
            fp16 var_3489_to_fp16 = const()[name = string("op_3489_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3490_cast_fp16 = mul(x = var_3488_cast_fp16, y = var_3489_to_fp16)[name = string("op_3490_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3492_cast_fp16 = add(x = var_3486_cast_fp16, y = var_3490_cast_fp16)[name = string("op_3492_cast_fp16")];
            fp32 var_3493_epsilon_0 = const()[name = string("op_3493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3493_cast_fp16 = rsqrt(epsilon = var_3493_epsilon_0, x = var_3492_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_57_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3493_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27416768)))];
            tensor<fp16, [1, 32, 512]> var_3495_cast_fp16 = mul(x = out_57_cast_fp16, y = tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3495_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_7_cast_fp16 = add(x = input_45_cast_fp16, y = var_3495_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<fp16, [1]> tm_layers_3_layer_scalar_to_fp16 = const()[name = string("tm_layers_3_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 32, 512]> x_73_cast_fp16 = mul(x = h_7_cast_fp16, y = tm_layers_3_layer_scalar_to_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3499_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_3499_cast_fp16")];
            bool var_3502 = const()[name = string("op_3502"), val = bool(true)];
            tensor<int32, [1]> var_3503_axes_0 = const()[name = string("op_3503_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3503_cast_fp16 = reduce_max(axes = var_3503_axes_0, keep_dims = var_3502, x = var_3499_cast_fp16)[name = string("op_3503_cast_fp16")];
            fp16 var_3504_to_fp16 = const()[name = string("op_3504_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_67_cast_fp16 = maximum(x = var_3503_cast_fp16, y = var_3504_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_67_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_67_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3507_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<int32, [1]> var_3512_axes_0 = const()[name = string("op_3512_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3512_keep_dims_0 = const()[name = string("op_3512_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3512_cast_fp16 = reduce_mean(axes = var_3512_axes_0, keep_dims = var_3512_keep_dims_0, x = var_3507_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3513_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_3513_cast_fp16")];
            fp32 var_3514_epsilon_0 = const()[name = string("op_3514_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3514_cast_fp16 = inverse(epsilon = var_3514_epsilon_0, x = var_3513_cast_fp16)[name = string("op_3514_cast_fp16")];
            fp16 var_3515_to_fp16 = const()[name = string("op_3515_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3516_cast_fp16 = mul(x = var_3514_cast_fp16, y = var_3515_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3518_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp32 var_3519_epsilon_0 = const()[name = string("op_3519_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3519_cast_fp16 = rsqrt(epsilon = var_3519_epsilon_0, x = var_3518_cast_fp16)[name = string("op_3519_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_59_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3519_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27417856)))];
            tensor<fp16, [1, 32, 512]> input_49_cast_fp16 = mul(x = out_59_cast_fp16, y = tm_layers_4_input_layernorm_weight_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27418944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27943296))))[name = string("tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3529 = const()[name = string("op_3529"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_3530_cast_fp16 = reshape(shape = var_3529, x = linear_37_cast_fp16)[name = string("op_3530_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27945408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28207616))))[name = string("tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3541 = const()[name = string("op_3541"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_3542_cast_fp16 = reshape(shape = var_3541, x = linear_38_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28208704))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28470912))))[name = string("tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_3553 = const()[name = string("op_3553"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_3554_cast_fp16 = reshape(shape = var_3553, x = linear_39_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<int32, [4]> x_79_perm_0 = const()[name = string("x_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3530_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 4, 32, 256]> var_3558_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3558_cast_fp16")];
            bool var_3561 = const()[name = string("op_3561"), val = bool(true)];
            tensor<int32, [1]> var_3562_axes_0 = const()[name = string("op_3562_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_3562_cast_fp16 = reduce_max(axes = var_3562_axes_0, keep_dims = var_3561, x = var_3558_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp16 var_3563_to_fp16 = const()[name = string("op_3563_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_69_cast_fp16 = maximum(x = var_3562_cast_fp16, y = var_3563_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_69_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_69_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_3566_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3566_cast_fp16")];
            tensor<int32, [1]> var_3571_axes_0 = const()[name = string("op_3571_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3571_keep_dims_0 = const()[name = string("op_3571_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_3571_cast_fp16 = reduce_mean(axes = var_3571_axes_0, keep_dims = var_3571_keep_dims_0, x = var_3566_cast_fp16)[name = string("op_3571_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_3572_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3572_cast_fp16")];
            fp32 var_3573_epsilon_0 = const()[name = string("op_3573_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_3573_cast_fp16 = inverse(epsilon = var_3573_epsilon_0, x = var_3572_cast_fp16)[name = string("op_3573_cast_fp16")];
            fp16 var_3574_to_fp16 = const()[name = string("op_3574_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_3575_cast_fp16 = mul(x = var_3573_cast_fp16, y = var_3574_to_fp16)[name = string("op_3575_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_3577_cast_fp16 = add(x = var_3571_cast_fp16, y = var_3575_cast_fp16)[name = string("op_3577_cast_fp16")];
            fp32 var_3578_epsilon_0 = const()[name = string("op_3578_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_3578_cast_fp16 = rsqrt(epsilon = var_3578_epsilon_0, x = var_3577_cast_fp16)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_61_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3578_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [256]> tm_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472000)))];
            tensor<fp16, [1, 4, 32, 256]> q_17_cast_fp16 = mul(x = out_61_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3542_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2, 32, 256]> var_3581_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3581_cast_fp16")];
            bool var_3584 = const()[name = string("op_3584"), val = bool(true)];
            tensor<int32, [1]> var_3585_axes_0 = const()[name = string("op_3585_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_3585_cast_fp16 = reduce_max(axes = var_3585_axes_0, keep_dims = var_3584, x = var_3581_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp16 var_3586_to_fp16 = const()[name = string("op_3586_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_71_cast_fp16 = maximum(x = var_3585_cast_fp16, y = var_3586_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_71_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_71_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_3589_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3589_cast_fp16")];
            tensor<int32, [1]> var_3594_axes_0 = const()[name = string("op_3594_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3594_keep_dims_0 = const()[name = string("op_3594_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_3594_cast_fp16 = reduce_mean(axes = var_3594_axes_0, keep_dims = var_3594_keep_dims_0, x = var_3589_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_3595_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp32 var_3596_epsilon_0 = const()[name = string("op_3596_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_3596_cast_fp16 = inverse(epsilon = var_3596_epsilon_0, x = var_3595_cast_fp16)[name = string("op_3596_cast_fp16")];
            fp16 var_3597_to_fp16 = const()[name = string("op_3597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_3598_cast_fp16 = mul(x = var_3596_cast_fp16, y = var_3597_to_fp16)[name = string("op_3598_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_3600_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3598_cast_fp16)[name = string("op_3600_cast_fp16")];
            fp32 var_3601_epsilon_0 = const()[name = string("op_3601_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_3601_cast_fp16 = rsqrt(epsilon = var_3601_epsilon_0, x = var_3600_cast_fp16)[name = string("op_3601_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_63_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3601_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472576)))];
            tensor<fp16, [1, 2, 32, 256]> k_25_cast_fp16 = mul(x = out_63_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_79_cast_fp16 = transpose(perm = x_79_perm_0, x = var_3554_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 2, 32, 256]> var_3604_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3604_cast_fp16")];
            bool var_3607 = const()[name = string("op_3607"), val = bool(true)];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_3607, x = var_3604_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_73_cast_fp16 = maximum(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_73_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_73_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_3612_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<int32, [1]> var_3617_axes_0 = const()[name = string("op_3617_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3617_keep_dims_0 = const()[name = string("op_3617_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_3617_cast_fp16 = reduce_mean(axes = var_3617_axes_0, keep_dims = var_3617_keep_dims_0, x = var_3612_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_3618_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_3619_cast_fp16 = inverse(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            fp16 var_3620_to_fp16 = const()[name = string("op_3620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_3621_cast_fp16 = mul(x = var_3619_cast_fp16, y = var_3620_to_fp16)[name = string("op_3621_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_3623_cast_fp16 = add(x = var_3617_cast_fp16, y = var_3621_cast_fp16)[name = string("op_3623_cast_fp16")];
            fp32 var_3624_epsilon_0 = const()[name = string("op_3624_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_3624_cast_fp16 = rsqrt(epsilon = var_3624_epsilon_0, x = var_3623_cast_fp16)[name = string("op_3624_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_17_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3624_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_3626_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3626_cast_fp16")];
            tensor<int32, [4]> var_3631_begin_0 = const()[name = string("op_3631_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3631_end_0 = const()[name = string("op_3631_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_3631_end_mask_0 = const()[name = string("op_3631_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_3631_cast_fp16 = slice_by_index(begin = var_3631_begin_0, end = var_3631_end_0, end_mask = var_3631_end_mask_0, x = q_17_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_3632_cast_fp16 = mul(x = var_3631_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> var_3637_begin_0 = const()[name = string("op_3637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3637_end_0 = const()[name = string("op_3637_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_3637_end_mask_0 = const()[name = string("op_3637_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_3637_cast_fp16 = slice_by_index(begin = var_3637_begin_0, end = var_3637_end_0, end_mask = var_3637_end_mask_0, x = q_17_cast_fp16)[name = string("op_3637_cast_fp16")];
            int32 var_3639 = const()[name = string("op_3639"), val = int32(-1)];
            bool var_3640_interleave_0 = const()[name = string("op_3640_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_3640_cast_fp16 = concat(axis = var_3639, interleave = var_3640_interleave_0, values = (var_3632_cast_fp16, var_3637_cast_fp16))[name = string("op_3640_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_3641_cast_fp16 = mul(x = var_3640_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3641_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_19_cast_fp16 = add(x = var_3626_cast_fp16, y = var_3641_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_3644_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = k_25_cast_fp16)[name = string("op_3649_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_3650_cast_fp16 = mul(x = var_3649_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            tensor<int32, [4]> var_3655_begin_0 = const()[name = string("op_3655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3655_end_0 = const()[name = string("op_3655_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_3655_end_mask_0 = const()[name = string("op_3655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_3655_cast_fp16 = slice_by_index(begin = var_3655_begin_0, end = var_3655_end_0, end_mask = var_3655_end_mask_0, x = k_25_cast_fp16)[name = string("op_3655_cast_fp16")];
            int32 var_3657 = const()[name = string("op_3657"), val = int32(-1)];
            bool var_3658_interleave_0 = const()[name = string("op_3658_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_3658_cast_fp16 = concat(axis = var_3657, interleave = var_3658_interleave_0, values = (var_3650_cast_fp16, var_3655_cast_fp16))[name = string("op_3658_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_3659_cast_fp16 = mul(x = var_3658_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3659_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_27_cast_fp16 = add(x = var_3644_cast_fp16, y = var_3659_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = k_27_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [4, 1, 32, 256]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
            tensor<int32, [5]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_16_cast_fp16 = reshape(shape = concat_16, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [4, 1, 32, 256]> reshape_17_cast_fp16 = reshape(shape = concat_17, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = v_17_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [4, 1, 32, 256]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
            tensor<int32, [5]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_18_cast_fp16 = reshape(shape = concat_18, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [4, 1, 32, 256]> reshape_19_cast_fp16 = reshape(shape = concat_19, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3673_transpose_x_0 = const()[name = string("op_3673_transpose_x_0"), val = bool(false)];
            bool var_3673_transpose_y_0 = const()[name = string("op_3673_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 4, 32, 32]> var_3673_cast_fp16 = matmul(transpose_x = var_3673_transpose_x_0, transpose_y = var_3673_transpose_y_0, x = q_19_cast_fp16, y = transpose_100_cast_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_3675_cast_fp16 = add(x = var_3673_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3675_cast_fp16")];
            int32 var_3676 = const()[name = string("op_3676"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_73_cast_fp16 = softmax(axis = var_3676, x = var_3675_cast_fp16)[name = string("w_73_cast_fp16")];
            bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)];
            bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 4, 32, 256]> var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = w_73_cast_fp16, y = v_19_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<int32, [4]> var_3682_perm_0 = const()[name = string("op_3682_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3686 = const()[name = string("op_3686"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_3682_cast_fp16 = transpose(perm = var_3682_perm_0, x = var_3679_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 32, 1024]> input_51_cast_fp16 = reshape(shape = var_3686, x = var_3682_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28473152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28997504))))[name = string("tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3691_cast_fp16 = abs(x = linear_40_cast_fp16)[name = string("op_3691_cast_fp16")];
            bool var_3694 = const()[name = string("op_3694"), val = bool(true)];
            tensor<int32, [1]> var_3695_axes_0 = const()[name = string("op_3695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3695_cast_fp16 = reduce_max(axes = var_3695_axes_0, keep_dims = var_3694, x = var_3691_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 var_3696_to_fp16 = const()[name = string("op_3696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_75_cast_fp16 = maximum(x = var_3695_cast_fp16, y = var_3696_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_75_cast_fp16 = real_div(x = linear_40_cast_fp16, y = s_75_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3699_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3699_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3704_keep_dims_0 = const()[name = string("op_3704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3704_cast_fp16 = reduce_mean(axes = var_3704_axes_0, keep_dims = var_3704_keep_dims_0, x = var_3699_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3705_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3705_cast_fp16")];
            fp32 var_3706_epsilon_0 = const()[name = string("op_3706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3706_cast_fp16 = inverse(epsilon = var_3706_epsilon_0, x = var_3705_cast_fp16)[name = string("op_3706_cast_fp16")];
            fp16 var_3707_to_fp16 = const()[name = string("op_3707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3708_cast_fp16 = mul(x = var_3706_cast_fp16, y = var_3707_to_fp16)[name = string("op_3708_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3710_cast_fp16 = add(x = var_3704_cast_fp16, y = var_3708_cast_fp16)[name = string("op_3710_cast_fp16")];
            fp32 var_3711_epsilon_0 = const()[name = string("op_3711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3711_cast_fp16 = rsqrt(epsilon = var_3711_epsilon_0, x = var_3710_cast_fp16)[name = string("op_3711_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3711_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28998592)))];
            tensor<fp16, [1, 32, 512]> var_3713_cast_fp16 = mul(x = out_65_cast_fp16, y = tm_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = var_3713_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3716_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3716_cast_fp16")];
            bool var_3719 = const()[name = string("op_3719"), val = bool(true)];
            tensor<int32, [1]> var_3720_axes_0 = const()[name = string("op_3720_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3720_cast_fp16 = reduce_max(axes = var_3720_axes_0, keep_dims = var_3719, x = var_3716_cast_fp16)[name = string("op_3720_cast_fp16")];
            fp16 var_3721_to_fp16 = const()[name = string("op_3721_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_77_cast_fp16 = maximum(x = var_3720_cast_fp16, y = var_3721_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_77_cast_fp16 = real_div(x = x_83_cast_fp16, y = s_77_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3724_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3729_axes_0 = const()[name = string("op_3729_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3729_keep_dims_0 = const()[name = string("op_3729_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3729_cast_fp16 = reduce_mean(axes = var_3729_axes_0, keep_dims = var_3729_keep_dims_0, x = var_3724_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3730_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3730_cast_fp16")];
            fp32 var_3731_epsilon_0 = const()[name = string("op_3731_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3731_cast_fp16 = inverse(epsilon = var_3731_epsilon_0, x = var_3730_cast_fp16)[name = string("op_3731_cast_fp16")];
            fp16 var_3732_to_fp16 = const()[name = string("op_3732_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3733_cast_fp16 = mul(x = var_3731_cast_fp16, y = var_3732_to_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3735_cast_fp16 = add(x = var_3729_cast_fp16, y = var_3733_cast_fp16)[name = string("op_3735_cast_fp16")];
            fp32 var_3736_epsilon_0 = const()[name = string("op_3736_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3736_cast_fp16 = rsqrt(epsilon = var_3736_epsilon_0, x = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_67_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3736_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28999680)))];
            tensor<fp16, [1, 32, 512]> input_53_cast_fp16 = mul(x = out_67_cast_fp16, y = tm_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29000768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30049408))))[name = string("tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_41_cast_fp16")];
            string var_3743_mode_0 = const()[name = string("op_3743_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_3743_cast_fp16 = gelu(mode = var_3743_mode_0, x = linear_41_cast_fp16)[name = string("op_3743_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30053568))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31102208))))[name = string("tm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_42_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_55_cast_fp16 = mul(x = var_3743_cast_fp16, y = linear_42_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31106368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32155008))))[name = string("tm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3751_cast_fp16 = abs(x = linear_43_cast_fp16)[name = string("op_3751_cast_fp16")];
            bool var_3754 = const()[name = string("op_3754"), val = bool(true)];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_3754, x = var_3751_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 var_3756_to_fp16 = const()[name = string("op_3756_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_79_cast_fp16 = maximum(x = var_3755_cast_fp16, y = var_3756_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_79_cast_fp16 = real_div(x = linear_43_cast_fp16, y = s_79_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3759_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<int32, [1]> var_3764_axes_0 = const()[name = string("op_3764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3764_keep_dims_0 = const()[name = string("op_3764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3764_cast_fp16 = reduce_mean(axes = var_3764_axes_0, keep_dims = var_3764_keep_dims_0, x = var_3759_cast_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3765_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3766_cast_fp16 = inverse(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3768_cast_fp16 = mul(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("op_3768_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3770_cast_fp16 = add(x = var_3764_cast_fp16, y = var_3768_cast_fp16)[name = string("op_3770_cast_fp16")];
            fp32 var_3771_epsilon_0 = const()[name = string("op_3771_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3771_cast_fp16 = rsqrt(epsilon = var_3771_epsilon_0, x = var_3770_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_69_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3771_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32156096)))];
            tensor<fp16, [1, 32, 512]> var_3773_cast_fp16 = mul(x = out_69_cast_fp16, y = tm_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3773_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_57_cast_fp16 = add(x = x_83_cast_fp16, y = var_3773_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32157184))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32419392))))[name = string("tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_44_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_44_cast_fp16")];
            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_87_cast_fp16 = gelu(mode = x_87_mode_0, x = linear_44_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [4]> var_3793_begin_0 = const()[name = string("op_3793_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_3793_end_0 = const()[name = string("op_3793_end_0"), val = tensor<int32, [4]>([1, 32, 5, 512])];
            tensor<bool, [4]> var_3793_end_mask_0 = const()[name = string("op_3793_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3793_squeeze_mask_0 = const()[name = string("op_3793_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_3793_cast_fp16 = slice_by_index(begin = var_3793_begin_0, end = var_3793_end_0, end_mask = var_3793_end_mask_0, squeeze_mask = var_3793_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_59_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3793_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32420480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32682688))))[name = string("tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3803_cast_fp16 = abs(x = linear_45_cast_fp16)[name = string("op_3803_cast_fp16")];
            bool var_3806 = const()[name = string("op_3806"), val = bool(true)];
            tensor<int32, [1]> var_3807_axes_0 = const()[name = string("op_3807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3807_cast_fp16 = reduce_max(axes = var_3807_axes_0, keep_dims = var_3806, x = var_3803_cast_fp16)[name = string("op_3807_cast_fp16")];
            fp16 var_3808_to_fp16 = const()[name = string("op_3808_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_81_cast_fp16 = maximum(x = var_3807_cast_fp16, y = var_3808_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_81_cast_fp16 = real_div(x = linear_45_cast_fp16, y = s_81_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3811_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [1]> var_3816_axes_0 = const()[name = string("op_3816_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3816_keep_dims_0 = const()[name = string("op_3816_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3816_cast_fp16 = reduce_mean(axes = var_3816_axes_0, keep_dims = var_3816_keep_dims_0, x = var_3811_cast_fp16)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3817_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3817_cast_fp16")];
            fp32 var_3818_epsilon_0 = const()[name = string("op_3818_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3818_cast_fp16 = inverse(epsilon = var_3818_epsilon_0, x = var_3817_cast_fp16)[name = string("op_3818_cast_fp16")];
            fp16 var_3819_to_fp16 = const()[name = string("op_3819_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3820_cast_fp16 = mul(x = var_3818_cast_fp16, y = var_3819_to_fp16)[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3822_cast_fp16 = add(x = var_3816_cast_fp16, y = var_3820_cast_fp16)[name = string("op_3822_cast_fp16")];
            fp32 var_3823_epsilon_0 = const()[name = string("op_3823_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3823_cast_fp16 = rsqrt(epsilon = var_3823_epsilon_0, x = var_3822_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_71_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3823_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32683776)))];
            tensor<fp16, [1, 32, 512]> var_3825_cast_fp16 = mul(x = out_71_cast_fp16, y = tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3825_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_9_cast_fp16 = add(x = input_57_cast_fp16, y = var_3825_cast_fp16)[name = string("h_9_cast_fp16")];
            tensor<fp16, [1]> tm_layers_4_layer_scalar_to_fp16 = const()[name = string("tm_layers_4_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.62p-1])];
            tensor<fp16, [1, 32, 512]> x_91_cast_fp16 = mul(x = h_9_cast_fp16, y = tm_layers_4_layer_scalar_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3829_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3829_cast_fp16")];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<int32, [1]> var_3833_axes_0 = const()[name = string("op_3833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3833_cast_fp16 = reduce_max(axes = var_3833_axes_0, keep_dims = var_3832, x = var_3829_cast_fp16)[name = string("op_3833_cast_fp16")];
            fp16 var_3834_to_fp16 = const()[name = string("op_3834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_83_cast_fp16 = maximum(x = var_3833_cast_fp16, y = var_3834_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_83_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_83_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_3837_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3837_cast_fp16")];
            tensor<int32, [1]> var_3842_axes_0 = const()[name = string("op_3842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3842_keep_dims_0 = const()[name = string("op_3842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3842_cast_fp16 = reduce_mean(axes = var_3842_axes_0, keep_dims = var_3842_keep_dims_0, x = var_3837_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3843_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp32 var_3844_epsilon_0 = const()[name = string("op_3844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3844_cast_fp16 = inverse(epsilon = var_3844_epsilon_0, x = var_3843_cast_fp16)[name = string("op_3844_cast_fp16")];
            fp16 var_3845_to_fp16 = const()[name = string("op_3845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3846_cast_fp16 = mul(x = var_3844_cast_fp16, y = var_3845_to_fp16)[name = string("op_3846_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3848_cast_fp16 = add(x = var_3842_cast_fp16, y = var_3846_cast_fp16)[name = string("op_3848_cast_fp16")];
            fp32 var_3849_epsilon_0 = const()[name = string("op_3849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3849_cast_fp16 = rsqrt(epsilon = var_3849_epsilon_0, x = var_3848_cast_fp16)[name = string("op_3849_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_73_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3849_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32684864)))];
            tensor<fp16, [1, 32, 512]> input_61_cast_fp16 = mul(x = out_73_cast_fp16, y = tm_layers_5_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32685952))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33734592))))[name = string("tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_46_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [4]> var_3859 = const()[name = string("op_3859"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 32, 4, 512]> var_3860_cast_fp16 = reshape(shape = var_3859, x = linear_46_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<int32, [4]> x_93_perm_0 = const()[name = string("x_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33738752))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34000960))))[name = string("tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<int32, [4]> var_3871 = const()[name = string("op_3871"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 32, 1, 512]> var_3872_cast_fp16 = reshape(shape = var_3871, x = linear_47_cast_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [4]> x_95_perm_0 = const()[name = string("x_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34002048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34264256))))[name = string("tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [4]> var_3883 = const()[name = string("op_3883"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 32, 1, 512]> var_3884_cast_fp16 = reshape(shape = var_3883, x = linear_48_cast_fp16)[name = string("op_3884_cast_fp16")];
            tensor<int32, [4]> x_97_perm_0 = const()[name = string("x_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 512]> x_93_cast_fp16 = transpose(perm = x_93_perm_0, x = var_3860_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 4, 32, 512]> var_3888_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3888_cast_fp16")];
            bool var_3891 = const()[name = string("op_3891"), val = bool(true)];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_3892_cast_fp16 = reduce_max(axes = var_3892_axes_0, keep_dims = var_3891, x = var_3888_cast_fp16)[name = string("op_3892_cast_fp16")];
            fp16 var_3893_to_fp16 = const()[name = string("op_3893_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_85_cast_fp16 = maximum(x = var_3892_cast_fp16, y = var_3893_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> y_85_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_85_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_3896_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3896_cast_fp16")];
            tensor<int32, [1]> var_3901_axes_0 = const()[name = string("op_3901_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3901_keep_dims_0 = const()[name = string("op_3901_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_3901_cast_fp16 = reduce_mean(axes = var_3901_axes_0, keep_dims = var_3901_keep_dims_0, x = var_3896_cast_fp16)[name = string("op_3901_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_3902_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3902_cast_fp16")];
            fp32 var_3903_epsilon_0 = const()[name = string("op_3903_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_3903_cast_fp16 = inverse(epsilon = var_3903_epsilon_0, x = var_3902_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_3905_cast_fp16 = mul(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_3907_cast_fp16 = add(x = var_3901_cast_fp16, y = var_3905_cast_fp16)[name = string("op_3907_cast_fp16")];
            fp32 var_3908_epsilon_0 = const()[name = string("op_3908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_3908_cast_fp16 = rsqrt(epsilon = var_3908_epsilon_0, x = var_3907_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> out_75_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3908_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34265344)))];
            tensor<fp16, [1, 4, 32, 512]> q_21_cast_fp16 = mul(x = out_75_cast_fp16, y = tm_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> x_95_cast_fp16 = transpose(perm = x_95_perm_0, x = var_3872_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1, 32, 512]> var_3911_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 32, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 32, 1]> s_87_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> y_87_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_87_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_3919_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 32, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_3925_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 32, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 32, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 32, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> out_77_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3931_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34266432)))];
            tensor<fp16, [1, 1, 32, 512]> k_31_cast_fp16 = mul(x = out_77_cast_fp16, y = tm_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = var_3884_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 1, 32, 512]> var_3934_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3934_cast_fp16")];
            bool var_3937 = const()[name = string("op_3937"), val = bool(true)];
            tensor<int32, [1]> var_3938_axes_0 = const()[name = string("op_3938_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 32, 1]> var_3938_cast_fp16 = reduce_max(axes = var_3938_axes_0, keep_dims = var_3937, x = var_3934_cast_fp16)[name = string("op_3938_cast_fp16")];
            fp16 var_3939_to_fp16 = const()[name = string("op_3939_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 32, 1]> s_89_cast_fp16 = maximum(x = var_3938_cast_fp16, y = var_3939_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> y_89_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_89_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_3942_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3942_cast_fp16")];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3947_keep_dims_0 = const()[name = string("op_3947_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 32, 1]> var_3947_cast_fp16 = reduce_mean(axes = var_3947_axes_0, keep_dims = var_3947_keep_dims_0, x = var_3942_cast_fp16)[name = string("op_3947_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_3948_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3948_cast_fp16")];
            fp32 var_3949_epsilon_0 = const()[name = string("op_3949_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 32, 1]> var_3949_cast_fp16 = inverse(epsilon = var_3949_epsilon_0, x = var_3948_cast_fp16)[name = string("op_3949_cast_fp16")];
            fp16 var_3950_to_fp16 = const()[name = string("op_3950_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 32, 1]> var_3951_cast_fp16 = mul(x = var_3949_cast_fp16, y = var_3950_to_fp16)[name = string("op_3951_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_3953_cast_fp16 = add(x = var_3947_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3953_cast_fp16")];
            fp32 var_3954_epsilon_0 = const()[name = string("op_3954_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 32, 1]> var_3954_cast_fp16 = rsqrt(epsilon = var_3954_epsilon_0, x = var_3953_cast_fp16)[name = string("op_3954_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> v_21_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3954_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<fp16, [32, 512]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [32, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137443712))), scale = tensor<fp16, [32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137435200))))[name = string("cos_to_fp16_quantized")];
            tensor<fp16, [1, 4, 32, 512]> var_3956_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3956_cast_fp16")];
            tensor<int32, [4]> var_3961_begin_0 = const()[name = string("op_3961_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3961_end_0 = const()[name = string("op_3961_end_0"), val = tensor<int32, [4]>([1, 4, 32, 512])];
            tensor<bool, [4]> var_3961_end_mask_0 = const()[name = string("op_3961_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 256]> var_3961_cast_fp16 = slice_by_index(begin = var_3961_begin_0, end = var_3961_end_0, end_mask = var_3961_end_mask_0, x = q_21_cast_fp16)[name = string("op_3961_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 256]> var_3962_cast_fp16 = mul(x = var_3961_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_3962_cast_fp16")];
            tensor<int32, [4]> var_3967_begin_0 = const()[name = string("op_3967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3967_end_0 = const()[name = string("op_3967_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_3967_end_mask_0 = const()[name = string("op_3967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 256]> var_3967_cast_fp16 = slice_by_index(begin = var_3967_begin_0, end = var_3967_end_0, end_mask = var_3967_end_mask_0, x = q_21_cast_fp16)[name = string("op_3967_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(-1)];
            bool var_3970_interleave_0 = const()[name = string("op_3970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 512]> var_3970_cast_fp16 = concat(axis = var_3969, interleave = var_3970_interleave_0, values = (var_3962_cast_fp16, var_3967_cast_fp16))[name = string("op_3970_cast_fp16")];
            tensor<fp16, [32, 512]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [32, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137460160))), scale = tensor<fp16, [32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137476608))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [1, 4, 32, 512]> var_3971_cast_fp16 = mul(x = var_3970_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3971_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> q_23_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3971_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_3974_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3974_cast_fp16")];
            tensor<int32, [4]> var_3979_begin_0 = const()[name = string("op_3979_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3979_end_0 = const()[name = string("op_3979_end_0"), val = tensor<int32, [4]>([1, 1, 32, 512])];
            tensor<bool, [4]> var_3979_end_mask_0 = const()[name = string("op_3979_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 32, 256]> var_3979_cast_fp16 = slice_by_index(begin = var_3979_begin_0, end = var_3979_end_0, end_mask = var_3979_end_mask_0, x = k_31_cast_fp16)[name = string("op_3979_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 32, 256]> var_3980_cast_fp16 = mul(x = var_3979_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_3980_cast_fp16")];
            tensor<int32, [4]> var_3985_begin_0 = const()[name = string("op_3985_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3985_end_0 = const()[name = string("op_3985_end_0"), val = tensor<int32, [4]>([1, 1, 32, 256])];
            tensor<bool, [4]> var_3985_end_mask_0 = const()[name = string("op_3985_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 32, 256]> var_3985_cast_fp16 = slice_by_index(begin = var_3985_begin_0, end = var_3985_end_0, end_mask = var_3985_end_mask_0, x = k_31_cast_fp16)[name = string("op_3985_cast_fp16")];
            int32 var_3987 = const()[name = string("op_3987"), val = int32(-1)];
            bool var_3988_interleave_0 = const()[name = string("op_3988_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 32, 512]> var_3988_cast_fp16 = concat(axis = var_3987, interleave = var_3988_interleave_0, values = (var_3980_cast_fp16, var_3985_cast_fp16))[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_3989_cast_fp16 = mul(x = var_3988_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> k_33_cast_fp16 = add(x = var_3974_cast_fp16, y = var_3989_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 32, 512]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = k_33_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [4, 1, 32, 512]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([4, 1, 1, 32, 512])];
            tensor<fp16, [4, 1, 1, 32, 512]> reshape_20_cast_fp16 = reshape(shape = concat_20, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 32, 512])];
            tensor<fp16, [1, 4, 1, 32, 512]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [4, 1, 32, 512]> reshape_21_cast_fp16 = reshape(shape = concat_21, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 32, 512]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = v_21_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [4, 1, 32, 512]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([4, 1, 1, 32, 512])];
            tensor<fp16, [4, 1, 1, 32, 512]> reshape_22_cast_fp16 = reshape(shape = concat_22, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 32, 512])];
            tensor<fp16, [1, 4, 1, 32, 512]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [4, 1, 32, 512]> reshape_23_cast_fp16 = reshape(shape = concat_23, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4003_transpose_x_0 = const()[name = string("op_4003_transpose_x_0"), val = bool(false)];
            bool var_4003_transpose_y_0 = const()[name = string("op_4003_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 32]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 4, 32, 32]> var_4003_cast_fp16 = matmul(transpose_x = var_4003_transpose_x_0, transpose_y = var_4003_transpose_y_0, x = q_23_cast_fp16, y = transpose_101_cast_fp16)[name = string("op_4003_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_4005_cast_fp16 = add(x = var_4003_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4005_cast_fp16")];
            int32 var_4006 = const()[name = string("op_4006"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_89_cast_fp16 = softmax(axis = var_4006, x = var_4005_cast_fp16)[name = string("w_89_cast_fp16")];
            bool var_4009_transpose_x_0 = const()[name = string("op_4009_transpose_x_0"), val = bool(false)];
            bool var_4009_transpose_y_0 = const()[name = string("op_4009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 512]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 4, 32, 512]> var_4009_cast_fp16 = matmul(transpose_x = var_4009_transpose_x_0, transpose_y = var_4009_transpose_y_0, x = w_89_cast_fp16, y = v_23_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4012_perm_0 = const()[name = string("op_4012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4016 = const()[name = string("op_4016"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 32, 4, 512]> var_4012_cast_fp16 = transpose(perm = var_4012_perm_0, x = var_4009_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 32, 2048]> input_63_cast_fp16 = reshape(shape = var_4016, x = var_4012_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34267520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35316160))))[name = string("tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4021_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_4021_cast_fp16")];
            bool var_4024 = const()[name = string("op_4024"), val = bool(true)];
            tensor<int32, [1]> var_4025_axes_0 = const()[name = string("op_4025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4025_cast_fp16 = reduce_max(axes = var_4025_axes_0, keep_dims = var_4024, x = var_4021_cast_fp16)[name = string("op_4025_cast_fp16")];
            fp16 var_4026_to_fp16 = const()[name = string("op_4026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_91_cast_fp16 = maximum(x = var_4025_cast_fp16, y = var_4026_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_91_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_91_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4029_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<int32, [1]> var_4034_axes_0 = const()[name = string("op_4034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4034_keep_dims_0 = const()[name = string("op_4034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4034_cast_fp16 = reduce_mean(axes = var_4034_axes_0, keep_dims = var_4034_keep_dims_0, x = var_4029_cast_fp16)[name = string("op_4034_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4035_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_4035_cast_fp16")];
            fp32 var_4036_epsilon_0 = const()[name = string("op_4036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4036_cast_fp16 = inverse(epsilon = var_4036_epsilon_0, x = var_4035_cast_fp16)[name = string("op_4036_cast_fp16")];
            fp16 var_4037_to_fp16 = const()[name = string("op_4037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4038_cast_fp16 = mul(x = var_4036_cast_fp16, y = var_4037_to_fp16)[name = string("op_4038_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4040_cast_fp16 = add(x = var_4034_cast_fp16, y = var_4038_cast_fp16)[name = string("op_4040_cast_fp16")];
            fp32 var_4041_epsilon_0 = const()[name = string("op_4041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4041_cast_fp16 = rsqrt(epsilon = var_4041_epsilon_0, x = var_4040_cast_fp16)[name = string("op_4041_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_79_cast_fp16 = mul(x = y_91_cast_fp16, y = var_4041_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35317248)))];
            tensor<fp16, [1, 32, 512]> var_4043_cast_fp16 = mul(x = out_79_cast_fp16, y = tm_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_4043_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_101_cast_fp16 = add(x = x_91_cast_fp16, y = var_4043_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4046_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_4046_cast_fp16")];
            bool var_4049 = const()[name = string("op_4049"), val = bool(true)];
            tensor<int32, [1]> var_4050_axes_0 = const()[name = string("op_4050_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4050_cast_fp16 = reduce_max(axes = var_4050_axes_0, keep_dims = var_4049, x = var_4046_cast_fp16)[name = string("op_4050_cast_fp16")];
            fp16 var_4051_to_fp16 = const()[name = string("op_4051_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_93_cast_fp16 = maximum(x = var_4050_cast_fp16, y = var_4051_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_93_cast_fp16 = real_div(x = x_101_cast_fp16, y = s_93_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4054_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [1]> var_4059_axes_0 = const()[name = string("op_4059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4059_keep_dims_0 = const()[name = string("op_4059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4059_cast_fp16 = reduce_mean(axes = var_4059_axes_0, keep_dims = var_4059_keep_dims_0, x = var_4054_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4060_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_4060_cast_fp16")];
            fp32 var_4061_epsilon_0 = const()[name = string("op_4061_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4061_cast_fp16 = inverse(epsilon = var_4061_epsilon_0, x = var_4060_cast_fp16)[name = string("op_4061_cast_fp16")];
            fp16 var_4062_to_fp16 = const()[name = string("op_4062_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4063_cast_fp16 = mul(x = var_4061_cast_fp16, y = var_4062_to_fp16)[name = string("op_4063_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4065_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4063_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp32 var_4066_epsilon_0 = const()[name = string("op_4066_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4066_cast_fp16 = rsqrt(epsilon = var_4066_epsilon_0, x = var_4065_cast_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_81_cast_fp16 = mul(x = y_93_cast_fp16, y = var_4066_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35318336)))];
            tensor<fp16, [1, 32, 512]> input_65_cast_fp16 = mul(x = out_81_cast_fp16, y = tm_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35319424))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36368064))))[name = string("tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_50_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_50_cast_fp16")];
            string var_4073_mode_0 = const()[name = string("op_4073_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_4073_cast_fp16 = gelu(mode = var_4073_mode_0, x = linear_50_cast_fp16)[name = string("op_4073_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36372224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37420864))))[name = string("tm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_51_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_67_cast_fp16 = mul(x = var_4073_cast_fp16, y = linear_51_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37425024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38473664))))[name = string("tm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4081_cast_fp16 = abs(x = linear_52_cast_fp16)[name = string("op_4081_cast_fp16")];
            bool var_4084 = const()[name = string("op_4084"), val = bool(true)];
            tensor<int32, [1]> var_4085_axes_0 = const()[name = string("op_4085_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4085_cast_fp16 = reduce_max(axes = var_4085_axes_0, keep_dims = var_4084, x = var_4081_cast_fp16)[name = string("op_4085_cast_fp16")];
            fp16 var_4086_to_fp16 = const()[name = string("op_4086_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_95_cast_fp16 = maximum(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_95_cast_fp16 = real_div(x = linear_52_cast_fp16, y = s_95_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4089_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_4089_cast_fp16")];
            tensor<int32, [1]> var_4094_axes_0 = const()[name = string("op_4094_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4094_keep_dims_0 = const()[name = string("op_4094_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4094_cast_fp16 = reduce_mean(axes = var_4094_axes_0, keep_dims = var_4094_keep_dims_0, x = var_4089_cast_fp16)[name = string("op_4094_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4095_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_4095_cast_fp16")];
            fp32 var_4096_epsilon_0 = const()[name = string("op_4096_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4096_cast_fp16 = inverse(epsilon = var_4096_epsilon_0, x = var_4095_cast_fp16)[name = string("op_4096_cast_fp16")];
            fp16 var_4097_to_fp16 = const()[name = string("op_4097_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4098_cast_fp16 = mul(x = var_4096_cast_fp16, y = var_4097_to_fp16)[name = string("op_4098_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4100_cast_fp16 = add(x = var_4094_cast_fp16, y = var_4098_cast_fp16)[name = string("op_4100_cast_fp16")];
            fp32 var_4101_epsilon_0 = const()[name = string("op_4101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4101_cast_fp16 = rsqrt(epsilon = var_4101_epsilon_0, x = var_4100_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_83_cast_fp16 = mul(x = y_95_cast_fp16, y = var_4101_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38474752)))];
            tensor<fp16, [1, 32, 512]> var_4103_cast_fp16 = mul(x = out_83_cast_fp16, y = tm_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_69_cast_fp16 = add(x = x_101_cast_fp16, y = var_4103_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38475840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38738048))))[name = string("tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_53_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_53_cast_fp16")];
            string x_105_mode_0 = const()[name = string("x_105_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_105_cast_fp16 = gelu(mode = x_105_mode_0, x = linear_53_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [4]> var_4123_begin_0 = const()[name = string("op_4123_begin_0"), val = tensor<int32, [4]>([0, 0, 5, 0])];
            tensor<int32, [4]> var_4123_end_0 = const()[name = string("op_4123_end_0"), val = tensor<int32, [4]>([1, 32, 6, 512])];
            tensor<bool, [4]> var_4123_end_mask_0 = const()[name = string("op_4123_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4123_squeeze_mask_0 = const()[name = string("op_4123_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_4123_cast_fp16 = slice_by_index(begin = var_4123_begin_0, end = var_4123_end_0, end_mask = var_4123_end_mask_0, squeeze_mask = var_4123_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_71_cast_fp16 = mul(x = x_105_cast_fp16, y = var_4123_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38739136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39001344))))[name = string("tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4133_cast_fp16 = abs(x = linear_54_cast_fp16)[name = string("op_4133_cast_fp16")];
            bool var_4136 = const()[name = string("op_4136"), val = bool(true)];
            tensor<int32, [1]> var_4137_axes_0 = const()[name = string("op_4137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4137_cast_fp16 = reduce_max(axes = var_4137_axes_0, keep_dims = var_4136, x = var_4133_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 var_4138_to_fp16 = const()[name = string("op_4138_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_97_cast_fp16 = maximum(x = var_4137_cast_fp16, y = var_4138_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_97_cast_fp16 = real_div(x = linear_54_cast_fp16, y = s_97_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4141_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_4141_cast_fp16")];
            tensor<int32, [1]> var_4146_axes_0 = const()[name = string("op_4146_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4146_keep_dims_0 = const()[name = string("op_4146_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4146_cast_fp16 = reduce_mean(axes = var_4146_axes_0, keep_dims = var_4146_keep_dims_0, x = var_4141_cast_fp16)[name = string("op_4146_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4147_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_4147_cast_fp16")];
            fp32 var_4148_epsilon_0 = const()[name = string("op_4148_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4148_cast_fp16 = inverse(epsilon = var_4148_epsilon_0, x = var_4147_cast_fp16)[name = string("op_4148_cast_fp16")];
            fp16 var_4149_to_fp16 = const()[name = string("op_4149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4150_cast_fp16 = mul(x = var_4148_cast_fp16, y = var_4149_to_fp16)[name = string("op_4150_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4152_cast_fp16 = add(x = var_4146_cast_fp16, y = var_4150_cast_fp16)[name = string("op_4152_cast_fp16")];
            fp32 var_4153_epsilon_0 = const()[name = string("op_4153_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4153_cast_fp16 = rsqrt(epsilon = var_4153_epsilon_0, x = var_4152_cast_fp16)[name = string("op_4153_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_85_cast_fp16 = mul(x = y_97_cast_fp16, y = var_4153_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39002432)))];
            tensor<fp16, [1, 32, 512]> var_4155_cast_fp16 = mul(x = out_85_cast_fp16, y = tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4155_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_11_cast_fp16 = add(x = input_69_cast_fp16, y = var_4155_cast_fp16)[name = string("h_11_cast_fp16")];
            tensor<fp16, [1]> tm_layers_5_layer_scalar_to_fp16 = const()[name = string("tm_layers_5_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-2])];
            tensor<fp16, [1, 32, 512]> x_109_cast_fp16 = mul(x = h_11_cast_fp16, y = tm_layers_5_layer_scalar_to_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4159_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_4159_cast_fp16")];
            bool var_4162 = const()[name = string("op_4162"), val = bool(true)];
            tensor<int32, [1]> var_4163_axes_0 = const()[name = string("op_4163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4163_cast_fp16 = reduce_max(axes = var_4163_axes_0, keep_dims = var_4162, x = var_4159_cast_fp16)[name = string("op_4163_cast_fp16")];
            fp16 var_4164_to_fp16 = const()[name = string("op_4164_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_99_cast_fp16 = maximum(x = var_4163_cast_fp16, y = var_4164_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_99_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_99_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4167_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_4167_cast_fp16")];
            tensor<int32, [1]> var_4172_axes_0 = const()[name = string("op_4172_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4172_keep_dims_0 = const()[name = string("op_4172_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4172_cast_fp16 = reduce_mean(axes = var_4172_axes_0, keep_dims = var_4172_keep_dims_0, x = var_4167_cast_fp16)[name = string("op_4172_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4173_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4174_cast_fp16 = inverse(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            fp16 var_4175_to_fp16 = const()[name = string("op_4175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4176_cast_fp16 = mul(x = var_4174_cast_fp16, y = var_4175_to_fp16)[name = string("op_4176_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4178_cast_fp16 = add(x = var_4172_cast_fp16, y = var_4176_cast_fp16)[name = string("op_4178_cast_fp16")];
            fp32 var_4179_epsilon_0 = const()[name = string("op_4179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4179_cast_fp16 = rsqrt(epsilon = var_4179_epsilon_0, x = var_4178_cast_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_87_cast_fp16 = mul(x = y_99_cast_fp16, y = var_4179_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39003520)))];
            tensor<fp16, [1, 32, 512]> input_73_cast_fp16 = mul(x = out_87_cast_fp16, y = tm_layers_6_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39004608))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39528960))))[name = string("tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [4]> var_4189 = const()[name = string("op_4189"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_4190_cast_fp16 = reshape(shape = var_4189, x = linear_55_cast_fp16)[name = string("op_4190_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39531072))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39793280))))[name = string("tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_4201 = const()[name = string("op_4201"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_4202_cast_fp16 = reshape(shape = var_4201, x = linear_56_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39794368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40056576))))[name = string("tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_57_cast_fp16)[name = string("op_4214_cast_fp16")];
            tensor<int32, [4]> x_115_perm_0 = const()[name = string("x_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4190_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 4, 32, 256]> var_4218_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_4218_cast_fp16")];
            bool var_4221 = const()[name = string("op_4221"), val = bool(true)];
            tensor<int32, [1]> var_4222_axes_0 = const()[name = string("op_4222_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_4222_cast_fp16 = reduce_max(axes = var_4222_axes_0, keep_dims = var_4221, x = var_4218_cast_fp16)[name = string("op_4222_cast_fp16")];
            fp16 var_4223_to_fp16 = const()[name = string("op_4223_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_101_cast_fp16 = maximum(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_101_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_101_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4226_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [1]> var_4231_axes_0 = const()[name = string("op_4231_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4231_keep_dims_0 = const()[name = string("op_4231_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_4231_cast_fp16 = reduce_mean(axes = var_4231_axes_0, keep_dims = var_4231_keep_dims_0, x = var_4226_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_4232_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_4232_cast_fp16")];
            fp32 var_4233_epsilon_0 = const()[name = string("op_4233_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_4233_cast_fp16 = inverse(epsilon = var_4233_epsilon_0, x = var_4232_cast_fp16)[name = string("op_4233_cast_fp16")];
            fp16 var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_4235_cast_fp16 = mul(x = var_4233_cast_fp16, y = var_4234_to_fp16)[name = string("op_4235_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_4237_cast_fp16 = add(x = var_4231_cast_fp16, y = var_4235_cast_fp16)[name = string("op_4237_cast_fp16")];
            fp32 var_4238_epsilon_0 = const()[name = string("op_4238_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_4238_cast_fp16 = rsqrt(epsilon = var_4238_epsilon_0, x = var_4237_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_89_cast_fp16 = mul(x = y_101_cast_fp16, y = var_4238_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40057664)))];
            tensor<fp16, [1, 4, 32, 256]> q_25_cast_fp16 = mul(x = out_89_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4202_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 2, 32, 256]> var_4241_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_4241_cast_fp16")];
            bool var_4244 = const()[name = string("op_4244"), val = bool(true)];
            tensor<int32, [1]> var_4245_axes_0 = const()[name = string("op_4245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_4245_cast_fp16 = reduce_max(axes = var_4245_axes_0, keep_dims = var_4244, x = var_4241_cast_fp16)[name = string("op_4245_cast_fp16")];
            fp16 var_4246_to_fp16 = const()[name = string("op_4246_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_103_cast_fp16 = maximum(x = var_4245_cast_fp16, y = var_4246_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_103_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_103_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4249_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<int32, [1]> var_4254_axes_0 = const()[name = string("op_4254_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4254_keep_dims_0 = const()[name = string("op_4254_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_4254_cast_fp16 = reduce_mean(axes = var_4254_axes_0, keep_dims = var_4254_keep_dims_0, x = var_4249_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4255_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_4255_cast_fp16")];
            fp32 var_4256_epsilon_0 = const()[name = string("op_4256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_4256_cast_fp16 = inverse(epsilon = var_4256_epsilon_0, x = var_4255_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp16 var_4257_to_fp16 = const()[name = string("op_4257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_4258_cast_fp16 = mul(x = var_4256_cast_fp16, y = var_4257_to_fp16)[name = string("op_4258_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4260_cast_fp16 = add(x = var_4254_cast_fp16, y = var_4258_cast_fp16)[name = string("op_4260_cast_fp16")];
            fp32 var_4261_epsilon_0 = const()[name = string("op_4261_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_4261_cast_fp16 = rsqrt(epsilon = var_4261_epsilon_0, x = var_4260_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_91_cast_fp16 = mul(x = y_103_cast_fp16, y = var_4261_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058240)))];
            tensor<fp16, [1, 2, 32, 256]> k_37_cast_fp16 = mul(x = out_91_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_37_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_115_cast_fp16 = transpose(perm = x_115_perm_0, x = var_4214_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 2, 32, 256]> var_4264_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_4264_cast_fp16")];
            bool var_4267 = const()[name = string("op_4267"), val = bool(true)];
            tensor<int32, [1]> var_4268_axes_0 = const()[name = string("op_4268_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_4268_cast_fp16 = reduce_max(axes = var_4268_axes_0, keep_dims = var_4267, x = var_4264_cast_fp16)[name = string("op_4268_cast_fp16")];
            fp16 var_4269_to_fp16 = const()[name = string("op_4269_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_105_cast_fp16 = maximum(x = var_4268_cast_fp16, y = var_4269_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_105_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_105_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4272_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_4272_cast_fp16")];
            tensor<int32, [1]> var_4277_axes_0 = const()[name = string("op_4277_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4277_keep_dims_0 = const()[name = string("op_4277_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_4277_cast_fp16 = reduce_mean(axes = var_4277_axes_0, keep_dims = var_4277_keep_dims_0, x = var_4272_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4278_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp32 var_4279_epsilon_0 = const()[name = string("op_4279_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_4279_cast_fp16 = inverse(epsilon = var_4279_epsilon_0, x = var_4278_cast_fp16)[name = string("op_4279_cast_fp16")];
            fp16 var_4280_to_fp16 = const()[name = string("op_4280_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_4281_cast_fp16 = mul(x = var_4279_cast_fp16, y = var_4280_to_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4283_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4281_cast_fp16)[name = string("op_4283_cast_fp16")];
            fp32 var_4284_epsilon_0 = const()[name = string("op_4284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_4284_cast_fp16 = rsqrt(epsilon = var_4284_epsilon_0, x = var_4283_cast_fp16)[name = string("op_4284_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_25_cast_fp16 = mul(x = y_105_cast_fp16, y = var_4284_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4286_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4286_cast_fp16")];
            tensor<int32, [4]> var_4291_begin_0 = const()[name = string("op_4291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4291_end_0 = const()[name = string("op_4291_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_4291_end_mask_0 = const()[name = string("op_4291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_4291_cast_fp16 = slice_by_index(begin = var_4291_begin_0, end = var_4291_end_0, end_mask = var_4291_end_mask_0, x = q_25_cast_fp16)[name = string("op_4291_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_4292_cast_fp16 = mul(x = var_4291_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = q_25_cast_fp16)[name = string("op_4297_cast_fp16")];
            int32 var_4299 = const()[name = string("op_4299"), val = int32(-1)];
            bool var_4300_interleave_0 = const()[name = string("op_4300_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_4300_cast_fp16 = concat(axis = var_4299, interleave = var_4300_interleave_0, values = (var_4292_cast_fp16, var_4297_cast_fp16))[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4301_cast_fp16 = mul(x = var_4300_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4301_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_27_cast_fp16 = add(x = var_4286_cast_fp16, y = var_4301_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4304_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4304_cast_fp16")];
            tensor<int32, [4]> var_4309_begin_0 = const()[name = string("op_4309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4309_end_0 = const()[name = string("op_4309_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_4309_end_mask_0 = const()[name = string("op_4309_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_4309_cast_fp16 = slice_by_index(begin = var_4309_begin_0, end = var_4309_end_0, end_mask = var_4309_end_mask_0, x = k_37_cast_fp16)[name = string("op_4309_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_4310_cast_fp16 = mul(x = var_4309_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_4310_cast_fp16")];
            tensor<int32, [4]> var_4315_begin_0 = const()[name = string("op_4315_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4315_end_0 = const()[name = string("op_4315_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_4315_end_mask_0 = const()[name = string("op_4315_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_4315_cast_fp16 = slice_by_index(begin = var_4315_begin_0, end = var_4315_end_0, end_mask = var_4315_end_mask_0, x = k_37_cast_fp16)[name = string("op_4315_cast_fp16")];
            int32 var_4317 = const()[name = string("op_4317"), val = int32(-1)];
            bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_4318_cast_fp16 = concat(axis = var_4317, interleave = var_4318_interleave_0, values = (var_4310_cast_fp16, var_4315_cast_fp16))[name = string("op_4318_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4319_cast_fp16 = mul(x = var_4318_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_39_cast_fp16 = add(x = var_4304_cast_fp16, y = var_4319_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = k_39_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [4, 1, 32, 256]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
            tensor<int32, [5]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_24_cast_fp16 = reshape(shape = concat_24, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [4, 1, 32, 256]> reshape_25_cast_fp16 = reshape(shape = concat_25, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = v_25_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [4, 1, 32, 256]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
            tensor<int32, [5]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_26_cast_fp16 = reshape(shape = concat_26, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [4, 1, 32, 256]> reshape_27_cast_fp16 = reshape(shape = concat_27, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4333_transpose_x_0 = const()[name = string("op_4333_transpose_x_0"), val = bool(false)];
            bool var_4333_transpose_y_0 = const()[name = string("op_4333_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_102_cast_fp16 = transpose(perm = transpose_102_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 4, 32, 32]> var_4333_cast_fp16 = matmul(transpose_x = var_4333_transpose_x_0, transpose_y = var_4333_transpose_y_0, x = q_27_cast_fp16, y = transpose_102_cast_fp16)[name = string("op_4333_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_4335_cast_fp16 = add(x = var_4333_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4335_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_105_cast_fp16 = softmax(axis = var_4336, x = var_4335_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4339_transpose_x_0 = const()[name = string("op_4339_transpose_x_0"), val = bool(false)];
            bool var_4339_transpose_y_0 = const()[name = string("op_4339_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 4, 32, 256]> var_4339_cast_fp16 = matmul(transpose_x = var_4339_transpose_x_0, transpose_y = var_4339_transpose_y_0, x = w_105_cast_fp16, y = v_27_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<int32, [4]> var_4342_perm_0 = const()[name = string("op_4342_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4346 = const()[name = string("op_4346"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_4342_cast_fp16 = transpose(perm = var_4342_perm_0, x = var_4339_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 32, 1024]> input_75_cast_fp16 = reshape(shape = var_4346, x = var_4342_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40583168))))[name = string("tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4351_cast_fp16 = abs(x = linear_58_cast_fp16)[name = string("op_4351_cast_fp16")];
            bool var_4354 = const()[name = string("op_4354"), val = bool(true)];
            tensor<int32, [1]> var_4355_axes_0 = const()[name = string("op_4355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4355_cast_fp16 = reduce_max(axes = var_4355_axes_0, keep_dims = var_4354, x = var_4351_cast_fp16)[name = string("op_4355_cast_fp16")];
            fp16 var_4356_to_fp16 = const()[name = string("op_4356_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_107_cast_fp16 = maximum(x = var_4355_cast_fp16, y = var_4356_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_107_cast_fp16 = real_div(x = linear_58_cast_fp16, y = s_107_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4359_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<int32, [1]> var_4364_axes_0 = const()[name = string("op_4364_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4364_keep_dims_0 = const()[name = string("op_4364_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4364_cast_fp16 = reduce_mean(axes = var_4364_axes_0, keep_dims = var_4364_keep_dims_0, x = var_4359_cast_fp16)[name = string("op_4364_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4365_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_4365_cast_fp16")];
            fp32 var_4366_epsilon_0 = const()[name = string("op_4366_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4366_cast_fp16 = inverse(epsilon = var_4366_epsilon_0, x = var_4365_cast_fp16)[name = string("op_4366_cast_fp16")];
            fp16 var_4367_to_fp16 = const()[name = string("op_4367_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4368_cast_fp16 = mul(x = var_4366_cast_fp16, y = var_4367_to_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4370_cast_fp16 = add(x = var_4364_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            fp32 var_4371_epsilon_0 = const()[name = string("op_4371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4371_cast_fp16 = rsqrt(epsilon = var_4371_epsilon_0, x = var_4370_cast_fp16)[name = string("op_4371_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_93_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4371_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40584256)))];
            tensor<fp16, [1, 32, 512]> var_4373_cast_fp16 = mul(x = out_93_cast_fp16, y = tm_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_4373_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = var_4373_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4376_cast_fp16 = abs(x = x_119_cast_fp16)[name = string("op_4376_cast_fp16")];
            bool var_4379 = const()[name = string("op_4379"), val = bool(true)];
            tensor<int32, [1]> var_4380_axes_0 = const()[name = string("op_4380_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4380_cast_fp16 = reduce_max(axes = var_4380_axes_0, keep_dims = var_4379, x = var_4376_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 var_4381_to_fp16 = const()[name = string("op_4381_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_109_cast_fp16 = maximum(x = var_4380_cast_fp16, y = var_4381_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_109_cast_fp16 = real_div(x = x_119_cast_fp16, y = s_109_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4384_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4384_cast_fp16")];
            tensor<int32, [1]> var_4389_axes_0 = const()[name = string("op_4389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4389_keep_dims_0 = const()[name = string("op_4389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4389_cast_fp16 = reduce_mean(axes = var_4389_axes_0, keep_dims = var_4389_keep_dims_0, x = var_4384_cast_fp16)[name = string("op_4389_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4390_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_4390_cast_fp16")];
            fp32 var_4391_epsilon_0 = const()[name = string("op_4391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4391_cast_fp16 = inverse(epsilon = var_4391_epsilon_0, x = var_4390_cast_fp16)[name = string("op_4391_cast_fp16")];
            fp16 var_4392_to_fp16 = const()[name = string("op_4392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = string("op_4393_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4395_cast_fp16 = add(x = var_4389_cast_fp16, y = var_4393_cast_fp16)[name = string("op_4395_cast_fp16")];
            fp32 var_4396_epsilon_0 = const()[name = string("op_4396_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4396_cast_fp16 = rsqrt(epsilon = var_4396_epsilon_0, x = var_4395_cast_fp16)[name = string("op_4396_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_95_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4396_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40585344)))];
            tensor<fp16, [1, 32, 512]> input_77_cast_fp16 = mul(x = out_95_cast_fp16, y = tm_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40586432))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41635072))))[name = string("tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_59_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_59_cast_fp16")];
            string var_4403_mode_0 = const()[name = string("op_4403_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_4403_cast_fp16 = gelu(mode = var_4403_mode_0, x = linear_59_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41639232))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42687872))))[name = string("tm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_60_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_79_cast_fp16 = mul(x = var_4403_cast_fp16, y = linear_60_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42692032))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43740672))))[name = string("tm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4411_cast_fp16 = abs(x = linear_61_cast_fp16)[name = string("op_4411_cast_fp16")];
            bool var_4414 = const()[name = string("op_4414"), val = bool(true)];
            tensor<int32, [1]> var_4415_axes_0 = const()[name = string("op_4415_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4415_cast_fp16 = reduce_max(axes = var_4415_axes_0, keep_dims = var_4414, x = var_4411_cast_fp16)[name = string("op_4415_cast_fp16")];
            fp16 var_4416_to_fp16 = const()[name = string("op_4416_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_111_cast_fp16 = maximum(x = var_4415_cast_fp16, y = var_4416_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_111_cast_fp16 = real_div(x = linear_61_cast_fp16, y = s_111_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4419_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4419_cast_fp16")];
            tensor<int32, [1]> var_4424_axes_0 = const()[name = string("op_4424_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4424_keep_dims_0 = const()[name = string("op_4424_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4424_cast_fp16 = reduce_mean(axes = var_4424_axes_0, keep_dims = var_4424_keep_dims_0, x = var_4419_cast_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4425_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4426_cast_fp16 = inverse(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            fp16 var_4427_to_fp16 = const()[name = string("op_4427_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4428_cast_fp16 = mul(x = var_4426_cast_fp16, y = var_4427_to_fp16)[name = string("op_4428_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4430_cast_fp16 = add(x = var_4424_cast_fp16, y = var_4428_cast_fp16)[name = string("op_4430_cast_fp16")];
            fp32 var_4431_epsilon_0 = const()[name = string("op_4431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4431_cast_fp16 = rsqrt(epsilon = var_4431_epsilon_0, x = var_4430_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_97_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4431_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43741760)))];
            tensor<fp16, [1, 32, 512]> var_4433_cast_fp16 = mul(x = out_97_cast_fp16, y = tm_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4433_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_81_cast_fp16 = add(x = x_119_cast_fp16, y = var_4433_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43742848))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44005056))))[name = string("tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_62_cast_fp16")];
            string x_123_mode_0 = const()[name = string("x_123_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_123_cast_fp16 = gelu(mode = x_123_mode_0, x = linear_62_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [4]> var_4453_begin_0 = const()[name = string("op_4453_begin_0"), val = tensor<int32, [4]>([0, 0, 6, 0])];
            tensor<int32, [4]> var_4453_end_0 = const()[name = string("op_4453_end_0"), val = tensor<int32, [4]>([1, 32, 7, 512])];
            tensor<bool, [4]> var_4453_end_mask_0 = const()[name = string("op_4453_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4453_squeeze_mask_0 = const()[name = string("op_4453_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_4453_cast_fp16 = slice_by_index(begin = var_4453_begin_0, end = var_4453_end_0, end_mask = var_4453_end_mask_0, squeeze_mask = var_4453_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_83_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4453_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44006144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44268352))))[name = string("tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4463_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4463_cast_fp16")];
            bool var_4466 = const()[name = string("op_4466"), val = bool(true)];
            tensor<int32, [1]> var_4467_axes_0 = const()[name = string("op_4467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4467_cast_fp16 = reduce_max(axes = var_4467_axes_0, keep_dims = var_4466, x = var_4463_cast_fp16)[name = string("op_4467_cast_fp16")];
            fp16 var_4468_to_fp16 = const()[name = string("op_4468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_113_cast_fp16 = maximum(x = var_4467_cast_fp16, y = var_4468_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_113_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_113_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4471_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [1]> var_4476_axes_0 = const()[name = string("op_4476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4476_keep_dims_0 = const()[name = string("op_4476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4476_cast_fp16 = reduce_mean(axes = var_4476_axes_0, keep_dims = var_4476_keep_dims_0, x = var_4471_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4477_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_fp16)[name = string("op_4477_cast_fp16")];
            fp32 var_4478_epsilon_0 = const()[name = string("op_4478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4478_cast_fp16 = inverse(epsilon = var_4478_epsilon_0, x = var_4477_cast_fp16)[name = string("op_4478_cast_fp16")];
            fp16 var_4479_to_fp16 = const()[name = string("op_4479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4480_cast_fp16 = mul(x = var_4478_cast_fp16, y = var_4479_to_fp16)[name = string("op_4480_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4482_cast_fp16 = add(x = var_4476_cast_fp16, y = var_4480_cast_fp16)[name = string("op_4482_cast_fp16")];
            fp32 var_4483_epsilon_0 = const()[name = string("op_4483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4483_cast_fp16 = rsqrt(epsilon = var_4483_epsilon_0, x = var_4482_cast_fp16)[name = string("op_4483_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_99_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4483_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44269440)))];
            tensor<fp16, [1, 32, 512]> var_4485_cast_fp16 = mul(x = out_99_cast_fp16, y = tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_13_cast_fp16 = add(x = input_81_cast_fp16, y = var_4485_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<fp16, [1]> tm_layers_6_layer_scalar_to_fp16 = const()[name = string("tm_layers_6_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.cep-3])];
            tensor<fp16, [1, 32, 512]> x_127_cast_fp16 = mul(x = h_13_cast_fp16, y = tm_layers_6_layer_scalar_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4489_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_4489_cast_fp16")];
            bool var_4492 = const()[name = string("op_4492"), val = bool(true)];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4493_cast_fp16 = reduce_max(axes = var_4493_axes_0, keep_dims = var_4492, x = var_4489_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_115_cast_fp16 = maximum(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_115_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_115_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4497_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<int32, [1]> var_4502_axes_0 = const()[name = string("op_4502_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4502_keep_dims_0 = const()[name = string("op_4502_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4502_cast_fp16 = reduce_mean(axes = var_4502_axes_0, keep_dims = var_4502_keep_dims_0, x = var_4497_cast_fp16)[name = string("op_4502_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4503_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4504_cast_fp16 = inverse(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            fp16 var_4505_to_fp16 = const()[name = string("op_4505_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4506_cast_fp16 = mul(x = var_4504_cast_fp16, y = var_4505_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4508_cast_fp16 = add(x = var_4502_cast_fp16, y = var_4506_cast_fp16)[name = string("op_4508_cast_fp16")];
            fp32 var_4509_epsilon_0 = const()[name = string("op_4509_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4509_cast_fp16 = rsqrt(epsilon = var_4509_epsilon_0, x = var_4508_cast_fp16)[name = string("op_4509_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_101_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4509_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44270528)))];
            tensor<fp16, [1, 32, 512]> input_85_cast_fp16 = mul(x = out_101_cast_fp16, y = tm_layers_7_input_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44271616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44795968))))[name = string("tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4519 = const()[name = string("op_4519"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_4520_cast_fp16 = reshape(shape = var_4519, x = linear_64_cast_fp16)[name = string("op_4520_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44798080))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45060288))))[name = string("tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4531 = const()[name = string("op_4531"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_4532_cast_fp16 = reshape(shape = var_4531, x = linear_65_cast_fp16)[name = string("op_4532_cast_fp16")];
            tensor<int32, [4]> x_131_perm_0 = const()[name = string("x_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45061376))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45323584))))[name = string("tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4543 = const()[name = string("op_4543"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_4544_cast_fp16 = reshape(shape = var_4543, x = linear_66_cast_fp16)[name = string("op_4544_cast_fp16")];
            tensor<int32, [4]> x_133_perm_0 = const()[name = string("x_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_4520_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 4, 32, 256]> var_4548_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_4548_cast_fp16")];
            bool var_4551 = const()[name = string("op_4551"), val = bool(true)];
            tensor<int32, [1]> var_4552_axes_0 = const()[name = string("op_4552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_4552_cast_fp16 = reduce_max(axes = var_4552_axes_0, keep_dims = var_4551, x = var_4548_cast_fp16)[name = string("op_4552_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_117_cast_fp16 = maximum(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_117_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_117_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4556_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<int32, [1]> var_4561_axes_0 = const()[name = string("op_4561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4561_keep_dims_0 = const()[name = string("op_4561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_4561_cast_fp16 = reduce_mean(axes = var_4561_axes_0, keep_dims = var_4561_keep_dims_0, x = var_4556_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_4562_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4562_cast_fp16")];
            fp32 var_4563_epsilon_0 = const()[name = string("op_4563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_4563_cast_fp16 = inverse(epsilon = var_4563_epsilon_0, x = var_4562_cast_fp16)[name = string("op_4563_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_4565_cast_fp16 = mul(x = var_4563_cast_fp16, y = var_4564_to_fp16)[name = string("op_4565_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_4567_cast_fp16 = add(x = var_4561_cast_fp16, y = var_4565_cast_fp16)[name = string("op_4567_cast_fp16")];
            fp32 var_4568_epsilon_0 = const()[name = string("op_4568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_4568_cast_fp16 = rsqrt(epsilon = var_4568_epsilon_0, x = var_4567_cast_fp16)[name = string("op_4568_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_103_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4568_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_29_cast_fp16 = mul(x = out_103_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_131_cast_fp16 = transpose(perm = x_131_perm_0, x = var_4532_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2, 32, 256]> var_4571_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_4571_cast_fp16")];
            bool var_4574 = const()[name = string("op_4574"), val = bool(true)];
            tensor<int32, [1]> var_4575_axes_0 = const()[name = string("op_4575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_4575_cast_fp16 = reduce_max(axes = var_4575_axes_0, keep_dims = var_4574, x = var_4571_cast_fp16)[name = string("op_4575_cast_fp16")];
            fp16 var_4576_to_fp16 = const()[name = string("op_4576_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_119_cast_fp16 = maximum(x = var_4575_cast_fp16, y = var_4576_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_119_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_119_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4579_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4579_cast_fp16")];
            tensor<int32, [1]> var_4584_axes_0 = const()[name = string("op_4584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4584_keep_dims_0 = const()[name = string("op_4584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_4584_cast_fp16 = reduce_mean(axes = var_4584_axes_0, keep_dims = var_4584_keep_dims_0, x = var_4579_cast_fp16)[name = string("op_4584_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4585_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4585_cast_fp16")];
            fp32 var_4586_epsilon_0 = const()[name = string("op_4586_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_4586_cast_fp16 = inverse(epsilon = var_4586_epsilon_0, x = var_4585_cast_fp16)[name = string("op_4586_cast_fp16")];
            fp16 var_4587_to_fp16 = const()[name = string("op_4587_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_4588_cast_fp16 = mul(x = var_4586_cast_fp16, y = var_4587_to_fp16)[name = string("op_4588_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4590_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4588_cast_fp16)[name = string("op_4590_cast_fp16")];
            fp32 var_4591_epsilon_0 = const()[name = string("op_4591_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_4591_cast_fp16 = rsqrt(epsilon = var_4591_epsilon_0, x = var_4590_cast_fp16)[name = string("op_4591_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_105_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4591_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [256]> tm_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45324672)))];
            tensor<fp16, [1, 2, 32, 256]> k_43_cast_fp16 = mul(x = out_105_cast_fp16, y = tm_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_133_cast_fp16 = transpose(perm = x_133_perm_0, x = var_4544_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 2, 32, 256]> var_4594_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_4594_cast_fp16")];
            bool var_4597 = const()[name = string("op_4597"), val = bool(true)];
            tensor<int32, [1]> var_4598_axes_0 = const()[name = string("op_4598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_4598_cast_fp16 = reduce_max(axes = var_4598_axes_0, keep_dims = var_4597, x = var_4594_cast_fp16)[name = string("op_4598_cast_fp16")];
            fp16 var_4599_to_fp16 = const()[name = string("op_4599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_121_cast_fp16 = maximum(x = var_4598_cast_fp16, y = var_4599_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_121_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_121_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4602_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4602_cast_fp16")];
            tensor<int32, [1]> var_4607_axes_0 = const()[name = string("op_4607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4607_keep_dims_0 = const()[name = string("op_4607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_4607_cast_fp16 = reduce_mean(axes = var_4607_axes_0, keep_dims = var_4607_keep_dims_0, x = var_4602_cast_fp16)[name = string("op_4607_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4608_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4608_cast_fp16")];
            fp32 var_4609_epsilon_0 = const()[name = string("op_4609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_4609_cast_fp16 = inverse(epsilon = var_4609_epsilon_0, x = var_4608_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_4611_cast_fp16 = mul(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("op_4611_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4613_cast_fp16 = add(x = var_4607_cast_fp16, y = var_4611_cast_fp16)[name = string("op_4613_cast_fp16")];
            fp32 var_4614_epsilon_0 = const()[name = string("op_4614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_4614_cast_fp16 = rsqrt(epsilon = var_4614_epsilon_0, x = var_4613_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_29_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4614_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4616_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4616_cast_fp16")];
            tensor<int32, [4]> var_4621_begin_0 = const()[name = string("op_4621_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4621_end_0 = const()[name = string("op_4621_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_4621_end_mask_0 = const()[name = string("op_4621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_4621_cast_fp16 = slice_by_index(begin = var_4621_begin_0, end = var_4621_end_0, end_mask = var_4621_end_mask_0, x = q_29_cast_fp16)[name = string("op_4621_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_4622_cast_fp16 = mul(x = var_4621_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<int32, [4]> var_4627_begin_0 = const()[name = string("op_4627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4627_end_0 = const()[name = string("op_4627_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_4627_end_mask_0 = const()[name = string("op_4627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_4627_cast_fp16 = slice_by_index(begin = var_4627_begin_0, end = var_4627_end_0, end_mask = var_4627_end_mask_0, x = q_29_cast_fp16)[name = string("op_4627_cast_fp16")];
            int32 var_4629 = const()[name = string("op_4629"), val = int32(-1)];
            bool var_4630_interleave_0 = const()[name = string("op_4630_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_4630_cast_fp16 = concat(axis = var_4629, interleave = var_4630_interleave_0, values = (var_4622_cast_fp16, var_4627_cast_fp16))[name = string("op_4630_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4631_cast_fp16 = mul(x = var_4630_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4631_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_31_cast_fp16 = add(x = var_4616_cast_fp16, y = var_4631_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4634_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4634_cast_fp16")];
            tensor<int32, [4]> var_4639_begin_0 = const()[name = string("op_4639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4639_end_0 = const()[name = string("op_4639_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_4639_end_mask_0 = const()[name = string("op_4639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_4639_cast_fp16 = slice_by_index(begin = var_4639_begin_0, end = var_4639_end_0, end_mask = var_4639_end_mask_0, x = k_43_cast_fp16)[name = string("op_4639_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_4640_cast_fp16 = mul(x = var_4639_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4640_cast_fp16")];
            tensor<int32, [4]> var_4645_begin_0 = const()[name = string("op_4645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4645_end_0 = const()[name = string("op_4645_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_4645_end_mask_0 = const()[name = string("op_4645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_4645_cast_fp16 = slice_by_index(begin = var_4645_begin_0, end = var_4645_end_0, end_mask = var_4645_end_mask_0, x = k_43_cast_fp16)[name = string("op_4645_cast_fp16")];
            int32 var_4647 = const()[name = string("op_4647"), val = int32(-1)];
            bool var_4648_interleave_0 = const()[name = string("op_4648_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_4648_cast_fp16 = concat(axis = var_4647, interleave = var_4648_interleave_0, values = (var_4640_cast_fp16, var_4645_cast_fp16))[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4649_cast_fp16 = mul(x = var_4648_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4649_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_45_cast_fp16 = add(x = var_4634_cast_fp16, y = var_4649_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_14_reps_0 = const()[name = string("tile_14_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = k_45_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [4, 1, 32, 256]> tile_14_cast_fp16 = tile(reps = tile_14_reps_0, x = transpose_28_cast_fp16)[name = string("tile_14_cast_fp16")];
            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_28_cast_fp16 = reshape(shape = concat_28, x = tile_14_cast_fp16)[name = string("reshape_28_cast_fp16")];
            tensor<int32, [5]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_28_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [4, 1, 32, 256]> reshape_29_cast_fp16 = reshape(shape = concat_29, x = transpose_29_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_15_reps_0 = const()[name = string("tile_15_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = v_29_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [4, 1, 32, 256]> tile_15_cast_fp16 = tile(reps = tile_15_reps_0, x = transpose_30_cast_fp16)[name = string("tile_15_cast_fp16")];
            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_30_cast_fp16 = reshape(shape = concat_30, x = tile_15_cast_fp16)[name = string("reshape_30_cast_fp16")];
            tensor<int32, [5]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_30_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [4, 1, 32, 256]> reshape_31_cast_fp16 = reshape(shape = concat_31, x = transpose_31_cast_fp16)[name = string("reshape_31_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4663_transpose_x_0 = const()[name = string("op_4663_transpose_x_0"), val = bool(false)];
            bool var_4663_transpose_y_0 = const()[name = string("op_4663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 4, 32, 32]> var_4663_cast_fp16 = matmul(transpose_x = var_4663_transpose_x_0, transpose_y = var_4663_transpose_y_0, x = q_31_cast_fp16, y = transpose_103_cast_fp16)[name = string("op_4663_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_4665_cast_fp16 = add(x = var_4663_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4666 = const()[name = string("op_4666"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_121_cast_fp16 = softmax(axis = var_4666, x = var_4665_cast_fp16)[name = string("w_121_cast_fp16")];
            bool var_4669_transpose_x_0 = const()[name = string("op_4669_transpose_x_0"), val = bool(false)];
            bool var_4669_transpose_y_0 = const()[name = string("op_4669_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = reshape_31_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 4, 32, 256]> var_4669_cast_fp16 = matmul(transpose_x = var_4669_transpose_x_0, transpose_y = var_4669_transpose_y_0, x = w_121_cast_fp16, y = v_31_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<int32, [4]> var_4672_perm_0 = const()[name = string("op_4672_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4676 = const()[name = string("op_4676"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_4672_cast_fp16 = transpose(perm = var_4672_perm_0, x = var_4669_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 32, 1024]> input_87_cast_fp16 = reshape(shape = var_4676, x = var_4672_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45325248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45849600))))[name = string("tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4681_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4684 = const()[name = string("op_4684"), val = bool(true)];
            tensor<int32, [1]> var_4685_axes_0 = const()[name = string("op_4685_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4685_cast_fp16 = reduce_max(axes = var_4685_axes_0, keep_dims = var_4684, x = var_4681_cast_fp16)[name = string("op_4685_cast_fp16")];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_123_cast_fp16 = maximum(x = var_4685_cast_fp16, y = var_4686_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_123_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_123_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4689_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [1]> var_4694_axes_0 = const()[name = string("op_4694_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4694_keep_dims_0 = const()[name = string("op_4694_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4694_cast_fp16 = reduce_mean(axes = var_4694_axes_0, keep_dims = var_4694_keep_dims_0, x = var_4689_cast_fp16)[name = string("op_4694_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4695_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4695_cast_fp16")];
            fp32 var_4696_epsilon_0 = const()[name = string("op_4696_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4696_cast_fp16 = inverse(epsilon = var_4696_epsilon_0, x = var_4695_cast_fp16)[name = string("op_4696_cast_fp16")];
            fp16 var_4697_to_fp16 = const()[name = string("op_4697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4698_cast_fp16 = mul(x = var_4696_cast_fp16, y = var_4697_to_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4700_cast_fp16 = add(x = var_4694_cast_fp16, y = var_4698_cast_fp16)[name = string("op_4700_cast_fp16")];
            fp32 var_4701_epsilon_0 = const()[name = string("op_4701_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4701_cast_fp16 = rsqrt(epsilon = var_4701_epsilon_0, x = var_4700_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_107_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4701_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45850688)))];
            tensor<fp16, [1, 32, 512]> var_4703_cast_fp16 = mul(x = out_107_cast_fp16, y = tm_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4703_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_137_cast_fp16 = add(x = x_127_cast_fp16, y = var_4703_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4706_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_4706_cast_fp16")];
            bool var_4709 = const()[name = string("op_4709"), val = bool(true)];
            tensor<int32, [1]> var_4710_axes_0 = const()[name = string("op_4710_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4710_cast_fp16 = reduce_max(axes = var_4710_axes_0, keep_dims = var_4709, x = var_4706_cast_fp16)[name = string("op_4710_cast_fp16")];
            fp16 var_4711_to_fp16 = const()[name = string("op_4711_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_125_cast_fp16 = maximum(x = var_4710_cast_fp16, y = var_4711_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_125_cast_fp16 = real_div(x = x_137_cast_fp16, y = s_125_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4714_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4714_cast_fp16")];
            tensor<int32, [1]> var_4719_axes_0 = const()[name = string("op_4719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4719_keep_dims_0 = const()[name = string("op_4719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4719_cast_fp16 = reduce_mean(axes = var_4719_axes_0, keep_dims = var_4719_keep_dims_0, x = var_4714_cast_fp16)[name = string("op_4719_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4720_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4720_cast_fp16")];
            fp32 var_4721_epsilon_0 = const()[name = string("op_4721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4721_cast_fp16 = inverse(epsilon = var_4721_epsilon_0, x = var_4720_cast_fp16)[name = string("op_4721_cast_fp16")];
            fp16 var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4723_cast_fp16 = mul(x = var_4721_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4725_cast_fp16 = add(x = var_4719_cast_fp16, y = var_4723_cast_fp16)[name = string("op_4725_cast_fp16")];
            fp32 var_4726_epsilon_0 = const()[name = string("op_4726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4726_cast_fp16 = rsqrt(epsilon = var_4726_epsilon_0, x = var_4725_cast_fp16)[name = string("op_4726_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_109_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4726_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45851776)))];
            tensor<fp16, [1, 32, 512]> input_89_cast_fp16 = mul(x = out_109_cast_fp16, y = tm_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45852864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46901504))))[name = string("tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4733_mode_0 = const()[name = string("op_4733_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_4733_cast_fp16 = gelu(mode = var_4733_mode_0, x = linear_68_cast_fp16)[name = string("op_4733_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46905664))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47954304))))[name = string("tm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_91_cast_fp16 = mul(x = var_4733_cast_fp16, y = linear_69_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47958464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49007104))))[name = string("tm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4741_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4741_cast_fp16")];
            bool var_4744 = const()[name = string("op_4744"), val = bool(true)];
            tensor<int32, [1]> var_4745_axes_0 = const()[name = string("op_4745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4745_cast_fp16 = reduce_max(axes = var_4745_axes_0, keep_dims = var_4744, x = var_4741_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_127_cast_fp16 = maximum(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_127_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_127_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4749_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<int32, [1]> var_4754_axes_0 = const()[name = string("op_4754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4754_keep_dims_0 = const()[name = string("op_4754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4754_cast_fp16 = reduce_mean(axes = var_4754_axes_0, keep_dims = var_4754_keep_dims_0, x = var_4749_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4755_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4755_cast_fp16")];
            fp32 var_4756_epsilon_0 = const()[name = string("op_4756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4756_cast_fp16 = inverse(epsilon = var_4756_epsilon_0, x = var_4755_cast_fp16)[name = string("op_4756_cast_fp16")];
            fp16 var_4757_to_fp16 = const()[name = string("op_4757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4758_cast_fp16 = mul(x = var_4756_cast_fp16, y = var_4757_to_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4760_cast_fp16 = add(x = var_4754_cast_fp16, y = var_4758_cast_fp16)[name = string("op_4760_cast_fp16")];
            fp32 var_4761_epsilon_0 = const()[name = string("op_4761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4761_cast_fp16 = rsqrt(epsilon = var_4761_epsilon_0, x = var_4760_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_111_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4761_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49008192)))];
            tensor<fp16, [1, 32, 512]> var_4763_cast_fp16 = mul(x = out_111_cast_fp16, y = tm_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4763_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_93_cast_fp16 = add(x = x_137_cast_fp16, y = var_4763_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49009280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49271488))))[name = string("tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_71_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_71_cast_fp16")];
            string x_141_mode_0 = const()[name = string("x_141_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_141_cast_fp16 = gelu(mode = x_141_mode_0, x = linear_71_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_4783_begin_0 = const()[name = string("op_4783_begin_0"), val = tensor<int32, [4]>([0, 0, 7, 0])];
            tensor<int32, [4]> var_4783_end_0 = const()[name = string("op_4783_end_0"), val = tensor<int32, [4]>([1, 32, 8, 512])];
            tensor<bool, [4]> var_4783_end_mask_0 = const()[name = string("op_4783_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4783_squeeze_mask_0 = const()[name = string("op_4783_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_4783_cast_fp16 = slice_by_index(begin = var_4783_begin_0, end = var_4783_end_0, end_mask = var_4783_end_mask_0, squeeze_mask = var_4783_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_95_cast_fp16 = mul(x = x_141_cast_fp16, y = var_4783_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49272576))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49534784))))[name = string("tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4793_cast_fp16 = abs(x = linear_72_cast_fp16)[name = string("op_4793_cast_fp16")];
            bool var_4796 = const()[name = string("op_4796"), val = bool(true)];
            tensor<int32, [1]> var_4797_axes_0 = const()[name = string("op_4797_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4797_cast_fp16 = reduce_max(axes = var_4797_axes_0, keep_dims = var_4796, x = var_4793_cast_fp16)[name = string("op_4797_cast_fp16")];
            fp16 var_4798_to_fp16 = const()[name = string("op_4798_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_129_cast_fp16 = maximum(x = var_4797_cast_fp16, y = var_4798_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_129_cast_fp16 = real_div(x = linear_72_cast_fp16, y = s_129_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4801_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4801_cast_fp16")];
            tensor<int32, [1]> var_4806_axes_0 = const()[name = string("op_4806_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4806_keep_dims_0 = const()[name = string("op_4806_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4806_cast_fp16 = reduce_mean(axes = var_4806_axes_0, keep_dims = var_4806_keep_dims_0, x = var_4801_cast_fp16)[name = string("op_4806_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4807_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4807_cast_fp16")];
            fp32 var_4808_epsilon_0 = const()[name = string("op_4808_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4808_cast_fp16 = inverse(epsilon = var_4808_epsilon_0, x = var_4807_cast_fp16)[name = string("op_4808_cast_fp16")];
            fp16 var_4809_to_fp16 = const()[name = string("op_4809_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4810_cast_fp16 = mul(x = var_4808_cast_fp16, y = var_4809_to_fp16)[name = string("op_4810_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4812_cast_fp16 = add(x = var_4806_cast_fp16, y = var_4810_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp32 var_4813_epsilon_0 = const()[name = string("op_4813_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4813_cast_fp16 = rsqrt(epsilon = var_4813_epsilon_0, x = var_4812_cast_fp16)[name = string("op_4813_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_113_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4813_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49535872)))];
            tensor<fp16, [1, 32, 512]> var_4815_cast_fp16 = mul(x = out_113_cast_fp16, y = tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4815_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_15_cast_fp16 = add(x = input_93_cast_fp16, y = var_4815_cast_fp16)[name = string("h_15_cast_fp16")];
            tensor<fp16, [1]> tm_layers_7_layer_scalar_to_fp16 = const()[name = string("tm_layers_7_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-1])];
            tensor<fp16, [1, 32, 512]> x_145_cast_fp16 = mul(x = h_15_cast_fp16, y = tm_layers_7_layer_scalar_to_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4819_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4819_cast_fp16")];
            bool var_4822 = const()[name = string("op_4822"), val = bool(true)];
            tensor<int32, [1]> var_4823_axes_0 = const()[name = string("op_4823_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_4823_cast_fp16 = reduce_max(axes = var_4823_axes_0, keep_dims = var_4822, x = var_4819_cast_fp16)[name = string("op_4823_cast_fp16")];
            fp16 var_4824_to_fp16 = const()[name = string("op_4824_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_131_cast_fp16 = maximum(x = var_4823_cast_fp16, y = var_4824_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_131_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_131_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_4827_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [1]> var_4832_axes_0 = const()[name = string("op_4832_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4832_keep_dims_0 = const()[name = string("op_4832_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_4832_cast_fp16 = reduce_mean(axes = var_4832_axes_0, keep_dims = var_4832_keep_dims_0, x = var_4827_cast_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4833_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4833_cast_fp16")];
            fp32 var_4834_epsilon_0 = const()[name = string("op_4834_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_4834_cast_fp16 = inverse(epsilon = var_4834_epsilon_0, x = var_4833_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp16 var_4835_to_fp16 = const()[name = string("op_4835_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_4836_cast_fp16 = mul(x = var_4834_cast_fp16, y = var_4835_to_fp16)[name = string("op_4836_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_4838_cast_fp16 = add(x = var_4832_cast_fp16, y = var_4836_cast_fp16)[name = string("op_4838_cast_fp16")];
            fp32 var_4839_epsilon_0 = const()[name = string("op_4839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_4839_cast_fp16 = rsqrt(epsilon = var_4839_epsilon_0, x = var_4838_cast_fp16)[name = string("op_4839_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_115_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4839_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49536960)))];
            tensor<fp16, [1, 32, 512]> input_97_cast_fp16 = mul(x = out_115_cast_fp16, y = tm_layers_8_input_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49538048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50062400))))[name = string("tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4849 = const()[name = string("op_4849"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_4850_cast_fp16 = reshape(shape = var_4849, x = linear_73_cast_fp16)[name = string("op_4850_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50064512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50326720))))[name = string("tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [4]> var_4861 = const()[name = string("op_4861"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_4862_cast_fp16 = reshape(shape = var_4861, x = linear_74_cast_fp16)[name = string("op_4862_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50327808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50590016))))[name = string("tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [4]> var_4873 = const()[name = string("op_4873"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_4874_cast_fp16 = reshape(shape = var_4873, x = linear_75_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<int32, [4]> x_151_perm_0 = const()[name = string("x_151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_4850_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 4, 32, 256]> var_4878_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4878_cast_fp16")];
            bool var_4881 = const()[name = string("op_4881"), val = bool(true)];
            tensor<int32, [1]> var_4882_axes_0 = const()[name = string("op_4882_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_4882_cast_fp16 = reduce_max(axes = var_4882_axes_0, keep_dims = var_4881, x = var_4878_cast_fp16)[name = string("op_4882_cast_fp16")];
            fp16 var_4883_to_fp16 = const()[name = string("op_4883_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_133_cast_fp16 = maximum(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_133_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_133_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4886_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4886_cast_fp16")];
            tensor<int32, [1]> var_4891_axes_0 = const()[name = string("op_4891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4891_keep_dims_0 = const()[name = string("op_4891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_4891_cast_fp16 = reduce_mean(axes = var_4891_axes_0, keep_dims = var_4891_keep_dims_0, x = var_4886_cast_fp16)[name = string("op_4891_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_4892_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4892_cast_fp16")];
            fp32 var_4893_epsilon_0 = const()[name = string("op_4893_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_4893_cast_fp16 = inverse(epsilon = var_4893_epsilon_0, x = var_4892_cast_fp16)[name = string("op_4893_cast_fp16")];
            fp16 var_4894_to_fp16 = const()[name = string("op_4894_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_4895_cast_fp16 = mul(x = var_4893_cast_fp16, y = var_4894_to_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_4897_cast_fp16 = add(x = var_4891_cast_fp16, y = var_4895_cast_fp16)[name = string("op_4897_cast_fp16")];
            fp32 var_4898_epsilon_0 = const()[name = string("op_4898_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_4898_cast_fp16 = rsqrt(epsilon = var_4898_epsilon_0, x = var_4897_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_117_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4898_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591104)))];
            tensor<fp16, [1, 4, 32, 256]> q_33_cast_fp16 = mul(x = out_117_cast_fp16, y = tm_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_4862_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 2, 32, 256]> var_4901_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4901_cast_fp16")];
            bool var_4904 = const()[name = string("op_4904"), val = bool(true)];
            tensor<int32, [1]> var_4905_axes_0 = const()[name = string("op_4905_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_4905_cast_fp16 = reduce_max(axes = var_4905_axes_0, keep_dims = var_4904, x = var_4901_cast_fp16)[name = string("op_4905_cast_fp16")];
            fp16 var_4906_to_fp16 = const()[name = string("op_4906_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_135_cast_fp16 = maximum(x = var_4905_cast_fp16, y = var_4906_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_135_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_135_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4909_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [1]> var_4914_axes_0 = const()[name = string("op_4914_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4914_keep_dims_0 = const()[name = string("op_4914_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_4914_cast_fp16 = reduce_mean(axes = var_4914_axes_0, keep_dims = var_4914_keep_dims_0, x = var_4909_cast_fp16)[name = string("op_4914_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4915_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4915_cast_fp16")];
            fp32 var_4916_epsilon_0 = const()[name = string("op_4916_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_4916_cast_fp16 = inverse(epsilon = var_4916_epsilon_0, x = var_4915_cast_fp16)[name = string("op_4916_cast_fp16")];
            fp16 var_4917_to_fp16 = const()[name = string("op_4917_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_4918_cast_fp16 = mul(x = var_4916_cast_fp16, y = var_4917_to_fp16)[name = string("op_4918_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4920_cast_fp16 = add(x = var_4914_cast_fp16, y = var_4918_cast_fp16)[name = string("op_4920_cast_fp16")];
            fp32 var_4921_epsilon_0 = const()[name = string("op_4921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_4921_cast_fp16 = rsqrt(epsilon = var_4921_epsilon_0, x = var_4920_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_119_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4921_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591680)))];
            tensor<fp16, [1, 2, 32, 256]> k_49_cast_fp16 = mul(x = out_119_cast_fp16, y = tm_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_151_cast_fp16 = transpose(perm = x_151_perm_0, x = var_4874_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 2, 32, 256]> var_4924_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4924_cast_fp16")];
            bool var_4927 = const()[name = string("op_4927"), val = bool(true)];
            tensor<int32, [1]> var_4928_axes_0 = const()[name = string("op_4928_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_4928_cast_fp16 = reduce_max(axes = var_4928_axes_0, keep_dims = var_4927, x = var_4924_cast_fp16)[name = string("op_4928_cast_fp16")];
            fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_137_cast_fp16 = maximum(x = var_4928_cast_fp16, y = var_4929_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_137_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_137_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4932_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4932_cast_fp16")];
            tensor<int32, [1]> var_4937_axes_0 = const()[name = string("op_4937_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4937_keep_dims_0 = const()[name = string("op_4937_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_4937_cast_fp16 = reduce_mean(axes = var_4937_axes_0, keep_dims = var_4937_keep_dims_0, x = var_4932_cast_fp16)[name = string("op_4937_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4938_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4938_cast_fp16")];
            fp32 var_4939_epsilon_0 = const()[name = string("op_4939_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_4939_cast_fp16 = inverse(epsilon = var_4939_epsilon_0, x = var_4938_cast_fp16)[name = string("op_4939_cast_fp16")];
            fp16 var_4940_to_fp16 = const()[name = string("op_4940_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_4941_cast_fp16 = mul(x = var_4939_cast_fp16, y = var_4940_to_fp16)[name = string("op_4941_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_4943_cast_fp16 = add(x = var_4937_cast_fp16, y = var_4941_cast_fp16)[name = string("op_4943_cast_fp16")];
            fp32 var_4944_epsilon_0 = const()[name = string("op_4944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_4944_cast_fp16 = rsqrt(epsilon = var_4944_epsilon_0, x = var_4943_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_33_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4944_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4946_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = q_33_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_4952_cast_fp16 = mul(x = var_4951_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4957_begin_0 = const()[name = string("op_4957_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4957_end_0 = const()[name = string("op_4957_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_4957_end_mask_0 = const()[name = string("op_4957_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_4957_cast_fp16 = slice_by_index(begin = var_4957_begin_0, end = var_4957_end_0, end_mask = var_4957_end_mask_0, x = q_33_cast_fp16)[name = string("op_4957_cast_fp16")];
            int32 var_4959 = const()[name = string("op_4959"), val = int32(-1)];
            bool var_4960_interleave_0 = const()[name = string("op_4960_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_4960_cast_fp16 = concat(axis = var_4959, interleave = var_4960_interleave_0, values = (var_4952_cast_fp16, var_4957_cast_fp16))[name = string("op_4960_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_4961_cast_fp16 = mul(x = var_4960_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4961_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_35_cast_fp16 = add(x = var_4946_cast_fp16, y = var_4961_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4964_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4964_cast_fp16")];
            tensor<int32, [4]> var_4969_begin_0 = const()[name = string("op_4969_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4969_end_0 = const()[name = string("op_4969_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_4969_end_mask_0 = const()[name = string("op_4969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_4969_cast_fp16 = slice_by_index(begin = var_4969_begin_0, end = var_4969_end_0, end_mask = var_4969_end_mask_0, x = k_49_cast_fp16)[name = string("op_4969_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_4970_cast_fp16 = mul(x = var_4969_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_4970_cast_fp16")];
            tensor<int32, [4]> var_4975_begin_0 = const()[name = string("op_4975_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4975_end_0 = const()[name = string("op_4975_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_4975_end_mask_0 = const()[name = string("op_4975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_4975_cast_fp16 = slice_by_index(begin = var_4975_begin_0, end = var_4975_end_0, end_mask = var_4975_end_mask_0, x = k_49_cast_fp16)[name = string("op_4975_cast_fp16")];
            int32 var_4977 = const()[name = string("op_4977"), val = int32(-1)];
            bool var_4978_interleave_0 = const()[name = string("op_4978_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_4978_cast_fp16 = concat(axis = var_4977, interleave = var_4978_interleave_0, values = (var_4970_cast_fp16, var_4975_cast_fp16))[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_4979_cast_fp16 = mul(x = var_4978_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_51_cast_fp16 = add(x = var_4964_cast_fp16, y = var_4979_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_16_reps_0 = const()[name = string("tile_16_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = k_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [4, 1, 32, 256]> tile_16_cast_fp16 = tile(reps = tile_16_reps_0, x = transpose_32_cast_fp16)[name = string("tile_16_cast_fp16")];
            tensor<int32, [5]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_32_cast_fp16 = reshape(shape = concat_32, x = tile_16_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [5]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [4, 1, 32, 256]> reshape_33_cast_fp16 = reshape(shape = concat_33, x = transpose_33_cast_fp16)[name = string("reshape_33_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_17_reps_0 = const()[name = string("tile_17_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = v_33_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [4, 1, 32, 256]> tile_17_cast_fp16 = tile(reps = tile_17_reps_0, x = transpose_34_cast_fp16)[name = string("tile_17_cast_fp16")];
            tensor<int32, [5]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_34_cast_fp16 = reshape(shape = concat_34, x = tile_17_cast_fp16)[name = string("reshape_34_cast_fp16")];
            tensor<int32, [5]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = reshape_34_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [4, 1, 32, 256]> reshape_35_cast_fp16 = reshape(shape = concat_35, x = transpose_35_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = reshape_33_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 4, 32, 32]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = q_35_cast_fp16, y = transpose_104_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_4995_cast_fp16 = add(x = var_4993_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4995_cast_fp16")];
            int32 var_4996 = const()[name = string("op_4996"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_137_cast_fp16 = softmax(axis = var_4996, x = var_4995_cast_fp16)[name = string("w_137_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 4, 32, 256]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = w_137_cast_fp16, y = v_35_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5002_perm_0 = const()[name = string("op_5002_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5006 = const()[name = string("op_5006"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_5002_cast_fp16 = transpose(perm = var_5002_perm_0, x = var_4999_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 32, 1024]> input_99_cast_fp16 = reshape(shape = var_5006, x = var_5002_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50592256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51116608))))[name = string("tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5011_cast_fp16 = abs(x = linear_76_cast_fp16)[name = string("op_5011_cast_fp16")];
            bool var_5014 = const()[name = string("op_5014"), val = bool(true)];
            tensor<int32, [1]> var_5015_axes_0 = const()[name = string("op_5015_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5015_cast_fp16 = reduce_max(axes = var_5015_axes_0, keep_dims = var_5014, x = var_5011_cast_fp16)[name = string("op_5015_cast_fp16")];
            fp16 var_5016_to_fp16 = const()[name = string("op_5016_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_139_cast_fp16 = maximum(x = var_5015_cast_fp16, y = var_5016_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_139_cast_fp16 = real_div(x = linear_76_cast_fp16, y = s_139_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5019_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_5019_cast_fp16")];
            tensor<int32, [1]> var_5024_axes_0 = const()[name = string("op_5024_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5024_keep_dims_0 = const()[name = string("op_5024_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5024_cast_fp16 = reduce_mean(axes = var_5024_axes_0, keep_dims = var_5024_keep_dims_0, x = var_5019_cast_fp16)[name = string("op_5024_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5025_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_5025_cast_fp16")];
            fp32 var_5026_epsilon_0 = const()[name = string("op_5026_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5026_cast_fp16 = inverse(epsilon = var_5026_epsilon_0, x = var_5025_cast_fp16)[name = string("op_5026_cast_fp16")];
            fp16 var_5027_to_fp16 = const()[name = string("op_5027_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5028_cast_fp16 = mul(x = var_5026_cast_fp16, y = var_5027_to_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5030_cast_fp16 = add(x = var_5024_cast_fp16, y = var_5028_cast_fp16)[name = string("op_5030_cast_fp16")];
            fp32 var_5031_epsilon_0 = const()[name = string("op_5031_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5031_cast_fp16 = rsqrt(epsilon = var_5031_epsilon_0, x = var_5030_cast_fp16)[name = string("op_5031_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_121_cast_fp16 = mul(x = y_139_cast_fp16, y = var_5031_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51117696)))];
            tensor<fp16, [1, 32, 512]> var_5033_cast_fp16 = mul(x = out_121_cast_fp16, y = tm_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_5033_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = var_5033_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5036_cast_fp16 = abs(x = x_155_cast_fp16)[name = string("op_5036_cast_fp16")];
            bool var_5039 = const()[name = string("op_5039"), val = bool(true)];
            tensor<int32, [1]> var_5040_axes_0 = const()[name = string("op_5040_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5040_cast_fp16 = reduce_max(axes = var_5040_axes_0, keep_dims = var_5039, x = var_5036_cast_fp16)[name = string("op_5040_cast_fp16")];
            fp16 var_5041_to_fp16 = const()[name = string("op_5041_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_141_cast_fp16 = maximum(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_141_cast_fp16 = real_div(x = x_155_cast_fp16, y = s_141_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5044_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<int32, [1]> var_5049_axes_0 = const()[name = string("op_5049_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5049_keep_dims_0 = const()[name = string("op_5049_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5049_cast_fp16 = reduce_mean(axes = var_5049_axes_0, keep_dims = var_5049_keep_dims_0, x = var_5044_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5050_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_5050_cast_fp16")];
            fp32 var_5051_epsilon_0 = const()[name = string("op_5051_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5051_cast_fp16 = inverse(epsilon = var_5051_epsilon_0, x = var_5050_cast_fp16)[name = string("op_5051_cast_fp16")];
            fp16 var_5052_to_fp16 = const()[name = string("op_5052_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5053_cast_fp16 = mul(x = var_5051_cast_fp16, y = var_5052_to_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5055_cast_fp16 = add(x = var_5049_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            fp32 var_5056_epsilon_0 = const()[name = string("op_5056_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5056_cast_fp16 = rsqrt(epsilon = var_5056_epsilon_0, x = var_5055_cast_fp16)[name = string("op_5056_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_123_cast_fp16 = mul(x = y_141_cast_fp16, y = var_5056_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51118784)))];
            tensor<fp16, [1, 32, 512]> input_101_cast_fp16 = mul(x = out_123_cast_fp16, y = tm_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51119872))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52168512))))[name = string("tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_77_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_77_cast_fp16")];
            string var_5063_mode_0 = const()[name = string("op_5063_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_5063_cast_fp16 = gelu(mode = var_5063_mode_0, x = linear_77_cast_fp16)[name = string("op_5063_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52172672))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53221312))))[name = string("tm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_78_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_103_cast_fp16 = mul(x = var_5063_cast_fp16, y = linear_78_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53225472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54274112))))[name = string("tm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5071_cast_fp16 = abs(x = linear_79_cast_fp16)[name = string("op_5071_cast_fp16")];
            bool var_5074 = const()[name = string("op_5074"), val = bool(true)];
            tensor<int32, [1]> var_5075_axes_0 = const()[name = string("op_5075_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5075_cast_fp16 = reduce_max(axes = var_5075_axes_0, keep_dims = var_5074, x = var_5071_cast_fp16)[name = string("op_5075_cast_fp16")];
            fp16 var_5076_to_fp16 = const()[name = string("op_5076_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_143_cast_fp16 = maximum(x = var_5075_cast_fp16, y = var_5076_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_143_cast_fp16 = real_div(x = linear_79_cast_fp16, y = s_143_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5079_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_5079_cast_fp16")];
            tensor<int32, [1]> var_5084_axes_0 = const()[name = string("op_5084_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5084_keep_dims_0 = const()[name = string("op_5084_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5084_cast_fp16 = reduce_mean(axes = var_5084_axes_0, keep_dims = var_5084_keep_dims_0, x = var_5079_cast_fp16)[name = string("op_5084_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5085_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5086_cast_fp16 = inverse(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            fp16 var_5087_to_fp16 = const()[name = string("op_5087_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5088_cast_fp16 = mul(x = var_5086_cast_fp16, y = var_5087_to_fp16)[name = string("op_5088_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5090_cast_fp16 = add(x = var_5084_cast_fp16, y = var_5088_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp32 var_5091_epsilon_0 = const()[name = string("op_5091_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5091_cast_fp16 = rsqrt(epsilon = var_5091_epsilon_0, x = var_5090_cast_fp16)[name = string("op_5091_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_125_cast_fp16 = mul(x = y_143_cast_fp16, y = var_5091_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54275200)))];
            tensor<fp16, [1, 32, 512]> var_5093_cast_fp16 = mul(x = out_125_cast_fp16, y = tm_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5093_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_105_cast_fp16 = add(x = x_155_cast_fp16, y = var_5093_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54276288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54538496))))[name = string("tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_80_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_80_cast_fp16")];
            string x_159_mode_0 = const()[name = string("x_159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_159_cast_fp16 = gelu(mode = x_159_mode_0, x = linear_80_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [4]> var_5113_begin_0 = const()[name = string("op_5113_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_5113_end_0 = const()[name = string("op_5113_end_0"), val = tensor<int32, [4]>([1, 32, 9, 512])];
            tensor<bool, [4]> var_5113_end_mask_0 = const()[name = string("op_5113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5113_squeeze_mask_0 = const()[name = string("op_5113_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_5113_cast_fp16 = slice_by_index(begin = var_5113_begin_0, end = var_5113_end_0, end_mask = var_5113_end_mask_0, squeeze_mask = var_5113_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_107_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5113_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54539584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54801792))))[name = string("tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5123_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5123_cast_fp16")];
            bool var_5126 = const()[name = string("op_5126"), val = bool(true)];
            tensor<int32, [1]> var_5127_axes_0 = const()[name = string("op_5127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5127_cast_fp16 = reduce_max(axes = var_5127_axes_0, keep_dims = var_5126, x = var_5123_cast_fp16)[name = string("op_5127_cast_fp16")];
            fp16 var_5128_to_fp16 = const()[name = string("op_5128_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_145_cast_fp16 = maximum(x = var_5127_cast_fp16, y = var_5128_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_145_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_145_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5131_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5136_axes_0 = const()[name = string("op_5136_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5136_keep_dims_0 = const()[name = string("op_5136_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5136_cast_fp16 = reduce_mean(axes = var_5136_axes_0, keep_dims = var_5136_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5136_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5137_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5138_cast_fp16 = inverse(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp16 var_5139_to_fp16 = const()[name = string("op_5139_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5140_cast_fp16 = mul(x = var_5138_cast_fp16, y = var_5139_to_fp16)[name = string("op_5140_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5142_cast_fp16 = add(x = var_5136_cast_fp16, y = var_5140_cast_fp16)[name = string("op_5142_cast_fp16")];
            fp32 var_5143_epsilon_0 = const()[name = string("op_5143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5143_cast_fp16 = rsqrt(epsilon = var_5143_epsilon_0, x = var_5142_cast_fp16)[name = string("op_5143_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_127_cast_fp16 = mul(x = y_145_cast_fp16, y = var_5143_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54802880)))];
            tensor<fp16, [1, 32, 512]> var_5145_cast_fp16 = mul(x = out_127_cast_fp16, y = tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5145_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_17_cast_fp16 = add(x = input_105_cast_fp16, y = var_5145_cast_fp16)[name = string("h_17_cast_fp16")];
            tensor<fp16, [1]> tm_layers_8_layer_scalar_to_fp16 = const()[name = string("tm_layers_8_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bcp-1])];
            tensor<fp16, [1, 32, 512]> x_163_cast_fp16 = mul(x = h_17_cast_fp16, y = tm_layers_8_layer_scalar_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5149_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_5149_cast_fp16")];
            bool var_5152 = const()[name = string("op_5152"), val = bool(true)];
            tensor<int32, [1]> var_5153_axes_0 = const()[name = string("op_5153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5153_cast_fp16 = reduce_max(axes = var_5153_axes_0, keep_dims = var_5152, x = var_5149_cast_fp16)[name = string("op_5153_cast_fp16")];
            fp16 var_5154_to_fp16 = const()[name = string("op_5154_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_147_cast_fp16 = maximum(x = var_5153_cast_fp16, y = var_5154_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_147_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_147_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5157_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_5157_cast_fp16")];
            tensor<int32, [1]> var_5162_axes_0 = const()[name = string("op_5162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5162_keep_dims_0 = const()[name = string("op_5162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5162_cast_fp16 = reduce_mean(axes = var_5162_axes_0, keep_dims = var_5162_keep_dims_0, x = var_5157_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5163_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_5163_cast_fp16")];
            fp32 var_5164_epsilon_0 = const()[name = string("op_5164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5164_cast_fp16 = inverse(epsilon = var_5164_epsilon_0, x = var_5163_cast_fp16)[name = string("op_5164_cast_fp16")];
            fp16 var_5165_to_fp16 = const()[name = string("op_5165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5166_cast_fp16 = mul(x = var_5164_cast_fp16, y = var_5165_to_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5168_cast_fp16 = add(x = var_5162_cast_fp16, y = var_5166_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp32 var_5169_epsilon_0 = const()[name = string("op_5169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5169_cast_fp16 = rsqrt(epsilon = var_5169_epsilon_0, x = var_5168_cast_fp16)[name = string("op_5169_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_129_cast_fp16 = mul(x = y_147_cast_fp16, y = var_5169_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54803968)))];
            tensor<fp16, [1, 32, 512]> input_109_cast_fp16 = mul(x = out_129_cast_fp16, y = tm_layers_9_input_layernorm_weight_to_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54805056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55329408))))[name = string("tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [4]> var_5179 = const()[name = string("op_5179"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_5180_cast_fp16 = reshape(shape = var_5179, x = linear_82_cast_fp16)[name = string("op_5180_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55331520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55593728))))[name = string("tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [4]> var_5191 = const()[name = string("op_5191"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_5192_cast_fp16 = reshape(shape = var_5191, x = linear_83_cast_fp16)[name = string("op_5192_cast_fp16")];
            tensor<int32, [4]> x_167_perm_0 = const()[name = string("x_167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55594816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55857024))))[name = string("tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5203 = const()[name = string("op_5203"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_5204_cast_fp16 = reshape(shape = var_5203, x = linear_84_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_5180_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 4, 32, 256]> var_5208_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_5208_cast_fp16")];
            bool var_5211 = const()[name = string("op_5211"), val = bool(true)];
            tensor<int32, [1]> var_5212_axes_0 = const()[name = string("op_5212_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_5212_cast_fp16 = reduce_max(axes = var_5212_axes_0, keep_dims = var_5211, x = var_5208_cast_fp16)[name = string("op_5212_cast_fp16")];
            fp16 var_5213_to_fp16 = const()[name = string("op_5213_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_149_cast_fp16 = maximum(x = var_5212_cast_fp16, y = var_5213_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_149_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_149_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_5216_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<int32, [1]> var_5221_axes_0 = const()[name = string("op_5221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5221_keep_dims_0 = const()[name = string("op_5221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_5221_cast_fp16 = reduce_mean(axes = var_5221_axes_0, keep_dims = var_5221_keep_dims_0, x = var_5216_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_5222_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_5223_cast_fp16 = inverse(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            fp16 var_5224_to_fp16 = const()[name = string("op_5224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_5225_cast_fp16 = mul(x = var_5223_cast_fp16, y = var_5224_to_fp16)[name = string("op_5225_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_5227_cast_fp16 = add(x = var_5221_cast_fp16, y = var_5225_cast_fp16)[name = string("op_5227_cast_fp16")];
            fp32 var_5228_epsilon_0 = const()[name = string("op_5228_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_5228_cast_fp16 = rsqrt(epsilon = var_5228_epsilon_0, x = var_5227_cast_fp16)[name = string("op_5228_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_131_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5228_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = var_5192_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 2, 32, 256]> var_5231_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_5231_cast_fp16")];
            bool var_5234 = const()[name = string("op_5234"), val = bool(true)];
            tensor<int32, [1]> var_5235_axes_0 = const()[name = string("op_5235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_5235_cast_fp16 = reduce_max(axes = var_5235_axes_0, keep_dims = var_5234, x = var_5231_cast_fp16)[name = string("op_5235_cast_fp16")];
            fp16 var_5236_to_fp16 = const()[name = string("op_5236_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_151_cast_fp16 = maximum(x = var_5235_cast_fp16, y = var_5236_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_151_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_151_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_5239_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5239_cast_fp16")];
            tensor<int32, [1]> var_5244_axes_0 = const()[name = string("op_5244_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5244_keep_dims_0 = const()[name = string("op_5244_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_5244_cast_fp16 = reduce_mean(axes = var_5244_axes_0, keep_dims = var_5244_keep_dims_0, x = var_5239_cast_fp16)[name = string("op_5244_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_5245_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_5245_cast_fp16")];
            fp32 var_5246_epsilon_0 = const()[name = string("op_5246_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_5246_cast_fp16 = inverse(epsilon = var_5246_epsilon_0, x = var_5245_cast_fp16)[name = string("op_5246_cast_fp16")];
            fp16 var_5247_to_fp16 = const()[name = string("op_5247_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_5248_cast_fp16 = mul(x = var_5246_cast_fp16, y = var_5247_to_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_5250_cast_fp16 = add(x = var_5244_cast_fp16, y = var_5248_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp32 var_5251_epsilon_0 = const()[name = string("op_5251_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_5251_cast_fp16 = rsqrt(epsilon = var_5251_epsilon_0, x = var_5250_cast_fp16)[name = string("op_5251_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_133_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5251_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [256]> tm_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858112)))];
            tensor<fp16, [1, 2, 32, 256]> k_55_cast_fp16 = mul(x = out_133_cast_fp16, y = tm_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_5204_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2, 32, 256]> var_5254_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_5254_cast_fp16")];
            bool var_5257 = const()[name = string("op_5257"), val = bool(true)];
            tensor<int32, [1]> var_5258_axes_0 = const()[name = string("op_5258_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_5258_cast_fp16 = reduce_max(axes = var_5258_axes_0, keep_dims = var_5257, x = var_5254_cast_fp16)[name = string("op_5258_cast_fp16")];
            fp16 var_5259_to_fp16 = const()[name = string("op_5259_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_153_cast_fp16 = maximum(x = var_5258_cast_fp16, y = var_5259_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_153_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_153_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_5262_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5262_cast_fp16")];
            tensor<int32, [1]> var_5267_axes_0 = const()[name = string("op_5267_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5267_keep_dims_0 = const()[name = string("op_5267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_5267_cast_fp16 = reduce_mean(axes = var_5267_axes_0, keep_dims = var_5267_keep_dims_0, x = var_5262_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_5268_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5268_cast_fp16")];
            fp32 var_5269_epsilon_0 = const()[name = string("op_5269_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_5269_cast_fp16 = inverse(epsilon = var_5269_epsilon_0, x = var_5268_cast_fp16)[name = string("op_5269_cast_fp16")];
            fp16 var_5270_to_fp16 = const()[name = string("op_5270_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_5271_cast_fp16 = mul(x = var_5269_cast_fp16, y = var_5270_to_fp16)[name = string("op_5271_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_5273_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5271_cast_fp16)[name = string("op_5273_cast_fp16")];
            fp32 var_5274_epsilon_0 = const()[name = string("op_5274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_5274_cast_fp16 = rsqrt(epsilon = var_5274_epsilon_0, x = var_5273_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_37_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5274_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_5276_cast_fp16 = mul(x = out_131_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5276_cast_fp16")];
            tensor<int32, [4]> var_5281_begin_0 = const()[name = string("op_5281_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5281_end_0 = const()[name = string("op_5281_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_5281_end_mask_0 = const()[name = string("op_5281_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_5281_cast_fp16 = slice_by_index(begin = var_5281_begin_0, end = var_5281_end_0, end_mask = var_5281_end_mask_0, x = out_131_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_5282_cast_fp16 = mul(x = var_5281_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_5282_cast_fp16")];
            tensor<int32, [4]> var_5287_begin_0 = const()[name = string("op_5287_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5287_end_0 = const()[name = string("op_5287_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_5287_end_mask_0 = const()[name = string("op_5287_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_5287_cast_fp16 = slice_by_index(begin = var_5287_begin_0, end = var_5287_end_0, end_mask = var_5287_end_mask_0, x = out_131_cast_fp16)[name = string("op_5287_cast_fp16")];
            int32 var_5289 = const()[name = string("op_5289"), val = int32(-1)];
            bool var_5290_interleave_0 = const()[name = string("op_5290_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_5290_cast_fp16 = concat(axis = var_5289, interleave = var_5290_interleave_0, values = (var_5282_cast_fp16, var_5287_cast_fp16))[name = string("op_5290_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_5291_cast_fp16 = mul(x = var_5290_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5291_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_39_cast_fp16 = add(x = var_5276_cast_fp16, y = var_5291_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_5294_cast_fp16 = mul(x = k_55_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5294_cast_fp16")];
            tensor<int32, [4]> var_5299_begin_0 = const()[name = string("op_5299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5299_end_0 = const()[name = string("op_5299_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_5299_end_mask_0 = const()[name = string("op_5299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_5299_cast_fp16 = slice_by_index(begin = var_5299_begin_0, end = var_5299_end_0, end_mask = var_5299_end_mask_0, x = k_55_cast_fp16)[name = string("op_5299_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_5300_cast_fp16 = mul(x = var_5299_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_5300_cast_fp16")];
            tensor<int32, [4]> var_5305_begin_0 = const()[name = string("op_5305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5305_end_0 = const()[name = string("op_5305_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_5305_end_mask_0 = const()[name = string("op_5305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_5305_cast_fp16 = slice_by_index(begin = var_5305_begin_0, end = var_5305_end_0, end_mask = var_5305_end_mask_0, x = k_55_cast_fp16)[name = string("op_5305_cast_fp16")];
            int32 var_5307 = const()[name = string("op_5307"), val = int32(-1)];
            bool var_5308_interleave_0 = const()[name = string("op_5308_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_5308_cast_fp16 = concat(axis = var_5307, interleave = var_5308_interleave_0, values = (var_5300_cast_fp16, var_5305_cast_fp16))[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_5309_cast_fp16 = mul(x = var_5308_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_57_cast_fp16 = add(x = var_5294_cast_fp16, y = var_5309_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_18_reps_0 = const()[name = string("tile_18_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = k_57_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [4, 1, 32, 256]> tile_18_cast_fp16 = tile(reps = tile_18_reps_0, x = transpose_36_cast_fp16)[name = string("tile_18_cast_fp16")];
            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_36_cast_fp16 = reshape(shape = concat_36, x = tile_18_cast_fp16)[name = string("reshape_36_cast_fp16")];
            tensor<int32, [5]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = reshape_36_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [4, 1, 32, 256]> reshape_37_cast_fp16 = reshape(shape = concat_37, x = transpose_37_cast_fp16)[name = string("reshape_37_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_19_reps_0 = const()[name = string("tile_19_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = v_37_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [4, 1, 32, 256]> tile_19_cast_fp16 = tile(reps = tile_19_reps_0, x = transpose_38_cast_fp16)[name = string("tile_19_cast_fp16")];
            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_38_cast_fp16 = reshape(shape = concat_38, x = tile_19_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [5]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [4, 1, 32, 256]> reshape_39_cast_fp16 = reshape(shape = concat_39, x = transpose_39_cast_fp16)[name = string("reshape_39_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5323_transpose_x_0 = const()[name = string("op_5323_transpose_x_0"), val = bool(false)];
            bool var_5323_transpose_y_0 = const()[name = string("op_5323_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = reshape_37_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 4, 32, 32]> var_5323_cast_fp16 = matmul(transpose_x = var_5323_transpose_x_0, transpose_y = var_5323_transpose_y_0, x = q_39_cast_fp16, y = transpose_105_cast_fp16)[name = string("op_5323_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_5325_cast_fp16 = add(x = var_5323_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5325_cast_fp16")];
            int32 var_5326 = const()[name = string("op_5326"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_153_cast_fp16 = softmax(axis = var_5326, x = var_5325_cast_fp16)[name = string("w_153_cast_fp16")];
            bool var_5329_transpose_x_0 = const()[name = string("op_5329_transpose_x_0"), val = bool(false)];
            bool var_5329_transpose_y_0 = const()[name = string("op_5329_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = reshape_39_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 4, 32, 256]> var_5329_cast_fp16 = matmul(transpose_x = var_5329_transpose_x_0, transpose_y = var_5329_transpose_y_0, x = w_153_cast_fp16, y = v_39_cast_fp16)[name = string("op_5329_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = var_5329_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 32, 1024]> input_111_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56383040))))[name = string("tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5341_cast_fp16 = abs(x = linear_85_cast_fp16)[name = string("op_5341_cast_fp16")];
            bool var_5344 = const()[name = string("op_5344"), val = bool(true)];
            tensor<int32, [1]> var_5345_axes_0 = const()[name = string("op_5345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5345_cast_fp16 = reduce_max(axes = var_5345_axes_0, keep_dims = var_5344, x = var_5341_cast_fp16)[name = string("op_5345_cast_fp16")];
            fp16 var_5346_to_fp16 = const()[name = string("op_5346_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_155_cast_fp16 = maximum(x = var_5345_cast_fp16, y = var_5346_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_155_cast_fp16 = real_div(x = linear_85_cast_fp16, y = s_155_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5349_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5354_axes_0 = const()[name = string("op_5354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5354_keep_dims_0 = const()[name = string("op_5354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5354_cast_fp16 = reduce_mean(axes = var_5354_axes_0, keep_dims = var_5354_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5355_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5355_cast_fp16")];
            fp32 var_5356_epsilon_0 = const()[name = string("op_5356_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5356_cast_fp16 = inverse(epsilon = var_5356_epsilon_0, x = var_5355_cast_fp16)[name = string("op_5356_cast_fp16")];
            fp16 var_5357_to_fp16 = const()[name = string("op_5357_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5358_cast_fp16 = mul(x = var_5356_cast_fp16, y = var_5357_to_fp16)[name = string("op_5358_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5360_cast_fp16 = add(x = var_5354_cast_fp16, y = var_5358_cast_fp16)[name = string("op_5360_cast_fp16")];
            fp32 var_5361_epsilon_0 = const()[name = string("op_5361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5361_cast_fp16 = rsqrt(epsilon = var_5361_epsilon_0, x = var_5360_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_135_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5361_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56384128)))];
            tensor<fp16, [1, 32, 512]> var_5363_cast_fp16 = mul(x = out_135_cast_fp16, y = tm_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_5363_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_173_cast_fp16 = add(x = x_163_cast_fp16, y = var_5363_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5366_cast_fp16 = abs(x = x_173_cast_fp16)[name = string("op_5366_cast_fp16")];
            bool var_5369 = const()[name = string("op_5369"), val = bool(true)];
            tensor<int32, [1]> var_5370_axes_0 = const()[name = string("op_5370_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5370_cast_fp16 = reduce_max(axes = var_5370_axes_0, keep_dims = var_5369, x = var_5366_cast_fp16)[name = string("op_5370_cast_fp16")];
            fp16 var_5371_to_fp16 = const()[name = string("op_5371_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_157_cast_fp16 = maximum(x = var_5370_cast_fp16, y = var_5371_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_157_cast_fp16 = real_div(x = x_173_cast_fp16, y = s_157_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5374_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [1]> var_5379_axes_0 = const()[name = string("op_5379_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5379_keep_dims_0 = const()[name = string("op_5379_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5379_cast_fp16 = reduce_mean(axes = var_5379_axes_0, keep_dims = var_5379_keep_dims_0, x = var_5374_cast_fp16)[name = string("op_5379_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5380_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5381_cast_fp16 = inverse(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            fp16 var_5382_to_fp16 = const()[name = string("op_5382_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5383_cast_fp16 = mul(x = var_5381_cast_fp16, y = var_5382_to_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5385_cast_fp16 = add(x = var_5379_cast_fp16, y = var_5383_cast_fp16)[name = string("op_5385_cast_fp16")];
            fp32 var_5386_epsilon_0 = const()[name = string("op_5386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5386_cast_fp16 = rsqrt(epsilon = var_5386_epsilon_0, x = var_5385_cast_fp16)[name = string("op_5386_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_137_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5386_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56385216)))];
            tensor<fp16, [1, 32, 512]> input_113_cast_fp16 = mul(x = out_137_cast_fp16, y = tm_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56386304))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57434944))))[name = string("tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_86_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_86_cast_fp16")];
            string var_5393_mode_0 = const()[name = string("op_5393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_5393_cast_fp16 = gelu(mode = var_5393_mode_0, x = linear_86_cast_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57439104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58487744))))[name = string("tm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_87_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_115_cast_fp16 = mul(x = var_5393_cast_fp16, y = linear_87_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58491904))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59540544))))[name = string("tm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5401_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5401_cast_fp16")];
            bool var_5404 = const()[name = string("op_5404"), val = bool(true)];
            tensor<int32, [1]> var_5405_axes_0 = const()[name = string("op_5405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5405_cast_fp16 = reduce_max(axes = var_5405_axes_0, keep_dims = var_5404, x = var_5401_cast_fp16)[name = string("op_5405_cast_fp16")];
            fp16 var_5406_to_fp16 = const()[name = string("op_5406_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_159_cast_fp16 = maximum(x = var_5405_cast_fp16, y = var_5406_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_159_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_159_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5409_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5409_cast_fp16")];
            tensor<int32, [1]> var_5414_axes_0 = const()[name = string("op_5414_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5414_keep_dims_0 = const()[name = string("op_5414_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5414_cast_fp16 = reduce_mean(axes = var_5414_axes_0, keep_dims = var_5414_keep_dims_0, x = var_5409_cast_fp16)[name = string("op_5414_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5415_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5415_cast_fp16")];
            fp32 var_5416_epsilon_0 = const()[name = string("op_5416_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5416_cast_fp16 = inverse(epsilon = var_5416_epsilon_0, x = var_5415_cast_fp16)[name = string("op_5416_cast_fp16")];
            fp16 var_5417_to_fp16 = const()[name = string("op_5417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5418_cast_fp16 = mul(x = var_5416_cast_fp16, y = var_5417_to_fp16)[name = string("op_5418_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5420_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5418_cast_fp16)[name = string("op_5420_cast_fp16")];
            fp32 var_5421_epsilon_0 = const()[name = string("op_5421_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5421_cast_fp16 = rsqrt(epsilon = var_5421_epsilon_0, x = var_5420_cast_fp16)[name = string("op_5421_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_139_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5421_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59541632)))];
            tensor<fp16, [1, 32, 512]> var_5423_cast_fp16 = mul(x = out_139_cast_fp16, y = tm_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5423_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_117_cast_fp16 = add(x = x_173_cast_fp16, y = var_5423_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59542720))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59804928))))[name = string("tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_89_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_89_cast_fp16")];
            string x_177_mode_0 = const()[name = string("x_177_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_177_cast_fp16 = gelu(mode = x_177_mode_0, x = linear_89_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [4]> var_5443_begin_0 = const()[name = string("op_5443_begin_0"), val = tensor<int32, [4]>([0, 0, 9, 0])];
            tensor<int32, [4]> var_5443_end_0 = const()[name = string("op_5443_end_0"), val = tensor<int32, [4]>([1, 32, 10, 512])];
            tensor<bool, [4]> var_5443_end_mask_0 = const()[name = string("op_5443_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5443_squeeze_mask_0 = const()[name = string("op_5443_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_5443_cast_fp16 = slice_by_index(begin = var_5443_begin_0, end = var_5443_end_0, end_mask = var_5443_end_mask_0, squeeze_mask = var_5443_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5443_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_119_cast_fp16 = mul(x = x_177_cast_fp16, y = var_5443_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59806016))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60068224))))[name = string("tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5453_cast_fp16 = abs(x = linear_90_cast_fp16)[name = string("op_5453_cast_fp16")];
            bool var_5456 = const()[name = string("op_5456"), val = bool(true)];
            tensor<int32, [1]> var_5457_axes_0 = const()[name = string("op_5457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5457_cast_fp16 = reduce_max(axes = var_5457_axes_0, keep_dims = var_5456, x = var_5453_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_161_cast_fp16 = maximum(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_161_cast_fp16 = real_div(x = linear_90_cast_fp16, y = s_161_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5461_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<int32, [1]> var_5466_axes_0 = const()[name = string("op_5466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5466_keep_dims_0 = const()[name = string("op_5466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5466_cast_fp16 = reduce_mean(axes = var_5466_axes_0, keep_dims = var_5466_keep_dims_0, x = var_5461_cast_fp16)[name = string("op_5466_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5467_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5467_cast_fp16")];
            fp32 var_5468_epsilon_0 = const()[name = string("op_5468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5468_cast_fp16 = inverse(epsilon = var_5468_epsilon_0, x = var_5467_cast_fp16)[name = string("op_5468_cast_fp16")];
            fp16 var_5469_to_fp16 = const()[name = string("op_5469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5470_cast_fp16 = mul(x = var_5468_cast_fp16, y = var_5469_to_fp16)[name = string("op_5470_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5472_cast_fp16 = add(x = var_5466_cast_fp16, y = var_5470_cast_fp16)[name = string("op_5472_cast_fp16")];
            fp32 var_5473_epsilon_0 = const()[name = string("op_5473_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5473_cast_fp16 = rsqrt(epsilon = var_5473_epsilon_0, x = var_5472_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_141_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5473_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60069312)))];
            tensor<fp16, [1, 32, 512]> var_5475_cast_fp16 = mul(x = out_141_cast_fp16, y = tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5475_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_19_cast_fp16 = add(x = input_117_cast_fp16, y = var_5475_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<fp16, [1]> tm_layers_9_layer_scalar_to_fp16 = const()[name = string("tm_layers_9_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9ep-1])];
            tensor<fp16, [1, 32, 512]> x_181_cast_fp16 = mul(x = h_19_cast_fp16, y = tm_layers_9_layer_scalar_to_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5479_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_5479_cast_fp16")];
            bool var_5482 = const()[name = string("op_5482"), val = bool(true)];
            tensor<int32, [1]> var_5483_axes_0 = const()[name = string("op_5483_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5483_cast_fp16 = reduce_max(axes = var_5483_axes_0, keep_dims = var_5482, x = var_5479_cast_fp16)[name = string("op_5483_cast_fp16")];
            fp16 var_5484_to_fp16 = const()[name = string("op_5484_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_163_cast_fp16 = maximum(x = var_5483_cast_fp16, y = var_5484_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_163_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_163_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5487_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5492_axes_0 = const()[name = string("op_5492_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5492_keep_dims_0 = const()[name = string("op_5492_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5492_cast_fp16 = reduce_mean(axes = var_5492_axes_0, keep_dims = var_5492_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5493_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5493_cast_fp16")];
            fp32 var_5494_epsilon_0 = const()[name = string("op_5494_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5494_cast_fp16 = inverse(epsilon = var_5494_epsilon_0, x = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp16 var_5495_to_fp16 = const()[name = string("op_5495_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5496_cast_fp16 = mul(x = var_5494_cast_fp16, y = var_5495_to_fp16)[name = string("op_5496_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5498_cast_fp16 = add(x = var_5492_cast_fp16, y = var_5496_cast_fp16)[name = string("op_5498_cast_fp16")];
            fp32 var_5499_epsilon_0 = const()[name = string("op_5499_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5499_cast_fp16 = rsqrt(epsilon = var_5499_epsilon_0, x = var_5498_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_143_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5499_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60070400)))];
            tensor<fp16, [1, 32, 512]> input_121_cast_fp16 = mul(x = out_143_cast_fp16, y = tm_layers_10_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60071488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60595840))))[name = string("tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5509 = const()[name = string("op_5509"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_5510_cast_fp16 = reshape(shape = var_5509, x = linear_91_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60597952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60860160))))[name = string("tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5521 = const()[name = string("op_5521"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_5522_cast_fp16 = reshape(shape = var_5521, x = linear_92_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60861248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61123456))))[name = string("tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5533 = const()[name = string("op_5533"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_5534_cast_fp16 = reshape(shape = var_5533, x = linear_93_cast_fp16)[name = string("op_5534_cast_fp16")];
            tensor<int32, [4]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5510_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 4, 32, 256]> var_5538_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_5538_cast_fp16")];
            bool var_5541 = const()[name = string("op_5541"), val = bool(true)];
            tensor<int32, [1]> var_5542_axes_0 = const()[name = string("op_5542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_5542_cast_fp16 = reduce_max(axes = var_5542_axes_0, keep_dims = var_5541, x = var_5538_cast_fp16)[name = string("op_5542_cast_fp16")];
            fp16 var_5543_to_fp16 = const()[name = string("op_5543_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_165_cast_fp16 = maximum(x = var_5542_cast_fp16, y = var_5543_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_165_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_165_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_5546_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<int32, [1]> var_5551_axes_0 = const()[name = string("op_5551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5551_keep_dims_0 = const()[name = string("op_5551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_5551_cast_fp16 = reduce_mean(axes = var_5551_axes_0, keep_dims = var_5551_keep_dims_0, x = var_5546_cast_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_5552_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5552_cast_fp16")];
            fp32 var_5553_epsilon_0 = const()[name = string("op_5553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_5553_cast_fp16 = inverse(epsilon = var_5553_epsilon_0, x = var_5552_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp16 var_5554_to_fp16 = const()[name = string("op_5554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_5555_cast_fp16 = mul(x = var_5553_cast_fp16, y = var_5554_to_fp16)[name = string("op_5555_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_5557_cast_fp16 = add(x = var_5551_cast_fp16, y = var_5555_cast_fp16)[name = string("op_5557_cast_fp16")];
            fp32 var_5558_epsilon_0 = const()[name = string("op_5558_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_5558_cast_fp16 = rsqrt(epsilon = var_5558_epsilon_0, x = var_5557_cast_fp16)[name = string("op_5558_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_145_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5558_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61124544)))];
            tensor<fp16, [1, 4, 32, 256]> q_41_cast_fp16 = mul(x = out_145_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_41_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_5522_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2, 32, 256]> var_5561_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_5561_cast_fp16")];
            bool var_5564 = const()[name = string("op_5564"), val = bool(true)];
            tensor<int32, [1]> var_5565_axes_0 = const()[name = string("op_5565_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_5565_cast_fp16 = reduce_max(axes = var_5565_axes_0, keep_dims = var_5564, x = var_5561_cast_fp16)[name = string("op_5565_cast_fp16")];
            fp16 var_5566_to_fp16 = const()[name = string("op_5566_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_167_cast_fp16 = maximum(x = var_5565_cast_fp16, y = var_5566_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_167_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_167_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_5569_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5569_cast_fp16")];
            tensor<int32, [1]> var_5574_axes_0 = const()[name = string("op_5574_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5574_keep_dims_0 = const()[name = string("op_5574_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_5574_cast_fp16 = reduce_mean(axes = var_5574_axes_0, keep_dims = var_5574_keep_dims_0, x = var_5569_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_5575_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5575_cast_fp16")];
            fp32 var_5576_epsilon_0 = const()[name = string("op_5576_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_5576_cast_fp16 = inverse(epsilon = var_5576_epsilon_0, x = var_5575_cast_fp16)[name = string("op_5576_cast_fp16")];
            fp16 var_5577_to_fp16 = const()[name = string("op_5577_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_5578_cast_fp16 = mul(x = var_5576_cast_fp16, y = var_5577_to_fp16)[name = string("op_5578_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_5580_cast_fp16 = add(x = var_5574_cast_fp16, y = var_5578_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_5581_cast_fp16 = rsqrt(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_147_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5581_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [256]> tm_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125120)))];
            tensor<fp16, [1, 2, 32, 256]> k_61_cast_fp16 = mul(x = out_147_cast_fp16, y = tm_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = var_5534_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 2, 32, 256]> var_5584_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_5584_cast_fp16")];
            bool var_5587 = const()[name = string("op_5587"), val = bool(true)];
            tensor<int32, [1]> var_5588_axes_0 = const()[name = string("op_5588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_5588_cast_fp16 = reduce_max(axes = var_5588_axes_0, keep_dims = var_5587, x = var_5584_cast_fp16)[name = string("op_5588_cast_fp16")];
            fp16 var_5589_to_fp16 = const()[name = string("op_5589_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_169_cast_fp16 = maximum(x = var_5588_cast_fp16, y = var_5589_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_169_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_169_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_5592_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5592_cast_fp16")];
            tensor<int32, [1]> var_5597_axes_0 = const()[name = string("op_5597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5597_keep_dims_0 = const()[name = string("op_5597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_5597_cast_fp16 = reduce_mean(axes = var_5597_axes_0, keep_dims = var_5597_keep_dims_0, x = var_5592_cast_fp16)[name = string("op_5597_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_5598_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5598_cast_fp16")];
            fp32 var_5599_epsilon_0 = const()[name = string("op_5599_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_5599_cast_fp16 = inverse(epsilon = var_5599_epsilon_0, x = var_5598_cast_fp16)[name = string("op_5599_cast_fp16")];
            fp16 var_5600_to_fp16 = const()[name = string("op_5600_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_5601_cast_fp16 = mul(x = var_5599_cast_fp16, y = var_5600_to_fp16)[name = string("op_5601_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_5603_cast_fp16 = add(x = var_5597_cast_fp16, y = var_5601_cast_fp16)[name = string("op_5603_cast_fp16")];
            fp32 var_5604_epsilon_0 = const()[name = string("op_5604_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_5604_cast_fp16 = rsqrt(epsilon = var_5604_epsilon_0, x = var_5603_cast_fp16)[name = string("op_5604_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_41_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5604_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_5606_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5611_begin_0 = const()[name = string("op_5611_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5611_end_0 = const()[name = string("op_5611_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_5611_end_mask_0 = const()[name = string("op_5611_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_5611_cast_fp16 = slice_by_index(begin = var_5611_begin_0, end = var_5611_end_0, end_mask = var_5611_end_mask_0, x = q_41_cast_fp16)[name = string("op_5611_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_5612_cast_fp16 = mul(x = var_5611_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_5612_cast_fp16")];
            tensor<int32, [4]> var_5617_begin_0 = const()[name = string("op_5617_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5617_end_0 = const()[name = string("op_5617_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_5617_end_mask_0 = const()[name = string("op_5617_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_5617_cast_fp16 = slice_by_index(begin = var_5617_begin_0, end = var_5617_end_0, end_mask = var_5617_end_mask_0, x = q_41_cast_fp16)[name = string("op_5617_cast_fp16")];
            int32 var_5619 = const()[name = string("op_5619"), val = int32(-1)];
            bool var_5620_interleave_0 = const()[name = string("op_5620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_5620_cast_fp16 = concat(axis = var_5619, interleave = var_5620_interleave_0, values = (var_5612_cast_fp16, var_5617_cast_fp16))[name = string("op_5620_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_5621_cast_fp16 = mul(x = var_5620_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5621_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_43_cast_fp16 = add(x = var_5606_cast_fp16, y = var_5621_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_5624_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = k_61_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = k_61_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool var_5638_interleave_0 = const()[name = string("op_5638_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_5638_cast_fp16 = concat(axis = var_5637, interleave = var_5638_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("op_5638_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_5639_cast_fp16 = mul(x = var_5638_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_63_cast_fp16 = add(x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_20_reps_0 = const()[name = string("tile_20_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = k_63_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [4, 1, 32, 256]> tile_20_cast_fp16 = tile(reps = tile_20_reps_0, x = transpose_40_cast_fp16)[name = string("tile_20_cast_fp16")];
            tensor<int32, [5]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_40_cast_fp16 = reshape(shape = concat_40, x = tile_20_cast_fp16)[name = string("reshape_40_cast_fp16")];
            tensor<int32, [5]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = reshape_40_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [4, 1, 32, 256]> reshape_41_cast_fp16 = reshape(shape = concat_41, x = transpose_41_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_21_reps_0 = const()[name = string("tile_21_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = v_41_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [4, 1, 32, 256]> tile_21_cast_fp16 = tile(reps = tile_21_reps_0, x = transpose_42_cast_fp16)[name = string("tile_21_cast_fp16")];
            tensor<int32, [5]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_42_cast_fp16 = reshape(shape = concat_42, x = tile_21_cast_fp16)[name = string("reshape_42_cast_fp16")];
            tensor<int32, [5]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = reshape_42_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [4, 1, 32, 256]> reshape_43_cast_fp16 = reshape(shape = concat_43, x = transpose_43_cast_fp16)[name = string("reshape_43_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5653_transpose_x_0 = const()[name = string("op_5653_transpose_x_0"), val = bool(false)];
            bool var_5653_transpose_y_0 = const()[name = string("op_5653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 4, 32, 32]> var_5653_cast_fp16 = matmul(transpose_x = var_5653_transpose_x_0, transpose_y = var_5653_transpose_y_0, x = q_43_cast_fp16, y = transpose_106_cast_fp16)[name = string("op_5653_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_5655_cast_fp16 = add(x = var_5653_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5655_cast_fp16")];
            int32 var_5656 = const()[name = string("op_5656"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_169_cast_fp16 = softmax(axis = var_5656, x = var_5655_cast_fp16)[name = string("w_169_cast_fp16")];
            bool var_5659_transpose_x_0 = const()[name = string("op_5659_transpose_x_0"), val = bool(false)];
            bool var_5659_transpose_y_0 = const()[name = string("op_5659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = reshape_43_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 4, 32, 256]> var_5659_cast_fp16 = matmul(transpose_x = var_5659_transpose_x_0, transpose_y = var_5659_transpose_y_0, x = w_169_cast_fp16, y = v_43_cast_fp16)[name = string("op_5659_cast_fp16")];
            tensor<int32, [4]> var_5662_perm_0 = const()[name = string("op_5662_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5666 = const()[name = string("op_5666"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_5662_cast_fp16 = transpose(perm = var_5662_perm_0, x = var_5659_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 32, 1024]> input_123_cast_fp16 = reshape(shape = var_5666, x = var_5662_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61650048))))[name = string("tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5671_cast_fp16 = abs(x = linear_94_cast_fp16)[name = string("op_5671_cast_fp16")];
            bool var_5674 = const()[name = string("op_5674"), val = bool(true)];
            tensor<int32, [1]> var_5675_axes_0 = const()[name = string("op_5675_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5675_cast_fp16 = reduce_max(axes = var_5675_axes_0, keep_dims = var_5674, x = var_5671_cast_fp16)[name = string("op_5675_cast_fp16")];
            fp16 var_5676_to_fp16 = const()[name = string("op_5676_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_171_cast_fp16 = maximum(x = var_5675_cast_fp16, y = var_5676_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_171_cast_fp16 = real_div(x = linear_94_cast_fp16, y = s_171_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5679_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5679_cast_fp16")];
            tensor<int32, [1]> var_5684_axes_0 = const()[name = string("op_5684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5684_keep_dims_0 = const()[name = string("op_5684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5684_cast_fp16 = reduce_mean(axes = var_5684_axes_0, keep_dims = var_5684_keep_dims_0, x = var_5679_cast_fp16)[name = string("op_5684_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5685_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5685_cast_fp16")];
            fp32 var_5686_epsilon_0 = const()[name = string("op_5686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5686_cast_fp16 = inverse(epsilon = var_5686_epsilon_0, x = var_5685_cast_fp16)[name = string("op_5686_cast_fp16")];
            fp16 var_5687_to_fp16 = const()[name = string("op_5687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5688_cast_fp16 = mul(x = var_5686_cast_fp16, y = var_5687_to_fp16)[name = string("op_5688_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5690_cast_fp16 = add(x = var_5684_cast_fp16, y = var_5688_cast_fp16)[name = string("op_5690_cast_fp16")];
            fp32 var_5691_epsilon_0 = const()[name = string("op_5691_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5691_cast_fp16 = rsqrt(epsilon = var_5691_epsilon_0, x = var_5690_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_149_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5691_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61651136)))];
            tensor<fp16, [1, 32, 512]> var_5693_cast_fp16 = mul(x = out_149_cast_fp16, y = tm_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5693_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = var_5693_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5696_cast_fp16 = abs(x = x_191_cast_fp16)[name = string("op_5696_cast_fp16")];
            bool var_5699 = const()[name = string("op_5699"), val = bool(true)];
            tensor<int32, [1]> var_5700_axes_0 = const()[name = string("op_5700_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5700_cast_fp16 = reduce_max(axes = var_5700_axes_0, keep_dims = var_5699, x = var_5696_cast_fp16)[name = string("op_5700_cast_fp16")];
            fp16 var_5701_to_fp16 = const()[name = string("op_5701_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_173_cast_fp16 = maximum(x = var_5700_cast_fp16, y = var_5701_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_173_cast_fp16 = real_div(x = x_191_cast_fp16, y = s_173_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5704_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5704_cast_fp16")];
            tensor<int32, [1]> var_5709_axes_0 = const()[name = string("op_5709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5709_keep_dims_0 = const()[name = string("op_5709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5709_cast_fp16 = reduce_mean(axes = var_5709_axes_0, keep_dims = var_5709_keep_dims_0, x = var_5704_cast_fp16)[name = string("op_5709_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5710_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5710_cast_fp16")];
            fp32 var_5711_epsilon_0 = const()[name = string("op_5711_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5711_cast_fp16 = inverse(epsilon = var_5711_epsilon_0, x = var_5710_cast_fp16)[name = string("op_5711_cast_fp16")];
            fp16 var_5712_to_fp16 = const()[name = string("op_5712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = string("op_5713_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5715_cast_fp16 = add(x = var_5709_cast_fp16, y = var_5713_cast_fp16)[name = string("op_5715_cast_fp16")];
            fp32 var_5716_epsilon_0 = const()[name = string("op_5716_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5716_cast_fp16 = rsqrt(epsilon = var_5716_epsilon_0, x = var_5715_cast_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_151_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5716_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61652224)))];
            tensor<fp16, [1, 32, 512]> input_125_cast_fp16 = mul(x = out_151_cast_fp16, y = tm_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61653312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62701952))))[name = string("tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_95_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_95_cast_fp16")];
            string var_5723_mode_0 = const()[name = string("op_5723_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_5723_cast_fp16 = gelu(mode = var_5723_mode_0, x = linear_95_cast_fp16)[name = string("op_5723_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62706112))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63754752))))[name = string("tm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_127_cast_fp16 = mul(x = var_5723_cast_fp16, y = linear_96_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63758912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64807552))))[name = string("tm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5731_cast_fp16 = abs(x = linear_97_cast_fp16)[name = string("op_5731_cast_fp16")];
            bool var_5734 = const()[name = string("op_5734"), val = bool(true)];
            tensor<int32, [1]> var_5735_axes_0 = const()[name = string("op_5735_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5735_cast_fp16 = reduce_max(axes = var_5735_axes_0, keep_dims = var_5734, x = var_5731_cast_fp16)[name = string("op_5735_cast_fp16")];
            fp16 var_5736_to_fp16 = const()[name = string("op_5736_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_175_cast_fp16 = maximum(x = var_5735_cast_fp16, y = var_5736_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_175_cast_fp16 = real_div(x = linear_97_cast_fp16, y = s_175_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5739_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [1]> var_5744_axes_0 = const()[name = string("op_5744_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5744_keep_dims_0 = const()[name = string("op_5744_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5744_cast_fp16 = reduce_mean(axes = var_5744_axes_0, keep_dims = var_5744_keep_dims_0, x = var_5739_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5745_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5745_cast_fp16")];
            fp32 var_5746_epsilon_0 = const()[name = string("op_5746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5746_cast_fp16 = inverse(epsilon = var_5746_epsilon_0, x = var_5745_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5747_to_fp16 = const()[name = string("op_5747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5748_cast_fp16 = mul(x = var_5746_cast_fp16, y = var_5747_to_fp16)[name = string("op_5748_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5750_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5748_cast_fp16)[name = string("op_5750_cast_fp16")];
            fp32 var_5751_epsilon_0 = const()[name = string("op_5751_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5751_cast_fp16 = rsqrt(epsilon = var_5751_epsilon_0, x = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_153_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5751_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64808640)))];
            tensor<fp16, [1, 32, 512]> var_5753_cast_fp16 = mul(x = out_153_cast_fp16, y = tm_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_129_cast_fp16 = add(x = x_191_cast_fp16, y = var_5753_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64809728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65071936))))[name = string("tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_98_cast_fp16")];
            string x_195_mode_0 = const()[name = string("x_195_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_195_cast_fp16 = gelu(mode = x_195_mode_0, x = linear_98_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [4]> var_5773_begin_0 = const()[name = string("op_5773_begin_0"), val = tensor<int32, [4]>([0, 0, 10, 0])];
            tensor<int32, [4]> var_5773_end_0 = const()[name = string("op_5773_end_0"), val = tensor<int32, [4]>([1, 32, 11, 512])];
            tensor<bool, [4]> var_5773_end_mask_0 = const()[name = string("op_5773_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5773_squeeze_mask_0 = const()[name = string("op_5773_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_5773_cast_fp16 = slice_by_index(begin = var_5773_begin_0, end = var_5773_end_0, end_mask = var_5773_end_mask_0, squeeze_mask = var_5773_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_131_cast_fp16 = mul(x = x_195_cast_fp16, y = var_5773_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65073024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65335232))))[name = string("tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5783_cast_fp16 = abs(x = linear_99_cast_fp16)[name = string("op_5783_cast_fp16")];
            bool var_5786 = const()[name = string("op_5786"), val = bool(true)];
            tensor<int32, [1]> var_5787_axes_0 = const()[name = string("op_5787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5787_cast_fp16 = reduce_max(axes = var_5787_axes_0, keep_dims = var_5786, x = var_5783_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp16 var_5788_to_fp16 = const()[name = string("op_5788_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_177_cast_fp16 = maximum(x = var_5787_cast_fp16, y = var_5788_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_177_cast_fp16 = real_div(x = linear_99_cast_fp16, y = s_177_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5791_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5791_cast_fp16")];
            tensor<int32, [1]> var_5796_axes_0 = const()[name = string("op_5796_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5796_keep_dims_0 = const()[name = string("op_5796_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5796_cast_fp16 = reduce_mean(axes = var_5796_axes_0, keep_dims = var_5796_keep_dims_0, x = var_5791_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5797_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5797_cast_fp16")];
            fp32 var_5798_epsilon_0 = const()[name = string("op_5798_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5798_cast_fp16 = inverse(epsilon = var_5798_epsilon_0, x = var_5797_cast_fp16)[name = string("op_5798_cast_fp16")];
            fp16 var_5799_to_fp16 = const()[name = string("op_5799_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5800_cast_fp16 = mul(x = var_5798_cast_fp16, y = var_5799_to_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5802_cast_fp16 = add(x = var_5796_cast_fp16, y = var_5800_cast_fp16)[name = string("op_5802_cast_fp16")];
            fp32 var_5803_epsilon_0 = const()[name = string("op_5803_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5803_cast_fp16 = rsqrt(epsilon = var_5803_epsilon_0, x = var_5802_cast_fp16)[name = string("op_5803_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_155_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5803_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65336320)))];
            tensor<fp16, [1, 32, 512]> var_5805_cast_fp16 = mul(x = out_155_cast_fp16, y = tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5805_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_21_cast_fp16 = add(x = input_129_cast_fp16, y = var_5805_cast_fp16)[name = string("h_21_cast_fp16")];
            tensor<fp16, [1]> tm_layers_10_layer_scalar_to_fp16 = const()[name = string("tm_layers_10_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.1cp-1])];
            tensor<fp16, [1, 32, 512]> x_199_cast_fp16 = mul(x = h_21_cast_fp16, y = tm_layers_10_layer_scalar_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5809_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5809_cast_fp16")];
            bool var_5812 = const()[name = string("op_5812"), val = bool(true)];
            tensor<int32, [1]> var_5813_axes_0 = const()[name = string("op_5813_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_5813_cast_fp16 = reduce_max(axes = var_5813_axes_0, keep_dims = var_5812, x = var_5809_cast_fp16)[name = string("op_5813_cast_fp16")];
            fp16 var_5814_to_fp16 = const()[name = string("op_5814_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_179_cast_fp16 = maximum(x = var_5813_cast_fp16, y = var_5814_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_179_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_179_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_5817_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5817_cast_fp16")];
            tensor<int32, [1]> var_5822_axes_0 = const()[name = string("op_5822_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5822_keep_dims_0 = const()[name = string("op_5822_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5822_cast_fp16 = reduce_mean(axes = var_5822_axes_0, keep_dims = var_5822_keep_dims_0, x = var_5817_cast_fp16)[name = string("op_5822_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5823_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp32 var_5824_epsilon_0 = const()[name = string("op_5824_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_5824_cast_fp16 = inverse(epsilon = var_5824_epsilon_0, x = var_5823_cast_fp16)[name = string("op_5824_cast_fp16")];
            fp16 var_5825_to_fp16 = const()[name = string("op_5825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_5826_cast_fp16 = mul(x = var_5824_cast_fp16, y = var_5825_to_fp16)[name = string("op_5826_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_5828_cast_fp16 = add(x = var_5822_cast_fp16, y = var_5826_cast_fp16)[name = string("op_5828_cast_fp16")];
            fp32 var_5829_epsilon_0 = const()[name = string("op_5829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_5829_cast_fp16 = rsqrt(epsilon = var_5829_epsilon_0, x = var_5828_cast_fp16)[name = string("op_5829_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_157_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5829_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65337408)))];
            tensor<fp16, [1, 32, 512]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = tm_layers_11_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65338496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66387136))))[name = string("tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_100_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5839 = const()[name = string("op_5839"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 32, 4, 512]> var_5840_cast_fp16 = reshape(shape = var_5839, x = linear_100_cast_fp16)[name = string("op_5840_cast_fp16")];
            tensor<int32, [4]> x_201_perm_0 = const()[name = string("x_201_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66391296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66653504))))[name = string("tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_5851 = const()[name = string("op_5851"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 32, 1, 512]> var_5852_cast_fp16 = reshape(shape = var_5851, x = linear_101_cast_fp16)[name = string("op_5852_cast_fp16")];
            tensor<int32, [4]> x_203_perm_0 = const()[name = string("x_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66654592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66916800))))[name = string("tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 32, 1, 512]> var_5864_cast_fp16 = reshape(shape = var_5863, x = linear_102_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 512]> x_201_cast_fp16 = transpose(perm = x_201_perm_0, x = var_5840_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 4, 32, 512]> var_5868_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5868_cast_fp16")];
            bool var_5871 = const()[name = string("op_5871"), val = bool(true)];
            tensor<int32, [1]> var_5872_axes_0 = const()[name = string("op_5872_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_5872_cast_fp16 = reduce_max(axes = var_5872_axes_0, keep_dims = var_5871, x = var_5868_cast_fp16)[name = string("op_5872_cast_fp16")];
            fp16 var_5873_to_fp16 = const()[name = string("op_5873_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_181_cast_fp16 = maximum(x = var_5872_cast_fp16, y = var_5873_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> y_181_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_181_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_5876_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5876_cast_fp16")];
            tensor<int32, [1]> var_5881_axes_0 = const()[name = string("op_5881_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5881_keep_dims_0 = const()[name = string("op_5881_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_5881_cast_fp16 = reduce_mean(axes = var_5881_axes_0, keep_dims = var_5881_keep_dims_0, x = var_5876_cast_fp16)[name = string("op_5881_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_5882_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5882_cast_fp16")];
            fp32 var_5883_epsilon_0 = const()[name = string("op_5883_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_5883_cast_fp16 = inverse(epsilon = var_5883_epsilon_0, x = var_5882_cast_fp16)[name = string("op_5883_cast_fp16")];
            fp16 var_5884_to_fp16 = const()[name = string("op_5884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_5885_cast_fp16 = mul(x = var_5883_cast_fp16, y = var_5884_to_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_5887_cast_fp16 = add(x = var_5881_cast_fp16, y = var_5885_cast_fp16)[name = string("op_5887_cast_fp16")];
            fp32 var_5888_epsilon_0 = const()[name = string("op_5888_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_5888_cast_fp16 = rsqrt(epsilon = var_5888_epsilon_0, x = var_5887_cast_fp16)[name = string("op_5888_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> out_159_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5888_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66917888)))];
            tensor<fp16, [1, 4, 32, 512]> q_45_cast_fp16 = mul(x = out_159_cast_fp16, y = tm_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> x_203_cast_fp16 = transpose(perm = x_203_perm_0, x = var_5852_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1, 32, 512]> var_5891_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5891_cast_fp16")];
            bool var_5894 = const()[name = string("op_5894"), val = bool(true)];
            tensor<int32, [1]> var_5895_axes_0 = const()[name = string("op_5895_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 32, 1]> var_5895_cast_fp16 = reduce_max(axes = var_5895_axes_0, keep_dims = var_5894, x = var_5891_cast_fp16)[name = string("op_5895_cast_fp16")];
            fp16 var_5896_to_fp16 = const()[name = string("op_5896_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 32, 1]> s_183_cast_fp16 = maximum(x = var_5895_cast_fp16, y = var_5896_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> y_183_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_183_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_5899_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5899_cast_fp16")];
            tensor<int32, [1]> var_5904_axes_0 = const()[name = string("op_5904_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5904_keep_dims_0 = const()[name = string("op_5904_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 32, 1]> var_5904_cast_fp16 = reduce_mean(axes = var_5904_axes_0, keep_dims = var_5904_keep_dims_0, x = var_5899_cast_fp16)[name = string("op_5904_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_5905_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5905_cast_fp16")];
            fp32 var_5906_epsilon_0 = const()[name = string("op_5906_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 32, 1]> var_5906_cast_fp16 = inverse(epsilon = var_5906_epsilon_0, x = var_5905_cast_fp16)[name = string("op_5906_cast_fp16")];
            fp16 var_5907_to_fp16 = const()[name = string("op_5907_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 32, 1]> var_5908_cast_fp16 = mul(x = var_5906_cast_fp16, y = var_5907_to_fp16)[name = string("op_5908_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_5910_cast_fp16 = add(x = var_5904_cast_fp16, y = var_5908_cast_fp16)[name = string("op_5910_cast_fp16")];
            fp32 var_5911_epsilon_0 = const()[name = string("op_5911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 32, 1]> var_5911_cast_fp16 = rsqrt(epsilon = var_5911_epsilon_0, x = var_5910_cast_fp16)[name = string("op_5911_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> out_161_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5911_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66918976)))];
            tensor<fp16, [1, 1, 32, 512]> k_67_cast_fp16 = mul(x = out_161_cast_fp16, y = tm_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_5864_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1, 32, 512]> var_5914_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5914_cast_fp16")];
            bool var_5917 = const()[name = string("op_5917"), val = bool(true)];
            tensor<int32, [1]> var_5918_axes_0 = const()[name = string("op_5918_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 32, 1]> var_5918_cast_fp16 = reduce_max(axes = var_5918_axes_0, keep_dims = var_5917, x = var_5914_cast_fp16)[name = string("op_5918_cast_fp16")];
            fp16 var_5919_to_fp16 = const()[name = string("op_5919_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 32, 1]> s_185_cast_fp16 = maximum(x = var_5918_cast_fp16, y = var_5919_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> y_185_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_185_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_5922_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5922_cast_fp16")];
            tensor<int32, [1]> var_5927_axes_0 = const()[name = string("op_5927_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5927_keep_dims_0 = const()[name = string("op_5927_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 32, 1]> var_5927_cast_fp16 = reduce_mean(axes = var_5927_axes_0, keep_dims = var_5927_keep_dims_0, x = var_5922_cast_fp16)[name = string("op_5927_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_5928_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5928_cast_fp16")];
            fp32 var_5929_epsilon_0 = const()[name = string("op_5929_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 32, 1]> var_5929_cast_fp16 = inverse(epsilon = var_5929_epsilon_0, x = var_5928_cast_fp16)[name = string("op_5929_cast_fp16")];
            fp16 var_5930_to_fp16 = const()[name = string("op_5930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 32, 1]> var_5931_cast_fp16 = mul(x = var_5929_cast_fp16, y = var_5930_to_fp16)[name = string("op_5931_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_5933_cast_fp16 = add(x = var_5927_cast_fp16, y = var_5931_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp32 var_5934_epsilon_0 = const()[name = string("op_5934_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 32, 1]> var_5934_cast_fp16 = rsqrt(epsilon = var_5934_epsilon_0, x = var_5933_cast_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> v_45_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5934_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_5936_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5936_cast_fp16")];
            tensor<int32, [4]> var_5941_begin_0 = const()[name = string("op_5941_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5941_end_0 = const()[name = string("op_5941_end_0"), val = tensor<int32, [4]>([1, 4, 32, 512])];
            tensor<bool, [4]> var_5941_end_mask_0 = const()[name = string("op_5941_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 256]> var_5941_cast_fp16 = slice_by_index(begin = var_5941_begin_0, end = var_5941_end_0, end_mask = var_5941_end_mask_0, x = q_45_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 256]> var_5942_cast_fp16 = mul(x = var_5941_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_5942_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 256]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = q_45_cast_fp16)[name = string("op_5947_cast_fp16")];
            int32 var_5949 = const()[name = string("op_5949"), val = int32(-1)];
            bool var_5950_interleave_0 = const()[name = string("op_5950_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 512]> var_5950_cast_fp16 = concat(axis = var_5949, interleave = var_5950_interleave_0, values = (var_5942_cast_fp16, var_5947_cast_fp16))[name = string("op_5950_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_5951_cast_fp16 = mul(x = var_5950_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5951_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> q_47_cast_fp16 = add(x = var_5936_cast_fp16, y = var_5951_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_5954_cast_fp16 = mul(x = k_67_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5954_cast_fp16")];
            tensor<int32, [4]> var_5959_begin_0 = const()[name = string("op_5959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5959_end_0 = const()[name = string("op_5959_end_0"), val = tensor<int32, [4]>([1, 1, 32, 512])];
            tensor<bool, [4]> var_5959_end_mask_0 = const()[name = string("op_5959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 32, 256]> var_5959_cast_fp16 = slice_by_index(begin = var_5959_begin_0, end = var_5959_end_0, end_mask = var_5959_end_mask_0, x = k_67_cast_fp16)[name = string("op_5959_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 32, 256]> var_5960_cast_fp16 = mul(x = var_5959_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_5960_cast_fp16")];
            tensor<int32, [4]> var_5965_begin_0 = const()[name = string("op_5965_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5965_end_0 = const()[name = string("op_5965_end_0"), val = tensor<int32, [4]>([1, 1, 32, 256])];
            tensor<bool, [4]> var_5965_end_mask_0 = const()[name = string("op_5965_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 32, 256]> var_5965_cast_fp16 = slice_by_index(begin = var_5965_begin_0, end = var_5965_end_0, end_mask = var_5965_end_mask_0, x = k_67_cast_fp16)[name = string("op_5965_cast_fp16")];
            int32 var_5967 = const()[name = string("op_5967"), val = int32(-1)];
            bool var_5968_interleave_0 = const()[name = string("op_5968_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 32, 512]> var_5968_cast_fp16 = concat(axis = var_5967, interleave = var_5968_interleave_0, values = (var_5960_cast_fp16, var_5965_cast_fp16))[name = string("op_5968_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_5969_cast_fp16 = mul(x = var_5968_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5969_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> k_69_cast_fp16 = add(x = var_5954_cast_fp16, y = var_5969_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_22_reps_0 = const()[name = string("tile_22_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 32, 512]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = k_69_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [4, 1, 32, 512]> tile_22_cast_fp16 = tile(reps = tile_22_reps_0, x = transpose_44_cast_fp16)[name = string("tile_22_cast_fp16")];
            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([4, 1, 1, 32, 512])];
            tensor<fp16, [4, 1, 1, 32, 512]> reshape_44_cast_fp16 = reshape(shape = concat_44, x = tile_22_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [5]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 32, 512])];
            tensor<fp16, [1, 4, 1, 32, 512]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [4, 1, 32, 512]> reshape_45_cast_fp16 = reshape(shape = concat_45, x = transpose_45_cast_fp16)[name = string("reshape_45_cast_fp16")];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_23_reps_0 = const()[name = string("tile_23_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 32, 512]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = v_45_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [4, 1, 32, 512]> tile_23_cast_fp16 = tile(reps = tile_23_reps_0, x = transpose_46_cast_fp16)[name = string("tile_23_cast_fp16")];
            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([4, 1, 1, 32, 512])];
            tensor<fp16, [4, 1, 1, 32, 512]> reshape_46_cast_fp16 = reshape(shape = concat_46, x = tile_23_cast_fp16)[name = string("reshape_46_cast_fp16")];
            tensor<int32, [5]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 32, 512])];
            tensor<fp16, [1, 4, 1, 32, 512]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = reshape_46_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [4, 1, 32, 512]> reshape_47_cast_fp16 = reshape(shape = concat_47, x = transpose_47_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5983_transpose_x_0 = const()[name = string("op_5983_transpose_x_0"), val = bool(false)];
            bool var_5983_transpose_y_0 = const()[name = string("op_5983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 32]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = reshape_45_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 4, 32, 32]> var_5983_cast_fp16 = matmul(transpose_x = var_5983_transpose_x_0, transpose_y = var_5983_transpose_y_0, x = q_47_cast_fp16, y = transpose_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_5985_cast_fp16 = add(x = var_5983_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5985_cast_fp16")];
            int32 var_5986 = const()[name = string("op_5986"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_185_cast_fp16 = softmax(axis = var_5986, x = var_5985_cast_fp16)[name = string("w_185_cast_fp16")];
            bool var_5989_transpose_x_0 = const()[name = string("op_5989_transpose_x_0"), val = bool(false)];
            bool var_5989_transpose_y_0 = const()[name = string("op_5989_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 512]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 4, 32, 512]> var_5989_cast_fp16 = matmul(transpose_x = var_5989_transpose_x_0, transpose_y = var_5989_transpose_y_0, x = w_185_cast_fp16, y = v_47_cast_fp16)[name = string("op_5989_cast_fp16")];
            tensor<int32, [4]> var_5992_perm_0 = const()[name = string("op_5992_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5996 = const()[name = string("op_5996"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 32, 4, 512]> var_5992_cast_fp16 = transpose(perm = var_5992_perm_0, x = var_5989_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 32, 2048]> input_135_cast_fp16 = reshape(shape = var_5996, x = var_5992_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66920064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67968704))))[name = string("tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6001_cast_fp16 = abs(x = linear_103_cast_fp16)[name = string("op_6001_cast_fp16")];
            bool var_6004 = const()[name = string("op_6004"), val = bool(true)];
            tensor<int32, [1]> var_6005_axes_0 = const()[name = string("op_6005_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6005_cast_fp16 = reduce_max(axes = var_6005_axes_0, keep_dims = var_6004, x = var_6001_cast_fp16)[name = string("op_6005_cast_fp16")];
            fp16 var_6006_to_fp16 = const()[name = string("op_6006_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_187_cast_fp16 = maximum(x = var_6005_cast_fp16, y = var_6006_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_187_cast_fp16 = real_div(x = linear_103_cast_fp16, y = s_187_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6009_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_6009_cast_fp16")];
            tensor<int32, [1]> var_6014_axes_0 = const()[name = string("op_6014_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6014_keep_dims_0 = const()[name = string("op_6014_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6014_cast_fp16 = reduce_mean(axes = var_6014_axes_0, keep_dims = var_6014_keep_dims_0, x = var_6009_cast_fp16)[name = string("op_6014_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6015_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_6015_cast_fp16")];
            fp32 var_6016_epsilon_0 = const()[name = string("op_6016_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6016_cast_fp16 = inverse(epsilon = var_6016_epsilon_0, x = var_6015_cast_fp16)[name = string("op_6016_cast_fp16")];
            fp16 var_6017_to_fp16 = const()[name = string("op_6017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6018_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6020_cast_fp16 = add(x = var_6014_cast_fp16, y = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp32 var_6021_epsilon_0 = const()[name = string("op_6021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6021_cast_fp16 = rsqrt(epsilon = var_6021_epsilon_0, x = var_6020_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_163_cast_fp16 = mul(x = y_187_cast_fp16, y = var_6021_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969792)))];
            tensor<fp16, [1, 32, 512]> var_6023_cast_fp16 = mul(x = out_163_cast_fp16, y = tm_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_209_cast_fp16 = add(x = x_199_cast_fp16, y = var_6023_cast_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6026_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_6026_cast_fp16")];
            bool var_6029 = const()[name = string("op_6029"), val = bool(true)];
            tensor<int32, [1]> var_6030_axes_0 = const()[name = string("op_6030_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6030_cast_fp16 = reduce_max(axes = var_6030_axes_0, keep_dims = var_6029, x = var_6026_cast_fp16)[name = string("op_6030_cast_fp16")];
            fp16 var_6031_to_fp16 = const()[name = string("op_6031_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_189_cast_fp16 = maximum(x = var_6030_cast_fp16, y = var_6031_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_189_cast_fp16 = real_div(x = x_209_cast_fp16, y = s_189_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6034_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [1]> var_6039_axes_0 = const()[name = string("op_6039_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6039_keep_dims_0 = const()[name = string("op_6039_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6039_cast_fp16 = reduce_mean(axes = var_6039_axes_0, keep_dims = var_6039_keep_dims_0, x = var_6034_cast_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6040_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6041_cast_fp16 = inverse(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            fp16 var_6042_to_fp16 = const()[name = string("op_6042_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6043_cast_fp16 = mul(x = var_6041_cast_fp16, y = var_6042_to_fp16)[name = string("op_6043_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6045_cast_fp16 = add(x = var_6039_cast_fp16, y = var_6043_cast_fp16)[name = string("op_6045_cast_fp16")];
            fp32 var_6046_epsilon_0 = const()[name = string("op_6046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6046_cast_fp16 = rsqrt(epsilon = var_6046_epsilon_0, x = var_6045_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_165_cast_fp16 = mul(x = y_189_cast_fp16, y = var_6046_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67970880)))];
            tensor<fp16, [1, 32, 512]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = tm_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971968))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69020608))))[name = string("tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_104_cast_fp16")];
            string var_6053_mode_0 = const()[name = string("op_6053_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_6053_cast_fp16 = gelu(mode = var_6053_mode_0, x = linear_104_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70073408))))[name = string("tm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_105_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_139_cast_fp16 = mul(x = var_6053_cast_fp16, y = linear_105_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70077568))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71126208))))[name = string("tm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6061_cast_fp16 = abs(x = linear_106_cast_fp16)[name = string("op_6061_cast_fp16")];
            bool var_6064 = const()[name = string("op_6064"), val = bool(true)];
            tensor<int32, [1]> var_6065_axes_0 = const()[name = string("op_6065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6065_cast_fp16 = reduce_max(axes = var_6065_axes_0, keep_dims = var_6064, x = var_6061_cast_fp16)[name = string("op_6065_cast_fp16")];
            fp16 var_6066_to_fp16 = const()[name = string("op_6066_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_191_cast_fp16 = maximum(x = var_6065_cast_fp16, y = var_6066_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_191_cast_fp16 = real_div(x = linear_106_cast_fp16, y = s_191_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6069_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<int32, [1]> var_6074_axes_0 = const()[name = string("op_6074_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6074_keep_dims_0 = const()[name = string("op_6074_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6074_cast_fp16 = reduce_mean(axes = var_6074_axes_0, keep_dims = var_6074_keep_dims_0, x = var_6069_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6075_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_6075_cast_fp16")];
            fp32 var_6076_epsilon_0 = const()[name = string("op_6076_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6076_cast_fp16 = inverse(epsilon = var_6076_epsilon_0, x = var_6075_cast_fp16)[name = string("op_6076_cast_fp16")];
            fp16 var_6077_to_fp16 = const()[name = string("op_6077_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6078_cast_fp16 = mul(x = var_6076_cast_fp16, y = var_6077_to_fp16)[name = string("op_6078_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6080_cast_fp16 = add(x = var_6074_cast_fp16, y = var_6078_cast_fp16)[name = string("op_6080_cast_fp16")];
            fp32 var_6081_epsilon_0 = const()[name = string("op_6081_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6081_cast_fp16 = rsqrt(epsilon = var_6081_epsilon_0, x = var_6080_cast_fp16)[name = string("op_6081_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_167_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6081_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71127296)))];
            tensor<fp16, [1, 32, 512]> var_6083_cast_fp16 = mul(x = out_167_cast_fp16, y = tm_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6083_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_141_cast_fp16 = add(x = x_209_cast_fp16, y = var_6083_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71128384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71390592))))[name = string("tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_107_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_107_cast_fp16")];
            string x_213_mode_0 = const()[name = string("x_213_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_213_cast_fp16 = gelu(mode = x_213_mode_0, x = linear_107_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [4]> var_6103_begin_0 = const()[name = string("op_6103_begin_0"), val = tensor<int32, [4]>([0, 0, 11, 0])];
            tensor<int32, [4]> var_6103_end_0 = const()[name = string("op_6103_end_0"), val = tensor<int32, [4]>([1, 32, 12, 512])];
            tensor<bool, [4]> var_6103_end_mask_0 = const()[name = string("op_6103_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6103_squeeze_mask_0 = const()[name = string("op_6103_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_6103_cast_fp16 = slice_by_index(begin = var_6103_begin_0, end = var_6103_end_0, end_mask = var_6103_end_mask_0, squeeze_mask = var_6103_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6103_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_143_cast_fp16 = mul(x = x_213_cast_fp16, y = var_6103_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71391680))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71653888))))[name = string("tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6113_cast_fp16 = abs(x = linear_108_cast_fp16)[name = string("op_6113_cast_fp16")];
            bool var_6116 = const()[name = string("op_6116"), val = bool(true)];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6117_cast_fp16 = reduce_max(axes = var_6117_axes_0, keep_dims = var_6116, x = var_6113_cast_fp16)[name = string("op_6117_cast_fp16")];
            fp16 var_6118_to_fp16 = const()[name = string("op_6118_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_193_cast_fp16 = maximum(x = var_6117_cast_fp16, y = var_6118_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_193_cast_fp16 = real_div(x = linear_108_cast_fp16, y = s_193_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6121_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6121_cast_fp16")];
            tensor<int32, [1]> var_6126_axes_0 = const()[name = string("op_6126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6126_keep_dims_0 = const()[name = string("op_6126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6126_cast_fp16 = reduce_mean(axes = var_6126_axes_0, keep_dims = var_6126_keep_dims_0, x = var_6121_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6127_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_6127_cast_fp16")];
            fp32 var_6128_epsilon_0 = const()[name = string("op_6128_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6128_cast_fp16 = inverse(epsilon = var_6128_epsilon_0, x = var_6127_cast_fp16)[name = string("op_6128_cast_fp16")];
            fp16 var_6129_to_fp16 = const()[name = string("op_6129_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6130_cast_fp16 = mul(x = var_6128_cast_fp16, y = var_6129_to_fp16)[name = string("op_6130_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6132_cast_fp16 = add(x = var_6126_cast_fp16, y = var_6130_cast_fp16)[name = string("op_6132_cast_fp16")];
            fp32 var_6133_epsilon_0 = const()[name = string("op_6133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6133_cast_fp16 = rsqrt(epsilon = var_6133_epsilon_0, x = var_6132_cast_fp16)[name = string("op_6133_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_169_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6133_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71654976)))];
            tensor<fp16, [1, 32, 512]> var_6135_cast_fp16 = mul(x = out_169_cast_fp16, y = tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6135_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_23_cast_fp16 = add(x = input_141_cast_fp16, y = var_6135_cast_fp16)[name = string("h_23_cast_fp16")];
            tensor<fp16, [1]> tm_layers_11_layer_scalar_to_fp16 = const()[name = string("tm_layers_11_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.86p-2])];
            tensor<fp16, [1, 32, 512]> x_217_cast_fp16 = mul(x = h_23_cast_fp16, y = tm_layers_11_layer_scalar_to_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6139_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_195_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_195_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_195_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6153_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_171_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71656064)))];
            tensor<fp16, [1, 32, 512]> input_145_cast_fp16 = mul(x = out_171_cast_fp16, y = tm_layers_12_input_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71657152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72181504))))[name = string("tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<int32, [4]> var_6169 = const()[name = string("op_6169"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_6170_cast_fp16 = reshape(shape = var_6169, x = linear_109_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72183616))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72445824))))[name = string("tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [4]> var_6181 = const()[name = string("op_6181"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_6182_cast_fp16 = reshape(shape = var_6181, x = linear_110_cast_fp16)[name = string("op_6182_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72446912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72709120))))[name = string("tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<int32, [4]> var_6193 = const()[name = string("op_6193"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_6194_cast_fp16 = reshape(shape = var_6193, x = linear_111_cast_fp16)[name = string("op_6194_cast_fp16")];
            tensor<int32, [4]> x_223_perm_0 = const()[name = string("x_223_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6170_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 4, 32, 256]> var_6198_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_6198_cast_fp16")];
            bool var_6201 = const()[name = string("op_6201"), val = bool(true)];
            tensor<int32, [1]> var_6202_axes_0 = const()[name = string("op_6202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_6202_cast_fp16 = reduce_max(axes = var_6202_axes_0, keep_dims = var_6201, x = var_6198_cast_fp16)[name = string("op_6202_cast_fp16")];
            fp16 var_6203_to_fp16 = const()[name = string("op_6203_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_197_cast_fp16 = maximum(x = var_6202_cast_fp16, y = var_6203_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_197_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_197_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6206_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6206_cast_fp16")];
            tensor<int32, [1]> var_6211_axes_0 = const()[name = string("op_6211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6211_keep_dims_0 = const()[name = string("op_6211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_6211_cast_fp16 = reduce_mean(axes = var_6211_axes_0, keep_dims = var_6211_keep_dims_0, x = var_6206_cast_fp16)[name = string("op_6211_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_6212_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6212_cast_fp16")];
            fp32 var_6213_epsilon_0 = const()[name = string("op_6213_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_6213_cast_fp16 = inverse(epsilon = var_6213_epsilon_0, x = var_6212_cast_fp16)[name = string("op_6213_cast_fp16")];
            fp16 var_6214_to_fp16 = const()[name = string("op_6214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_6215_cast_fp16 = mul(x = var_6213_cast_fp16, y = var_6214_to_fp16)[name = string("op_6215_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_6217_cast_fp16 = add(x = var_6211_cast_fp16, y = var_6215_cast_fp16)[name = string("op_6217_cast_fp16")];
            fp32 var_6218_epsilon_0 = const()[name = string("op_6218_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_6218_cast_fp16 = rsqrt(epsilon = var_6218_epsilon_0, x = var_6217_cast_fp16)[name = string("op_6218_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_173_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6218_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_49_cast_fp16 = mul(x = out_173_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_49_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6182_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 2, 32, 256]> var_6221_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_6221_cast_fp16")];
            bool var_6224 = const()[name = string("op_6224"), val = bool(true)];
            tensor<int32, [1]> var_6225_axes_0 = const()[name = string("op_6225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_6225_cast_fp16 = reduce_max(axes = var_6225_axes_0, keep_dims = var_6224, x = var_6221_cast_fp16)[name = string("op_6225_cast_fp16")];
            fp16 var_6226_to_fp16 = const()[name = string("op_6226_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_199_cast_fp16 = maximum(x = var_6225_cast_fp16, y = var_6226_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_199_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_199_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6229_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6229_cast_fp16")];
            tensor<int32, [1]> var_6234_axes_0 = const()[name = string("op_6234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6234_keep_dims_0 = const()[name = string("op_6234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_6234_cast_fp16 = reduce_mean(axes = var_6234_axes_0, keep_dims = var_6234_keep_dims_0, x = var_6229_cast_fp16)[name = string("op_6234_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6235_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6235_cast_fp16")];
            fp32 var_6236_epsilon_0 = const()[name = string("op_6236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_6236_cast_fp16 = inverse(epsilon = var_6236_epsilon_0, x = var_6235_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_6238_cast_fp16 = mul(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6240_cast_fp16 = add(x = var_6234_cast_fp16, y = var_6238_cast_fp16)[name = string("op_6240_cast_fp16")];
            fp32 var_6241_epsilon_0 = const()[name = string("op_6241_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_6241_cast_fp16 = rsqrt(epsilon = var_6241_epsilon_0, x = var_6240_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_175_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6241_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_73_cast_fp16 = mul(x = out_175_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_223_cast_fp16 = transpose(perm = x_223_perm_0, x = var_6194_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2, 32, 256]> var_6244_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_6244_cast_fp16")];
            bool var_6247 = const()[name = string("op_6247"), val = bool(true)];
            tensor<int32, [1]> var_6248_axes_0 = const()[name = string("op_6248_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_6248_cast_fp16 = reduce_max(axes = var_6248_axes_0, keep_dims = var_6247, x = var_6244_cast_fp16)[name = string("op_6248_cast_fp16")];
            fp16 var_6249_to_fp16 = const()[name = string("op_6249_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_201_cast_fp16 = maximum(x = var_6248_cast_fp16, y = var_6249_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_201_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_201_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6252_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6252_cast_fp16")];
            tensor<int32, [1]> var_6257_axes_0 = const()[name = string("op_6257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6257_keep_dims_0 = const()[name = string("op_6257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_6257_cast_fp16 = reduce_mean(axes = var_6257_axes_0, keep_dims = var_6257_keep_dims_0, x = var_6252_cast_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6258_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6258_cast_fp16")];
            fp32 var_6259_epsilon_0 = const()[name = string("op_6259_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_6259_cast_fp16 = inverse(epsilon = var_6259_epsilon_0, x = var_6258_cast_fp16)[name = string("op_6259_cast_fp16")];
            fp16 var_6260_to_fp16 = const()[name = string("op_6260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_6261_cast_fp16 = mul(x = var_6259_cast_fp16, y = var_6260_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6263_cast_fp16 = add(x = var_6257_cast_fp16, y = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp32 var_6264_epsilon_0 = const()[name = string("op_6264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_6264_cast_fp16 = rsqrt(epsilon = var_6264_epsilon_0, x = var_6263_cast_fp16)[name = string("op_6264_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_49_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6264_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6266_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6266_cast_fp16")];
            tensor<int32, [4]> var_6271_begin_0 = const()[name = string("op_6271_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6271_end_0 = const()[name = string("op_6271_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_6271_end_mask_0 = const()[name = string("op_6271_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_6271_cast_fp16 = slice_by_index(begin = var_6271_begin_0, end = var_6271_end_0, end_mask = var_6271_end_mask_0, x = q_49_cast_fp16)[name = string("op_6271_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_6272_cast_fp16 = mul(x = var_6271_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_6272_cast_fp16")];
            tensor<int32, [4]> var_6277_begin_0 = const()[name = string("op_6277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6277_end_0 = const()[name = string("op_6277_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_6277_end_mask_0 = const()[name = string("op_6277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_6277_cast_fp16 = slice_by_index(begin = var_6277_begin_0, end = var_6277_end_0, end_mask = var_6277_end_mask_0, x = q_49_cast_fp16)[name = string("op_6277_cast_fp16")];
            int32 var_6279 = const()[name = string("op_6279"), val = int32(-1)];
            bool var_6280_interleave_0 = const()[name = string("op_6280_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_6280_cast_fp16 = concat(axis = var_6279, interleave = var_6280_interleave_0, values = (var_6272_cast_fp16, var_6277_cast_fp16))[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6281_cast_fp16 = mul(x = var_6280_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6281_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_51_cast_fp16 = add(x = var_6266_cast_fp16, y = var_6281_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6284_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6284_cast_fp16")];
            tensor<int32, [4]> var_6289_begin_0 = const()[name = string("op_6289_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6289_end_0 = const()[name = string("op_6289_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_6289_end_mask_0 = const()[name = string("op_6289_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_6289_cast_fp16 = slice_by_index(begin = var_6289_begin_0, end = var_6289_end_0, end_mask = var_6289_end_mask_0, x = k_73_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_6290_cast_fp16 = mul(x = var_6289_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_6290_cast_fp16")];
            tensor<int32, [4]> var_6295_begin_0 = const()[name = string("op_6295_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6295_end_0 = const()[name = string("op_6295_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_6295_end_mask_0 = const()[name = string("op_6295_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_6295_cast_fp16 = slice_by_index(begin = var_6295_begin_0, end = var_6295_end_0, end_mask = var_6295_end_mask_0, x = k_73_cast_fp16)[name = string("op_6295_cast_fp16")];
            int32 var_6297 = const()[name = string("op_6297"), val = int32(-1)];
            bool var_6298_interleave_0 = const()[name = string("op_6298_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_6298_cast_fp16 = concat(axis = var_6297, interleave = var_6298_interleave_0, values = (var_6290_cast_fp16, var_6295_cast_fp16))[name = string("op_6298_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6299_cast_fp16 = mul(x = var_6298_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6299_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_75_cast_fp16 = add(x = var_6284_cast_fp16, y = var_6299_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_24_reps_0 = const()[name = string("tile_24_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = k_75_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [4, 1, 32, 256]> tile_24_cast_fp16 = tile(reps = tile_24_reps_0, x = transpose_48_cast_fp16)[name = string("tile_24_cast_fp16")];
            tensor<int32, [5]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_48_cast_fp16 = reshape(shape = concat_48, x = tile_24_cast_fp16)[name = string("reshape_48_cast_fp16")];
            tensor<int32, [5]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = reshape_48_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [4, 1, 32, 256]> reshape_49_cast_fp16 = reshape(shape = concat_49, x = transpose_49_cast_fp16)[name = string("reshape_49_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_25_reps_0 = const()[name = string("tile_25_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = v_49_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [4, 1, 32, 256]> tile_25_cast_fp16 = tile(reps = tile_25_reps_0, x = transpose_50_cast_fp16)[name = string("tile_25_cast_fp16")];
            tensor<int32, [5]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_50_cast_fp16 = reshape(shape = concat_50, x = tile_25_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [5]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [4, 1, 32, 256]> reshape_51_cast_fp16 = reshape(shape = concat_51, x = transpose_51_cast_fp16)[name = string("reshape_51_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6313_transpose_x_0 = const()[name = string("op_6313_transpose_x_0"), val = bool(false)];
            bool var_6313_transpose_y_0 = const()[name = string("op_6313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = reshape_49_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 4, 32, 32]> var_6313_cast_fp16 = matmul(transpose_x = var_6313_transpose_x_0, transpose_y = var_6313_transpose_y_0, x = q_51_cast_fp16, y = transpose_108_cast_fp16)[name = string("op_6313_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_6315_cast_fp16 = add(x = var_6313_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6315_cast_fp16")];
            int32 var_6316 = const()[name = string("op_6316"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_201_cast_fp16 = softmax(axis = var_6316, x = var_6315_cast_fp16)[name = string("w_201_cast_fp16")];
            bool var_6319_transpose_x_0 = const()[name = string("op_6319_transpose_x_0"), val = bool(false)];
            bool var_6319_transpose_y_0 = const()[name = string("op_6319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = reshape_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 4, 32, 256]> var_6319_cast_fp16 = matmul(transpose_x = var_6319_transpose_x_0, transpose_y = var_6319_transpose_y_0, x = w_201_cast_fp16, y = v_51_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<int32, [4]> var_6322_perm_0 = const()[name = string("op_6322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6326 = const()[name = string("op_6326"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_6322_cast_fp16 = transpose(perm = var_6322_perm_0, x = var_6319_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 32, 1024]> input_147_cast_fp16 = reshape(shape = var_6326, x = var_6322_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72710208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73234560))))[name = string("tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6331_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6331_cast_fp16")];
            bool var_6334 = const()[name = string("op_6334"), val = bool(true)];
            tensor<int32, [1]> var_6335_axes_0 = const()[name = string("op_6335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6335_cast_fp16 = reduce_max(axes = var_6335_axes_0, keep_dims = var_6334, x = var_6331_cast_fp16)[name = string("op_6335_cast_fp16")];
            fp16 var_6336_to_fp16 = const()[name = string("op_6336_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_203_cast_fp16 = maximum(x = var_6335_cast_fp16, y = var_6336_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_203_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_203_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6339_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6339_cast_fp16")];
            tensor<int32, [1]> var_6344_axes_0 = const()[name = string("op_6344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6344_keep_dims_0 = const()[name = string("op_6344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6344_cast_fp16 = reduce_mean(axes = var_6344_axes_0, keep_dims = var_6344_keep_dims_0, x = var_6339_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6345_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6345_cast_fp16")];
            fp32 var_6346_epsilon_0 = const()[name = string("op_6346_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6346_cast_fp16 = inverse(epsilon = var_6346_epsilon_0, x = var_6345_cast_fp16)[name = string("op_6346_cast_fp16")];
            fp16 var_6347_to_fp16 = const()[name = string("op_6347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6348_cast_fp16 = mul(x = var_6346_cast_fp16, y = var_6347_to_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6350_cast_fp16 = add(x = var_6344_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            fp32 var_6351_epsilon_0 = const()[name = string("op_6351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6351_cast_fp16 = rsqrt(epsilon = var_6351_epsilon_0, x = var_6350_cast_fp16)[name = string("op_6351_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_177_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6351_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73235648)))];
            tensor<fp16, [1, 32, 512]> var_6353_cast_fp16 = mul(x = out_177_cast_fp16, y = tm_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = var_6353_cast_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6356_cast_fp16 = abs(x = x_227_cast_fp16)[name = string("op_6356_cast_fp16")];
            bool var_6359 = const()[name = string("op_6359"), val = bool(true)];
            tensor<int32, [1]> var_6360_axes_0 = const()[name = string("op_6360_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6360_cast_fp16 = reduce_max(axes = var_6360_axes_0, keep_dims = var_6359, x = var_6356_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_205_cast_fp16 = maximum(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_205_cast_fp16 = real_div(x = x_227_cast_fp16, y = s_205_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6364_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<int32, [1]> var_6369_axes_0 = const()[name = string("op_6369_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6369_keep_dims_0 = const()[name = string("op_6369_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6369_cast_fp16 = reduce_mean(axes = var_6369_axes_0, keep_dims = var_6369_keep_dims_0, x = var_6364_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6370_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6370_cast_fp16")];
            fp32 var_6371_epsilon_0 = const()[name = string("op_6371_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6371_cast_fp16 = inverse(epsilon = var_6371_epsilon_0, x = var_6370_cast_fp16)[name = string("op_6371_cast_fp16")];
            fp16 var_6372_to_fp16 = const()[name = string("op_6372_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6373_cast_fp16 = mul(x = var_6371_cast_fp16, y = var_6372_to_fp16)[name = string("op_6373_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6375_cast_fp16 = add(x = var_6369_cast_fp16, y = var_6373_cast_fp16)[name = string("op_6375_cast_fp16")];
            fp32 var_6376_epsilon_0 = const()[name = string("op_6376_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6376_cast_fp16 = rsqrt(epsilon = var_6376_epsilon_0, x = var_6375_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_179_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6376_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73236736)))];
            tensor<fp16, [1, 32, 512]> input_149_cast_fp16 = mul(x = out_179_cast_fp16, y = tm_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73237824))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74286464))))[name = string("tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_113_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_113_cast_fp16")];
            string var_6383_mode_0 = const()[name = string("op_6383_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_6383_cast_fp16 = gelu(mode = var_6383_mode_0, x = linear_113_cast_fp16)[name = string("op_6383_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74290624))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339264))))[name = string("tm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_114_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_151_cast_fp16 = mul(x = var_6383_cast_fp16, y = linear_114_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75343424))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76392064))))[name = string("tm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6391_cast_fp16 = abs(x = linear_115_cast_fp16)[name = string("op_6391_cast_fp16")];
            bool var_6394 = const()[name = string("op_6394"), val = bool(true)];
            tensor<int32, [1]> var_6395_axes_0 = const()[name = string("op_6395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6395_cast_fp16 = reduce_max(axes = var_6395_axes_0, keep_dims = var_6394, x = var_6391_cast_fp16)[name = string("op_6395_cast_fp16")];
            fp16 var_6396_to_fp16 = const()[name = string("op_6396_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_207_cast_fp16 = maximum(x = var_6395_cast_fp16, y = var_6396_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_207_cast_fp16 = real_div(x = linear_115_cast_fp16, y = s_207_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6399_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6399_cast_fp16")];
            tensor<int32, [1]> var_6404_axes_0 = const()[name = string("op_6404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6404_keep_dims_0 = const()[name = string("op_6404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6404_cast_fp16 = reduce_mean(axes = var_6404_axes_0, keep_dims = var_6404_keep_dims_0, x = var_6399_cast_fp16)[name = string("op_6404_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6405_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6405_cast_fp16")];
            fp32 var_6406_epsilon_0 = const()[name = string("op_6406_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6406_cast_fp16 = inverse(epsilon = var_6406_epsilon_0, x = var_6405_cast_fp16)[name = string("op_6406_cast_fp16")];
            fp16 var_6407_to_fp16 = const()[name = string("op_6407_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6408_cast_fp16 = mul(x = var_6406_cast_fp16, y = var_6407_to_fp16)[name = string("op_6408_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6410_cast_fp16 = add(x = var_6404_cast_fp16, y = var_6408_cast_fp16)[name = string("op_6410_cast_fp16")];
            fp32 var_6411_epsilon_0 = const()[name = string("op_6411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6411_cast_fp16 = rsqrt(epsilon = var_6411_epsilon_0, x = var_6410_cast_fp16)[name = string("op_6411_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_181_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6411_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76393152)))];
            tensor<fp16, [1, 32, 512]> var_6413_cast_fp16 = mul(x = out_181_cast_fp16, y = tm_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6413_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_153_cast_fp16 = add(x = x_227_cast_fp16, y = var_6413_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76394240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76656448))))[name = string("tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_116_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_116_cast_fp16")];
            string x_231_mode_0 = const()[name = string("x_231_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_231_cast_fp16 = gelu(mode = x_231_mode_0, x = linear_116_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_6433_begin_0 = const()[name = string("op_6433_begin_0"), val = tensor<int32, [4]>([0, 0, 12, 0])];
            tensor<int32, [4]> var_6433_end_0 = const()[name = string("op_6433_end_0"), val = tensor<int32, [4]>([1, 32, 13, 512])];
            tensor<bool, [4]> var_6433_end_mask_0 = const()[name = string("op_6433_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6433_squeeze_mask_0 = const()[name = string("op_6433_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_6433_cast_fp16 = slice_by_index(begin = var_6433_begin_0, end = var_6433_end_0, end_mask = var_6433_end_mask_0, squeeze_mask = var_6433_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6433_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_155_cast_fp16 = mul(x = x_231_cast_fp16, y = var_6433_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76657536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76919744))))[name = string("tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6443_cast_fp16 = abs(x = linear_117_cast_fp16)[name = string("op_6443_cast_fp16")];
            bool var_6446 = const()[name = string("op_6446"), val = bool(true)];
            tensor<int32, [1]> var_6447_axes_0 = const()[name = string("op_6447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6447_cast_fp16 = reduce_max(axes = var_6447_axes_0, keep_dims = var_6446, x = var_6443_cast_fp16)[name = string("op_6447_cast_fp16")];
            fp16 var_6448_to_fp16 = const()[name = string("op_6448_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_209_cast_fp16 = maximum(x = var_6447_cast_fp16, y = var_6448_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_209_cast_fp16 = real_div(x = linear_117_cast_fp16, y = s_209_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6451_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6451_cast_fp16")];
            tensor<int32, [1]> var_6456_axes_0 = const()[name = string("op_6456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6456_keep_dims_0 = const()[name = string("op_6456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6456_cast_fp16 = reduce_mean(axes = var_6456_axes_0, keep_dims = var_6456_keep_dims_0, x = var_6451_cast_fp16)[name = string("op_6456_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6457_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6457_cast_fp16")];
            fp32 var_6458_epsilon_0 = const()[name = string("op_6458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6458_cast_fp16 = inverse(epsilon = var_6458_epsilon_0, x = var_6457_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6459_to_fp16 = const()[name = string("op_6459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6460_cast_fp16 = mul(x = var_6458_cast_fp16, y = var_6459_to_fp16)[name = string("op_6460_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6462_cast_fp16 = add(x = var_6456_cast_fp16, y = var_6460_cast_fp16)[name = string("op_6462_cast_fp16")];
            fp32 var_6463_epsilon_0 = const()[name = string("op_6463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6463_cast_fp16 = rsqrt(epsilon = var_6463_epsilon_0, x = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_183_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6463_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76920832)))];
            tensor<fp16, [1, 32, 512]> var_6465_cast_fp16 = mul(x = out_183_cast_fp16, y = tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6465_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_25_cast_fp16 = add(x = input_153_cast_fp16, y = var_6465_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<fp16, [1]> tm_layers_12_layer_scalar_to_fp16 = const()[name = string("tm_layers_12_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.8cp-1])];
            tensor<fp16, [1, 32, 512]> x_235_cast_fp16 = mul(x = h_25_cast_fp16, y = tm_layers_12_layer_scalar_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6469_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_6469_cast_fp16")];
            bool var_6472 = const()[name = string("op_6472"), val = bool(true)];
            tensor<int32, [1]> var_6473_axes_0 = const()[name = string("op_6473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6473_cast_fp16 = reduce_max(axes = var_6473_axes_0, keep_dims = var_6472, x = var_6469_cast_fp16)[name = string("op_6473_cast_fp16")];
            fp16 var_6474_to_fp16 = const()[name = string("op_6474_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_211_cast_fp16 = maximum(x = var_6473_cast_fp16, y = var_6474_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_211_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_211_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6477_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6477_cast_fp16")];
            tensor<int32, [1]> var_6482_axes_0 = const()[name = string("op_6482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6482_keep_dims_0 = const()[name = string("op_6482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6482_cast_fp16 = reduce_mean(axes = var_6482_axes_0, keep_dims = var_6482_keep_dims_0, x = var_6477_cast_fp16)[name = string("op_6482_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6483_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6483_cast_fp16")];
            fp32 var_6484_epsilon_0 = const()[name = string("op_6484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6484_cast_fp16 = inverse(epsilon = var_6484_epsilon_0, x = var_6483_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp16 var_6485_to_fp16 = const()[name = string("op_6485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6486_cast_fp16 = mul(x = var_6484_cast_fp16, y = var_6485_to_fp16)[name = string("op_6486_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6488_cast_fp16 = add(x = var_6482_cast_fp16, y = var_6486_cast_fp16)[name = string("op_6488_cast_fp16")];
            fp32 var_6489_epsilon_0 = const()[name = string("op_6489_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6489_cast_fp16 = rsqrt(epsilon = var_6489_epsilon_0, x = var_6488_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_185_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6489_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76921920)))];
            tensor<fp16, [1, 32, 512]> input_157_cast_fp16 = mul(x = out_185_cast_fp16, y = tm_layers_13_input_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76923008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77447360))))[name = string("tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<int32, [4]> var_6499 = const()[name = string("op_6499"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_6500_cast_fp16 = reshape(shape = var_6499, x = linear_118_cast_fp16)[name = string("op_6500_cast_fp16")];
            tensor<int32, [4]> x_237_perm_0 = const()[name = string("x_237_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77449472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77711680))))[name = string("tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6511 = const()[name = string("op_6511"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_6512_cast_fp16 = reshape(shape = var_6511, x = linear_119_cast_fp16)[name = string("op_6512_cast_fp16")];
            tensor<int32, [4]> x_239_perm_0 = const()[name = string("x_239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77712768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77974976))))[name = string("tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6523 = const()[name = string("op_6523"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_6524_cast_fp16 = reshape(shape = var_6523, x = linear_120_cast_fp16)[name = string("op_6524_cast_fp16")];
            tensor<int32, [4]> x_241_perm_0 = const()[name = string("x_241_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_237_cast_fp16 = transpose(perm = x_237_perm_0, x = var_6500_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 4, 32, 256]> var_6528_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_6528_cast_fp16")];
            bool var_6531 = const()[name = string("op_6531"), val = bool(true)];
            tensor<int32, [1]> var_6532_axes_0 = const()[name = string("op_6532_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_6532_cast_fp16 = reduce_max(axes = var_6532_axes_0, keep_dims = var_6531, x = var_6528_cast_fp16)[name = string("op_6532_cast_fp16")];
            fp16 var_6533_to_fp16 = const()[name = string("op_6533_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_213_cast_fp16 = maximum(x = var_6532_cast_fp16, y = var_6533_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_213_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_213_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6536_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6536_cast_fp16")];
            tensor<int32, [1]> var_6541_axes_0 = const()[name = string("op_6541_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6541_keep_dims_0 = const()[name = string("op_6541_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_6541_cast_fp16 = reduce_mean(axes = var_6541_axes_0, keep_dims = var_6541_keep_dims_0, x = var_6536_cast_fp16)[name = string("op_6541_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_6542_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp32 var_6543_epsilon_0 = const()[name = string("op_6543_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_6543_cast_fp16 = inverse(epsilon = var_6543_epsilon_0, x = var_6542_cast_fp16)[name = string("op_6543_cast_fp16")];
            fp16 var_6544_to_fp16 = const()[name = string("op_6544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_6545_cast_fp16 = mul(x = var_6543_cast_fp16, y = var_6544_to_fp16)[name = string("op_6545_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_6547_cast_fp16 = add(x = var_6541_cast_fp16, y = var_6545_cast_fp16)[name = string("op_6547_cast_fp16")];
            fp32 var_6548_epsilon_0 = const()[name = string("op_6548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_6548_cast_fp16 = rsqrt(epsilon = var_6548_epsilon_0, x = var_6547_cast_fp16)[name = string("op_6548_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_187_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6548_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976064)))];
            tensor<fp16, [1, 4, 32, 256]> q_53_cast_fp16 = mul(x = out_187_cast_fp16, y = tm_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("q_53_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_239_cast_fp16 = transpose(perm = x_239_perm_0, x = var_6512_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2, 32, 256]> var_6551_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_6551_cast_fp16")];
            bool var_6554 = const()[name = string("op_6554"), val = bool(true)];
            tensor<int32, [1]> var_6555_axes_0 = const()[name = string("op_6555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_6555_cast_fp16 = reduce_max(axes = var_6555_axes_0, keep_dims = var_6554, x = var_6551_cast_fp16)[name = string("op_6555_cast_fp16")];
            fp16 var_6556_to_fp16 = const()[name = string("op_6556_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_215_cast_fp16 = maximum(x = var_6555_cast_fp16, y = var_6556_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_215_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_215_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6559_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6559_cast_fp16")];
            tensor<int32, [1]> var_6564_axes_0 = const()[name = string("op_6564_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6564_keep_dims_0 = const()[name = string("op_6564_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_6564_cast_fp16 = reduce_mean(axes = var_6564_axes_0, keep_dims = var_6564_keep_dims_0, x = var_6559_cast_fp16)[name = string("op_6564_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6565_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6565_cast_fp16")];
            fp32 var_6566_epsilon_0 = const()[name = string("op_6566_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_6566_cast_fp16 = inverse(epsilon = var_6566_epsilon_0, x = var_6565_cast_fp16)[name = string("op_6566_cast_fp16")];
            fp16 var_6567_to_fp16 = const()[name = string("op_6567_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_6568_cast_fp16 = mul(x = var_6566_cast_fp16, y = var_6567_to_fp16)[name = string("op_6568_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6570_cast_fp16 = add(x = var_6564_cast_fp16, y = var_6568_cast_fp16)[name = string("op_6570_cast_fp16")];
            fp32 var_6571_epsilon_0 = const()[name = string("op_6571_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_6571_cast_fp16 = rsqrt(epsilon = var_6571_epsilon_0, x = var_6570_cast_fp16)[name = string("op_6571_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_189_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6571_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976640)))];
            tensor<fp16, [1, 2, 32, 256]> k_79_cast_fp16 = mul(x = out_189_cast_fp16, y = tm_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("k_79_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_241_cast_fp16 = transpose(perm = x_241_perm_0, x = var_6524_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 2, 32, 256]> var_6574_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_6574_cast_fp16")];
            bool var_6577 = const()[name = string("op_6577"), val = bool(true)];
            tensor<int32, [1]> var_6578_axes_0 = const()[name = string("op_6578_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_6578_cast_fp16 = reduce_max(axes = var_6578_axes_0, keep_dims = var_6577, x = var_6574_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_217_cast_fp16 = maximum(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_217_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_217_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6582_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6587_axes_0 = const()[name = string("op_6587_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6587_keep_dims_0 = const()[name = string("op_6587_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_6587_cast_fp16 = reduce_mean(axes = var_6587_axes_0, keep_dims = var_6587_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6588_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6588_cast_fp16")];
            fp32 var_6589_epsilon_0 = const()[name = string("op_6589_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_6589_cast_fp16 = inverse(epsilon = var_6589_epsilon_0, x = var_6588_cast_fp16)[name = string("op_6589_cast_fp16")];
            fp16 var_6590_to_fp16 = const()[name = string("op_6590_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_6591_cast_fp16 = mul(x = var_6589_cast_fp16, y = var_6590_to_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6593_cast_fp16 = add(x = var_6587_cast_fp16, y = var_6591_cast_fp16)[name = string("op_6593_cast_fp16")];
            fp32 var_6594_epsilon_0 = const()[name = string("op_6594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_6594_cast_fp16 = rsqrt(epsilon = var_6594_epsilon_0, x = var_6593_cast_fp16)[name = string("op_6594_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_53_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6594_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6596_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6596_cast_fp16")];
            tensor<int32, [4]> var_6601_begin_0 = const()[name = string("op_6601_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6601_end_0 = const()[name = string("op_6601_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_6601_end_mask_0 = const()[name = string("op_6601_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_6601_cast_fp16 = slice_by_index(begin = var_6601_begin_0, end = var_6601_end_0, end_mask = var_6601_end_mask_0, x = q_53_cast_fp16)[name = string("op_6601_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_6602_cast_fp16 = mul(x = var_6601_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_6602_cast_fp16")];
            tensor<int32, [4]> var_6607_begin_0 = const()[name = string("op_6607_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6607_end_0 = const()[name = string("op_6607_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_6607_end_mask_0 = const()[name = string("op_6607_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_6607_cast_fp16 = slice_by_index(begin = var_6607_begin_0, end = var_6607_end_0, end_mask = var_6607_end_mask_0, x = q_53_cast_fp16)[name = string("op_6607_cast_fp16")];
            int32 var_6609 = const()[name = string("op_6609"), val = int32(-1)];
            bool var_6610_interleave_0 = const()[name = string("op_6610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_6610_cast_fp16 = concat(axis = var_6609, interleave = var_6610_interleave_0, values = (var_6602_cast_fp16, var_6607_cast_fp16))[name = string("op_6610_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6611_cast_fp16 = mul(x = var_6610_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_55_cast_fp16 = add(x = var_6596_cast_fp16, y = var_6611_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6614_cast_fp16 = mul(x = k_79_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = k_79_cast_fp16)[name = string("op_6619_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_6620_cast_fp16 = mul(x = var_6619_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_6620_cast_fp16")];
            tensor<int32, [4]> var_6625_begin_0 = const()[name = string("op_6625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6625_end_0 = const()[name = string("op_6625_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_6625_end_mask_0 = const()[name = string("op_6625_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_6625_cast_fp16 = slice_by_index(begin = var_6625_begin_0, end = var_6625_end_0, end_mask = var_6625_end_mask_0, x = k_79_cast_fp16)[name = string("op_6625_cast_fp16")];
            int32 var_6627 = const()[name = string("op_6627"), val = int32(-1)];
            bool var_6628_interleave_0 = const()[name = string("op_6628_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_6628_cast_fp16 = concat(axis = var_6627, interleave = var_6628_interleave_0, values = (var_6620_cast_fp16, var_6625_cast_fp16))[name = string("op_6628_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6629_cast_fp16 = mul(x = var_6628_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6629_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_81_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6629_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_26_reps_0 = const()[name = string("tile_26_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = k_81_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [4, 1, 32, 256]> tile_26_cast_fp16 = tile(reps = tile_26_reps_0, x = transpose_52_cast_fp16)[name = string("tile_26_cast_fp16")];
            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_52_cast_fp16 = reshape(shape = concat_52, x = tile_26_cast_fp16)[name = string("reshape_52_cast_fp16")];
            tensor<int32, [5]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = reshape_52_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [4, 1, 32, 256]> reshape_53_cast_fp16 = reshape(shape = concat_53, x = transpose_53_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_27_reps_0 = const()[name = string("tile_27_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = v_53_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [4, 1, 32, 256]> tile_27_cast_fp16 = tile(reps = tile_27_reps_0, x = transpose_54_cast_fp16)[name = string("tile_27_cast_fp16")];
            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_54_cast_fp16 = reshape(shape = concat_54, x = tile_27_cast_fp16)[name = string("reshape_54_cast_fp16")];
            tensor<int32, [5]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = reshape_54_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [4, 1, 32, 256]> reshape_55_cast_fp16 = reshape(shape = concat_55, x = transpose_55_cast_fp16)[name = string("reshape_55_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6643_transpose_x_0 = const()[name = string("op_6643_transpose_x_0"), val = bool(false)];
            bool var_6643_transpose_y_0 = const()[name = string("op_6643_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 4, 32, 32]> var_6643_cast_fp16 = matmul(transpose_x = var_6643_transpose_x_0, transpose_y = var_6643_transpose_y_0, x = q_55_cast_fp16, y = transpose_109_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_6645_cast_fp16 = add(x = var_6643_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6646 = const()[name = string("op_6646"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_217_cast_fp16 = softmax(axis = var_6646, x = var_6645_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6649_transpose_x_0 = const()[name = string("op_6649_transpose_x_0"), val = bool(false)];
            bool var_6649_transpose_y_0 = const()[name = string("op_6649_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = reshape_55_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 4, 32, 256]> var_6649_cast_fp16 = matmul(transpose_x = var_6649_transpose_x_0, transpose_y = var_6649_transpose_y_0, x = w_217_cast_fp16, y = v_55_cast_fp16)[name = string("op_6649_cast_fp16")];
            tensor<int32, [4]> var_6652_perm_0 = const()[name = string("op_6652_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6656 = const()[name = string("op_6656"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_6652_cast_fp16 = transpose(perm = var_6652_perm_0, x = var_6649_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 32, 1024]> input_159_cast_fp16 = reshape(shape = var_6656, x = var_6652_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77977216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78501568))))[name = string("tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6661_cast_fp16 = abs(x = linear_121_cast_fp16)[name = string("op_6661_cast_fp16")];
            bool var_6664 = const()[name = string("op_6664"), val = bool(true)];
            tensor<int32, [1]> var_6665_axes_0 = const()[name = string("op_6665_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6665_cast_fp16 = reduce_max(axes = var_6665_axes_0, keep_dims = var_6664, x = var_6661_cast_fp16)[name = string("op_6665_cast_fp16")];
            fp16 var_6666_to_fp16 = const()[name = string("op_6666_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_219_cast_fp16 = maximum(x = var_6665_cast_fp16, y = var_6666_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_219_cast_fp16 = real_div(x = linear_121_cast_fp16, y = s_219_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6669_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6669_cast_fp16")];
            tensor<int32, [1]> var_6674_axes_0 = const()[name = string("op_6674_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6674_keep_dims_0 = const()[name = string("op_6674_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6674_cast_fp16 = reduce_mean(axes = var_6674_axes_0, keep_dims = var_6674_keep_dims_0, x = var_6669_cast_fp16)[name = string("op_6674_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6675_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6675_cast_fp16")];
            fp32 var_6676_epsilon_0 = const()[name = string("op_6676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6676_cast_fp16 = inverse(epsilon = var_6676_epsilon_0, x = var_6675_cast_fp16)[name = string("op_6676_cast_fp16")];
            fp16 var_6677_to_fp16 = const()[name = string("op_6677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6678_cast_fp16 = mul(x = var_6676_cast_fp16, y = var_6677_to_fp16)[name = string("op_6678_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6680_cast_fp16 = add(x = var_6674_cast_fp16, y = var_6678_cast_fp16)[name = string("op_6680_cast_fp16")];
            fp32 var_6681_epsilon_0 = const()[name = string("op_6681_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6681_cast_fp16 = rsqrt(epsilon = var_6681_epsilon_0, x = var_6680_cast_fp16)[name = string("op_6681_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_191_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6681_cast_fp16)[name = string("out_191_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78502656)))];
            tensor<fp16, [1, 32, 512]> var_6683_cast_fp16 = mul(x = out_191_cast_fp16, y = tm_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6683_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_245_cast_fp16 = add(x = x_235_cast_fp16, y = var_6683_cast_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6686_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_6686_cast_fp16")];
            bool var_6689 = const()[name = string("op_6689"), val = bool(true)];
            tensor<int32, [1]> var_6690_axes_0 = const()[name = string("op_6690_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6690_cast_fp16 = reduce_max(axes = var_6690_axes_0, keep_dims = var_6689, x = var_6686_cast_fp16)[name = string("op_6690_cast_fp16")];
            fp16 var_6691_to_fp16 = const()[name = string("op_6691_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_221_cast_fp16 = maximum(x = var_6690_cast_fp16, y = var_6691_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_221_cast_fp16 = real_div(x = x_245_cast_fp16, y = s_221_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6694_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6694_cast_fp16")];
            tensor<int32, [1]> var_6699_axes_0 = const()[name = string("op_6699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6699_keep_dims_0 = const()[name = string("op_6699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6699_cast_fp16 = reduce_mean(axes = var_6699_axes_0, keep_dims = var_6699_keep_dims_0, x = var_6694_cast_fp16)[name = string("op_6699_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6700_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6700_cast_fp16")];
            fp32 var_6701_epsilon_0 = const()[name = string("op_6701_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6701_cast_fp16 = inverse(epsilon = var_6701_epsilon_0, x = var_6700_cast_fp16)[name = string("op_6701_cast_fp16")];
            fp16 var_6702_to_fp16 = const()[name = string("op_6702_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6703_cast_fp16 = mul(x = var_6701_cast_fp16, y = var_6702_to_fp16)[name = string("op_6703_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6705_cast_fp16 = add(x = var_6699_cast_fp16, y = var_6703_cast_fp16)[name = string("op_6705_cast_fp16")];
            fp32 var_6706_epsilon_0 = const()[name = string("op_6706_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6706_cast_fp16 = rsqrt(epsilon = var_6706_epsilon_0, x = var_6705_cast_fp16)[name = string("op_6706_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_193_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6706_cast_fp16)[name = string("out_193_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78503744)))];
            tensor<fp16, [1, 32, 512]> input_161_cast_fp16 = mul(x = out_193_cast_fp16, y = tm_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78504832))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79553472))))[name = string("tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_122_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_122_cast_fp16")];
            string var_6713_mode_0 = const()[name = string("op_6713_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_6713_cast_fp16 = gelu(mode = var_6713_mode_0, x = linear_122_cast_fp16)[name = string("op_6713_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79557632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80606272))))[name = string("tm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_123_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_163_cast_fp16 = mul(x = var_6713_cast_fp16, y = linear_123_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80610432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81659072))))[name = string("tm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6721_cast_fp16 = abs(x = linear_124_cast_fp16)[name = string("op_6721_cast_fp16")];
            bool var_6724 = const()[name = string("op_6724"), val = bool(true)];
            tensor<int32, [1]> var_6725_axes_0 = const()[name = string("op_6725_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6725_cast_fp16 = reduce_max(axes = var_6725_axes_0, keep_dims = var_6724, x = var_6721_cast_fp16)[name = string("op_6725_cast_fp16")];
            fp16 var_6726_to_fp16 = const()[name = string("op_6726_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_223_cast_fp16 = maximum(x = var_6725_cast_fp16, y = var_6726_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_223_cast_fp16 = real_div(x = linear_124_cast_fp16, y = s_223_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6729_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [1]> var_6734_axes_0 = const()[name = string("op_6734_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6734_keep_dims_0 = const()[name = string("op_6734_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6734_cast_fp16 = reduce_mean(axes = var_6734_axes_0, keep_dims = var_6734_keep_dims_0, x = var_6729_cast_fp16)[name = string("op_6734_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6735_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6735_cast_fp16")];
            fp32 var_6736_epsilon_0 = const()[name = string("op_6736_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6736_cast_fp16 = inverse(epsilon = var_6736_epsilon_0, x = var_6735_cast_fp16)[name = string("op_6736_cast_fp16")];
            fp16 var_6737_to_fp16 = const()[name = string("op_6737_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6738_cast_fp16 = mul(x = var_6736_cast_fp16, y = var_6737_to_fp16)[name = string("op_6738_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6740_cast_fp16 = add(x = var_6734_cast_fp16, y = var_6738_cast_fp16)[name = string("op_6740_cast_fp16")];
            fp32 var_6741_epsilon_0 = const()[name = string("op_6741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6741_cast_fp16 = rsqrt(epsilon = var_6741_epsilon_0, x = var_6740_cast_fp16)[name = string("op_6741_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_195_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6741_cast_fp16)[name = string("out_195_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81660160)))];
            tensor<fp16, [1, 32, 512]> var_6743_cast_fp16 = mul(x = out_195_cast_fp16, y = tm_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6743_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_165_cast_fp16 = add(x = x_245_cast_fp16, y = var_6743_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81661248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81923456))))[name = string("tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_125_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_125_cast_fp16")];
            string x_249_mode_0 = const()[name = string("x_249_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_249_cast_fp16 = gelu(mode = x_249_mode_0, x = linear_125_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<int32, [4]> var_6763_begin_0 = const()[name = string("op_6763_begin_0"), val = tensor<int32, [4]>([0, 0, 13, 0])];
            tensor<int32, [4]> var_6763_end_0 = const()[name = string("op_6763_end_0"), val = tensor<int32, [4]>([1, 32, 14, 512])];
            tensor<bool, [4]> var_6763_end_mask_0 = const()[name = string("op_6763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6763_squeeze_mask_0 = const()[name = string("op_6763_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_6763_cast_fp16 = slice_by_index(begin = var_6763_begin_0, end = var_6763_end_0, end_mask = var_6763_end_mask_0, squeeze_mask = var_6763_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_167_cast_fp16 = mul(x = x_249_cast_fp16, y = var_6763_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81924544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82186752))))[name = string("tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6773_cast_fp16 = abs(x = linear_126_cast_fp16)[name = string("op_6773_cast_fp16")];
            bool var_6776 = const()[name = string("op_6776"), val = bool(true)];
            tensor<int32, [1]> var_6777_axes_0 = const()[name = string("op_6777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6777_cast_fp16 = reduce_max(axes = var_6777_axes_0, keep_dims = var_6776, x = var_6773_cast_fp16)[name = string("op_6777_cast_fp16")];
            fp16 var_6778_to_fp16 = const()[name = string("op_6778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_225_cast_fp16 = maximum(x = var_6777_cast_fp16, y = var_6778_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_225_cast_fp16 = real_div(x = linear_126_cast_fp16, y = s_225_cast_fp16)[name = string("y_225_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6781_cast_fp16 = mul(x = y_225_cast_fp16, y = y_225_cast_fp16)[name = string("op_6781_cast_fp16")];
            tensor<int32, [1]> var_6786_axes_0 = const()[name = string("op_6786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6786_keep_dims_0 = const()[name = string("op_6786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6786_cast_fp16 = reduce_mean(axes = var_6786_axes_0, keep_dims = var_6786_keep_dims_0, x = var_6781_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6787_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6787_cast_fp16")];
            fp32 var_6788_epsilon_0 = const()[name = string("op_6788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6788_cast_fp16 = inverse(epsilon = var_6788_epsilon_0, x = var_6787_cast_fp16)[name = string("op_6788_cast_fp16")];
            fp16 var_6789_to_fp16 = const()[name = string("op_6789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6790_cast_fp16 = mul(x = var_6788_cast_fp16, y = var_6789_to_fp16)[name = string("op_6790_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6792_cast_fp16 = add(x = var_6786_cast_fp16, y = var_6790_cast_fp16)[name = string("op_6792_cast_fp16")];
            fp32 var_6793_epsilon_0 = const()[name = string("op_6793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6793_cast_fp16 = rsqrt(epsilon = var_6793_epsilon_0, x = var_6792_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_197_cast_fp16 = mul(x = y_225_cast_fp16, y = var_6793_cast_fp16)[name = string("out_197_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82187840)))];
            tensor<fp16, [1, 32, 512]> var_6795_cast_fp16 = mul(x = out_197_cast_fp16, y = tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6795_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_27_cast_fp16 = add(x = input_165_cast_fp16, y = var_6795_cast_fp16)[name = string("h_27_cast_fp16")];
            tensor<fp16, [1]> tm_layers_13_layer_scalar_to_fp16 = const()[name = string("tm_layers_13_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9cp-1])];
            tensor<fp16, [1, 32, 512]> x_253_cast_fp16 = mul(x = h_27_cast_fp16, y = tm_layers_13_layer_scalar_to_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6799_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6799_cast_fp16")];
            bool var_6802 = const()[name = string("op_6802"), val = bool(true)];
            tensor<int32, [1]> var_6803_axes_0 = const()[name = string("op_6803_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6803_cast_fp16 = reduce_max(axes = var_6803_axes_0, keep_dims = var_6802, x = var_6799_cast_fp16)[name = string("op_6803_cast_fp16")];
            fp16 var_6804_to_fp16 = const()[name = string("op_6804_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_227_cast_fp16 = maximum(x = var_6803_cast_fp16, y = var_6804_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_227_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_227_cast_fp16)[name = string("y_227_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6807_cast_fp16 = mul(x = y_227_cast_fp16, y = y_227_cast_fp16)[name = string("op_6807_cast_fp16")];
            tensor<int32, [1]> var_6812_axes_0 = const()[name = string("op_6812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6812_keep_dims_0 = const()[name = string("op_6812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_6812_cast_fp16 = reduce_mean(axes = var_6812_axes_0, keep_dims = var_6812_keep_dims_0, x = var_6807_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6813_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6813_cast_fp16")];
            fp32 var_6814_epsilon_0 = const()[name = string("op_6814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_6814_cast_fp16 = inverse(epsilon = var_6814_epsilon_0, x = var_6813_cast_fp16)[name = string("op_6814_cast_fp16")];
            fp16 var_6815_to_fp16 = const()[name = string("op_6815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_6816_cast_fp16 = mul(x = var_6814_cast_fp16, y = var_6815_to_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_6818_cast_fp16 = add(x = var_6812_cast_fp16, y = var_6816_cast_fp16)[name = string("op_6818_cast_fp16")];
            fp32 var_6819_epsilon_0 = const()[name = string("op_6819_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_6819_cast_fp16 = rsqrt(epsilon = var_6819_epsilon_0, x = var_6818_cast_fp16)[name = string("op_6819_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_199_cast_fp16 = mul(x = y_227_cast_fp16, y = var_6819_cast_fp16)[name = string("out_199_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82188928)))];
            tensor<fp16, [1, 32, 512]> input_169_cast_fp16 = mul(x = out_199_cast_fp16, y = tm_layers_14_input_layernorm_weight_to_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82190016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82714368))))[name = string("tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6829 = const()[name = string("op_6829"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_6830_cast_fp16 = reshape(shape = var_6829, x = linear_127_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82716480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82978688))))[name = string("tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6841 = const()[name = string("op_6841"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_6842_cast_fp16 = reshape(shape = var_6841, x = linear_128_cast_fp16)[name = string("op_6842_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82979776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83241984))))[name = string("tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_6853 = const()[name = string("op_6853"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_6854_cast_fp16 = reshape(shape = var_6853, x = linear_129_cast_fp16)[name = string("op_6854_cast_fp16")];
            tensor<int32, [4]> x_259_perm_0 = const()[name = string("x_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_6830_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 4, 32, 256]> var_6858_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6858_cast_fp16")];
            bool var_6861 = const()[name = string("op_6861"), val = bool(true)];
            tensor<int32, [1]> var_6862_axes_0 = const()[name = string("op_6862_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_6862_cast_fp16 = reduce_max(axes = var_6862_axes_0, keep_dims = var_6861, x = var_6858_cast_fp16)[name = string("op_6862_cast_fp16")];
            fp16 var_6863_to_fp16 = const()[name = string("op_6863_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_229_cast_fp16 = maximum(x = var_6862_cast_fp16, y = var_6863_to_fp16)[name = string("s_229_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_229_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_229_cast_fp16)[name = string("y_229_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6866_cast_fp16 = mul(x = y_229_cast_fp16, y = y_229_cast_fp16)[name = string("op_6866_cast_fp16")];
            tensor<int32, [1]> var_6871_axes_0 = const()[name = string("op_6871_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6871_keep_dims_0 = const()[name = string("op_6871_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_6871_cast_fp16 = reduce_mean(axes = var_6871_axes_0, keep_dims = var_6871_keep_dims_0, x = var_6866_cast_fp16)[name = string("op_6871_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_6872_cast_fp16 = mul(x = s_229_cast_fp16, y = s_229_cast_fp16)[name = string("op_6872_cast_fp16")];
            fp32 var_6873_epsilon_0 = const()[name = string("op_6873_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_6873_cast_fp16 = inverse(epsilon = var_6873_epsilon_0, x = var_6872_cast_fp16)[name = string("op_6873_cast_fp16")];
            fp16 var_6874_to_fp16 = const()[name = string("op_6874_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_6875_cast_fp16 = mul(x = var_6873_cast_fp16, y = var_6874_to_fp16)[name = string("op_6875_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_6877_cast_fp16 = add(x = var_6871_cast_fp16, y = var_6875_cast_fp16)[name = string("op_6877_cast_fp16")];
            fp32 var_6878_epsilon_0 = const()[name = string("op_6878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_6878_cast_fp16 = rsqrt(epsilon = var_6878_epsilon_0, x = var_6877_cast_fp16)[name = string("op_6878_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_201_cast_fp16 = mul(x = y_229_cast_fp16, y = var_6878_cast_fp16)[name = string("out_201_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243072)))];
            tensor<fp16, [1, 4, 32, 256]> q_57_cast_fp16 = mul(x = out_201_cast_fp16, y = tm_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("q_57_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_6842_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, 32, 256]> var_6881_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6884 = const()[name = string("op_6884"), val = bool(true)];
            tensor<int32, [1]> var_6885_axes_0 = const()[name = string("op_6885_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_6885_cast_fp16 = reduce_max(axes = var_6885_axes_0, keep_dims = var_6884, x = var_6881_cast_fp16)[name = string("op_6885_cast_fp16")];
            fp16 var_6886_to_fp16 = const()[name = string("op_6886_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_231_cast_fp16 = maximum(x = var_6885_cast_fp16, y = var_6886_to_fp16)[name = string("s_231_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_231_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_231_cast_fp16)[name = string("y_231_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6889_cast_fp16 = mul(x = y_231_cast_fp16, y = y_231_cast_fp16)[name = string("op_6889_cast_fp16")];
            tensor<int32, [1]> var_6894_axes_0 = const()[name = string("op_6894_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6894_keep_dims_0 = const()[name = string("op_6894_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_6894_cast_fp16 = reduce_mean(axes = var_6894_axes_0, keep_dims = var_6894_keep_dims_0, x = var_6889_cast_fp16)[name = string("op_6894_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6895_cast_fp16 = mul(x = s_231_cast_fp16, y = s_231_cast_fp16)[name = string("op_6895_cast_fp16")];
            fp32 var_6896_epsilon_0 = const()[name = string("op_6896_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_6896_cast_fp16 = inverse(epsilon = var_6896_epsilon_0, x = var_6895_cast_fp16)[name = string("op_6896_cast_fp16")];
            fp16 var_6897_to_fp16 = const()[name = string("op_6897_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_6898_cast_fp16 = mul(x = var_6896_cast_fp16, y = var_6897_to_fp16)[name = string("op_6898_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6900_cast_fp16 = add(x = var_6894_cast_fp16, y = var_6898_cast_fp16)[name = string("op_6900_cast_fp16")];
            fp32 var_6901_epsilon_0 = const()[name = string("op_6901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_6901_cast_fp16 = rsqrt(epsilon = var_6901_epsilon_0, x = var_6900_cast_fp16)[name = string("op_6901_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_203_cast_fp16 = mul(x = y_231_cast_fp16, y = var_6901_cast_fp16)[name = string("out_203_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243648)))];
            tensor<fp16, [1, 2, 32, 256]> k_85_cast_fp16 = mul(x = out_203_cast_fp16, y = tm_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("k_85_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_259_cast_fp16 = transpose(perm = x_259_perm_0, x = var_6854_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 2, 32, 256]> var_6904_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6904_cast_fp16")];
            bool var_6907 = const()[name = string("op_6907"), val = bool(true)];
            tensor<int32, [1]> var_6908_axes_0 = const()[name = string("op_6908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_6908_cast_fp16 = reduce_max(axes = var_6908_axes_0, keep_dims = var_6907, x = var_6904_cast_fp16)[name = string("op_6908_cast_fp16")];
            fp16 var_6909_to_fp16 = const()[name = string("op_6909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_233_cast_fp16 = maximum(x = var_6908_cast_fp16, y = var_6909_to_fp16)[name = string("s_233_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_233_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_233_cast_fp16)[name = string("y_233_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6912_cast_fp16 = mul(x = y_233_cast_fp16, y = y_233_cast_fp16)[name = string("op_6912_cast_fp16")];
            tensor<int32, [1]> var_6917_axes_0 = const()[name = string("op_6917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6917_keep_dims_0 = const()[name = string("op_6917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_6917_cast_fp16 = reduce_mean(axes = var_6917_axes_0, keep_dims = var_6917_keep_dims_0, x = var_6912_cast_fp16)[name = string("op_6917_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6918_cast_fp16 = mul(x = s_233_cast_fp16, y = s_233_cast_fp16)[name = string("op_6918_cast_fp16")];
            fp32 var_6919_epsilon_0 = const()[name = string("op_6919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_6919_cast_fp16 = inverse(epsilon = var_6919_epsilon_0, x = var_6918_cast_fp16)[name = string("op_6919_cast_fp16")];
            fp16 var_6920_to_fp16 = const()[name = string("op_6920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_6921_cast_fp16 = mul(x = var_6919_cast_fp16, y = var_6920_to_fp16)[name = string("op_6921_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_6923_cast_fp16 = add(x = var_6917_cast_fp16, y = var_6921_cast_fp16)[name = string("op_6923_cast_fp16")];
            fp32 var_6924_epsilon_0 = const()[name = string("op_6924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_6924_cast_fp16 = rsqrt(epsilon = var_6924_epsilon_0, x = var_6923_cast_fp16)[name = string("op_6924_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_57_cast_fp16 = mul(x = y_233_cast_fp16, y = var_6924_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6926_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6926_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = q_57_cast_fp16)[name = string("op_6931_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_6932_cast_fp16 = mul(x = var_6931_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_6932_cast_fp16")];
            tensor<int32, [4]> var_6937_begin_0 = const()[name = string("op_6937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6937_end_0 = const()[name = string("op_6937_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_6937_end_mask_0 = const()[name = string("op_6937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_6937_cast_fp16 = slice_by_index(begin = var_6937_begin_0, end = var_6937_end_0, end_mask = var_6937_end_mask_0, x = q_57_cast_fp16)[name = string("op_6937_cast_fp16")];
            int32 var_6939 = const()[name = string("op_6939"), val = int32(-1)];
            bool var_6940_interleave_0 = const()[name = string("op_6940_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_6940_cast_fp16 = concat(axis = var_6939, interleave = var_6940_interleave_0, values = (var_6932_cast_fp16, var_6937_cast_fp16))[name = string("op_6940_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_6941_cast_fp16 = mul(x = var_6940_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6941_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_59_cast_fp16 = add(x = var_6926_cast_fp16, y = var_6941_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6944_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6944_cast_fp16")];
            tensor<int32, [4]> var_6949_begin_0 = const()[name = string("op_6949_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6949_end_0 = const()[name = string("op_6949_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_6949_end_mask_0 = const()[name = string("op_6949_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_6949_cast_fp16 = slice_by_index(begin = var_6949_begin_0, end = var_6949_end_0, end_mask = var_6949_end_mask_0, x = k_85_cast_fp16)[name = string("op_6949_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_6950_cast_fp16 = mul(x = var_6949_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_6950_cast_fp16")];
            tensor<int32, [4]> var_6955_begin_0 = const()[name = string("op_6955_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6955_end_0 = const()[name = string("op_6955_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_6955_end_mask_0 = const()[name = string("op_6955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_6955_cast_fp16 = slice_by_index(begin = var_6955_begin_0, end = var_6955_end_0, end_mask = var_6955_end_mask_0, x = k_85_cast_fp16)[name = string("op_6955_cast_fp16")];
            int32 var_6957 = const()[name = string("op_6957"), val = int32(-1)];
            bool var_6958_interleave_0 = const()[name = string("op_6958_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_6958_cast_fp16 = concat(axis = var_6957, interleave = var_6958_interleave_0, values = (var_6950_cast_fp16, var_6955_cast_fp16))[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_6959_cast_fp16 = mul(x = var_6958_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_87_cast_fp16 = add(x = var_6944_cast_fp16, y = var_6959_cast_fp16)[name = string("k_87_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_28_reps_0 = const()[name = string("tile_28_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = k_87_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 1, 32, 256]> tile_28_cast_fp16 = tile(reps = tile_28_reps_0, x = transpose_56_cast_fp16)[name = string("tile_28_cast_fp16")];
            tensor<int32, [5]> concat_56 = const()[name = string("concat_56"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_56_cast_fp16 = reshape(shape = concat_56, x = tile_28_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [5]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_57 = const()[name = string("concat_57"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 1, 32, 256]> reshape_57_cast_fp16 = reshape(shape = concat_57, x = transpose_57_cast_fp16)[name = string("reshape_57_cast_fp16")];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_29_reps_0 = const()[name = string("tile_29_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = v_57_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 1, 32, 256]> tile_29_cast_fp16 = tile(reps = tile_29_reps_0, x = transpose_58_cast_fp16)[name = string("tile_29_cast_fp16")];
            tensor<int32, [5]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_58_cast_fp16 = reshape(shape = concat_58, x = tile_29_cast_fp16)[name = string("reshape_58_cast_fp16")];
            tensor<int32, [5]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = reshape_58_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 1, 32, 256]> reshape_59_cast_fp16 = reshape(shape = concat_59, x = transpose_59_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6973_transpose_x_0 = const()[name = string("op_6973_transpose_x_0"), val = bool(false)];
            bool var_6973_transpose_y_0 = const()[name = string("op_6973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_110_cast_fp16 = transpose(perm = transpose_110_perm_0, x = reshape_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 4, 32, 32]> var_6973_cast_fp16 = matmul(transpose_x = var_6973_transpose_x_0, transpose_y = var_6973_transpose_y_0, x = q_59_cast_fp16, y = transpose_110_cast_fp16)[name = string("op_6973_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_6975_cast_fp16 = add(x = var_6973_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6976 = const()[name = string("op_6976"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_233_cast_fp16 = softmax(axis = var_6976, x = var_6975_cast_fp16)[name = string("w_233_cast_fp16")];
            bool var_6979_transpose_x_0 = const()[name = string("op_6979_transpose_x_0"), val = bool(false)];
            bool var_6979_transpose_y_0 = const()[name = string("op_6979_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 4, 32, 256]> var_6979_cast_fp16 = matmul(transpose_x = var_6979_transpose_x_0, transpose_y = var_6979_transpose_y_0, x = w_233_cast_fp16, y = v_59_cast_fp16)[name = string("op_6979_cast_fp16")];
            tensor<int32, [4]> var_6982_perm_0 = const()[name = string("op_6982_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6986 = const()[name = string("op_6986"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_6982_cast_fp16 = transpose(perm = var_6982_perm_0, x = var_6979_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 32, 1024]> input_171_cast_fp16 = reshape(shape = var_6986, x = var_6982_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83244224))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83768576))))[name = string("tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6991_cast_fp16 = abs(x = linear_130_cast_fp16)[name = string("op_6991_cast_fp16")];
            bool var_6994 = const()[name = string("op_6994"), val = bool(true)];
            tensor<int32, [1]> var_6995_axes_0 = const()[name = string("op_6995_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_6995_cast_fp16 = reduce_max(axes = var_6995_axes_0, keep_dims = var_6994, x = var_6991_cast_fp16)[name = string("op_6995_cast_fp16")];
            fp16 var_6996_to_fp16 = const()[name = string("op_6996_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_235_cast_fp16 = maximum(x = var_6995_cast_fp16, y = var_6996_to_fp16)[name = string("s_235_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_235_cast_fp16 = real_div(x = linear_130_cast_fp16, y = s_235_cast_fp16)[name = string("y_235_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_6999_cast_fp16 = mul(x = y_235_cast_fp16, y = y_235_cast_fp16)[name = string("op_6999_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_6999_cast_fp16)[name = string("op_7004_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7005_cast_fp16 = mul(x = s_235_cast_fp16, y = s_235_cast_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7010_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7010_cast_fp16")];
            fp32 var_7011_epsilon_0 = const()[name = string("op_7011_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7011_cast_fp16 = rsqrt(epsilon = var_7011_epsilon_0, x = var_7010_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_205_cast_fp16 = mul(x = y_235_cast_fp16, y = var_7011_cast_fp16)[name = string("out_205_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83769664)))];
            tensor<fp16, [1, 32, 512]> var_7013_cast_fp16 = mul(x = out_205_cast_fp16, y = tm_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_7013_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = var_7013_cast_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7016_cast_fp16 = abs(x = x_263_cast_fp16)[name = string("op_7016_cast_fp16")];
            bool var_7019 = const()[name = string("op_7019"), val = bool(true)];
            tensor<int32, [1]> var_7020_axes_0 = const()[name = string("op_7020_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7020_cast_fp16 = reduce_max(axes = var_7020_axes_0, keep_dims = var_7019, x = var_7016_cast_fp16)[name = string("op_7020_cast_fp16")];
            fp16 var_7021_to_fp16 = const()[name = string("op_7021_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_237_cast_fp16 = maximum(x = var_7020_cast_fp16, y = var_7021_to_fp16)[name = string("s_237_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_237_cast_fp16 = real_div(x = x_263_cast_fp16, y = s_237_cast_fp16)[name = string("y_237_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7024_cast_fp16 = mul(x = y_237_cast_fp16, y = y_237_cast_fp16)[name = string("op_7024_cast_fp16")];
            tensor<int32, [1]> var_7029_axes_0 = const()[name = string("op_7029_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7029_keep_dims_0 = const()[name = string("op_7029_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7029_cast_fp16 = reduce_mean(axes = var_7029_axes_0, keep_dims = var_7029_keep_dims_0, x = var_7024_cast_fp16)[name = string("op_7029_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7030_cast_fp16 = mul(x = s_237_cast_fp16, y = s_237_cast_fp16)[name = string("op_7030_cast_fp16")];
            fp32 var_7031_epsilon_0 = const()[name = string("op_7031_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7031_cast_fp16 = inverse(epsilon = var_7031_epsilon_0, x = var_7030_cast_fp16)[name = string("op_7031_cast_fp16")];
            fp16 var_7032_to_fp16 = const()[name = string("op_7032_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7033_cast_fp16 = mul(x = var_7031_cast_fp16, y = var_7032_to_fp16)[name = string("op_7033_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7035_cast_fp16 = add(x = var_7029_cast_fp16, y = var_7033_cast_fp16)[name = string("op_7035_cast_fp16")];
            fp32 var_7036_epsilon_0 = const()[name = string("op_7036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7036_cast_fp16 = rsqrt(epsilon = var_7036_epsilon_0, x = var_7035_cast_fp16)[name = string("op_7036_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_207_cast_fp16 = mul(x = y_237_cast_fp16, y = var_7036_cast_fp16)[name = string("out_207_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83770752)))];
            tensor<fp16, [1, 32, 512]> input_173_cast_fp16 = mul(x = out_207_cast_fp16, y = tm_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83771840))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84820480))))[name = string("tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_131_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_131_cast_fp16")];
            string var_7043_mode_0 = const()[name = string("op_7043_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_7043_cast_fp16 = gelu(mode = var_7043_mode_0, x = linear_131_cast_fp16)[name = string("op_7043_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84824640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85873280))))[name = string("tm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_132_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_175_cast_fp16 = mul(x = var_7043_cast_fp16, y = linear_132_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85877440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86926080))))[name = string("tm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7051_cast_fp16 = abs(x = linear_133_cast_fp16)[name = string("op_7051_cast_fp16")];
            bool var_7054 = const()[name = string("op_7054"), val = bool(true)];
            tensor<int32, [1]> var_7055_axes_0 = const()[name = string("op_7055_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7055_cast_fp16 = reduce_max(axes = var_7055_axes_0, keep_dims = var_7054, x = var_7051_cast_fp16)[name = string("op_7055_cast_fp16")];
            fp16 var_7056_to_fp16 = const()[name = string("op_7056_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_239_cast_fp16 = maximum(x = var_7055_cast_fp16, y = var_7056_to_fp16)[name = string("s_239_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_239_cast_fp16 = real_div(x = linear_133_cast_fp16, y = s_239_cast_fp16)[name = string("y_239_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7059_cast_fp16 = mul(x = y_239_cast_fp16, y = y_239_cast_fp16)[name = string("op_7059_cast_fp16")];
            tensor<int32, [1]> var_7064_axes_0 = const()[name = string("op_7064_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7064_keep_dims_0 = const()[name = string("op_7064_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7064_cast_fp16 = reduce_mean(axes = var_7064_axes_0, keep_dims = var_7064_keep_dims_0, x = var_7059_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7065_cast_fp16 = mul(x = s_239_cast_fp16, y = s_239_cast_fp16)[name = string("op_7065_cast_fp16")];
            fp32 var_7066_epsilon_0 = const()[name = string("op_7066_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7066_cast_fp16 = inverse(epsilon = var_7066_epsilon_0, x = var_7065_cast_fp16)[name = string("op_7066_cast_fp16")];
            fp16 var_7067_to_fp16 = const()[name = string("op_7067_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7068_cast_fp16 = mul(x = var_7066_cast_fp16, y = var_7067_to_fp16)[name = string("op_7068_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7070_cast_fp16 = add(x = var_7064_cast_fp16, y = var_7068_cast_fp16)[name = string("op_7070_cast_fp16")];
            fp32 var_7071_epsilon_0 = const()[name = string("op_7071_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7071_cast_fp16 = rsqrt(epsilon = var_7071_epsilon_0, x = var_7070_cast_fp16)[name = string("op_7071_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_209_cast_fp16 = mul(x = y_239_cast_fp16, y = var_7071_cast_fp16)[name = string("out_209_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86927168)))];
            tensor<fp16, [1, 32, 512]> var_7073_cast_fp16 = mul(x = out_209_cast_fp16, y = tm_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7073_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_177_cast_fp16 = add(x = x_263_cast_fp16, y = var_7073_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86928256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87190464))))[name = string("tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_134_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_134_cast_fp16")];
            string x_267_mode_0 = const()[name = string("x_267_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_267_cast_fp16 = gelu(mode = x_267_mode_0, x = linear_134_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<int32, [4]> var_7093_begin_0 = const()[name = string("op_7093_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
            tensor<int32, [4]> var_7093_end_0 = const()[name = string("op_7093_end_0"), val = tensor<int32, [4]>([1, 32, 15, 512])];
            tensor<bool, [4]> var_7093_end_mask_0 = const()[name = string("op_7093_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7093_squeeze_mask_0 = const()[name = string("op_7093_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_7093_cast_fp16 = slice_by_index(begin = var_7093_begin_0, end = var_7093_end_0, end_mask = var_7093_end_mask_0, squeeze_mask = var_7093_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7093_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_179_cast_fp16 = mul(x = x_267_cast_fp16, y = var_7093_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87191552))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87453760))))[name = string("tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7103_cast_fp16 = abs(x = linear_135_cast_fp16)[name = string("op_7103_cast_fp16")];
            bool var_7106 = const()[name = string("op_7106"), val = bool(true)];
            tensor<int32, [1]> var_7107_axes_0 = const()[name = string("op_7107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7107_cast_fp16 = reduce_max(axes = var_7107_axes_0, keep_dims = var_7106, x = var_7103_cast_fp16)[name = string("op_7107_cast_fp16")];
            fp16 var_7108_to_fp16 = const()[name = string("op_7108_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_241_cast_fp16 = maximum(x = var_7107_cast_fp16, y = var_7108_to_fp16)[name = string("s_241_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_241_cast_fp16 = real_div(x = linear_135_cast_fp16, y = s_241_cast_fp16)[name = string("y_241_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7111_cast_fp16 = mul(x = y_241_cast_fp16, y = y_241_cast_fp16)[name = string("op_7111_cast_fp16")];
            tensor<int32, [1]> var_7116_axes_0 = const()[name = string("op_7116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7116_keep_dims_0 = const()[name = string("op_7116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7116_cast_fp16 = reduce_mean(axes = var_7116_axes_0, keep_dims = var_7116_keep_dims_0, x = var_7111_cast_fp16)[name = string("op_7116_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7117_cast_fp16 = mul(x = s_241_cast_fp16, y = s_241_cast_fp16)[name = string("op_7117_cast_fp16")];
            fp32 var_7118_epsilon_0 = const()[name = string("op_7118_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7118_cast_fp16 = inverse(epsilon = var_7118_epsilon_0, x = var_7117_cast_fp16)[name = string("op_7118_cast_fp16")];
            fp16 var_7119_to_fp16 = const()[name = string("op_7119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7120_cast_fp16 = mul(x = var_7118_cast_fp16, y = var_7119_to_fp16)[name = string("op_7120_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7122_cast_fp16 = add(x = var_7116_cast_fp16, y = var_7120_cast_fp16)[name = string("op_7122_cast_fp16")];
            fp32 var_7123_epsilon_0 = const()[name = string("op_7123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7123_cast_fp16 = rsqrt(epsilon = var_7123_epsilon_0, x = var_7122_cast_fp16)[name = string("op_7123_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_211_cast_fp16 = mul(x = y_241_cast_fp16, y = var_7123_cast_fp16)[name = string("out_211_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87454848)))];
            tensor<fp16, [1, 32, 512]> var_7125_cast_fp16 = mul(x = out_211_cast_fp16, y = tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_29_cast_fp16 = add(x = input_177_cast_fp16, y = var_7125_cast_fp16)[name = string("h_29_cast_fp16")];
            tensor<fp16, [1]> tm_layers_14_layer_scalar_to_fp16 = const()[name = string("tm_layers_14_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1p-1])];
            tensor<fp16, [1, 32, 512]> x_271_cast_fp16 = mul(x = h_29_cast_fp16, y = tm_layers_14_layer_scalar_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7129_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_7129_cast_fp16")];
            bool var_7132 = const()[name = string("op_7132"), val = bool(true)];
            tensor<int32, [1]> var_7133_axes_0 = const()[name = string("op_7133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7133_cast_fp16 = reduce_max(axes = var_7133_axes_0, keep_dims = var_7132, x = var_7129_cast_fp16)[name = string("op_7133_cast_fp16")];
            fp16 var_7134_to_fp16 = const()[name = string("op_7134_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_243_cast_fp16 = maximum(x = var_7133_cast_fp16, y = var_7134_to_fp16)[name = string("s_243_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_243_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_243_cast_fp16)[name = string("y_243_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7137_cast_fp16 = mul(x = y_243_cast_fp16, y = y_243_cast_fp16)[name = string("op_7137_cast_fp16")];
            tensor<int32, [1]> var_7142_axes_0 = const()[name = string("op_7142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7142_keep_dims_0 = const()[name = string("op_7142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7142_cast_fp16 = reduce_mean(axes = var_7142_axes_0, keep_dims = var_7142_keep_dims_0, x = var_7137_cast_fp16)[name = string("op_7142_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7143_cast_fp16 = mul(x = s_243_cast_fp16, y = s_243_cast_fp16)[name = string("op_7143_cast_fp16")];
            fp32 var_7144_epsilon_0 = const()[name = string("op_7144_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7144_cast_fp16 = inverse(epsilon = var_7144_epsilon_0, x = var_7143_cast_fp16)[name = string("op_7144_cast_fp16")];
            fp16 var_7145_to_fp16 = const()[name = string("op_7145_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7146_cast_fp16 = mul(x = var_7144_cast_fp16, y = var_7145_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7148_cast_fp16 = add(x = var_7142_cast_fp16, y = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp32 var_7149_epsilon_0 = const()[name = string("op_7149_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7149_cast_fp16 = rsqrt(epsilon = var_7149_epsilon_0, x = var_7148_cast_fp16)[name = string("op_7149_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_213_cast_fp16 = mul(x = y_243_cast_fp16, y = var_7149_cast_fp16)[name = string("out_213_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87455936)))];
            tensor<fp16, [1, 32, 512]> input_181_cast_fp16 = mul(x = out_213_cast_fp16, y = tm_layers_15_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87457024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87981376))))[name = string("tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<int32, [4]> var_7159 = const()[name = string("op_7159"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_7160_cast_fp16 = reshape(shape = var_7159, x = linear_136_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<int32, [4]> x_273_perm_0 = const()[name = string("x_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87983488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88245696))))[name = string("tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<int32, [4]> var_7171 = const()[name = string("op_7171"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_7172_cast_fp16 = reshape(shape = var_7171, x = linear_137_cast_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [4]> x_275_perm_0 = const()[name = string("x_275_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88246784))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88508992))))[name = string("tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_7183 = const()[name = string("op_7183"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_7184_cast_fp16 = reshape(shape = var_7183, x = linear_138_cast_fp16)[name = string("op_7184_cast_fp16")];
            tensor<int32, [4]> x_277_perm_0 = const()[name = string("x_277_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_273_cast_fp16 = transpose(perm = x_273_perm_0, x = var_7160_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 4, 32, 256]> var_7188_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_7188_cast_fp16")];
            bool var_7191 = const()[name = string("op_7191"), val = bool(true)];
            tensor<int32, [1]> var_7192_axes_0 = const()[name = string("op_7192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_7192_cast_fp16 = reduce_max(axes = var_7192_axes_0, keep_dims = var_7191, x = var_7188_cast_fp16)[name = string("op_7192_cast_fp16")];
            fp16 var_7193_to_fp16 = const()[name = string("op_7193_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_245_cast_fp16 = maximum(x = var_7192_cast_fp16, y = var_7193_to_fp16)[name = string("s_245_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_245_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_245_cast_fp16)[name = string("y_245_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_7196_cast_fp16 = mul(x = y_245_cast_fp16, y = y_245_cast_fp16)[name = string("op_7196_cast_fp16")];
            tensor<int32, [1]> var_7201_axes_0 = const()[name = string("op_7201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7201_keep_dims_0 = const()[name = string("op_7201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_7201_cast_fp16 = reduce_mean(axes = var_7201_axes_0, keep_dims = var_7201_keep_dims_0, x = var_7196_cast_fp16)[name = string("op_7201_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_7202_cast_fp16 = mul(x = s_245_cast_fp16, y = s_245_cast_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_7203_cast_fp16 = inverse(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            fp16 var_7204_to_fp16 = const()[name = string("op_7204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_7205_cast_fp16 = mul(x = var_7203_cast_fp16, y = var_7204_to_fp16)[name = string("op_7205_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_7207_cast_fp16 = add(x = var_7201_cast_fp16, y = var_7205_cast_fp16)[name = string("op_7207_cast_fp16")];
            fp32 var_7208_epsilon_0 = const()[name = string("op_7208_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_7208_cast_fp16 = rsqrt(epsilon = var_7208_epsilon_0, x = var_7207_cast_fp16)[name = string("op_7208_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_215_cast_fp16 = mul(x = y_245_cast_fp16, y = var_7208_cast_fp16)[name = string("out_215_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_61_cast_fp16 = mul(x = out_215_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_61_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_275_cast_fp16 = transpose(perm = x_275_perm_0, x = var_7172_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 2, 32, 256]> var_7211_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_7211_cast_fp16")];
            bool var_7214 = const()[name = string("op_7214"), val = bool(true)];
            tensor<int32, [1]> var_7215_axes_0 = const()[name = string("op_7215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_7215_cast_fp16 = reduce_max(axes = var_7215_axes_0, keep_dims = var_7214, x = var_7211_cast_fp16)[name = string("op_7215_cast_fp16")];
            fp16 var_7216_to_fp16 = const()[name = string("op_7216_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_247_cast_fp16 = maximum(x = var_7215_cast_fp16, y = var_7216_to_fp16)[name = string("s_247_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_247_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_247_cast_fp16)[name = string("y_247_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_7219_cast_fp16 = mul(x = y_247_cast_fp16, y = y_247_cast_fp16)[name = string("op_7219_cast_fp16")];
            tensor<int32, [1]> var_7224_axes_0 = const()[name = string("op_7224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7224_keep_dims_0 = const()[name = string("op_7224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_7224_cast_fp16 = reduce_mean(axes = var_7224_axes_0, keep_dims = var_7224_keep_dims_0, x = var_7219_cast_fp16)[name = string("op_7224_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_7225_cast_fp16 = mul(x = s_247_cast_fp16, y = s_247_cast_fp16)[name = string("op_7225_cast_fp16")];
            fp32 var_7226_epsilon_0 = const()[name = string("op_7226_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_7226_cast_fp16 = inverse(epsilon = var_7226_epsilon_0, x = var_7225_cast_fp16)[name = string("op_7226_cast_fp16")];
            fp16 var_7227_to_fp16 = const()[name = string("op_7227_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_7228_cast_fp16 = mul(x = var_7226_cast_fp16, y = var_7227_to_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_7230_cast_fp16 = add(x = var_7224_cast_fp16, y = var_7228_cast_fp16)[name = string("op_7230_cast_fp16")];
            fp32 var_7231_epsilon_0 = const()[name = string("op_7231_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_7231_cast_fp16 = rsqrt(epsilon = var_7231_epsilon_0, x = var_7230_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_217_cast_fp16 = mul(x = y_247_cast_fp16, y = var_7231_cast_fp16)[name = string("out_217_cast_fp16")];
            tensor<fp16, [256]> tm_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510080)))];
            tensor<fp16, [1, 2, 32, 256]> k_91_cast_fp16 = mul(x = out_217_cast_fp16, y = tm_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_277_cast_fp16 = transpose(perm = x_277_perm_0, x = var_7184_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2, 32, 256]> var_7234_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_7234_cast_fp16")];
            bool var_7237 = const()[name = string("op_7237"), val = bool(true)];
            tensor<int32, [1]> var_7238_axes_0 = const()[name = string("op_7238_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_7238_cast_fp16 = reduce_max(axes = var_7238_axes_0, keep_dims = var_7237, x = var_7234_cast_fp16)[name = string("op_7238_cast_fp16")];
            fp16 var_7239_to_fp16 = const()[name = string("op_7239_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_249_cast_fp16 = maximum(x = var_7238_cast_fp16, y = var_7239_to_fp16)[name = string("s_249_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_249_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_249_cast_fp16)[name = string("y_249_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_7242_cast_fp16 = mul(x = y_249_cast_fp16, y = y_249_cast_fp16)[name = string("op_7242_cast_fp16")];
            tensor<int32, [1]> var_7247_axes_0 = const()[name = string("op_7247_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7247_keep_dims_0 = const()[name = string("op_7247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_7247_cast_fp16 = reduce_mean(axes = var_7247_axes_0, keep_dims = var_7247_keep_dims_0, x = var_7242_cast_fp16)[name = string("op_7247_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_7248_cast_fp16 = mul(x = s_249_cast_fp16, y = s_249_cast_fp16)[name = string("op_7248_cast_fp16")];
            fp32 var_7249_epsilon_0 = const()[name = string("op_7249_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_7249_cast_fp16 = inverse(epsilon = var_7249_epsilon_0, x = var_7248_cast_fp16)[name = string("op_7249_cast_fp16")];
            fp16 var_7250_to_fp16 = const()[name = string("op_7250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_7251_cast_fp16 = mul(x = var_7249_cast_fp16, y = var_7250_to_fp16)[name = string("op_7251_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_7253_cast_fp16 = add(x = var_7247_cast_fp16, y = var_7251_cast_fp16)[name = string("op_7253_cast_fp16")];
            fp32 var_7254_epsilon_0 = const()[name = string("op_7254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_7254_cast_fp16 = rsqrt(epsilon = var_7254_epsilon_0, x = var_7253_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_61_cast_fp16 = mul(x = y_249_cast_fp16, y = var_7254_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_7256_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7256_cast_fp16")];
            tensor<int32, [4]> var_7261_begin_0 = const()[name = string("op_7261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7261_end_0 = const()[name = string("op_7261_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_7261_end_mask_0 = const()[name = string("op_7261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_7261_cast_fp16 = slice_by_index(begin = var_7261_begin_0, end = var_7261_end_0, end_mask = var_7261_end_mask_0, x = q_61_cast_fp16)[name = string("op_7261_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_7262_cast_fp16 = mul(x = var_7261_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_7262_cast_fp16")];
            tensor<int32, [4]> var_7267_begin_0 = const()[name = string("op_7267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7267_end_0 = const()[name = string("op_7267_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_7267_end_mask_0 = const()[name = string("op_7267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_7267_cast_fp16 = slice_by_index(begin = var_7267_begin_0, end = var_7267_end_0, end_mask = var_7267_end_mask_0, x = q_61_cast_fp16)[name = string("op_7267_cast_fp16")];
            int32 var_7269 = const()[name = string("op_7269"), val = int32(-1)];
            bool var_7270_interleave_0 = const()[name = string("op_7270_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_7270_cast_fp16 = concat(axis = var_7269, interleave = var_7270_interleave_0, values = (var_7262_cast_fp16, var_7267_cast_fp16))[name = string("op_7270_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_7271_cast_fp16 = mul(x = var_7270_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7271_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_63_cast_fp16 = add(x = var_7256_cast_fp16, y = var_7271_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_7274_cast_fp16 = mul(x = k_91_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7274_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, x = k_91_cast_fp16)[name = string("op_7279_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_7280_cast_fp16 = mul(x = var_7279_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_7280_cast_fp16")];
            tensor<int32, [4]> var_7285_begin_0 = const()[name = string("op_7285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7285_end_0 = const()[name = string("op_7285_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_7285_end_mask_0 = const()[name = string("op_7285_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, x = k_91_cast_fp16)[name = string("op_7285_cast_fp16")];
            int32 var_7287 = const()[name = string("op_7287"), val = int32(-1)];
            bool var_7288_interleave_0 = const()[name = string("op_7288_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_7288_cast_fp16 = concat(axis = var_7287, interleave = var_7288_interleave_0, values = (var_7280_cast_fp16, var_7285_cast_fp16))[name = string("op_7288_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_7289_cast_fp16 = mul(x = var_7288_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7289_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_93_cast_fp16 = add(x = var_7274_cast_fp16, y = var_7289_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_30_reps_0 = const()[name = string("tile_30_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = k_93_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 1, 32, 256]> tile_30_cast_fp16 = tile(reps = tile_30_reps_0, x = transpose_60_cast_fp16)[name = string("tile_30_cast_fp16")];
            tensor<int32, [5]> concat_60 = const()[name = string("concat_60"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_60_cast_fp16 = reshape(shape = concat_60, x = tile_30_cast_fp16)[name = string("reshape_60_cast_fp16")];
            tensor<int32, [5]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = reshape_60_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 1, 32, 256]> reshape_61_cast_fp16 = reshape(shape = concat_61, x = transpose_61_cast_fp16)[name = string("reshape_61_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_31_reps_0 = const()[name = string("tile_31_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = v_61_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 1, 32, 256]> tile_31_cast_fp16 = tile(reps = tile_31_reps_0, x = transpose_62_cast_fp16)[name = string("tile_31_cast_fp16")];
            tensor<int32, [5]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_62_cast_fp16 = reshape(shape = concat_62, x = tile_31_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [5]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 1, 32, 256]> reshape_63_cast_fp16 = reshape(shape = concat_63, x = transpose_63_cast_fp16)[name = string("reshape_63_cast_fp16")];
            tensor<int32, [4]> v_63_perm_0 = const()[name = string("v_63_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7303_transpose_x_0 = const()[name = string("op_7303_transpose_x_0"), val = bool(false)];
            bool var_7303_transpose_y_0 = const()[name = string("op_7303_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = reshape_61_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 4, 32, 32]> var_7303_cast_fp16 = matmul(transpose_x = var_7303_transpose_x_0, transpose_y = var_7303_transpose_y_0, x = q_63_cast_fp16, y = transpose_111_cast_fp16)[name = string("op_7303_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_7305_cast_fp16 = add(x = var_7303_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7305_cast_fp16")];
            int32 var_7306 = const()[name = string("op_7306"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_249_cast_fp16 = softmax(axis = var_7306, x = var_7305_cast_fp16)[name = string("w_249_cast_fp16")];
            bool var_7309_transpose_x_0 = const()[name = string("op_7309_transpose_x_0"), val = bool(false)];
            bool var_7309_transpose_y_0 = const()[name = string("op_7309_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = reshape_63_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 4, 32, 256]> var_7309_cast_fp16 = matmul(transpose_x = var_7309_transpose_x_0, transpose_y = var_7309_transpose_y_0, x = w_249_cast_fp16, y = v_63_cast_fp16)[name = string("op_7309_cast_fp16")];
            tensor<int32, [4]> var_7312_perm_0 = const()[name = string("op_7312_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7316 = const()[name = string("op_7316"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_7312_cast_fp16 = transpose(perm = var_7312_perm_0, x = var_7309_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 32, 1024]> input_183_cast_fp16 = reshape(shape = var_7316, x = var_7312_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510656))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89035008))))[name = string("tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7321_cast_fp16 = abs(x = linear_139_cast_fp16)[name = string("op_7321_cast_fp16")];
            bool var_7324 = const()[name = string("op_7324"), val = bool(true)];
            tensor<int32, [1]> var_7325_axes_0 = const()[name = string("op_7325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7325_cast_fp16 = reduce_max(axes = var_7325_axes_0, keep_dims = var_7324, x = var_7321_cast_fp16)[name = string("op_7325_cast_fp16")];
            fp16 var_7326_to_fp16 = const()[name = string("op_7326_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_251_cast_fp16 = maximum(x = var_7325_cast_fp16, y = var_7326_to_fp16)[name = string("s_251_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_251_cast_fp16 = real_div(x = linear_139_cast_fp16, y = s_251_cast_fp16)[name = string("y_251_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7329_cast_fp16 = mul(x = y_251_cast_fp16, y = y_251_cast_fp16)[name = string("op_7329_cast_fp16")];
            tensor<int32, [1]> var_7334_axes_0 = const()[name = string("op_7334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7334_keep_dims_0 = const()[name = string("op_7334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7334_cast_fp16 = reduce_mean(axes = var_7334_axes_0, keep_dims = var_7334_keep_dims_0, x = var_7329_cast_fp16)[name = string("op_7334_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7335_cast_fp16 = mul(x = s_251_cast_fp16, y = s_251_cast_fp16)[name = string("op_7335_cast_fp16")];
            fp32 var_7336_epsilon_0 = const()[name = string("op_7336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7336_cast_fp16 = inverse(epsilon = var_7336_epsilon_0, x = var_7335_cast_fp16)[name = string("op_7336_cast_fp16")];
            fp16 var_7337_to_fp16 = const()[name = string("op_7337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7338_cast_fp16 = mul(x = var_7336_cast_fp16, y = var_7337_to_fp16)[name = string("op_7338_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7340_cast_fp16 = add(x = var_7334_cast_fp16, y = var_7338_cast_fp16)[name = string("op_7340_cast_fp16")];
            fp32 var_7341_epsilon_0 = const()[name = string("op_7341_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7341_cast_fp16 = rsqrt(epsilon = var_7341_epsilon_0, x = var_7340_cast_fp16)[name = string("op_7341_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_219_cast_fp16 = mul(x = y_251_cast_fp16, y = var_7341_cast_fp16)[name = string("out_219_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89036096)))];
            tensor<fp16, [1, 32, 512]> var_7343_cast_fp16 = mul(x = out_219_cast_fp16, y = tm_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_7343_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_281_cast_fp16 = add(x = x_271_cast_fp16, y = var_7343_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7346_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_7346_cast_fp16")];
            bool var_7349 = const()[name = string("op_7349"), val = bool(true)];
            tensor<int32, [1]> var_7350_axes_0 = const()[name = string("op_7350_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7350_cast_fp16 = reduce_max(axes = var_7350_axes_0, keep_dims = var_7349, x = var_7346_cast_fp16)[name = string("op_7350_cast_fp16")];
            fp16 var_7351_to_fp16 = const()[name = string("op_7351_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_253_cast_fp16 = maximum(x = var_7350_cast_fp16, y = var_7351_to_fp16)[name = string("s_253_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_253_cast_fp16 = real_div(x = x_281_cast_fp16, y = s_253_cast_fp16)[name = string("y_253_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7354_cast_fp16 = mul(x = y_253_cast_fp16, y = y_253_cast_fp16)[name = string("op_7354_cast_fp16")];
            tensor<int32, [1]> var_7359_axes_0 = const()[name = string("op_7359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7359_keep_dims_0 = const()[name = string("op_7359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7359_cast_fp16 = reduce_mean(axes = var_7359_axes_0, keep_dims = var_7359_keep_dims_0, x = var_7354_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7360_cast_fp16 = mul(x = s_253_cast_fp16, y = s_253_cast_fp16)[name = string("op_7360_cast_fp16")];
            fp32 var_7361_epsilon_0 = const()[name = string("op_7361_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7361_cast_fp16 = inverse(epsilon = var_7361_epsilon_0, x = var_7360_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7362_to_fp16 = const()[name = string("op_7362_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7363_cast_fp16 = mul(x = var_7361_cast_fp16, y = var_7362_to_fp16)[name = string("op_7363_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7365_cast_fp16 = add(x = var_7359_cast_fp16, y = var_7363_cast_fp16)[name = string("op_7365_cast_fp16")];
            fp32 var_7366_epsilon_0 = const()[name = string("op_7366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7366_cast_fp16 = rsqrt(epsilon = var_7366_epsilon_0, x = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_221_cast_fp16 = mul(x = y_253_cast_fp16, y = var_7366_cast_fp16)[name = string("out_221_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89037184)))];
            tensor<fp16, [1, 32, 512]> input_185_cast_fp16 = mul(x = out_221_cast_fp16, y = tm_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89038272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90086912))))[name = string("tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_140_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_140_cast_fp16")];
            string var_7373_mode_0 = const()[name = string("op_7373_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_7373_cast_fp16 = gelu(mode = var_7373_mode_0, x = linear_140_cast_fp16)[name = string("op_7373_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90091072))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91139712))))[name = string("tm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_141_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_187_cast_fp16 = mul(x = var_7373_cast_fp16, y = linear_141_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91143872))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92192512))))[name = string("tm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7381_cast_fp16 = abs(x = linear_142_cast_fp16)[name = string("op_7381_cast_fp16")];
            bool var_7384 = const()[name = string("op_7384"), val = bool(true)];
            tensor<int32, [1]> var_7385_axes_0 = const()[name = string("op_7385_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7385_cast_fp16 = reduce_max(axes = var_7385_axes_0, keep_dims = var_7384, x = var_7381_cast_fp16)[name = string("op_7385_cast_fp16")];
            fp16 var_7386_to_fp16 = const()[name = string("op_7386_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_255_cast_fp16 = maximum(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = string("s_255_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_255_cast_fp16 = real_div(x = linear_142_cast_fp16, y = s_255_cast_fp16)[name = string("y_255_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7389_cast_fp16 = mul(x = y_255_cast_fp16, y = y_255_cast_fp16)[name = string("op_7389_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7389_cast_fp16)[name = string("op_7394_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7395_cast_fp16 = mul(x = s_255_cast_fp16, y = s_255_cast_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7400_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7400_cast_fp16")];
            fp32 var_7401_epsilon_0 = const()[name = string("op_7401_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7401_cast_fp16 = rsqrt(epsilon = var_7401_epsilon_0, x = var_7400_cast_fp16)[name = string("op_7401_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_223_cast_fp16 = mul(x = y_255_cast_fp16, y = var_7401_cast_fp16)[name = string("out_223_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92193600)))];
            tensor<fp16, [1, 32, 512]> var_7403_cast_fp16 = mul(x = out_223_cast_fp16, y = tm_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7403_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_189_cast_fp16 = add(x = x_281_cast_fp16, y = var_7403_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92194688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92456896))))[name = string("tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_143_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_143_cast_fp16")];
            string x_285_mode_0 = const()[name = string("x_285_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_285_cast_fp16 = gelu(mode = x_285_mode_0, x = linear_143_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [4]> var_7423_begin_0 = const()[name = string("op_7423_begin_0"), val = tensor<int32, [4]>([0, 0, 15, 0])];
            tensor<int32, [4]> var_7423_end_0 = const()[name = string("op_7423_end_0"), val = tensor<int32, [4]>([1, 32, 16, 512])];
            tensor<bool, [4]> var_7423_end_mask_0 = const()[name = string("op_7423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7423_squeeze_mask_0 = const()[name = string("op_7423_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_7423_cast_fp16 = slice_by_index(begin = var_7423_begin_0, end = var_7423_end_0, end_mask = var_7423_end_mask_0, squeeze_mask = var_7423_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7423_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_191_cast_fp16 = mul(x = x_285_cast_fp16, y = var_7423_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92457984))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92720192))))[name = string("tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7433_cast_fp16 = abs(x = linear_144_cast_fp16)[name = string("op_7433_cast_fp16")];
            bool var_7436 = const()[name = string("op_7436"), val = bool(true)];
            tensor<int32, [1]> var_7437_axes_0 = const()[name = string("op_7437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7437_cast_fp16 = reduce_max(axes = var_7437_axes_0, keep_dims = var_7436, x = var_7433_cast_fp16)[name = string("op_7437_cast_fp16")];
            fp16 var_7438_to_fp16 = const()[name = string("op_7438_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_257_cast_fp16 = maximum(x = var_7437_cast_fp16, y = var_7438_to_fp16)[name = string("s_257_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_257_cast_fp16 = real_div(x = linear_144_cast_fp16, y = s_257_cast_fp16)[name = string("y_257_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7441_cast_fp16 = mul(x = y_257_cast_fp16, y = y_257_cast_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7446_axes_0 = const()[name = string("op_7446_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7446_keep_dims_0 = const()[name = string("op_7446_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7446_cast_fp16 = reduce_mean(axes = var_7446_axes_0, keep_dims = var_7446_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7447_cast_fp16 = mul(x = s_257_cast_fp16, y = s_257_cast_fp16)[name = string("op_7447_cast_fp16")];
            fp32 var_7448_epsilon_0 = const()[name = string("op_7448_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7448_cast_fp16 = inverse(epsilon = var_7448_epsilon_0, x = var_7447_cast_fp16)[name = string("op_7448_cast_fp16")];
            fp16 var_7449_to_fp16 = const()[name = string("op_7449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7450_cast_fp16 = mul(x = var_7448_cast_fp16, y = var_7449_to_fp16)[name = string("op_7450_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7452_cast_fp16 = add(x = var_7446_cast_fp16, y = var_7450_cast_fp16)[name = string("op_7452_cast_fp16")];
            fp32 var_7453_epsilon_0 = const()[name = string("op_7453_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7453_cast_fp16 = rsqrt(epsilon = var_7453_epsilon_0, x = var_7452_cast_fp16)[name = string("op_7453_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_225_cast_fp16 = mul(x = y_257_cast_fp16, y = var_7453_cast_fp16)[name = string("out_225_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92721280)))];
            tensor<fp16, [1, 32, 512]> var_7455_cast_fp16 = mul(x = out_225_cast_fp16, y = tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7455_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_31_cast_fp16 = add(x = input_189_cast_fp16, y = var_7455_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<fp16, [1]> tm_layers_15_layer_scalar_to_fp16 = const()[name = string("tm_layers_15_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7ap-1])];
            tensor<fp16, [1, 32, 512]> x_289_cast_fp16 = mul(x = h_31_cast_fp16, y = tm_layers_15_layer_scalar_to_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7459_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_7459_cast_fp16")];
            bool var_7462 = const()[name = string("op_7462"), val = bool(true)];
            tensor<int32, [1]> var_7463_axes_0 = const()[name = string("op_7463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7463_cast_fp16 = reduce_max(axes = var_7463_axes_0, keep_dims = var_7462, x = var_7459_cast_fp16)[name = string("op_7463_cast_fp16")];
            fp16 var_7464_to_fp16 = const()[name = string("op_7464_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_259_cast_fp16 = maximum(x = var_7463_cast_fp16, y = var_7464_to_fp16)[name = string("s_259_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_259_cast_fp16 = real_div(x = x_289_cast_fp16, y = s_259_cast_fp16)[name = string("y_259_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7467_cast_fp16 = mul(x = y_259_cast_fp16, y = y_259_cast_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7472_axes_0 = const()[name = string("op_7472_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7472_keep_dims_0 = const()[name = string("op_7472_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7472_cast_fp16 = reduce_mean(axes = var_7472_axes_0, keep_dims = var_7472_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7473_cast_fp16 = mul(x = s_259_cast_fp16, y = s_259_cast_fp16)[name = string("op_7473_cast_fp16")];
            fp32 var_7474_epsilon_0 = const()[name = string("op_7474_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7474_cast_fp16 = inverse(epsilon = var_7474_epsilon_0, x = var_7473_cast_fp16)[name = string("op_7474_cast_fp16")];
            fp16 var_7475_to_fp16 = const()[name = string("op_7475_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7476_cast_fp16 = mul(x = var_7474_cast_fp16, y = var_7475_to_fp16)[name = string("op_7476_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7478_cast_fp16 = add(x = var_7472_cast_fp16, y = var_7476_cast_fp16)[name = string("op_7478_cast_fp16")];
            fp32 var_7479_epsilon_0 = const()[name = string("op_7479_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7479_cast_fp16 = rsqrt(epsilon = var_7479_epsilon_0, x = var_7478_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_227_cast_fp16 = mul(x = y_259_cast_fp16, y = var_7479_cast_fp16)[name = string("out_227_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92722368)))];
            tensor<fp16, [1, 32, 512]> input_193_cast_fp16 = mul(x = out_227_cast_fp16, y = tm_layers_16_input_layernorm_weight_to_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92723456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93247808))))[name = string("tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<int32, [4]> var_7489 = const()[name = string("op_7489"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_7490_cast_fp16 = reshape(shape = var_7489, x = linear_145_cast_fp16)[name = string("op_7490_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93249920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93512128))))[name = string("tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<int32, [4]> var_7501 = const()[name = string("op_7501"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_7502_cast_fp16 = reshape(shape = var_7501, x = linear_146_cast_fp16)[name = string("op_7502_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93513216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93775424))))[name = string("tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7513 = const()[name = string("op_7513"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_7514_cast_fp16 = reshape(shape = var_7513, x = linear_147_cast_fp16)[name = string("op_7514_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_7490_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 4, 32, 256]> var_7518_cast_fp16 = abs(x = x_291_cast_fp16)[name = string("op_7518_cast_fp16")];
            bool var_7521 = const()[name = string("op_7521"), val = bool(true)];
            tensor<int32, [1]> var_7522_axes_0 = const()[name = string("op_7522_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_7522_cast_fp16 = reduce_max(axes = var_7522_axes_0, keep_dims = var_7521, x = var_7518_cast_fp16)[name = string("op_7522_cast_fp16")];
            fp16 var_7523_to_fp16 = const()[name = string("op_7523_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_261_cast_fp16 = maximum(x = var_7522_cast_fp16, y = var_7523_to_fp16)[name = string("s_261_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_261_cast_fp16 = real_div(x = x_291_cast_fp16, y = s_261_cast_fp16)[name = string("y_261_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_7526_cast_fp16 = mul(x = y_261_cast_fp16, y = y_261_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<int32, [1]> var_7531_axes_0 = const()[name = string("op_7531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7531_keep_dims_0 = const()[name = string("op_7531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_7531_cast_fp16 = reduce_mean(axes = var_7531_axes_0, keep_dims = var_7531_keep_dims_0, x = var_7526_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_7532_cast_fp16 = mul(x = s_261_cast_fp16, y = s_261_cast_fp16)[name = string("op_7532_cast_fp16")];
            fp32 var_7533_epsilon_0 = const()[name = string("op_7533_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_7533_cast_fp16 = inverse(epsilon = var_7533_epsilon_0, x = var_7532_cast_fp16)[name = string("op_7533_cast_fp16")];
            fp16 var_7534_to_fp16 = const()[name = string("op_7534_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_7535_cast_fp16 = mul(x = var_7533_cast_fp16, y = var_7534_to_fp16)[name = string("op_7535_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_7537_cast_fp16 = add(x = var_7531_cast_fp16, y = var_7535_cast_fp16)[name = string("op_7537_cast_fp16")];
            fp32 var_7538_epsilon_0 = const()[name = string("op_7538_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_7538_cast_fp16 = rsqrt(epsilon = var_7538_epsilon_0, x = var_7537_cast_fp16)[name = string("op_7538_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_229_cast_fp16 = mul(x = y_261_cast_fp16, y = var_7538_cast_fp16)[name = string("out_229_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_65_cast_fp16 = mul(x = out_229_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_65_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_7502_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 32, 256]> var_7541_cast_fp16 = abs(x = x_293_cast_fp16)[name = string("op_7541_cast_fp16")];
            bool var_7544 = const()[name = string("op_7544"), val = bool(true)];
            tensor<int32, [1]> var_7545_axes_0 = const()[name = string("op_7545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_7545_cast_fp16 = reduce_max(axes = var_7545_axes_0, keep_dims = var_7544, x = var_7541_cast_fp16)[name = string("op_7545_cast_fp16")];
            fp16 var_7546_to_fp16 = const()[name = string("op_7546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_263_cast_fp16 = maximum(x = var_7545_cast_fp16, y = var_7546_to_fp16)[name = string("s_263_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_263_cast_fp16 = real_div(x = x_293_cast_fp16, y = s_263_cast_fp16)[name = string("y_263_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_7549_cast_fp16 = mul(x = y_263_cast_fp16, y = y_263_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<int32, [1]> var_7554_axes_0 = const()[name = string("op_7554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7554_keep_dims_0 = const()[name = string("op_7554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_7554_cast_fp16 = reduce_mean(axes = var_7554_axes_0, keep_dims = var_7554_keep_dims_0, x = var_7549_cast_fp16)[name = string("op_7554_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_7555_cast_fp16 = mul(x = s_263_cast_fp16, y = s_263_cast_fp16)[name = string("op_7555_cast_fp16")];
            fp32 var_7556_epsilon_0 = const()[name = string("op_7556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_7556_cast_fp16 = inverse(epsilon = var_7556_epsilon_0, x = var_7555_cast_fp16)[name = string("op_7556_cast_fp16")];
            fp16 var_7557_to_fp16 = const()[name = string("op_7557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_7558_cast_fp16 = mul(x = var_7556_cast_fp16, y = var_7557_to_fp16)[name = string("op_7558_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_7560_cast_fp16 = add(x = var_7554_cast_fp16, y = var_7558_cast_fp16)[name = string("op_7560_cast_fp16")];
            fp32 var_7561_epsilon_0 = const()[name = string("op_7561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_7561_cast_fp16 = rsqrt(epsilon = var_7561_epsilon_0, x = var_7560_cast_fp16)[name = string("op_7561_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_231_cast_fp16 = mul(x = y_263_cast_fp16, y = var_7561_cast_fp16)[name = string("out_231_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_97_cast_fp16 = mul(x = out_231_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_97_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_7514_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 32, 256]> var_7564_cast_fp16 = abs(x = x_295_cast_fp16)[name = string("op_7564_cast_fp16")];
            bool var_7567 = const()[name = string("op_7567"), val = bool(true)];
            tensor<int32, [1]> var_7568_axes_0 = const()[name = string("op_7568_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_7568_cast_fp16 = reduce_max(axes = var_7568_axes_0, keep_dims = var_7567, x = var_7564_cast_fp16)[name = string("op_7568_cast_fp16")];
            fp16 var_7569_to_fp16 = const()[name = string("op_7569_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_265_cast_fp16 = maximum(x = var_7568_cast_fp16, y = var_7569_to_fp16)[name = string("s_265_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_265_cast_fp16 = real_div(x = x_295_cast_fp16, y = s_265_cast_fp16)[name = string("y_265_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_7572_cast_fp16 = mul(x = y_265_cast_fp16, y = y_265_cast_fp16)[name = string("op_7572_cast_fp16")];
            tensor<int32, [1]> var_7577_axes_0 = const()[name = string("op_7577_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7577_keep_dims_0 = const()[name = string("op_7577_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_7577_cast_fp16 = reduce_mean(axes = var_7577_axes_0, keep_dims = var_7577_keep_dims_0, x = var_7572_cast_fp16)[name = string("op_7577_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_7578_cast_fp16 = mul(x = s_265_cast_fp16, y = s_265_cast_fp16)[name = string("op_7578_cast_fp16")];
            fp32 var_7579_epsilon_0 = const()[name = string("op_7579_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_7579_cast_fp16 = inverse(epsilon = var_7579_epsilon_0, x = var_7578_cast_fp16)[name = string("op_7579_cast_fp16")];
            fp16 var_7580_to_fp16 = const()[name = string("op_7580_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_7581_cast_fp16 = mul(x = var_7579_cast_fp16, y = var_7580_to_fp16)[name = string("op_7581_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_7583_cast_fp16 = add(x = var_7577_cast_fp16, y = var_7581_cast_fp16)[name = string("op_7583_cast_fp16")];
            fp32 var_7584_epsilon_0 = const()[name = string("op_7584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_7584_cast_fp16 = rsqrt(epsilon = var_7584_epsilon_0, x = var_7583_cast_fp16)[name = string("op_7584_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_65_cast_fp16 = mul(x = y_265_cast_fp16, y = var_7584_cast_fp16)[name = string("v_65_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_7586_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7586_cast_fp16")];
            tensor<int32, [4]> var_7591_begin_0 = const()[name = string("op_7591_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7591_end_0 = const()[name = string("op_7591_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_7591_end_mask_0 = const()[name = string("op_7591_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_7591_cast_fp16 = slice_by_index(begin = var_7591_begin_0, end = var_7591_end_0, end_mask = var_7591_end_mask_0, x = q_65_cast_fp16)[name = string("op_7591_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_7592_cast_fp16 = mul(x = var_7591_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_7592_cast_fp16")];
            tensor<int32, [4]> var_7597_begin_0 = const()[name = string("op_7597_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7597_end_0 = const()[name = string("op_7597_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_7597_end_mask_0 = const()[name = string("op_7597_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, x = q_65_cast_fp16)[name = string("op_7597_cast_fp16")];
            int32 var_7599 = const()[name = string("op_7599"), val = int32(-1)];
            bool var_7600_interleave_0 = const()[name = string("op_7600_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_7600_cast_fp16 = concat(axis = var_7599, interleave = var_7600_interleave_0, values = (var_7592_cast_fp16, var_7597_cast_fp16))[name = string("op_7600_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_7601_cast_fp16 = mul(x = var_7600_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7601_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_67_cast_fp16 = add(x = var_7586_cast_fp16, y = var_7601_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_7604_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7604_cast_fp16")];
            tensor<int32, [4]> var_7609_begin_0 = const()[name = string("op_7609_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7609_end_0 = const()[name = string("op_7609_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_7609_end_mask_0 = const()[name = string("op_7609_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_7609_cast_fp16 = slice_by_index(begin = var_7609_begin_0, end = var_7609_end_0, end_mask = var_7609_end_mask_0, x = k_97_cast_fp16)[name = string("op_7609_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_7610_cast_fp16 = mul(x = var_7609_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_7610_cast_fp16")];
            tensor<int32, [4]> var_7615_begin_0 = const()[name = string("op_7615_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7615_end_0 = const()[name = string("op_7615_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_7615_end_mask_0 = const()[name = string("op_7615_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_7615_cast_fp16 = slice_by_index(begin = var_7615_begin_0, end = var_7615_end_0, end_mask = var_7615_end_mask_0, x = k_97_cast_fp16)[name = string("op_7615_cast_fp16")];
            int32 var_7617 = const()[name = string("op_7617"), val = int32(-1)];
            bool var_7618_interleave_0 = const()[name = string("op_7618_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_7618_cast_fp16 = concat(axis = var_7617, interleave = var_7618_interleave_0, values = (var_7610_cast_fp16, var_7615_cast_fp16))[name = string("op_7618_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_7619_cast_fp16 = mul(x = var_7618_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7619_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_99_cast_fp16 = add(x = var_7604_cast_fp16, y = var_7619_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_32_reps_0 = const()[name = string("tile_32_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = k_99_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 1, 32, 256]> tile_32_cast_fp16 = tile(reps = tile_32_reps_0, x = transpose_64_cast_fp16)[name = string("tile_32_cast_fp16")];
            tensor<int32, [5]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_64_cast_fp16 = reshape(shape = concat_64, x = tile_32_cast_fp16)[name = string("reshape_64_cast_fp16")];
            tensor<int32, [5]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = reshape_64_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 1, 32, 256]> reshape_65_cast_fp16 = reshape(shape = concat_65, x = transpose_65_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_33_reps_0 = const()[name = string("tile_33_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = v_65_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 1, 32, 256]> tile_33_cast_fp16 = tile(reps = tile_33_reps_0, x = transpose_66_cast_fp16)[name = string("tile_33_cast_fp16")];
            tensor<int32, [5]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_66_cast_fp16 = reshape(shape = concat_66, x = tile_33_cast_fp16)[name = string("reshape_66_cast_fp16")];
            tensor<int32, [5]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = reshape_66_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 1, 32, 256]> reshape_67_cast_fp16 = reshape(shape = concat_67, x = transpose_67_cast_fp16)[name = string("reshape_67_cast_fp16")];
            tensor<int32, [4]> v_67_perm_0 = const()[name = string("v_67_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7633_transpose_x_0 = const()[name = string("op_7633_transpose_x_0"), val = bool(false)];
            bool var_7633_transpose_y_0 = const()[name = string("op_7633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_112_cast_fp16 = transpose(perm = transpose_112_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 4, 32, 32]> var_7633_cast_fp16 = matmul(transpose_x = var_7633_transpose_x_0, transpose_y = var_7633_transpose_y_0, x = q_67_cast_fp16, y = transpose_112_cast_fp16)[name = string("op_7633_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_7635_cast_fp16 = add(x = var_7633_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7635_cast_fp16")];
            int32 var_7636 = const()[name = string("op_7636"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_265_cast_fp16 = softmax(axis = var_7636, x = var_7635_cast_fp16)[name = string("w_265_cast_fp16")];
            bool var_7639_transpose_x_0 = const()[name = string("op_7639_transpose_x_0"), val = bool(false)];
            bool var_7639_transpose_y_0 = const()[name = string("op_7639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = reshape_67_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 4, 32, 256]> var_7639_cast_fp16 = matmul(transpose_x = var_7639_transpose_x_0, transpose_y = var_7639_transpose_y_0, x = w_265_cast_fp16, y = v_67_cast_fp16)[name = string("op_7639_cast_fp16")];
            tensor<int32, [4]> var_7642_perm_0 = const()[name = string("op_7642_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7646 = const()[name = string("op_7646"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_7642_cast_fp16 = transpose(perm = var_7642_perm_0, x = var_7639_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 32, 1024]> input_195_cast_fp16 = reshape(shape = var_7646, x = var_7642_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93776512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94300864))))[name = string("tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7651_cast_fp16 = abs(x = linear_148_cast_fp16)[name = string("op_7651_cast_fp16")];
            bool var_7654 = const()[name = string("op_7654"), val = bool(true)];
            tensor<int32, [1]> var_7655_axes_0 = const()[name = string("op_7655_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7655_cast_fp16 = reduce_max(axes = var_7655_axes_0, keep_dims = var_7654, x = var_7651_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp16 var_7656_to_fp16 = const()[name = string("op_7656_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_267_cast_fp16 = maximum(x = var_7655_cast_fp16, y = var_7656_to_fp16)[name = string("s_267_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_267_cast_fp16 = real_div(x = linear_148_cast_fp16, y = s_267_cast_fp16)[name = string("y_267_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7659_cast_fp16 = mul(x = y_267_cast_fp16, y = y_267_cast_fp16)[name = string("op_7659_cast_fp16")];
            tensor<int32, [1]> var_7664_axes_0 = const()[name = string("op_7664_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7664_keep_dims_0 = const()[name = string("op_7664_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7664_cast_fp16 = reduce_mean(axes = var_7664_axes_0, keep_dims = var_7664_keep_dims_0, x = var_7659_cast_fp16)[name = string("op_7664_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7665_cast_fp16 = mul(x = s_267_cast_fp16, y = s_267_cast_fp16)[name = string("op_7665_cast_fp16")];
            fp32 var_7666_epsilon_0 = const()[name = string("op_7666_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7666_cast_fp16 = inverse(epsilon = var_7666_epsilon_0, x = var_7665_cast_fp16)[name = string("op_7666_cast_fp16")];
            fp16 var_7667_to_fp16 = const()[name = string("op_7667_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7668_cast_fp16 = mul(x = var_7666_cast_fp16, y = var_7667_to_fp16)[name = string("op_7668_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7670_cast_fp16 = add(x = var_7664_cast_fp16, y = var_7668_cast_fp16)[name = string("op_7670_cast_fp16")];
            fp32 var_7671_epsilon_0 = const()[name = string("op_7671_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7671_cast_fp16 = rsqrt(epsilon = var_7671_epsilon_0, x = var_7670_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_233_cast_fp16 = mul(x = y_267_cast_fp16, y = var_7671_cast_fp16)[name = string("out_233_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94301952)))];
            tensor<fp16, [1, 32, 512]> var_7673_cast_fp16 = mul(x = out_233_cast_fp16, y = tm_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("op_7673_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = var_7673_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7676_cast_fp16 = abs(x = x_299_cast_fp16)[name = string("op_7676_cast_fp16")];
            bool var_7679 = const()[name = string("op_7679"), val = bool(true)];
            tensor<int32, [1]> var_7680_axes_0 = const()[name = string("op_7680_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7680_cast_fp16 = reduce_max(axes = var_7680_axes_0, keep_dims = var_7679, x = var_7676_cast_fp16)[name = string("op_7680_cast_fp16")];
            fp16 var_7681_to_fp16 = const()[name = string("op_7681_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_269_cast_fp16 = maximum(x = var_7680_cast_fp16, y = var_7681_to_fp16)[name = string("s_269_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_269_cast_fp16 = real_div(x = x_299_cast_fp16, y = s_269_cast_fp16)[name = string("y_269_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7684_cast_fp16 = mul(x = y_269_cast_fp16, y = y_269_cast_fp16)[name = string("op_7684_cast_fp16")];
            tensor<int32, [1]> var_7689_axes_0 = const()[name = string("op_7689_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7689_keep_dims_0 = const()[name = string("op_7689_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7689_cast_fp16 = reduce_mean(axes = var_7689_axes_0, keep_dims = var_7689_keep_dims_0, x = var_7684_cast_fp16)[name = string("op_7689_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7690_cast_fp16 = mul(x = s_269_cast_fp16, y = s_269_cast_fp16)[name = string("op_7690_cast_fp16")];
            fp32 var_7691_epsilon_0 = const()[name = string("op_7691_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7691_cast_fp16 = inverse(epsilon = var_7691_epsilon_0, x = var_7690_cast_fp16)[name = string("op_7691_cast_fp16")];
            fp16 var_7692_to_fp16 = const()[name = string("op_7692_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7693_cast_fp16 = mul(x = var_7691_cast_fp16, y = var_7692_to_fp16)[name = string("op_7693_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7695_cast_fp16 = add(x = var_7689_cast_fp16, y = var_7693_cast_fp16)[name = string("op_7695_cast_fp16")];
            fp32 var_7696_epsilon_0 = const()[name = string("op_7696_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7696_cast_fp16 = rsqrt(epsilon = var_7696_epsilon_0, x = var_7695_cast_fp16)[name = string("op_7696_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_235_cast_fp16 = mul(x = y_269_cast_fp16, y = var_7696_cast_fp16)[name = string("out_235_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94303040)))];
            tensor<fp16, [1, 32, 512]> input_197_cast_fp16 = mul(x = out_235_cast_fp16, y = tm_layers_16_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94304128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95352768))))[name = string("tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_149_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_149_cast_fp16")];
            string var_7703_mode_0 = const()[name = string("op_7703_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_7703_cast_fp16 = gelu(mode = var_7703_mode_0, x = linear_149_cast_fp16)[name = string("op_7703_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95356928))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96405568))))[name = string("tm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_150_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_199_cast_fp16 = mul(x = var_7703_cast_fp16, y = linear_150_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96409728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97458368))))[name = string("tm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7711_cast_fp16 = abs(x = linear_151_cast_fp16)[name = string("op_7711_cast_fp16")];
            bool var_7714 = const()[name = string("op_7714"), val = bool(true)];
            tensor<int32, [1]> var_7715_axes_0 = const()[name = string("op_7715_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7715_cast_fp16 = reduce_max(axes = var_7715_axes_0, keep_dims = var_7714, x = var_7711_cast_fp16)[name = string("op_7715_cast_fp16")];
            fp16 var_7716_to_fp16 = const()[name = string("op_7716_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_271_cast_fp16 = maximum(x = var_7715_cast_fp16, y = var_7716_to_fp16)[name = string("s_271_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_271_cast_fp16 = real_div(x = linear_151_cast_fp16, y = s_271_cast_fp16)[name = string("y_271_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7719_cast_fp16 = mul(x = y_271_cast_fp16, y = y_271_cast_fp16)[name = string("op_7719_cast_fp16")];
            tensor<int32, [1]> var_7724_axes_0 = const()[name = string("op_7724_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7724_keep_dims_0 = const()[name = string("op_7724_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7724_cast_fp16 = reduce_mean(axes = var_7724_axes_0, keep_dims = var_7724_keep_dims_0, x = var_7719_cast_fp16)[name = string("op_7724_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7725_cast_fp16 = mul(x = s_271_cast_fp16, y = s_271_cast_fp16)[name = string("op_7725_cast_fp16")];
            fp32 var_7726_epsilon_0 = const()[name = string("op_7726_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7726_cast_fp16 = inverse(epsilon = var_7726_epsilon_0, x = var_7725_cast_fp16)[name = string("op_7726_cast_fp16")];
            fp16 var_7727_to_fp16 = const()[name = string("op_7727_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7728_cast_fp16 = mul(x = var_7726_cast_fp16, y = var_7727_to_fp16)[name = string("op_7728_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7730_cast_fp16 = add(x = var_7724_cast_fp16, y = var_7728_cast_fp16)[name = string("op_7730_cast_fp16")];
            fp32 var_7731_epsilon_0 = const()[name = string("op_7731_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7731_cast_fp16 = rsqrt(epsilon = var_7731_epsilon_0, x = var_7730_cast_fp16)[name = string("op_7731_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_237_cast_fp16 = mul(x = y_271_cast_fp16, y = var_7731_cast_fp16)[name = string("out_237_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97459456)))];
            tensor<fp16, [1, 32, 512]> var_7733_cast_fp16 = mul(x = out_237_cast_fp16, y = tm_layers_16_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7733_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_201_cast_fp16 = add(x = x_299_cast_fp16, y = var_7733_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97460544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97722752))))[name = string("tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_152_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_152_cast_fp16")];
            string x_303_mode_0 = const()[name = string("x_303_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_303_cast_fp16 = gelu(mode = x_303_mode_0, x = linear_152_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [4]> var_7753_begin_0 = const()[name = string("op_7753_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7753_end_0 = const()[name = string("op_7753_end_0"), val = tensor<int32, [4]>([1, 32, 17, 512])];
            tensor<bool, [4]> var_7753_end_mask_0 = const()[name = string("op_7753_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7753_squeeze_mask_0 = const()[name = string("op_7753_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_7753_cast_fp16 = slice_by_index(begin = var_7753_begin_0, end = var_7753_end_0, end_mask = var_7753_end_mask_0, squeeze_mask = var_7753_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7753_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_203_cast_fp16 = mul(x = x_303_cast_fp16, y = var_7753_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97723840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97986048))))[name = string("tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7763_cast_fp16 = abs(x = linear_153_cast_fp16)[name = string("op_7763_cast_fp16")];
            bool var_7766 = const()[name = string("op_7766"), val = bool(true)];
            tensor<int32, [1]> var_7767_axes_0 = const()[name = string("op_7767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7767_cast_fp16 = reduce_max(axes = var_7767_axes_0, keep_dims = var_7766, x = var_7763_cast_fp16)[name = string("op_7767_cast_fp16")];
            fp16 var_7768_to_fp16 = const()[name = string("op_7768_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_273_cast_fp16 = maximum(x = var_7767_cast_fp16, y = var_7768_to_fp16)[name = string("s_273_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_273_cast_fp16 = real_div(x = linear_153_cast_fp16, y = s_273_cast_fp16)[name = string("y_273_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7771_cast_fp16 = mul(x = y_273_cast_fp16, y = y_273_cast_fp16)[name = string("op_7771_cast_fp16")];
            tensor<int32, [1]> var_7776_axes_0 = const()[name = string("op_7776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7776_keep_dims_0 = const()[name = string("op_7776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7776_cast_fp16 = reduce_mean(axes = var_7776_axes_0, keep_dims = var_7776_keep_dims_0, x = var_7771_cast_fp16)[name = string("op_7776_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7777_cast_fp16 = mul(x = s_273_cast_fp16, y = s_273_cast_fp16)[name = string("op_7777_cast_fp16")];
            fp32 var_7778_epsilon_0 = const()[name = string("op_7778_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7778_cast_fp16 = inverse(epsilon = var_7778_epsilon_0, x = var_7777_cast_fp16)[name = string("op_7778_cast_fp16")];
            fp16 var_7779_to_fp16 = const()[name = string("op_7779_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7780_cast_fp16 = mul(x = var_7778_cast_fp16, y = var_7779_to_fp16)[name = string("op_7780_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7782_cast_fp16 = add(x = var_7776_cast_fp16, y = var_7780_cast_fp16)[name = string("op_7782_cast_fp16")];
            fp32 var_7783_epsilon_0 = const()[name = string("op_7783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7783_cast_fp16 = rsqrt(epsilon = var_7783_epsilon_0, x = var_7782_cast_fp16)[name = string("op_7783_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_239_cast_fp16 = mul(x = y_273_cast_fp16, y = var_7783_cast_fp16)[name = string("out_239_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97987136)))];
            tensor<fp16, [1, 32, 512]> var_7785_cast_fp16 = mul(x = out_239_cast_fp16, y = tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_33_cast_fp16 = add(x = input_201_cast_fp16, y = var_7785_cast_fp16)[name = string("h_33_cast_fp16")];
            tensor<fp16, [1]> tm_layers_16_layer_scalar_to_fp16 = const()[name = string("tm_layers_16_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7p-2])];
            tensor<fp16, [1, 32, 512]> x_307_cast_fp16 = mul(x = h_33_cast_fp16, y = tm_layers_16_layer_scalar_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7789_cast_fp16 = abs(x = x_307_cast_fp16)[name = string("op_7789_cast_fp16")];
            bool var_7792 = const()[name = string("op_7792"), val = bool(true)];
            tensor<int32, [1]> var_7793_axes_0 = const()[name = string("op_7793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7793_cast_fp16 = reduce_max(axes = var_7793_axes_0, keep_dims = var_7792, x = var_7789_cast_fp16)[name = string("op_7793_cast_fp16")];
            fp16 var_7794_to_fp16 = const()[name = string("op_7794_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_275_cast_fp16 = maximum(x = var_7793_cast_fp16, y = var_7794_to_fp16)[name = string("s_275_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_275_cast_fp16 = real_div(x = x_307_cast_fp16, y = s_275_cast_fp16)[name = string("y_275_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7797_cast_fp16 = mul(x = y_275_cast_fp16, y = y_275_cast_fp16)[name = string("op_7797_cast_fp16")];
            tensor<int32, [1]> var_7802_axes_0 = const()[name = string("op_7802_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7802_keep_dims_0 = const()[name = string("op_7802_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7802_cast_fp16 = reduce_mean(axes = var_7802_axes_0, keep_dims = var_7802_keep_dims_0, x = var_7797_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7803_cast_fp16 = mul(x = s_275_cast_fp16, y = s_275_cast_fp16)[name = string("op_7803_cast_fp16")];
            fp32 var_7804_epsilon_0 = const()[name = string("op_7804_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7804_cast_fp16 = inverse(epsilon = var_7804_epsilon_0, x = var_7803_cast_fp16)[name = string("op_7804_cast_fp16")];
            fp16 var_7805_to_fp16 = const()[name = string("op_7805_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7806_cast_fp16 = mul(x = var_7804_cast_fp16, y = var_7805_to_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7808_cast_fp16 = add(x = var_7802_cast_fp16, y = var_7806_cast_fp16)[name = string("op_7808_cast_fp16")];
            fp32 var_7809_epsilon_0 = const()[name = string("op_7809_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_7809_cast_fp16 = rsqrt(epsilon = var_7809_epsilon_0, x = var_7808_cast_fp16)[name = string("op_7809_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_241_cast_fp16 = mul(x = y_275_cast_fp16, y = var_7809_cast_fp16)[name = string("out_241_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97988224)))];
            tensor<fp16, [1, 32, 512]> input_205_cast_fp16 = mul(x = out_241_cast_fp16, y = tm_layers_17_input_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97989312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99037952))))[name = string("tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_154_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_7819 = const()[name = string("op_7819"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 32, 4, 512]> var_7820_cast_fp16 = reshape(shape = var_7819, x = linear_154_cast_fp16)[name = string("op_7820_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99042112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99304320))))[name = string("tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_7831 = const()[name = string("op_7831"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 32, 1, 512]> var_7832_cast_fp16 = reshape(shape = var_7831, x = linear_155_cast_fp16)[name = string("op_7832_cast_fp16")];
            tensor<int32, [4]> x_311_perm_0 = const()[name = string("x_311_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99305408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99567616))))[name = string("tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_7843 = const()[name = string("op_7843"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 32, 1, 512]> var_7844_cast_fp16 = reshape(shape = var_7843, x = linear_156_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<int32, [4]> x_313_perm_0 = const()[name = string("x_313_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 512]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_7820_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 4, 32, 512]> var_7848_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_7848_cast_fp16")];
            bool var_7851 = const()[name = string("op_7851"), val = bool(true)];
            tensor<int32, [1]> var_7852_axes_0 = const()[name = string("op_7852_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_7852_cast_fp16 = reduce_max(axes = var_7852_axes_0, keep_dims = var_7851, x = var_7848_cast_fp16)[name = string("op_7852_cast_fp16")];
            fp16 var_7853_to_fp16 = const()[name = string("op_7853_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_277_cast_fp16 = maximum(x = var_7852_cast_fp16, y = var_7853_to_fp16)[name = string("s_277_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> y_277_cast_fp16 = real_div(x = x_309_cast_fp16, y = s_277_cast_fp16)[name = string("y_277_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_7856_cast_fp16 = mul(x = y_277_cast_fp16, y = y_277_cast_fp16)[name = string("op_7856_cast_fp16")];
            tensor<int32, [1]> var_7861_axes_0 = const()[name = string("op_7861_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7861_keep_dims_0 = const()[name = string("op_7861_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_7861_cast_fp16 = reduce_mean(axes = var_7861_axes_0, keep_dims = var_7861_keep_dims_0, x = var_7856_cast_fp16)[name = string("op_7861_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_7862_cast_fp16 = mul(x = s_277_cast_fp16, y = s_277_cast_fp16)[name = string("op_7862_cast_fp16")];
            fp32 var_7863_epsilon_0 = const()[name = string("op_7863_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_7863_cast_fp16 = inverse(epsilon = var_7863_epsilon_0, x = var_7862_cast_fp16)[name = string("op_7863_cast_fp16")];
            fp16 var_7864_to_fp16 = const()[name = string("op_7864_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_7865_cast_fp16 = mul(x = var_7863_cast_fp16, y = var_7864_to_fp16)[name = string("op_7865_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_7867_cast_fp16 = add(x = var_7861_cast_fp16, y = var_7865_cast_fp16)[name = string("op_7867_cast_fp16")];
            fp32 var_7868_epsilon_0 = const()[name = string("op_7868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_7868_cast_fp16 = rsqrt(epsilon = var_7868_epsilon_0, x = var_7867_cast_fp16)[name = string("op_7868_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> out_243_cast_fp16 = mul(x = y_277_cast_fp16, y = var_7868_cast_fp16)[name = string("out_243_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99568704)))];
            tensor<fp16, [1, 4, 32, 512]> q_69_cast_fp16 = mul(x = out_243_cast_fp16, y = tm_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("q_69_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> x_311_cast_fp16 = transpose(perm = x_311_perm_0, x = var_7832_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 1, 32, 512]> var_7871_cast_fp16 = abs(x = x_311_cast_fp16)[name = string("op_7871_cast_fp16")];
            bool var_7874 = const()[name = string("op_7874"), val = bool(true)];
            tensor<int32, [1]> var_7875_axes_0 = const()[name = string("op_7875_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 32, 1]> var_7875_cast_fp16 = reduce_max(axes = var_7875_axes_0, keep_dims = var_7874, x = var_7871_cast_fp16)[name = string("op_7875_cast_fp16")];
            fp16 var_7876_to_fp16 = const()[name = string("op_7876_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 32, 1]> s_279_cast_fp16 = maximum(x = var_7875_cast_fp16, y = var_7876_to_fp16)[name = string("s_279_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> y_279_cast_fp16 = real_div(x = x_311_cast_fp16, y = s_279_cast_fp16)[name = string("y_279_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_7879_cast_fp16 = mul(x = y_279_cast_fp16, y = y_279_cast_fp16)[name = string("op_7879_cast_fp16")];
            tensor<int32, [1]> var_7884_axes_0 = const()[name = string("op_7884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7884_keep_dims_0 = const()[name = string("op_7884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 32, 1]> var_7884_cast_fp16 = reduce_mean(axes = var_7884_axes_0, keep_dims = var_7884_keep_dims_0, x = var_7879_cast_fp16)[name = string("op_7884_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_7885_cast_fp16 = mul(x = s_279_cast_fp16, y = s_279_cast_fp16)[name = string("op_7885_cast_fp16")];
            fp32 var_7886_epsilon_0 = const()[name = string("op_7886_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 32, 1]> var_7886_cast_fp16 = inverse(epsilon = var_7886_epsilon_0, x = var_7885_cast_fp16)[name = string("op_7886_cast_fp16")];
            fp16 var_7887_to_fp16 = const()[name = string("op_7887_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 32, 1]> var_7888_cast_fp16 = mul(x = var_7886_cast_fp16, y = var_7887_to_fp16)[name = string("op_7888_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_7890_cast_fp16 = add(x = var_7884_cast_fp16, y = var_7888_cast_fp16)[name = string("op_7890_cast_fp16")];
            fp32 var_7891_epsilon_0 = const()[name = string("op_7891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 32, 1]> var_7891_cast_fp16 = rsqrt(epsilon = var_7891_epsilon_0, x = var_7890_cast_fp16)[name = string("op_7891_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> out_245_cast_fp16 = mul(x = y_279_cast_fp16, y = var_7891_cast_fp16)[name = string("out_245_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99569792)))];
            tensor<fp16, [1, 1, 32, 512]> k_103_cast_fp16 = mul(x = out_245_cast_fp16, y = tm_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("k_103_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> x_313_cast_fp16 = transpose(perm = x_313_perm_0, x = var_7844_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1, 32, 512]> var_7894_cast_fp16 = abs(x = x_313_cast_fp16)[name = string("op_7894_cast_fp16")];
            bool var_7897 = const()[name = string("op_7897"), val = bool(true)];
            tensor<int32, [1]> var_7898_axes_0 = const()[name = string("op_7898_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 32, 1]> var_7898_cast_fp16 = reduce_max(axes = var_7898_axes_0, keep_dims = var_7897, x = var_7894_cast_fp16)[name = string("op_7898_cast_fp16")];
            fp16 var_7899_to_fp16 = const()[name = string("op_7899_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 32, 1]> s_281_cast_fp16 = maximum(x = var_7898_cast_fp16, y = var_7899_to_fp16)[name = string("s_281_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> y_281_cast_fp16 = real_div(x = x_313_cast_fp16, y = s_281_cast_fp16)[name = string("y_281_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_7902_cast_fp16 = mul(x = y_281_cast_fp16, y = y_281_cast_fp16)[name = string("op_7902_cast_fp16")];
            tensor<int32, [1]> var_7907_axes_0 = const()[name = string("op_7907_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7907_keep_dims_0 = const()[name = string("op_7907_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 32, 1]> var_7907_cast_fp16 = reduce_mean(axes = var_7907_axes_0, keep_dims = var_7907_keep_dims_0, x = var_7902_cast_fp16)[name = string("op_7907_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_7908_cast_fp16 = mul(x = s_281_cast_fp16, y = s_281_cast_fp16)[name = string("op_7908_cast_fp16")];
            fp32 var_7909_epsilon_0 = const()[name = string("op_7909_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 32, 1]> var_7909_cast_fp16 = inverse(epsilon = var_7909_epsilon_0, x = var_7908_cast_fp16)[name = string("op_7909_cast_fp16")];
            fp16 var_7910_to_fp16 = const()[name = string("op_7910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 32, 1]> var_7911_cast_fp16 = mul(x = var_7909_cast_fp16, y = var_7910_to_fp16)[name = string("op_7911_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_7913_cast_fp16 = add(x = var_7907_cast_fp16, y = var_7911_cast_fp16)[name = string("op_7913_cast_fp16")];
            fp32 var_7914_epsilon_0 = const()[name = string("op_7914_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 32, 1]> var_7914_cast_fp16 = rsqrt(epsilon = var_7914_epsilon_0, x = var_7913_cast_fp16)[name = string("op_7914_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> v_69_cast_fp16 = mul(x = y_281_cast_fp16, y = var_7914_cast_fp16)[name = string("v_69_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_7916_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7916_cast_fp16")];
            tensor<int32, [4]> var_7921_begin_0 = const()[name = string("op_7921_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7921_end_0 = const()[name = string("op_7921_end_0"), val = tensor<int32, [4]>([1, 4, 32, 512])];
            tensor<bool, [4]> var_7921_end_mask_0 = const()[name = string("op_7921_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 256]> var_7921_cast_fp16 = slice_by_index(begin = var_7921_begin_0, end = var_7921_end_0, end_mask = var_7921_end_mask_0, x = q_69_cast_fp16)[name = string("op_7921_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 256]> var_7922_cast_fp16 = mul(x = var_7921_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 256]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = q_69_cast_fp16)[name = string("op_7927_cast_fp16")];
            int32 var_7929 = const()[name = string("op_7929"), val = int32(-1)];
            bool var_7930_interleave_0 = const()[name = string("op_7930_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 512]> var_7930_cast_fp16 = concat(axis = var_7929, interleave = var_7930_interleave_0, values = (var_7922_cast_fp16, var_7927_cast_fp16))[name = string("op_7930_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_7931_cast_fp16 = mul(x = var_7930_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7931_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> q_71_cast_fp16 = add(x = var_7916_cast_fp16, y = var_7931_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_7934_cast_fp16 = mul(x = k_103_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7934_cast_fp16")];
            tensor<int32, [4]> var_7939_begin_0 = const()[name = string("op_7939_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7939_end_0 = const()[name = string("op_7939_end_0"), val = tensor<int32, [4]>([1, 1, 32, 512])];
            tensor<bool, [4]> var_7939_end_mask_0 = const()[name = string("op_7939_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 32, 256]> var_7939_cast_fp16 = slice_by_index(begin = var_7939_begin_0, end = var_7939_end_0, end_mask = var_7939_end_mask_0, x = k_103_cast_fp16)[name = string("op_7939_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 32, 256]> var_7940_cast_fp16 = mul(x = var_7939_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_7940_cast_fp16")];
            tensor<int32, [4]> var_7945_begin_0 = const()[name = string("op_7945_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7945_end_0 = const()[name = string("op_7945_end_0"), val = tensor<int32, [4]>([1, 1, 32, 256])];
            tensor<bool, [4]> var_7945_end_mask_0 = const()[name = string("op_7945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 32, 256]> var_7945_cast_fp16 = slice_by_index(begin = var_7945_begin_0, end = var_7945_end_0, end_mask = var_7945_end_mask_0, x = k_103_cast_fp16)[name = string("op_7945_cast_fp16")];
            int32 var_7947 = const()[name = string("op_7947"), val = int32(-1)];
            bool var_7948_interleave_0 = const()[name = string("op_7948_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 32, 512]> var_7948_cast_fp16 = concat(axis = var_7947, interleave = var_7948_interleave_0, values = (var_7940_cast_fp16, var_7945_cast_fp16))[name = string("op_7948_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_7949_cast_fp16 = mul(x = var_7948_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> k_105_cast_fp16 = add(x = var_7934_cast_fp16, y = var_7949_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_34_reps_0 = const()[name = string("tile_34_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 32, 512]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = k_105_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 1, 32, 512]> tile_34_cast_fp16 = tile(reps = tile_34_reps_0, x = transpose_68_cast_fp16)[name = string("tile_34_cast_fp16")];
            tensor<int32, [5]> concat_68 = const()[name = string("concat_68"), val = tensor<int32, [5]>([4, 1, 1, 32, 512])];
            tensor<fp16, [4, 1, 1, 32, 512]> reshape_68_cast_fp16 = reshape(shape = concat_68, x = tile_34_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [5]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([-1, 1, 32, 512])];
            tensor<fp16, [1, 4, 1, 32, 512]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 1, 32, 512]> reshape_69_cast_fp16 = reshape(shape = concat_69, x = transpose_69_cast_fp16)[name = string("reshape_69_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_35_reps_0 = const()[name = string("tile_35_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 32, 512]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = v_69_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 1, 32, 512]> tile_35_cast_fp16 = tile(reps = tile_35_reps_0, x = transpose_70_cast_fp16)[name = string("tile_35_cast_fp16")];
            tensor<int32, [5]> concat_70 = const()[name = string("concat_70"), val = tensor<int32, [5]>([4, 1, 1, 32, 512])];
            tensor<fp16, [4, 1, 1, 32, 512]> reshape_70_cast_fp16 = reshape(shape = concat_70, x = tile_35_cast_fp16)[name = string("reshape_70_cast_fp16")];
            tensor<int32, [5]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([-1, 1, 32, 512])];
            tensor<fp16, [1, 4, 1, 32, 512]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = reshape_70_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 1, 32, 512]> reshape_71_cast_fp16 = reshape(shape = concat_71, x = transpose_71_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7963_transpose_x_0 = const()[name = string("op_7963_transpose_x_0"), val = bool(false)];
            bool var_7963_transpose_y_0 = const()[name = string("op_7963_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 32]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = reshape_69_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 4, 32, 32]> var_7963_cast_fp16 = matmul(transpose_x = var_7963_transpose_x_0, transpose_y = var_7963_transpose_y_0, x = q_71_cast_fp16, y = transpose_113_cast_fp16)[name = string("op_7963_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_7965_cast_fp16 = add(x = var_7963_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7965_cast_fp16")];
            int32 var_7966 = const()[name = string("op_7966"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_281_cast_fp16 = softmax(axis = var_7966, x = var_7965_cast_fp16)[name = string("w_281_cast_fp16")];
            bool var_7969_transpose_x_0 = const()[name = string("op_7969_transpose_x_0"), val = bool(false)];
            bool var_7969_transpose_y_0 = const()[name = string("op_7969_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 512]> v_71_cast_fp16 = transpose(perm = v_71_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 4, 32, 512]> var_7969_cast_fp16 = matmul(transpose_x = var_7969_transpose_x_0, transpose_y = var_7969_transpose_y_0, x = w_281_cast_fp16, y = v_71_cast_fp16)[name = string("op_7969_cast_fp16")];
            tensor<int32, [4]> var_7972_perm_0 = const()[name = string("op_7972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7976 = const()[name = string("op_7976"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 32, 4, 512]> var_7972_cast_fp16 = transpose(perm = var_7972_perm_0, x = var_7969_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 32, 2048]> input_207_cast_fp16 = reshape(shape = var_7976, x = var_7972_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99570880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619520))))[name = string("tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7981_cast_fp16 = abs(x = linear_157_cast_fp16)[name = string("op_7981_cast_fp16")];
            bool var_7984 = const()[name = string("op_7984"), val = bool(true)];
            tensor<int32, [1]> var_7985_axes_0 = const()[name = string("op_7985_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_7985_cast_fp16 = reduce_max(axes = var_7985_axes_0, keep_dims = var_7984, x = var_7981_cast_fp16)[name = string("op_7985_cast_fp16")];
            fp16 var_7986_to_fp16 = const()[name = string("op_7986_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_283_cast_fp16 = maximum(x = var_7985_cast_fp16, y = var_7986_to_fp16)[name = string("s_283_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_283_cast_fp16 = real_div(x = linear_157_cast_fp16, y = s_283_cast_fp16)[name = string("y_283_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_7989_cast_fp16 = mul(x = y_283_cast_fp16, y = y_283_cast_fp16)[name = string("op_7989_cast_fp16")];
            tensor<int32, [1]> var_7994_axes_0 = const()[name = string("op_7994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7994_keep_dims_0 = const()[name = string("op_7994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7994_cast_fp16 = reduce_mean(axes = var_7994_axes_0, keep_dims = var_7994_keep_dims_0, x = var_7989_cast_fp16)[name = string("op_7994_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_7995_cast_fp16 = mul(x = s_283_cast_fp16, y = s_283_cast_fp16)[name = string("op_7995_cast_fp16")];
            fp32 var_7996_epsilon_0 = const()[name = string("op_7996_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_7996_cast_fp16 = inverse(epsilon = var_7996_epsilon_0, x = var_7995_cast_fp16)[name = string("op_7996_cast_fp16")];
            fp16 var_7997_to_fp16 = const()[name = string("op_7997_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_7998_cast_fp16 = mul(x = var_7996_cast_fp16, y = var_7997_to_fp16)[name = string("op_7998_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8000_cast_fp16 = add(x = var_7994_cast_fp16, y = var_7998_cast_fp16)[name = string("op_8000_cast_fp16")];
            fp32 var_8001_epsilon_0 = const()[name = string("op_8001_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8001_cast_fp16 = rsqrt(epsilon = var_8001_epsilon_0, x = var_8000_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_247_cast_fp16 = mul(x = y_283_cast_fp16, y = var_8001_cast_fp16)[name = string("out_247_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100620608)))];
            tensor<fp16, [1, 32, 512]> var_8003_cast_fp16 = mul(x = out_247_cast_fp16, y = tm_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("op_8003_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_317_cast_fp16 = add(x = x_307_cast_fp16, y = var_8003_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8006_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_8006_cast_fp16")];
            bool var_8009 = const()[name = string("op_8009"), val = bool(true)];
            tensor<int32, [1]> var_8010_axes_0 = const()[name = string("op_8010_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8010_cast_fp16 = reduce_max(axes = var_8010_axes_0, keep_dims = var_8009, x = var_8006_cast_fp16)[name = string("op_8010_cast_fp16")];
            fp16 var_8011_to_fp16 = const()[name = string("op_8011_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_285_cast_fp16 = maximum(x = var_8010_cast_fp16, y = var_8011_to_fp16)[name = string("s_285_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_285_cast_fp16 = real_div(x = x_317_cast_fp16, y = s_285_cast_fp16)[name = string("y_285_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8014_cast_fp16 = mul(x = y_285_cast_fp16, y = y_285_cast_fp16)[name = string("op_8014_cast_fp16")];
            tensor<int32, [1]> var_8019_axes_0 = const()[name = string("op_8019_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8019_keep_dims_0 = const()[name = string("op_8019_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8019_cast_fp16 = reduce_mean(axes = var_8019_axes_0, keep_dims = var_8019_keep_dims_0, x = var_8014_cast_fp16)[name = string("op_8019_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8020_cast_fp16 = mul(x = s_285_cast_fp16, y = s_285_cast_fp16)[name = string("op_8020_cast_fp16")];
            fp32 var_8021_epsilon_0 = const()[name = string("op_8021_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8021_cast_fp16 = inverse(epsilon = var_8021_epsilon_0, x = var_8020_cast_fp16)[name = string("op_8021_cast_fp16")];
            fp16 var_8022_to_fp16 = const()[name = string("op_8022_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8023_cast_fp16 = mul(x = var_8021_cast_fp16, y = var_8022_to_fp16)[name = string("op_8023_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8025_cast_fp16 = add(x = var_8019_cast_fp16, y = var_8023_cast_fp16)[name = string("op_8025_cast_fp16")];
            fp32 var_8026_epsilon_0 = const()[name = string("op_8026_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8026_cast_fp16 = rsqrt(epsilon = var_8026_epsilon_0, x = var_8025_cast_fp16)[name = string("op_8026_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_249_cast_fp16 = mul(x = y_285_cast_fp16, y = var_8026_cast_fp16)[name = string("out_249_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621696)))];
            tensor<fp16, [1, 32, 512]> input_209_cast_fp16 = mul(x = out_249_cast_fp16, y = tm_layers_17_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100622784))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101671424))))[name = string("tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_158_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_158_cast_fp16")];
            string var_8033_mode_0 = const()[name = string("op_8033_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_8033_cast_fp16 = gelu(mode = var_8033_mode_0, x = linear_158_cast_fp16)[name = string("op_8033_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101675584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102724224))))[name = string("tm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_159_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_211_cast_fp16 = mul(x = var_8033_cast_fp16, y = linear_159_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102728384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103777024))))[name = string("tm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8041_cast_fp16 = abs(x = linear_160_cast_fp16)[name = string("op_8041_cast_fp16")];
            bool var_8044 = const()[name = string("op_8044"), val = bool(true)];
            tensor<int32, [1]> var_8045_axes_0 = const()[name = string("op_8045_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8045_cast_fp16 = reduce_max(axes = var_8045_axes_0, keep_dims = var_8044, x = var_8041_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp16 var_8046_to_fp16 = const()[name = string("op_8046_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_287_cast_fp16 = maximum(x = var_8045_cast_fp16, y = var_8046_to_fp16)[name = string("s_287_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_287_cast_fp16 = real_div(x = linear_160_cast_fp16, y = s_287_cast_fp16)[name = string("y_287_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8049_cast_fp16 = mul(x = y_287_cast_fp16, y = y_287_cast_fp16)[name = string("op_8049_cast_fp16")];
            tensor<int32, [1]> var_8054_axes_0 = const()[name = string("op_8054_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8054_keep_dims_0 = const()[name = string("op_8054_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8054_cast_fp16 = reduce_mean(axes = var_8054_axes_0, keep_dims = var_8054_keep_dims_0, x = var_8049_cast_fp16)[name = string("op_8054_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8055_cast_fp16 = mul(x = s_287_cast_fp16, y = s_287_cast_fp16)[name = string("op_8055_cast_fp16")];
            fp32 var_8056_epsilon_0 = const()[name = string("op_8056_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8056_cast_fp16 = inverse(epsilon = var_8056_epsilon_0, x = var_8055_cast_fp16)[name = string("op_8056_cast_fp16")];
            fp16 var_8057_to_fp16 = const()[name = string("op_8057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8058_cast_fp16 = mul(x = var_8056_cast_fp16, y = var_8057_to_fp16)[name = string("op_8058_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8060_cast_fp16 = add(x = var_8054_cast_fp16, y = var_8058_cast_fp16)[name = string("op_8060_cast_fp16")];
            fp32 var_8061_epsilon_0 = const()[name = string("op_8061_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8061_cast_fp16 = rsqrt(epsilon = var_8061_epsilon_0, x = var_8060_cast_fp16)[name = string("op_8061_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_251_cast_fp16 = mul(x = y_287_cast_fp16, y = var_8061_cast_fp16)[name = string("out_251_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103778112)))];
            tensor<fp16, [1, 32, 512]> var_8063_cast_fp16 = mul(x = out_251_cast_fp16, y = tm_layers_17_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_213_cast_fp16 = add(x = x_317_cast_fp16, y = var_8063_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103779200))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041408))))[name = string("tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_161_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_161_cast_fp16")];
            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_321_cast_fp16 = gelu(mode = x_321_mode_0, x = linear_161_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [4]> var_8083_begin_0 = const()[name = string("op_8083_begin_0"), val = tensor<int32, [4]>([0, 0, 17, 0])];
            tensor<int32, [4]> var_8083_end_0 = const()[name = string("op_8083_end_0"), val = tensor<int32, [4]>([1, 32, 18, 512])];
            tensor<bool, [4]> var_8083_end_mask_0 = const()[name = string("op_8083_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8083_squeeze_mask_0 = const()[name = string("op_8083_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_8083_cast_fp16 = slice_by_index(begin = var_8083_begin_0, end = var_8083_end_0, end_mask = var_8083_end_mask_0, squeeze_mask = var_8083_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8083_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_215_cast_fp16 = mul(x = x_321_cast_fp16, y = var_8083_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104042496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104304704))))[name = string("tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8093_cast_fp16 = abs(x = linear_162_cast_fp16)[name = string("op_8093_cast_fp16")];
            bool var_8096 = const()[name = string("op_8096"), val = bool(true)];
            tensor<int32, [1]> var_8097_axes_0 = const()[name = string("op_8097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8097_cast_fp16 = reduce_max(axes = var_8097_axes_0, keep_dims = var_8096, x = var_8093_cast_fp16)[name = string("op_8097_cast_fp16")];
            fp16 var_8098_to_fp16 = const()[name = string("op_8098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_289_cast_fp16 = maximum(x = var_8097_cast_fp16, y = var_8098_to_fp16)[name = string("s_289_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_289_cast_fp16 = real_div(x = linear_162_cast_fp16, y = s_289_cast_fp16)[name = string("y_289_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8101_cast_fp16 = mul(x = y_289_cast_fp16, y = y_289_cast_fp16)[name = string("op_8101_cast_fp16")];
            tensor<int32, [1]> var_8106_axes_0 = const()[name = string("op_8106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8106_keep_dims_0 = const()[name = string("op_8106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8106_cast_fp16 = reduce_mean(axes = var_8106_axes_0, keep_dims = var_8106_keep_dims_0, x = var_8101_cast_fp16)[name = string("op_8106_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8107_cast_fp16 = mul(x = s_289_cast_fp16, y = s_289_cast_fp16)[name = string("op_8107_cast_fp16")];
            fp32 var_8108_epsilon_0 = const()[name = string("op_8108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8108_cast_fp16 = inverse(epsilon = var_8108_epsilon_0, x = var_8107_cast_fp16)[name = string("op_8108_cast_fp16")];
            fp16 var_8109_to_fp16 = const()[name = string("op_8109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8110_cast_fp16 = mul(x = var_8108_cast_fp16, y = var_8109_to_fp16)[name = string("op_8110_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8112_cast_fp16 = add(x = var_8106_cast_fp16, y = var_8110_cast_fp16)[name = string("op_8112_cast_fp16")];
            fp32 var_8113_epsilon_0 = const()[name = string("op_8113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8113_cast_fp16 = rsqrt(epsilon = var_8113_epsilon_0, x = var_8112_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_253_cast_fp16 = mul(x = y_289_cast_fp16, y = var_8113_cast_fp16)[name = string("out_253_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104305792)))];
            tensor<fp16, [1, 32, 512]> var_8115_cast_fp16 = mul(x = out_253_cast_fp16, y = tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8115_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_35_cast_fp16 = add(x = input_213_cast_fp16, y = var_8115_cast_fp16)[name = string("h_35_cast_fp16")];
            tensor<fp16, [1]> tm_layers_17_layer_scalar_to_fp16 = const()[name = string("tm_layers_17_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e6p-1])];
            tensor<fp16, [1, 32, 512]> x_325_cast_fp16 = mul(x = h_35_cast_fp16, y = tm_layers_17_layer_scalar_to_fp16)[name = string("x_325_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8119_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_8119_cast_fp16")];
            bool var_8122 = const()[name = string("op_8122"), val = bool(true)];
            tensor<int32, [1]> var_8123_axes_0 = const()[name = string("op_8123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8123_cast_fp16 = reduce_max(axes = var_8123_axes_0, keep_dims = var_8122, x = var_8119_cast_fp16)[name = string("op_8123_cast_fp16")];
            fp16 var_8124_to_fp16 = const()[name = string("op_8124_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_291_cast_fp16 = maximum(x = var_8123_cast_fp16, y = var_8124_to_fp16)[name = string("s_291_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_291_cast_fp16 = real_div(x = x_325_cast_fp16, y = s_291_cast_fp16)[name = string("y_291_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8127_cast_fp16 = mul(x = y_291_cast_fp16, y = y_291_cast_fp16)[name = string("op_8127_cast_fp16")];
            tensor<int32, [1]> var_8132_axes_0 = const()[name = string("op_8132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8132_keep_dims_0 = const()[name = string("op_8132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8132_cast_fp16 = reduce_mean(axes = var_8132_axes_0, keep_dims = var_8132_keep_dims_0, x = var_8127_cast_fp16)[name = string("op_8132_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8133_cast_fp16 = mul(x = s_291_cast_fp16, y = s_291_cast_fp16)[name = string("op_8133_cast_fp16")];
            fp32 var_8134_epsilon_0 = const()[name = string("op_8134_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8134_cast_fp16 = inverse(epsilon = var_8134_epsilon_0, x = var_8133_cast_fp16)[name = string("op_8134_cast_fp16")];
            fp16 var_8135_to_fp16 = const()[name = string("op_8135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = string("op_8136_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8138_cast_fp16 = add(x = var_8132_cast_fp16, y = var_8136_cast_fp16)[name = string("op_8138_cast_fp16")];
            fp32 var_8139_epsilon_0 = const()[name = string("op_8139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8139_cast_fp16 = rsqrt(epsilon = var_8139_epsilon_0, x = var_8138_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_255_cast_fp16 = mul(x = y_291_cast_fp16, y = var_8139_cast_fp16)[name = string("out_255_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104306880)))];
            tensor<fp16, [1, 32, 512]> input_217_cast_fp16 = mul(x = out_255_cast_fp16, y = tm_layers_18_input_layernorm_weight_to_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104307968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104832320))))[name = string("tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8149 = const()[name = string("op_8149"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_8150_cast_fp16 = reshape(shape = var_8149, x = linear_163_cast_fp16)[name = string("op_8150_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104834432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105096640))))[name = string("tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [4]> var_8161 = const()[name = string("op_8161"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_8162_cast_fp16 = reshape(shape = var_8161, x = linear_164_cast_fp16)[name = string("op_8162_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105097728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105359936))))[name = string("tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<int32, [4]> var_8173 = const()[name = string("op_8173"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_8174_cast_fp16 = reshape(shape = var_8173, x = linear_165_cast_fp16)[name = string("op_8174_cast_fp16")];
            tensor<int32, [4]> x_331_perm_0 = const()[name = string("x_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_8150_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 4, 32, 256]> var_8178_cast_fp16 = abs(x = x_327_cast_fp16)[name = string("op_8178_cast_fp16")];
            bool var_8181 = const()[name = string("op_8181"), val = bool(true)];
            tensor<int32, [1]> var_8182_axes_0 = const()[name = string("op_8182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_8182_cast_fp16 = reduce_max(axes = var_8182_axes_0, keep_dims = var_8181, x = var_8178_cast_fp16)[name = string("op_8182_cast_fp16")];
            fp16 var_8183_to_fp16 = const()[name = string("op_8183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_293_cast_fp16 = maximum(x = var_8182_cast_fp16, y = var_8183_to_fp16)[name = string("s_293_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_293_cast_fp16 = real_div(x = x_327_cast_fp16, y = s_293_cast_fp16)[name = string("y_293_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8186_cast_fp16 = mul(x = y_293_cast_fp16, y = y_293_cast_fp16)[name = string("op_8186_cast_fp16")];
            tensor<int32, [1]> var_8191_axes_0 = const()[name = string("op_8191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8191_keep_dims_0 = const()[name = string("op_8191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_8191_cast_fp16 = reduce_mean(axes = var_8191_axes_0, keep_dims = var_8191_keep_dims_0, x = var_8186_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_8192_cast_fp16 = mul(x = s_293_cast_fp16, y = s_293_cast_fp16)[name = string("op_8192_cast_fp16")];
            fp32 var_8193_epsilon_0 = const()[name = string("op_8193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_8193_cast_fp16 = inverse(epsilon = var_8193_epsilon_0, x = var_8192_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_8197_cast_fp16 = add(x = var_8191_cast_fp16, y = var_8195_cast_fp16)[name = string("op_8197_cast_fp16")];
            fp32 var_8198_epsilon_0 = const()[name = string("op_8198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_8198_cast_fp16 = rsqrt(epsilon = var_8198_epsilon_0, x = var_8197_cast_fp16)[name = string("op_8198_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_257_cast_fp16 = mul(x = y_293_cast_fp16, y = var_8198_cast_fp16)[name = string("out_257_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_73_cast_fp16 = mul(x = out_257_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_73_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_8162_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 32, 256]> var_8201_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_8201_cast_fp16")];
            bool var_8204 = const()[name = string("op_8204"), val = bool(true)];
            tensor<int32, [1]> var_8205_axes_0 = const()[name = string("op_8205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_8205_cast_fp16 = reduce_max(axes = var_8205_axes_0, keep_dims = var_8204, x = var_8201_cast_fp16)[name = string("op_8205_cast_fp16")];
            fp16 var_8206_to_fp16 = const()[name = string("op_8206_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_295_cast_fp16 = maximum(x = var_8205_cast_fp16, y = var_8206_to_fp16)[name = string("s_295_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_295_cast_fp16 = real_div(x = x_329_cast_fp16, y = s_295_cast_fp16)[name = string("y_295_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8209_cast_fp16 = mul(x = y_295_cast_fp16, y = y_295_cast_fp16)[name = string("op_8209_cast_fp16")];
            tensor<int32, [1]> var_8214_axes_0 = const()[name = string("op_8214_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8214_keep_dims_0 = const()[name = string("op_8214_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_8214_cast_fp16 = reduce_mean(axes = var_8214_axes_0, keep_dims = var_8214_keep_dims_0, x = var_8209_cast_fp16)[name = string("op_8214_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8215_cast_fp16 = mul(x = s_295_cast_fp16, y = s_295_cast_fp16)[name = string("op_8215_cast_fp16")];
            fp32 var_8216_epsilon_0 = const()[name = string("op_8216_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_8216_cast_fp16 = inverse(epsilon = var_8216_epsilon_0, x = var_8215_cast_fp16)[name = string("op_8216_cast_fp16")];
            fp16 var_8217_to_fp16 = const()[name = string("op_8217_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_8218_cast_fp16 = mul(x = var_8216_cast_fp16, y = var_8217_to_fp16)[name = string("op_8218_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8220_cast_fp16 = add(x = var_8214_cast_fp16, y = var_8218_cast_fp16)[name = string("op_8220_cast_fp16")];
            fp32 var_8221_epsilon_0 = const()[name = string("op_8221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_8221_cast_fp16 = rsqrt(epsilon = var_8221_epsilon_0, x = var_8220_cast_fp16)[name = string("op_8221_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_259_cast_fp16 = mul(x = y_295_cast_fp16, y = var_8221_cast_fp16)[name = string("out_259_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_109_cast_fp16 = mul(x = out_259_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_109_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_331_cast_fp16 = transpose(perm = x_331_perm_0, x = var_8174_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 32, 256]> var_8224_cast_fp16 = abs(x = x_331_cast_fp16)[name = string("op_8224_cast_fp16")];
            bool var_8227 = const()[name = string("op_8227"), val = bool(true)];
            tensor<int32, [1]> var_8228_axes_0 = const()[name = string("op_8228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_8228_cast_fp16 = reduce_max(axes = var_8228_axes_0, keep_dims = var_8227, x = var_8224_cast_fp16)[name = string("op_8228_cast_fp16")];
            fp16 var_8229_to_fp16 = const()[name = string("op_8229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_297_cast_fp16 = maximum(x = var_8228_cast_fp16, y = var_8229_to_fp16)[name = string("s_297_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_297_cast_fp16 = real_div(x = x_331_cast_fp16, y = s_297_cast_fp16)[name = string("y_297_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8232_cast_fp16 = mul(x = y_297_cast_fp16, y = y_297_cast_fp16)[name = string("op_8232_cast_fp16")];
            tensor<int32, [1]> var_8237_axes_0 = const()[name = string("op_8237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8237_keep_dims_0 = const()[name = string("op_8237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_8237_cast_fp16 = reduce_mean(axes = var_8237_axes_0, keep_dims = var_8237_keep_dims_0, x = var_8232_cast_fp16)[name = string("op_8237_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8238_cast_fp16 = mul(x = s_297_cast_fp16, y = s_297_cast_fp16)[name = string("op_8238_cast_fp16")];
            fp32 var_8239_epsilon_0 = const()[name = string("op_8239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_8239_cast_fp16 = inverse(epsilon = var_8239_epsilon_0, x = var_8238_cast_fp16)[name = string("op_8239_cast_fp16")];
            fp16 var_8240_to_fp16 = const()[name = string("op_8240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_8241_cast_fp16 = mul(x = var_8239_cast_fp16, y = var_8240_to_fp16)[name = string("op_8241_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8243_cast_fp16 = add(x = var_8237_cast_fp16, y = var_8241_cast_fp16)[name = string("op_8243_cast_fp16")];
            fp32 var_8244_epsilon_0 = const()[name = string("op_8244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_8244_cast_fp16 = rsqrt(epsilon = var_8244_epsilon_0, x = var_8243_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_73_cast_fp16 = mul(x = y_297_cast_fp16, y = var_8244_cast_fp16)[name = string("v_73_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8246_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8246_cast_fp16")];
            tensor<int32, [4]> var_8251_begin_0 = const()[name = string("op_8251_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8251_end_0 = const()[name = string("op_8251_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_8251_end_mask_0 = const()[name = string("op_8251_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_8251_cast_fp16 = slice_by_index(begin = var_8251_begin_0, end = var_8251_end_0, end_mask = var_8251_end_mask_0, x = q_73_cast_fp16)[name = string("op_8251_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_8252_cast_fp16 = mul(x = var_8251_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8257_begin_0 = const()[name = string("op_8257_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8257_end_0 = const()[name = string("op_8257_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_8257_end_mask_0 = const()[name = string("op_8257_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_8257_cast_fp16 = slice_by_index(begin = var_8257_begin_0, end = var_8257_end_0, end_mask = var_8257_end_mask_0, x = q_73_cast_fp16)[name = string("op_8257_cast_fp16")];
            int32 var_8259 = const()[name = string("op_8259"), val = int32(-1)];
            bool var_8260_interleave_0 = const()[name = string("op_8260_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_8260_cast_fp16 = concat(axis = var_8259, interleave = var_8260_interleave_0, values = (var_8252_cast_fp16, var_8257_cast_fp16))[name = string("op_8260_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8261_cast_fp16 = mul(x = var_8260_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8261_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_75_cast_fp16 = add(x = var_8246_cast_fp16, y = var_8261_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8264_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8264_cast_fp16")];
            tensor<int32, [4]> var_8269_begin_0 = const()[name = string("op_8269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8269_end_0 = const()[name = string("op_8269_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_8269_end_mask_0 = const()[name = string("op_8269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_8269_cast_fp16 = slice_by_index(begin = var_8269_begin_0, end = var_8269_end_0, end_mask = var_8269_end_mask_0, x = k_109_cast_fp16)[name = string("op_8269_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_8270_cast_fp16 = mul(x = var_8269_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_8270_cast_fp16")];
            tensor<int32, [4]> var_8275_begin_0 = const()[name = string("op_8275_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8275_end_0 = const()[name = string("op_8275_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_8275_end_mask_0 = const()[name = string("op_8275_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_8275_cast_fp16 = slice_by_index(begin = var_8275_begin_0, end = var_8275_end_0, end_mask = var_8275_end_mask_0, x = k_109_cast_fp16)[name = string("op_8275_cast_fp16")];
            int32 var_8277 = const()[name = string("op_8277"), val = int32(-1)];
            bool var_8278_interleave_0 = const()[name = string("op_8278_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_8278_cast_fp16 = concat(axis = var_8277, interleave = var_8278_interleave_0, values = (var_8270_cast_fp16, var_8275_cast_fp16))[name = string("op_8278_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8279_cast_fp16 = mul(x = var_8278_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8279_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_111_cast_fp16 = add(x = var_8264_cast_fp16, y = var_8279_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_36_reps_0 = const()[name = string("tile_36_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_72_cast_fp16 = transpose(perm = transpose_72_perm_0, x = k_111_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 1, 32, 256]> tile_36_cast_fp16 = tile(reps = tile_36_reps_0, x = transpose_72_cast_fp16)[name = string("tile_36_cast_fp16")];
            tensor<int32, [5]> concat_72 = const()[name = string("concat_72"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_72_cast_fp16 = reshape(shape = concat_72, x = tile_36_cast_fp16)[name = string("reshape_72_cast_fp16")];
            tensor<int32, [5]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_73 = const()[name = string("concat_73"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = reshape_72_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 1, 32, 256]> reshape_73_cast_fp16 = reshape(shape = concat_73, x = transpose_73_cast_fp16)[name = string("reshape_73_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_37_reps_0 = const()[name = string("tile_37_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = v_73_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 1, 32, 256]> tile_37_cast_fp16 = tile(reps = tile_37_reps_0, x = transpose_74_cast_fp16)[name = string("tile_37_cast_fp16")];
            tensor<int32, [5]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_74_cast_fp16 = reshape(shape = concat_74, x = tile_37_cast_fp16)[name = string("reshape_74_cast_fp16")];
            tensor<int32, [5]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = reshape_74_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 1, 32, 256]> reshape_75_cast_fp16 = reshape(shape = concat_75, x = transpose_75_cast_fp16)[name = string("reshape_75_cast_fp16")];
            tensor<int32, [4]> v_75_perm_0 = const()[name = string("v_75_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8293_transpose_x_0 = const()[name = string("op_8293_transpose_x_0"), val = bool(false)];
            bool var_8293_transpose_y_0 = const()[name = string("op_8293_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = reshape_73_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 4, 32, 32]> var_8293_cast_fp16 = matmul(transpose_x = var_8293_transpose_x_0, transpose_y = var_8293_transpose_y_0, x = q_75_cast_fp16, y = transpose_114_cast_fp16)[name = string("op_8293_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_8295_cast_fp16 = add(x = var_8293_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8295_cast_fp16")];
            int32 var_8296 = const()[name = string("op_8296"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_297_cast_fp16 = softmax(axis = var_8296, x = var_8295_cast_fp16)[name = string("w_297_cast_fp16")];
            bool var_8299_transpose_x_0 = const()[name = string("op_8299_transpose_x_0"), val = bool(false)];
            bool var_8299_transpose_y_0 = const()[name = string("op_8299_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_75_cast_fp16 = transpose(perm = v_75_perm_0, x = reshape_75_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 4, 32, 256]> var_8299_cast_fp16 = matmul(transpose_x = var_8299_transpose_x_0, transpose_y = var_8299_transpose_y_0, x = w_297_cast_fp16, y = v_75_cast_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [4]> var_8302_perm_0 = const()[name = string("op_8302_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8306 = const()[name = string("op_8306"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_8302_cast_fp16 = transpose(perm = var_8302_perm_0, x = var_8299_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 32, 1024]> input_219_cast_fp16 = reshape(shape = var_8306, x = var_8302_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105361024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105885376))))[name = string("tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8311_cast_fp16 = abs(x = linear_166_cast_fp16)[name = string("op_8311_cast_fp16")];
            bool var_8314 = const()[name = string("op_8314"), val = bool(true)];
            tensor<int32, [1]> var_8315_axes_0 = const()[name = string("op_8315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8315_cast_fp16 = reduce_max(axes = var_8315_axes_0, keep_dims = var_8314, x = var_8311_cast_fp16)[name = string("op_8315_cast_fp16")];
            fp16 var_8316_to_fp16 = const()[name = string("op_8316_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_299_cast_fp16 = maximum(x = var_8315_cast_fp16, y = var_8316_to_fp16)[name = string("s_299_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_299_cast_fp16 = real_div(x = linear_166_cast_fp16, y = s_299_cast_fp16)[name = string("y_299_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8319_cast_fp16 = mul(x = y_299_cast_fp16, y = y_299_cast_fp16)[name = string("op_8319_cast_fp16")];
            tensor<int32, [1]> var_8324_axes_0 = const()[name = string("op_8324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8324_keep_dims_0 = const()[name = string("op_8324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8324_cast_fp16 = reduce_mean(axes = var_8324_axes_0, keep_dims = var_8324_keep_dims_0, x = var_8319_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8325_cast_fp16 = mul(x = s_299_cast_fp16, y = s_299_cast_fp16)[name = string("op_8325_cast_fp16")];
            fp32 var_8326_epsilon_0 = const()[name = string("op_8326_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8326_cast_fp16 = inverse(epsilon = var_8326_epsilon_0, x = var_8325_cast_fp16)[name = string("op_8326_cast_fp16")];
            fp16 var_8327_to_fp16 = const()[name = string("op_8327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8328_cast_fp16 = mul(x = var_8326_cast_fp16, y = var_8327_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8330_cast_fp16 = add(x = var_8324_cast_fp16, y = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_261_cast_fp16 = mul(x = y_299_cast_fp16, y = var_8331_cast_fp16)[name = string("out_261_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105886464)))];
            tensor<fp16, [1, 32, 512]> var_8333_cast_fp16 = mul(x = out_261_cast_fp16, y = tm_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("op_8333_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = var_8333_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8336_cast_fp16 = abs(x = x_335_cast_fp16)[name = string("op_8336_cast_fp16")];
            bool var_8339 = const()[name = string("op_8339"), val = bool(true)];
            tensor<int32, [1]> var_8340_axes_0 = const()[name = string("op_8340_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8340_cast_fp16 = reduce_max(axes = var_8340_axes_0, keep_dims = var_8339, x = var_8336_cast_fp16)[name = string("op_8340_cast_fp16")];
            fp16 var_8341_to_fp16 = const()[name = string("op_8341_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_301_cast_fp16 = maximum(x = var_8340_cast_fp16, y = var_8341_to_fp16)[name = string("s_301_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_301_cast_fp16 = real_div(x = x_335_cast_fp16, y = s_301_cast_fp16)[name = string("y_301_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8344_cast_fp16 = mul(x = y_301_cast_fp16, y = y_301_cast_fp16)[name = string("op_8344_cast_fp16")];
            tensor<int32, [1]> var_8349_axes_0 = const()[name = string("op_8349_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8349_keep_dims_0 = const()[name = string("op_8349_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8349_cast_fp16 = reduce_mean(axes = var_8349_axes_0, keep_dims = var_8349_keep_dims_0, x = var_8344_cast_fp16)[name = string("op_8349_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8350_cast_fp16 = mul(x = s_301_cast_fp16, y = s_301_cast_fp16)[name = string("op_8350_cast_fp16")];
            fp32 var_8351_epsilon_0 = const()[name = string("op_8351_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8351_cast_fp16 = inverse(epsilon = var_8351_epsilon_0, x = var_8350_cast_fp16)[name = string("op_8351_cast_fp16")];
            fp16 var_8352_to_fp16 = const()[name = string("op_8352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8353_cast_fp16 = mul(x = var_8351_cast_fp16, y = var_8352_to_fp16)[name = string("op_8353_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8355_cast_fp16 = add(x = var_8349_cast_fp16, y = var_8353_cast_fp16)[name = string("op_8355_cast_fp16")];
            fp32 var_8356_epsilon_0 = const()[name = string("op_8356_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8356_cast_fp16 = rsqrt(epsilon = var_8356_epsilon_0, x = var_8355_cast_fp16)[name = string("op_8356_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_263_cast_fp16 = mul(x = y_301_cast_fp16, y = var_8356_cast_fp16)[name = string("out_263_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105887552)))];
            tensor<fp16, [1, 32, 512]> input_221_cast_fp16 = mul(x = out_263_cast_fp16, y = tm_layers_18_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105888640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106937280))))[name = string("tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_167_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_167_cast_fp16")];
            string var_8363_mode_0 = const()[name = string("op_8363_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_8363_cast_fp16 = gelu(mode = var_8363_mode_0, x = linear_167_cast_fp16)[name = string("op_8363_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106941440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107990080))))[name = string("tm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_168_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_223_cast_fp16 = mul(x = var_8363_cast_fp16, y = linear_168_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107994240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109042880))))[name = string("tm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8371_cast_fp16 = abs(x = linear_169_cast_fp16)[name = string("op_8371_cast_fp16")];
            bool var_8374 = const()[name = string("op_8374"), val = bool(true)];
            tensor<int32, [1]> var_8375_axes_0 = const()[name = string("op_8375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8375_cast_fp16 = reduce_max(axes = var_8375_axes_0, keep_dims = var_8374, x = var_8371_cast_fp16)[name = string("op_8375_cast_fp16")];
            fp16 var_8376_to_fp16 = const()[name = string("op_8376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_303_cast_fp16 = maximum(x = var_8375_cast_fp16, y = var_8376_to_fp16)[name = string("s_303_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_303_cast_fp16 = real_div(x = linear_169_cast_fp16, y = s_303_cast_fp16)[name = string("y_303_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8379_cast_fp16 = mul(x = y_303_cast_fp16, y = y_303_cast_fp16)[name = string("op_8379_cast_fp16")];
            tensor<int32, [1]> var_8384_axes_0 = const()[name = string("op_8384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8384_keep_dims_0 = const()[name = string("op_8384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8384_cast_fp16 = reduce_mean(axes = var_8384_axes_0, keep_dims = var_8384_keep_dims_0, x = var_8379_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8385_cast_fp16 = mul(x = s_303_cast_fp16, y = s_303_cast_fp16)[name = string("op_8385_cast_fp16")];
            fp32 var_8386_epsilon_0 = const()[name = string("op_8386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8386_cast_fp16 = inverse(epsilon = var_8386_epsilon_0, x = var_8385_cast_fp16)[name = string("op_8386_cast_fp16")];
            fp16 var_8387_to_fp16 = const()[name = string("op_8387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8388_cast_fp16 = mul(x = var_8386_cast_fp16, y = var_8387_to_fp16)[name = string("op_8388_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8390_cast_fp16 = add(x = var_8384_cast_fp16, y = var_8388_cast_fp16)[name = string("op_8390_cast_fp16")];
            fp32 var_8391_epsilon_0 = const()[name = string("op_8391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8391_cast_fp16 = rsqrt(epsilon = var_8391_epsilon_0, x = var_8390_cast_fp16)[name = string("op_8391_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_265_cast_fp16 = mul(x = y_303_cast_fp16, y = var_8391_cast_fp16)[name = string("out_265_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109043968)))];
            tensor<fp16, [1, 32, 512]> var_8393_cast_fp16 = mul(x = out_265_cast_fp16, y = tm_layers_18_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8393_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_225_cast_fp16 = add(x = x_335_cast_fp16, y = var_8393_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109045056))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109307264))))[name = string("tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_170_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_170_cast_fp16")];
            string x_339_mode_0 = const()[name = string("x_339_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_339_cast_fp16 = gelu(mode = x_339_mode_0, x = linear_170_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [4]> var_8413_begin_0 = const()[name = string("op_8413_begin_0"), val = tensor<int32, [4]>([0, 0, 18, 0])];
            tensor<int32, [4]> var_8413_end_0 = const()[name = string("op_8413_end_0"), val = tensor<int32, [4]>([1, 32, 19, 512])];
            tensor<bool, [4]> var_8413_end_mask_0 = const()[name = string("op_8413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8413_squeeze_mask_0 = const()[name = string("op_8413_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_8413_cast_fp16 = slice_by_index(begin = var_8413_begin_0, end = var_8413_end_0, end_mask = var_8413_end_mask_0, squeeze_mask = var_8413_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_227_cast_fp16 = mul(x = x_339_cast_fp16, y = var_8413_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109308352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109570560))))[name = string("tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8423_cast_fp16 = abs(x = linear_171_cast_fp16)[name = string("op_8423_cast_fp16")];
            bool var_8426 = const()[name = string("op_8426"), val = bool(true)];
            tensor<int32, [1]> var_8427_axes_0 = const()[name = string("op_8427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8427_cast_fp16 = reduce_max(axes = var_8427_axes_0, keep_dims = var_8426, x = var_8423_cast_fp16)[name = string("op_8427_cast_fp16")];
            fp16 var_8428_to_fp16 = const()[name = string("op_8428_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_305_cast_fp16 = maximum(x = var_8427_cast_fp16, y = var_8428_to_fp16)[name = string("s_305_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_305_cast_fp16 = real_div(x = linear_171_cast_fp16, y = s_305_cast_fp16)[name = string("y_305_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8431_cast_fp16 = mul(x = y_305_cast_fp16, y = y_305_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<int32, [1]> var_8436_axes_0 = const()[name = string("op_8436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8436_keep_dims_0 = const()[name = string("op_8436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8436_cast_fp16 = reduce_mean(axes = var_8436_axes_0, keep_dims = var_8436_keep_dims_0, x = var_8431_cast_fp16)[name = string("op_8436_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8437_cast_fp16 = mul(x = s_305_cast_fp16, y = s_305_cast_fp16)[name = string("op_8437_cast_fp16")];
            fp32 var_8438_epsilon_0 = const()[name = string("op_8438_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8438_cast_fp16 = inverse(epsilon = var_8438_epsilon_0, x = var_8437_cast_fp16)[name = string("op_8438_cast_fp16")];
            fp16 var_8439_to_fp16 = const()[name = string("op_8439_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8440_cast_fp16 = mul(x = var_8438_cast_fp16, y = var_8439_to_fp16)[name = string("op_8440_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8442_cast_fp16 = add(x = var_8436_cast_fp16, y = var_8440_cast_fp16)[name = string("op_8442_cast_fp16")];
            fp32 var_8443_epsilon_0 = const()[name = string("op_8443_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8443_cast_fp16 = rsqrt(epsilon = var_8443_epsilon_0, x = var_8442_cast_fp16)[name = string("op_8443_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_267_cast_fp16 = mul(x = y_305_cast_fp16, y = var_8443_cast_fp16)[name = string("out_267_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109571648)))];
            tensor<fp16, [1, 32, 512]> var_8445_cast_fp16 = mul(x = out_267_cast_fp16, y = tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8445_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_37_cast_fp16 = add(x = input_225_cast_fp16, y = var_8445_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<fp16, [1]> tm_layers_18_layer_scalar_to_fp16 = const()[name = string("tm_layers_18_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.68p-1])];
            tensor<fp16, [1, 32, 512]> x_343_cast_fp16 = mul(x = h_37_cast_fp16, y = tm_layers_18_layer_scalar_to_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8449_cast_fp16 = abs(x = x_343_cast_fp16)[name = string("op_8449_cast_fp16")];
            bool var_8452 = const()[name = string("op_8452"), val = bool(true)];
            tensor<int32, [1]> var_8453_axes_0 = const()[name = string("op_8453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8453_cast_fp16 = reduce_max(axes = var_8453_axes_0, keep_dims = var_8452, x = var_8449_cast_fp16)[name = string("op_8453_cast_fp16")];
            fp16 var_8454_to_fp16 = const()[name = string("op_8454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_307_cast_fp16 = maximum(x = var_8453_cast_fp16, y = var_8454_to_fp16)[name = string("s_307_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_307_cast_fp16 = real_div(x = x_343_cast_fp16, y = s_307_cast_fp16)[name = string("y_307_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8457_cast_fp16 = mul(x = y_307_cast_fp16, y = y_307_cast_fp16)[name = string("op_8457_cast_fp16")];
            tensor<int32, [1]> var_8462_axes_0 = const()[name = string("op_8462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8462_keep_dims_0 = const()[name = string("op_8462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8462_cast_fp16 = reduce_mean(axes = var_8462_axes_0, keep_dims = var_8462_keep_dims_0, x = var_8457_cast_fp16)[name = string("op_8462_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8463_cast_fp16 = mul(x = s_307_cast_fp16, y = s_307_cast_fp16)[name = string("op_8463_cast_fp16")];
            fp32 var_8464_epsilon_0 = const()[name = string("op_8464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8464_cast_fp16 = inverse(epsilon = var_8464_epsilon_0, x = var_8463_cast_fp16)[name = string("op_8464_cast_fp16")];
            fp16 var_8465_to_fp16 = const()[name = string("op_8465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8466_cast_fp16 = mul(x = var_8464_cast_fp16, y = var_8465_to_fp16)[name = string("op_8466_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8468_cast_fp16 = add(x = var_8462_cast_fp16, y = var_8466_cast_fp16)[name = string("op_8468_cast_fp16")];
            fp32 var_8469_epsilon_0 = const()[name = string("op_8469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8469_cast_fp16 = rsqrt(epsilon = var_8469_epsilon_0, x = var_8468_cast_fp16)[name = string("op_8469_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_269_cast_fp16 = mul(x = y_307_cast_fp16, y = var_8469_cast_fp16)[name = string("out_269_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109572736)))];
            tensor<fp16, [1, 32, 512]> input_229_cast_fp16 = mul(x = out_269_cast_fp16, y = tm_layers_19_input_layernorm_weight_to_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109573824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110098176))))[name = string("tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [4]> var_8479 = const()[name = string("op_8479"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_8480_cast_fp16 = reshape(shape = var_8479, x = linear_172_cast_fp16)[name = string("op_8480_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110100288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110362496))))[name = string("tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<int32, [4]> var_8491 = const()[name = string("op_8491"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_8492_cast_fp16 = reshape(shape = var_8491, x = linear_173_cast_fp16)[name = string("op_8492_cast_fp16")];
            tensor<int32, [4]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110363584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110625792))))[name = string("tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<int32, [4]> var_8503 = const()[name = string("op_8503"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_8504_cast_fp16 = reshape(shape = var_8503, x = linear_174_cast_fp16)[name = string("op_8504_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_8480_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 4, 32, 256]> var_8508_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_8508_cast_fp16")];
            bool var_8511 = const()[name = string("op_8511"), val = bool(true)];
            tensor<int32, [1]> var_8512_axes_0 = const()[name = string("op_8512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_8512_cast_fp16 = reduce_max(axes = var_8512_axes_0, keep_dims = var_8511, x = var_8508_cast_fp16)[name = string("op_8512_cast_fp16")];
            fp16 var_8513_to_fp16 = const()[name = string("op_8513_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_309_cast_fp16 = maximum(x = var_8512_cast_fp16, y = var_8513_to_fp16)[name = string("s_309_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_309_cast_fp16 = real_div(x = x_345_cast_fp16, y = s_309_cast_fp16)[name = string("y_309_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8516_cast_fp16 = mul(x = y_309_cast_fp16, y = y_309_cast_fp16)[name = string("op_8516_cast_fp16")];
            tensor<int32, [1]> var_8521_axes_0 = const()[name = string("op_8521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8521_keep_dims_0 = const()[name = string("op_8521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_8521_cast_fp16 = reduce_mean(axes = var_8521_axes_0, keep_dims = var_8521_keep_dims_0, x = var_8516_cast_fp16)[name = string("op_8521_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_8522_cast_fp16 = mul(x = s_309_cast_fp16, y = s_309_cast_fp16)[name = string("op_8522_cast_fp16")];
            fp32 var_8523_epsilon_0 = const()[name = string("op_8523_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_8523_cast_fp16 = inverse(epsilon = var_8523_epsilon_0, x = var_8522_cast_fp16)[name = string("op_8523_cast_fp16")];
            fp16 var_8524_to_fp16 = const()[name = string("op_8524_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_8525_cast_fp16 = mul(x = var_8523_cast_fp16, y = var_8524_to_fp16)[name = string("op_8525_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_8527_cast_fp16 = add(x = var_8521_cast_fp16, y = var_8525_cast_fp16)[name = string("op_8527_cast_fp16")];
            fp32 var_8528_epsilon_0 = const()[name = string("op_8528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_8528_cast_fp16 = rsqrt(epsilon = var_8528_epsilon_0, x = var_8527_cast_fp16)[name = string("op_8528_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_271_cast_fp16 = mul(x = y_309_cast_fp16, y = var_8528_cast_fp16)[name = string("out_271_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_77_cast_fp16 = mul(x = out_271_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_77_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = var_8492_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 32, 256]> var_8531_cast_fp16 = abs(x = x_347_cast_fp16)[name = string("op_8531_cast_fp16")];
            bool var_8534 = const()[name = string("op_8534"), val = bool(true)];
            tensor<int32, [1]> var_8535_axes_0 = const()[name = string("op_8535_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_8535_cast_fp16 = reduce_max(axes = var_8535_axes_0, keep_dims = var_8534, x = var_8531_cast_fp16)[name = string("op_8535_cast_fp16")];
            fp16 var_8536_to_fp16 = const()[name = string("op_8536_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_311_cast_fp16 = maximum(x = var_8535_cast_fp16, y = var_8536_to_fp16)[name = string("s_311_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_311_cast_fp16 = real_div(x = x_347_cast_fp16, y = s_311_cast_fp16)[name = string("y_311_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8539_cast_fp16 = mul(x = y_311_cast_fp16, y = y_311_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<int32, [1]> var_8544_axes_0 = const()[name = string("op_8544_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8544_keep_dims_0 = const()[name = string("op_8544_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_8544_cast_fp16 = reduce_mean(axes = var_8544_axes_0, keep_dims = var_8544_keep_dims_0, x = var_8539_cast_fp16)[name = string("op_8544_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8545_cast_fp16 = mul(x = s_311_cast_fp16, y = s_311_cast_fp16)[name = string("op_8545_cast_fp16")];
            fp32 var_8546_epsilon_0 = const()[name = string("op_8546_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_8546_cast_fp16 = inverse(epsilon = var_8546_epsilon_0, x = var_8545_cast_fp16)[name = string("op_8546_cast_fp16")];
            fp16 var_8547_to_fp16 = const()[name = string("op_8547_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_8548_cast_fp16 = mul(x = var_8546_cast_fp16, y = var_8547_to_fp16)[name = string("op_8548_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8550_cast_fp16 = add(x = var_8544_cast_fp16, y = var_8548_cast_fp16)[name = string("op_8550_cast_fp16")];
            fp32 var_8551_epsilon_0 = const()[name = string("op_8551_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_8551_cast_fp16 = rsqrt(epsilon = var_8551_epsilon_0, x = var_8550_cast_fp16)[name = string("op_8551_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_273_cast_fp16 = mul(x = y_311_cast_fp16, y = var_8551_cast_fp16)[name = string("out_273_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_115_cast_fp16 = mul(x = out_273_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_115_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_8504_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 2, 32, 256]> var_8554_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_8554_cast_fp16")];
            bool var_8557 = const()[name = string("op_8557"), val = bool(true)];
            tensor<int32, [1]> var_8558_axes_0 = const()[name = string("op_8558_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_8558_cast_fp16 = reduce_max(axes = var_8558_axes_0, keep_dims = var_8557, x = var_8554_cast_fp16)[name = string("op_8558_cast_fp16")];
            fp16 var_8559_to_fp16 = const()[name = string("op_8559_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_313_cast_fp16 = maximum(x = var_8558_cast_fp16, y = var_8559_to_fp16)[name = string("s_313_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_313_cast_fp16 = real_div(x = x_349_cast_fp16, y = s_313_cast_fp16)[name = string("y_313_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8562_cast_fp16 = mul(x = y_313_cast_fp16, y = y_313_cast_fp16)[name = string("op_8562_cast_fp16")];
            tensor<int32, [1]> var_8567_axes_0 = const()[name = string("op_8567_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8567_keep_dims_0 = const()[name = string("op_8567_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_8567_cast_fp16 = reduce_mean(axes = var_8567_axes_0, keep_dims = var_8567_keep_dims_0, x = var_8562_cast_fp16)[name = string("op_8567_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8568_cast_fp16 = mul(x = s_313_cast_fp16, y = s_313_cast_fp16)[name = string("op_8568_cast_fp16")];
            fp32 var_8569_epsilon_0 = const()[name = string("op_8569_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_8569_cast_fp16 = inverse(epsilon = var_8569_epsilon_0, x = var_8568_cast_fp16)[name = string("op_8569_cast_fp16")];
            fp16 var_8570_to_fp16 = const()[name = string("op_8570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_8571_cast_fp16 = mul(x = var_8569_cast_fp16, y = var_8570_to_fp16)[name = string("op_8571_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8573_cast_fp16 = add(x = var_8567_cast_fp16, y = var_8571_cast_fp16)[name = string("op_8573_cast_fp16")];
            fp32 var_8574_epsilon_0 = const()[name = string("op_8574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_8574_cast_fp16 = rsqrt(epsilon = var_8574_epsilon_0, x = var_8573_cast_fp16)[name = string("op_8574_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_77_cast_fp16 = mul(x = y_313_cast_fp16, y = var_8574_cast_fp16)[name = string("v_77_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8576_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8576_cast_fp16")];
            tensor<int32, [4]> var_8581_begin_0 = const()[name = string("op_8581_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8581_end_0 = const()[name = string("op_8581_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_8581_end_mask_0 = const()[name = string("op_8581_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_8581_cast_fp16 = slice_by_index(begin = var_8581_begin_0, end = var_8581_end_0, end_mask = var_8581_end_mask_0, x = q_77_cast_fp16)[name = string("op_8581_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_8582_cast_fp16 = mul(x = var_8581_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_8582_cast_fp16")];
            tensor<int32, [4]> var_8587_begin_0 = const()[name = string("op_8587_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8587_end_0 = const()[name = string("op_8587_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_8587_end_mask_0 = const()[name = string("op_8587_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_8587_cast_fp16 = slice_by_index(begin = var_8587_begin_0, end = var_8587_end_0, end_mask = var_8587_end_mask_0, x = q_77_cast_fp16)[name = string("op_8587_cast_fp16")];
            int32 var_8589 = const()[name = string("op_8589"), val = int32(-1)];
            bool var_8590_interleave_0 = const()[name = string("op_8590_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_8590_cast_fp16 = concat(axis = var_8589, interleave = var_8590_interleave_0, values = (var_8582_cast_fp16, var_8587_cast_fp16))[name = string("op_8590_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8591_cast_fp16 = mul(x = var_8590_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8591_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_79_cast_fp16 = add(x = var_8576_cast_fp16, y = var_8591_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8594_cast_fp16 = mul(x = k_115_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8594_cast_fp16")];
            tensor<int32, [4]> var_8599_begin_0 = const()[name = string("op_8599_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8599_end_0 = const()[name = string("op_8599_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_8599_end_mask_0 = const()[name = string("op_8599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_8599_cast_fp16 = slice_by_index(begin = var_8599_begin_0, end = var_8599_end_0, end_mask = var_8599_end_mask_0, x = k_115_cast_fp16)[name = string("op_8599_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_8600_cast_fp16 = mul(x = var_8599_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_8600_cast_fp16")];
            tensor<int32, [4]> var_8605_begin_0 = const()[name = string("op_8605_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8605_end_0 = const()[name = string("op_8605_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_8605_end_mask_0 = const()[name = string("op_8605_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_8605_cast_fp16 = slice_by_index(begin = var_8605_begin_0, end = var_8605_end_0, end_mask = var_8605_end_mask_0, x = k_115_cast_fp16)[name = string("op_8605_cast_fp16")];
            int32 var_8607 = const()[name = string("op_8607"), val = int32(-1)];
            bool var_8608_interleave_0 = const()[name = string("op_8608_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_8608_cast_fp16 = concat(axis = var_8607, interleave = var_8608_interleave_0, values = (var_8600_cast_fp16, var_8605_cast_fp16))[name = string("op_8608_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8609_cast_fp16 = mul(x = var_8608_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8609_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_117_cast_fp16 = add(x = var_8594_cast_fp16, y = var_8609_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_38_reps_0 = const()[name = string("tile_38_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = k_117_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 1, 32, 256]> tile_38_cast_fp16 = tile(reps = tile_38_reps_0, x = transpose_76_cast_fp16)[name = string("tile_38_cast_fp16")];
            tensor<int32, [5]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_76_cast_fp16 = reshape(shape = concat_76, x = tile_38_cast_fp16)[name = string("reshape_76_cast_fp16")];
            tensor<int32, [5]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_77 = const()[name = string("concat_77"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_77_cast_fp16 = transpose(perm = transpose_77_perm_0, x = reshape_76_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 1, 32, 256]> reshape_77_cast_fp16 = reshape(shape = concat_77, x = transpose_77_cast_fp16)[name = string("reshape_77_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_39_reps_0 = const()[name = string("tile_39_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_78_cast_fp16 = transpose(perm = transpose_78_perm_0, x = v_77_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 1, 32, 256]> tile_39_cast_fp16 = tile(reps = tile_39_reps_0, x = transpose_78_cast_fp16)[name = string("tile_39_cast_fp16")];
            tensor<int32, [5]> concat_78 = const()[name = string("concat_78"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_78_cast_fp16 = reshape(shape = concat_78, x = tile_39_cast_fp16)[name = string("reshape_78_cast_fp16")];
            tensor<int32, [5]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = reshape_78_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 1, 32, 256]> reshape_79_cast_fp16 = reshape(shape = concat_79, x = transpose_79_cast_fp16)[name = string("reshape_79_cast_fp16")];
            tensor<int32, [4]> v_79_perm_0 = const()[name = string("v_79_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8623_transpose_x_0 = const()[name = string("op_8623_transpose_x_0"), val = bool(false)];
            bool var_8623_transpose_y_0 = const()[name = string("op_8623_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = reshape_77_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4, 32, 32]> var_8623_cast_fp16 = matmul(transpose_x = var_8623_transpose_x_0, transpose_y = var_8623_transpose_y_0, x = q_79_cast_fp16, y = transpose_115_cast_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8625_cast_fp16")];
            int32 var_8626 = const()[name = string("op_8626"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_313_cast_fp16 = softmax(axis = var_8626, x = var_8625_cast_fp16)[name = string("w_313_cast_fp16")];
            bool var_8629_transpose_x_0 = const()[name = string("op_8629_transpose_x_0"), val = bool(false)];
            bool var_8629_transpose_y_0 = const()[name = string("op_8629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = reshape_79_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 4, 32, 256]> var_8629_cast_fp16 = matmul(transpose_x = var_8629_transpose_x_0, transpose_y = var_8629_transpose_y_0, x = w_313_cast_fp16, y = v_79_cast_fp16)[name = string("op_8629_cast_fp16")];
            tensor<int32, [4]> var_8632_perm_0 = const()[name = string("op_8632_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_8632_cast_fp16 = transpose(perm = var_8632_perm_0, x = var_8629_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 32, 1024]> input_231_cast_fp16 = reshape(shape = var_8636, x = var_8632_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110626880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111151232))))[name = string("tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8641_cast_fp16 = abs(x = linear_175_cast_fp16)[name = string("op_8641_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<int32, [1]> var_8645_axes_0 = const()[name = string("op_8645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8645_cast_fp16 = reduce_max(axes = var_8645_axes_0, keep_dims = var_8644, x = var_8641_cast_fp16)[name = string("op_8645_cast_fp16")];
            fp16 var_8646_to_fp16 = const()[name = string("op_8646_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_315_cast_fp16 = maximum(x = var_8645_cast_fp16, y = var_8646_to_fp16)[name = string("s_315_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_315_cast_fp16 = real_div(x = linear_175_cast_fp16, y = s_315_cast_fp16)[name = string("y_315_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8649_cast_fp16 = mul(x = y_315_cast_fp16, y = y_315_cast_fp16)[name = string("op_8649_cast_fp16")];
            tensor<int32, [1]> var_8654_axes_0 = const()[name = string("op_8654_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8654_keep_dims_0 = const()[name = string("op_8654_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8654_cast_fp16 = reduce_mean(axes = var_8654_axes_0, keep_dims = var_8654_keep_dims_0, x = var_8649_cast_fp16)[name = string("op_8654_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8655_cast_fp16 = mul(x = s_315_cast_fp16, y = s_315_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp32 var_8656_epsilon_0 = const()[name = string("op_8656_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8656_cast_fp16 = inverse(epsilon = var_8656_epsilon_0, x = var_8655_cast_fp16)[name = string("op_8656_cast_fp16")];
            fp16 var_8657_to_fp16 = const()[name = string("op_8657_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8658_cast_fp16 = mul(x = var_8656_cast_fp16, y = var_8657_to_fp16)[name = string("op_8658_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8660_cast_fp16 = add(x = var_8654_cast_fp16, y = var_8658_cast_fp16)[name = string("op_8660_cast_fp16")];
            fp32 var_8661_epsilon_0 = const()[name = string("op_8661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8661_cast_fp16 = rsqrt(epsilon = var_8661_epsilon_0, x = var_8660_cast_fp16)[name = string("op_8661_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_275_cast_fp16 = mul(x = y_315_cast_fp16, y = var_8661_cast_fp16)[name = string("out_275_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111152320)))];
            tensor<fp16, [1, 32, 512]> var_8663_cast_fp16 = mul(x = out_275_cast_fp16, y = tm_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("op_8663_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_353_cast_fp16 = add(x = x_343_cast_fp16, y = var_8663_cast_fp16)[name = string("x_353_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8666_cast_fp16 = abs(x = x_353_cast_fp16)[name = string("op_8666_cast_fp16")];
            bool var_8669 = const()[name = string("op_8669"), val = bool(true)];
            tensor<int32, [1]> var_8670_axes_0 = const()[name = string("op_8670_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8670_cast_fp16 = reduce_max(axes = var_8670_axes_0, keep_dims = var_8669, x = var_8666_cast_fp16)[name = string("op_8670_cast_fp16")];
            fp16 var_8671_to_fp16 = const()[name = string("op_8671_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_317_cast_fp16 = maximum(x = var_8670_cast_fp16, y = var_8671_to_fp16)[name = string("s_317_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_317_cast_fp16 = real_div(x = x_353_cast_fp16, y = s_317_cast_fp16)[name = string("y_317_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8674_cast_fp16 = mul(x = y_317_cast_fp16, y = y_317_cast_fp16)[name = string("op_8674_cast_fp16")];
            tensor<int32, [1]> var_8679_axes_0 = const()[name = string("op_8679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8679_keep_dims_0 = const()[name = string("op_8679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8679_cast_fp16 = reduce_mean(axes = var_8679_axes_0, keep_dims = var_8679_keep_dims_0, x = var_8674_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8680_cast_fp16 = mul(x = s_317_cast_fp16, y = s_317_cast_fp16)[name = string("op_8680_cast_fp16")];
            fp32 var_8681_epsilon_0 = const()[name = string("op_8681_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8681_cast_fp16 = inverse(epsilon = var_8681_epsilon_0, x = var_8680_cast_fp16)[name = string("op_8681_cast_fp16")];
            fp16 var_8682_to_fp16 = const()[name = string("op_8682_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8683_cast_fp16 = mul(x = var_8681_cast_fp16, y = var_8682_to_fp16)[name = string("op_8683_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8685_cast_fp16 = add(x = var_8679_cast_fp16, y = var_8683_cast_fp16)[name = string("op_8685_cast_fp16")];
            fp32 var_8686_epsilon_0 = const()[name = string("op_8686_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8686_cast_fp16 = rsqrt(epsilon = var_8686_epsilon_0, x = var_8685_cast_fp16)[name = string("op_8686_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_277_cast_fp16 = mul(x = y_317_cast_fp16, y = var_8686_cast_fp16)[name = string("out_277_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111153408)))];
            tensor<fp16, [1, 32, 512]> input_233_cast_fp16 = mul(x = out_277_cast_fp16, y = tm_layers_19_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111154496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112203136))))[name = string("tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_176_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_176_cast_fp16")];
            string var_8693_mode_0 = const()[name = string("op_8693_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_8693_cast_fp16 = gelu(mode = var_8693_mode_0, x = linear_176_cast_fp16)[name = string("op_8693_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112207296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113255936))))[name = string("tm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_177_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_235_cast_fp16 = mul(x = var_8693_cast_fp16, y = linear_177_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113260096))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114308736))))[name = string("tm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8701_cast_fp16 = abs(x = linear_178_cast_fp16)[name = string("op_8701_cast_fp16")];
            bool var_8704 = const()[name = string("op_8704"), val = bool(true)];
            tensor<int32, [1]> var_8705_axes_0 = const()[name = string("op_8705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8705_cast_fp16 = reduce_max(axes = var_8705_axes_0, keep_dims = var_8704, x = var_8701_cast_fp16)[name = string("op_8705_cast_fp16")];
            fp16 var_8706_to_fp16 = const()[name = string("op_8706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_319_cast_fp16 = maximum(x = var_8705_cast_fp16, y = var_8706_to_fp16)[name = string("s_319_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_319_cast_fp16 = real_div(x = linear_178_cast_fp16, y = s_319_cast_fp16)[name = string("y_319_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8709_cast_fp16 = mul(x = y_319_cast_fp16, y = y_319_cast_fp16)[name = string("op_8709_cast_fp16")];
            tensor<int32, [1]> var_8714_axes_0 = const()[name = string("op_8714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8714_keep_dims_0 = const()[name = string("op_8714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8714_cast_fp16 = reduce_mean(axes = var_8714_axes_0, keep_dims = var_8714_keep_dims_0, x = var_8709_cast_fp16)[name = string("op_8714_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8715_cast_fp16 = mul(x = s_319_cast_fp16, y = s_319_cast_fp16)[name = string("op_8715_cast_fp16")];
            fp32 var_8716_epsilon_0 = const()[name = string("op_8716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8716_cast_fp16 = inverse(epsilon = var_8716_epsilon_0, x = var_8715_cast_fp16)[name = string("op_8716_cast_fp16")];
            fp16 var_8717_to_fp16 = const()[name = string("op_8717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8718_cast_fp16 = mul(x = var_8716_cast_fp16, y = var_8717_to_fp16)[name = string("op_8718_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8720_cast_fp16 = add(x = var_8714_cast_fp16, y = var_8718_cast_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8721_cast_fp16 = rsqrt(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_279_cast_fp16 = mul(x = y_319_cast_fp16, y = var_8721_cast_fp16)[name = string("out_279_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114309824)))];
            tensor<fp16, [1, 32, 512]> var_8723_cast_fp16 = mul(x = out_279_cast_fp16, y = tm_layers_19_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_237_cast_fp16 = add(x = x_353_cast_fp16, y = var_8723_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114310912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114573120))))[name = string("tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_179_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_237_cast_fp16)[name = string("linear_179_cast_fp16")];
            string x_357_mode_0 = const()[name = string("x_357_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_357_cast_fp16 = gelu(mode = x_357_mode_0, x = linear_179_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [4]> var_8743_begin_0 = const()[name = string("op_8743_begin_0"), val = tensor<int32, [4]>([0, 0, 19, 0])];
            tensor<int32, [4]> var_8743_end_0 = const()[name = string("op_8743_end_0"), val = tensor<int32, [4]>([1, 32, 20, 512])];
            tensor<bool, [4]> var_8743_end_mask_0 = const()[name = string("op_8743_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8743_squeeze_mask_0 = const()[name = string("op_8743_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_8743_cast_fp16 = slice_by_index(begin = var_8743_begin_0, end = var_8743_end_0, end_mask = var_8743_end_mask_0, squeeze_mask = var_8743_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8743_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_239_cast_fp16 = mul(x = x_357_cast_fp16, y = var_8743_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114574208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114836416))))[name = string("tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8753_cast_fp16 = abs(x = linear_180_cast_fp16)[name = string("op_8753_cast_fp16")];
            bool var_8756 = const()[name = string("op_8756"), val = bool(true)];
            tensor<int32, [1]> var_8757_axes_0 = const()[name = string("op_8757_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8757_cast_fp16 = reduce_max(axes = var_8757_axes_0, keep_dims = var_8756, x = var_8753_cast_fp16)[name = string("op_8757_cast_fp16")];
            fp16 var_8758_to_fp16 = const()[name = string("op_8758_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_321_cast_fp16 = maximum(x = var_8757_cast_fp16, y = var_8758_to_fp16)[name = string("s_321_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_321_cast_fp16 = real_div(x = linear_180_cast_fp16, y = s_321_cast_fp16)[name = string("y_321_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8761_cast_fp16 = mul(x = y_321_cast_fp16, y = y_321_cast_fp16)[name = string("op_8761_cast_fp16")];
            tensor<int32, [1]> var_8766_axes_0 = const()[name = string("op_8766_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8766_keep_dims_0 = const()[name = string("op_8766_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8766_cast_fp16 = reduce_mean(axes = var_8766_axes_0, keep_dims = var_8766_keep_dims_0, x = var_8761_cast_fp16)[name = string("op_8766_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8767_cast_fp16 = mul(x = s_321_cast_fp16, y = s_321_cast_fp16)[name = string("op_8767_cast_fp16")];
            fp32 var_8768_epsilon_0 = const()[name = string("op_8768_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8768_cast_fp16 = inverse(epsilon = var_8768_epsilon_0, x = var_8767_cast_fp16)[name = string("op_8768_cast_fp16")];
            fp16 var_8769_to_fp16 = const()[name = string("op_8769_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8770_cast_fp16 = mul(x = var_8768_cast_fp16, y = var_8769_to_fp16)[name = string("op_8770_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8772_cast_fp16 = add(x = var_8766_cast_fp16, y = var_8770_cast_fp16)[name = string("op_8772_cast_fp16")];
            fp32 var_8773_epsilon_0 = const()[name = string("op_8773_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8773_cast_fp16 = rsqrt(epsilon = var_8773_epsilon_0, x = var_8772_cast_fp16)[name = string("op_8773_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_281_cast_fp16 = mul(x = y_321_cast_fp16, y = var_8773_cast_fp16)[name = string("out_281_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114837504)))];
            tensor<fp16, [1, 32, 512]> var_8775_cast_fp16 = mul(x = out_281_cast_fp16, y = tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_39_cast_fp16 = add(x = input_237_cast_fp16, y = var_8775_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<fp16, [1]> tm_layers_19_layer_scalar_to_fp16 = const()[name = string("tm_layers_19_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fp-1])];
            tensor<fp16, [1, 32, 512]> x_361_cast_fp16 = mul(x = h_39_cast_fp16, y = tm_layers_19_layer_scalar_to_fp16)[name = string("x_361_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8779_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_8779_cast_fp16")];
            bool var_8782 = const()[name = string("op_8782"), val = bool(true)];
            tensor<int32, [1]> var_8783_axes_0 = const()[name = string("op_8783_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8783_cast_fp16 = reduce_max(axes = var_8783_axes_0, keep_dims = var_8782, x = var_8779_cast_fp16)[name = string("op_8783_cast_fp16")];
            fp16 var_8784_to_fp16 = const()[name = string("op_8784_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_323_cast_fp16 = maximum(x = var_8783_cast_fp16, y = var_8784_to_fp16)[name = string("s_323_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_323_cast_fp16 = real_div(x = x_361_cast_fp16, y = s_323_cast_fp16)[name = string("y_323_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8787_cast_fp16 = mul(x = y_323_cast_fp16, y = y_323_cast_fp16)[name = string("op_8787_cast_fp16")];
            tensor<int32, [1]> var_8792_axes_0 = const()[name = string("op_8792_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8792_keep_dims_0 = const()[name = string("op_8792_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8792_cast_fp16 = reduce_mean(axes = var_8792_axes_0, keep_dims = var_8792_keep_dims_0, x = var_8787_cast_fp16)[name = string("op_8792_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8793_cast_fp16 = mul(x = s_323_cast_fp16, y = s_323_cast_fp16)[name = string("op_8793_cast_fp16")];
            fp32 var_8794_epsilon_0 = const()[name = string("op_8794_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8794_cast_fp16 = inverse(epsilon = var_8794_epsilon_0, x = var_8793_cast_fp16)[name = string("op_8794_cast_fp16")];
            fp16 var_8795_to_fp16 = const()[name = string("op_8795_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8796_cast_fp16 = mul(x = var_8794_cast_fp16, y = var_8795_to_fp16)[name = string("op_8796_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8798_cast_fp16 = add(x = var_8792_cast_fp16, y = var_8796_cast_fp16)[name = string("op_8798_cast_fp16")];
            fp32 var_8799_epsilon_0 = const()[name = string("op_8799_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8799_cast_fp16 = rsqrt(epsilon = var_8799_epsilon_0, x = var_8798_cast_fp16)[name = string("op_8799_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_283_cast_fp16 = mul(x = y_323_cast_fp16, y = var_8799_cast_fp16)[name = string("out_283_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114838592)))];
            tensor<fp16, [1, 32, 512]> input_241_cast_fp16 = mul(x = out_283_cast_fp16, y = tm_layers_20_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114839680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115364032))))[name = string("tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<int32, [4]> var_8809 = const()[name = string("op_8809"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_8810_cast_fp16 = reshape(shape = var_8809, x = linear_181_cast_fp16)[name = string("op_8810_cast_fp16")];
            tensor<int32, [4]> x_363_perm_0 = const()[name = string("x_363_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115366144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115628352))))[name = string("tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_8821 = const()[name = string("op_8821"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_8822_cast_fp16 = reshape(shape = var_8821, x = linear_182_cast_fp16)[name = string("op_8822_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115629440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115891648))))[name = string("tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_8833 = const()[name = string("op_8833"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_8834_cast_fp16 = reshape(shape = var_8833, x = linear_183_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<int32, [4]> x_367_perm_0 = const()[name = string("x_367_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_363_cast_fp16 = transpose(perm = x_363_perm_0, x = var_8810_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 4, 32, 256]> var_8838_cast_fp16 = abs(x = x_363_cast_fp16)[name = string("op_8838_cast_fp16")];
            bool var_8841 = const()[name = string("op_8841"), val = bool(true)];
            tensor<int32, [1]> var_8842_axes_0 = const()[name = string("op_8842_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_8842_cast_fp16 = reduce_max(axes = var_8842_axes_0, keep_dims = var_8841, x = var_8838_cast_fp16)[name = string("op_8842_cast_fp16")];
            fp16 var_8843_to_fp16 = const()[name = string("op_8843_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_325_cast_fp16 = maximum(x = var_8842_cast_fp16, y = var_8843_to_fp16)[name = string("s_325_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_325_cast_fp16 = real_div(x = x_363_cast_fp16, y = s_325_cast_fp16)[name = string("y_325_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8846_cast_fp16 = mul(x = y_325_cast_fp16, y = y_325_cast_fp16)[name = string("op_8846_cast_fp16")];
            tensor<int32, [1]> var_8851_axes_0 = const()[name = string("op_8851_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8851_keep_dims_0 = const()[name = string("op_8851_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_8851_cast_fp16 = reduce_mean(axes = var_8851_axes_0, keep_dims = var_8851_keep_dims_0, x = var_8846_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_8852_cast_fp16 = mul(x = s_325_cast_fp16, y = s_325_cast_fp16)[name = string("op_8852_cast_fp16")];
            fp32 var_8853_epsilon_0 = const()[name = string("op_8853_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_8853_cast_fp16 = inverse(epsilon = var_8853_epsilon_0, x = var_8852_cast_fp16)[name = string("op_8853_cast_fp16")];
            fp16 var_8854_to_fp16 = const()[name = string("op_8854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_8855_cast_fp16 = mul(x = var_8853_cast_fp16, y = var_8854_to_fp16)[name = string("op_8855_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_8857_cast_fp16 = add(x = var_8851_cast_fp16, y = var_8855_cast_fp16)[name = string("op_8857_cast_fp16")];
            fp32 var_8858_epsilon_0 = const()[name = string("op_8858_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_8858_cast_fp16 = rsqrt(epsilon = var_8858_epsilon_0, x = var_8857_cast_fp16)[name = string("op_8858_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_285_cast_fp16 = mul(x = y_325_cast_fp16, y = var_8858_cast_fp16)[name = string("out_285_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_81_cast_fp16 = mul(x = out_285_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_81_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_8822_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 32, 256]> var_8861_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_8861_cast_fp16")];
            bool var_8864 = const()[name = string("op_8864"), val = bool(true)];
            tensor<int32, [1]> var_8865_axes_0 = const()[name = string("op_8865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_8865_cast_fp16 = reduce_max(axes = var_8865_axes_0, keep_dims = var_8864, x = var_8861_cast_fp16)[name = string("op_8865_cast_fp16")];
            fp16 var_8866_to_fp16 = const()[name = string("op_8866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_327_cast_fp16 = maximum(x = var_8865_cast_fp16, y = var_8866_to_fp16)[name = string("s_327_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_327_cast_fp16 = real_div(x = x_365_cast_fp16, y = s_327_cast_fp16)[name = string("y_327_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8869_cast_fp16 = mul(x = y_327_cast_fp16, y = y_327_cast_fp16)[name = string("op_8869_cast_fp16")];
            tensor<int32, [1]> var_8874_axes_0 = const()[name = string("op_8874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8874_keep_dims_0 = const()[name = string("op_8874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_8874_cast_fp16 = reduce_mean(axes = var_8874_axes_0, keep_dims = var_8874_keep_dims_0, x = var_8869_cast_fp16)[name = string("op_8874_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8875_cast_fp16 = mul(x = s_327_cast_fp16, y = s_327_cast_fp16)[name = string("op_8875_cast_fp16")];
            fp32 var_8876_epsilon_0 = const()[name = string("op_8876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_8876_cast_fp16 = inverse(epsilon = var_8876_epsilon_0, x = var_8875_cast_fp16)[name = string("op_8876_cast_fp16")];
            fp16 var_8877_to_fp16 = const()[name = string("op_8877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_8878_cast_fp16 = mul(x = var_8876_cast_fp16, y = var_8877_to_fp16)[name = string("op_8878_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8880_cast_fp16 = add(x = var_8874_cast_fp16, y = var_8878_cast_fp16)[name = string("op_8880_cast_fp16")];
            fp32 var_8881_epsilon_0 = const()[name = string("op_8881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_8881_cast_fp16 = rsqrt(epsilon = var_8881_epsilon_0, x = var_8880_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_287_cast_fp16 = mul(x = y_327_cast_fp16, y = var_8881_cast_fp16)[name = string("out_287_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_121_cast_fp16 = mul(x = out_287_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_121_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_367_cast_fp16 = transpose(perm = x_367_perm_0, x = var_8834_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 32, 256]> var_8884_cast_fp16 = abs(x = x_367_cast_fp16)[name = string("op_8884_cast_fp16")];
            bool var_8887 = const()[name = string("op_8887"), val = bool(true)];
            tensor<int32, [1]> var_8888_axes_0 = const()[name = string("op_8888_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_8888_cast_fp16 = reduce_max(axes = var_8888_axes_0, keep_dims = var_8887, x = var_8884_cast_fp16)[name = string("op_8888_cast_fp16")];
            fp16 var_8889_to_fp16 = const()[name = string("op_8889_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_329_cast_fp16 = maximum(x = var_8888_cast_fp16, y = var_8889_to_fp16)[name = string("s_329_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_329_cast_fp16 = real_div(x = x_367_cast_fp16, y = s_329_cast_fp16)[name = string("y_329_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8892_cast_fp16 = mul(x = y_329_cast_fp16, y = y_329_cast_fp16)[name = string("op_8892_cast_fp16")];
            tensor<int32, [1]> var_8897_axes_0 = const()[name = string("op_8897_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8897_keep_dims_0 = const()[name = string("op_8897_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_8897_cast_fp16 = reduce_mean(axes = var_8897_axes_0, keep_dims = var_8897_keep_dims_0, x = var_8892_cast_fp16)[name = string("op_8897_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8898_cast_fp16 = mul(x = s_329_cast_fp16, y = s_329_cast_fp16)[name = string("op_8898_cast_fp16")];
            fp32 var_8899_epsilon_0 = const()[name = string("op_8899_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_8899_cast_fp16 = inverse(epsilon = var_8899_epsilon_0, x = var_8898_cast_fp16)[name = string("op_8899_cast_fp16")];
            fp16 var_8900_to_fp16 = const()[name = string("op_8900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_8901_cast_fp16 = mul(x = var_8899_cast_fp16, y = var_8900_to_fp16)[name = string("op_8901_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_8903_cast_fp16 = add(x = var_8897_cast_fp16, y = var_8901_cast_fp16)[name = string("op_8903_cast_fp16")];
            fp32 var_8904_epsilon_0 = const()[name = string("op_8904_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_8904_cast_fp16 = rsqrt(epsilon = var_8904_epsilon_0, x = var_8903_cast_fp16)[name = string("op_8904_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_81_cast_fp16 = mul(x = y_329_cast_fp16, y = var_8904_cast_fp16)[name = string("v_81_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8906_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8906_cast_fp16")];
            tensor<int32, [4]> var_8911_begin_0 = const()[name = string("op_8911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8911_end_0 = const()[name = string("op_8911_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_8911_end_mask_0 = const()[name = string("op_8911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_8911_cast_fp16 = slice_by_index(begin = var_8911_begin_0, end = var_8911_end_0, end_mask = var_8911_end_mask_0, x = q_81_cast_fp16)[name = string("op_8911_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_8912_cast_fp16 = mul(x = var_8911_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_8912_cast_fp16")];
            tensor<int32, [4]> var_8917_begin_0 = const()[name = string("op_8917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8917_end_0 = const()[name = string("op_8917_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_8917_end_mask_0 = const()[name = string("op_8917_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_8917_cast_fp16 = slice_by_index(begin = var_8917_begin_0, end = var_8917_end_0, end_mask = var_8917_end_mask_0, x = q_81_cast_fp16)[name = string("op_8917_cast_fp16")];
            int32 var_8919 = const()[name = string("op_8919"), val = int32(-1)];
            bool var_8920_interleave_0 = const()[name = string("op_8920_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_8920_cast_fp16 = concat(axis = var_8919, interleave = var_8920_interleave_0, values = (var_8912_cast_fp16, var_8917_cast_fp16))[name = string("op_8920_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_8921_cast_fp16 = mul(x = var_8920_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_83_cast_fp16 = add(x = var_8906_cast_fp16, y = var_8921_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8924_cast_fp16 = mul(x = k_121_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8924_cast_fp16")];
            tensor<int32, [4]> var_8929_begin_0 = const()[name = string("op_8929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8929_end_0 = const()[name = string("op_8929_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_8929_end_mask_0 = const()[name = string("op_8929_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, x = k_121_cast_fp16)[name = string("op_8929_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_8930_cast_fp16 = mul(x = var_8929_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_8930_cast_fp16")];
            tensor<int32, [4]> var_8935_begin_0 = const()[name = string("op_8935_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8935_end_0 = const()[name = string("op_8935_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_8935_end_mask_0 = const()[name = string("op_8935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, x = k_121_cast_fp16)[name = string("op_8935_cast_fp16")];
            int32 var_8937 = const()[name = string("op_8937"), val = int32(-1)];
            bool var_8938_interleave_0 = const()[name = string("op_8938_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_8938_cast_fp16 = concat(axis = var_8937, interleave = var_8938_interleave_0, values = (var_8930_cast_fp16, var_8935_cast_fp16))[name = string("op_8938_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_8939_cast_fp16 = mul(x = var_8938_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8939_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_123_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8939_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_40_reps_0 = const()[name = string("tile_40_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = k_123_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 1, 32, 256]> tile_40_cast_fp16 = tile(reps = tile_40_reps_0, x = transpose_80_cast_fp16)[name = string("tile_40_cast_fp16")];
            tensor<int32, [5]> concat_80 = const()[name = string("concat_80"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_80_cast_fp16 = reshape(shape = concat_80, x = tile_40_cast_fp16)[name = string("reshape_80_cast_fp16")];
            tensor<int32, [5]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_81 = const()[name = string("concat_81"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = reshape_80_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 1, 32, 256]> reshape_81_cast_fp16 = reshape(shape = concat_81, x = transpose_81_cast_fp16)[name = string("reshape_81_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_41_reps_0 = const()[name = string("tile_41_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_82_cast_fp16 = transpose(perm = transpose_82_perm_0, x = v_81_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 1, 32, 256]> tile_41_cast_fp16 = tile(reps = tile_41_reps_0, x = transpose_82_cast_fp16)[name = string("tile_41_cast_fp16")];
            tensor<int32, [5]> concat_82 = const()[name = string("concat_82"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_82_cast_fp16 = reshape(shape = concat_82, x = tile_41_cast_fp16)[name = string("reshape_82_cast_fp16")];
            tensor<int32, [5]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_83 = const()[name = string("concat_83"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = reshape_82_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 1, 32, 256]> reshape_83_cast_fp16 = reshape(shape = concat_83, x = transpose_83_cast_fp16)[name = string("reshape_83_cast_fp16")];
            tensor<int32, [4]> v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8953_transpose_x_0 = const()[name = string("op_8953_transpose_x_0"), val = bool(false)];
            bool var_8953_transpose_y_0 = const()[name = string("op_8953_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = reshape_81_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 4, 32, 32]> var_8953_cast_fp16 = matmul(transpose_x = var_8953_transpose_x_0, transpose_y = var_8953_transpose_y_0, x = q_83_cast_fp16, y = transpose_116_cast_fp16)[name = string("op_8953_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_8955_cast_fp16 = add(x = var_8953_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8955_cast_fp16")];
            int32 var_8956 = const()[name = string("op_8956"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_329_cast_fp16 = softmax(axis = var_8956, x = var_8955_cast_fp16)[name = string("w_329_cast_fp16")];
            bool var_8959_transpose_x_0 = const()[name = string("op_8959_transpose_x_0"), val = bool(false)];
            bool var_8959_transpose_y_0 = const()[name = string("op_8959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_83_cast_fp16 = transpose(perm = v_83_perm_0, x = reshape_83_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 4, 32, 256]> var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = w_329_cast_fp16, y = v_83_cast_fp16)[name = string("op_8959_cast_fp16")];
            tensor<int32, [4]> var_8962_perm_0 = const()[name = string("op_8962_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8966 = const()[name = string("op_8966"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_8962_cast_fp16 = transpose(perm = var_8962_perm_0, x = var_8959_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 32, 1024]> input_243_cast_fp16 = reshape(shape = var_8966, x = var_8962_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115892736))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116417088))))[name = string("tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8971_cast_fp16 = abs(x = linear_184_cast_fp16)[name = string("op_8971_cast_fp16")];
            bool var_8974 = const()[name = string("op_8974"), val = bool(true)];
            tensor<int32, [1]> var_8975_axes_0 = const()[name = string("op_8975_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_8975_cast_fp16 = reduce_max(axes = var_8975_axes_0, keep_dims = var_8974, x = var_8971_cast_fp16)[name = string("op_8975_cast_fp16")];
            fp16 var_8976_to_fp16 = const()[name = string("op_8976_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_331_cast_fp16 = maximum(x = var_8975_cast_fp16, y = var_8976_to_fp16)[name = string("s_331_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_331_cast_fp16 = real_div(x = linear_184_cast_fp16, y = s_331_cast_fp16)[name = string("y_331_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8979_cast_fp16 = mul(x = y_331_cast_fp16, y = y_331_cast_fp16)[name = string("op_8979_cast_fp16")];
            tensor<int32, [1]> var_8984_axes_0 = const()[name = string("op_8984_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8984_keep_dims_0 = const()[name = string("op_8984_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_8984_cast_fp16 = reduce_mean(axes = var_8984_axes_0, keep_dims = var_8984_keep_dims_0, x = var_8979_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8985_cast_fp16 = mul(x = s_331_cast_fp16, y = s_331_cast_fp16)[name = string("op_8985_cast_fp16")];
            fp32 var_8986_epsilon_0 = const()[name = string("op_8986_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_8986_cast_fp16 = inverse(epsilon = var_8986_epsilon_0, x = var_8985_cast_fp16)[name = string("op_8986_cast_fp16")];
            fp16 var_8987_to_fp16 = const()[name = string("op_8987_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_8988_cast_fp16 = mul(x = var_8986_cast_fp16, y = var_8987_to_fp16)[name = string("op_8988_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_8990_cast_fp16 = add(x = var_8984_cast_fp16, y = var_8988_cast_fp16)[name = string("op_8990_cast_fp16")];
            fp32 var_8991_epsilon_0 = const()[name = string("op_8991_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_8991_cast_fp16 = rsqrt(epsilon = var_8991_epsilon_0, x = var_8990_cast_fp16)[name = string("op_8991_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_289_cast_fp16 = mul(x = y_331_cast_fp16, y = var_8991_cast_fp16)[name = string("out_289_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116418176)))];
            tensor<fp16, [1, 32, 512]> var_8993_cast_fp16 = mul(x = out_289_cast_fp16, y = tm_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("op_8993_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_371_cast_fp16 = add(x = x_361_cast_fp16, y = var_8993_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_8996_cast_fp16 = abs(x = x_371_cast_fp16)[name = string("op_8996_cast_fp16")];
            bool var_8999 = const()[name = string("op_8999"), val = bool(true)];
            tensor<int32, [1]> var_9000_axes_0 = const()[name = string("op_9000_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9000_cast_fp16 = reduce_max(axes = var_9000_axes_0, keep_dims = var_8999, x = var_8996_cast_fp16)[name = string("op_9000_cast_fp16")];
            fp16 var_9001_to_fp16 = const()[name = string("op_9001_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_333_cast_fp16 = maximum(x = var_9000_cast_fp16, y = var_9001_to_fp16)[name = string("s_333_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_333_cast_fp16 = real_div(x = x_371_cast_fp16, y = s_333_cast_fp16)[name = string("y_333_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9004_cast_fp16 = mul(x = y_333_cast_fp16, y = y_333_cast_fp16)[name = string("op_9004_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9004_cast_fp16)[name = string("op_9009_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9010_cast_fp16 = mul(x = s_333_cast_fp16, y = s_333_cast_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9015_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9015_cast_fp16")];
            fp32 var_9016_epsilon_0 = const()[name = string("op_9016_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9016_cast_fp16 = rsqrt(epsilon = var_9016_epsilon_0, x = var_9015_cast_fp16)[name = string("op_9016_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_291_cast_fp16 = mul(x = y_333_cast_fp16, y = var_9016_cast_fp16)[name = string("out_291_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116419264)))];
            tensor<fp16, [1, 32, 512]> input_245_cast_fp16 = mul(x = out_291_cast_fp16, y = tm_layers_20_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116420352))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117468992))))[name = string("tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_185_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_185_cast_fp16")];
            string var_9023_mode_0 = const()[name = string("op_9023_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_9023_cast_fp16 = gelu(mode = var_9023_mode_0, x = linear_185_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117473152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118521792))))[name = string("tm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_186_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_247_cast_fp16 = mul(x = var_9023_cast_fp16, y = linear_186_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118525952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119574592))))[name = string("tm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9031_cast_fp16 = abs(x = linear_187_cast_fp16)[name = string("op_9031_cast_fp16")];
            bool var_9034 = const()[name = string("op_9034"), val = bool(true)];
            tensor<int32, [1]> var_9035_axes_0 = const()[name = string("op_9035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9035_cast_fp16 = reduce_max(axes = var_9035_axes_0, keep_dims = var_9034, x = var_9031_cast_fp16)[name = string("op_9035_cast_fp16")];
            fp16 var_9036_to_fp16 = const()[name = string("op_9036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_335_cast_fp16 = maximum(x = var_9035_cast_fp16, y = var_9036_to_fp16)[name = string("s_335_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_335_cast_fp16 = real_div(x = linear_187_cast_fp16, y = s_335_cast_fp16)[name = string("y_335_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9039_cast_fp16 = mul(x = y_335_cast_fp16, y = y_335_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [1]> var_9044_axes_0 = const()[name = string("op_9044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9044_keep_dims_0 = const()[name = string("op_9044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9044_cast_fp16 = reduce_mean(axes = var_9044_axes_0, keep_dims = var_9044_keep_dims_0, x = var_9039_cast_fp16)[name = string("op_9044_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9045_cast_fp16 = mul(x = s_335_cast_fp16, y = s_335_cast_fp16)[name = string("op_9045_cast_fp16")];
            fp32 var_9046_epsilon_0 = const()[name = string("op_9046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9046_cast_fp16 = inverse(epsilon = var_9046_epsilon_0, x = var_9045_cast_fp16)[name = string("op_9046_cast_fp16")];
            fp16 var_9047_to_fp16 = const()[name = string("op_9047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9048_cast_fp16 = mul(x = var_9046_cast_fp16, y = var_9047_to_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9050_cast_fp16 = add(x = var_9044_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9050_cast_fp16")];
            fp32 var_9051_epsilon_0 = const()[name = string("op_9051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9051_cast_fp16 = rsqrt(epsilon = var_9051_epsilon_0, x = var_9050_cast_fp16)[name = string("op_9051_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_293_cast_fp16 = mul(x = y_335_cast_fp16, y = var_9051_cast_fp16)[name = string("out_293_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119575680)))];
            tensor<fp16, [1, 32, 512]> var_9053_cast_fp16 = mul(x = out_293_cast_fp16, y = tm_layers_20_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9053_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_249_cast_fp16 = add(x = x_371_cast_fp16, y = var_9053_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119576768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119838976))))[name = string("tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_188_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_188_cast_fp16")];
            string x_375_mode_0 = const()[name = string("x_375_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_375_cast_fp16 = gelu(mode = x_375_mode_0, x = linear_188_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [4]> var_9073_begin_0 = const()[name = string("op_9073_begin_0"), val = tensor<int32, [4]>([0, 0, 20, 0])];
            tensor<int32, [4]> var_9073_end_0 = const()[name = string("op_9073_end_0"), val = tensor<int32, [4]>([1, 32, 21, 512])];
            tensor<bool, [4]> var_9073_end_mask_0 = const()[name = string("op_9073_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9073_squeeze_mask_0 = const()[name = string("op_9073_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_9073_cast_fp16 = slice_by_index(begin = var_9073_begin_0, end = var_9073_end_0, end_mask = var_9073_end_mask_0, squeeze_mask = var_9073_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_251_cast_fp16 = mul(x = x_375_cast_fp16, y = var_9073_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119840064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120102272))))[name = string("tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9083_cast_fp16 = abs(x = linear_189_cast_fp16)[name = string("op_9083_cast_fp16")];
            bool var_9086 = const()[name = string("op_9086"), val = bool(true)];
            tensor<int32, [1]> var_9087_axes_0 = const()[name = string("op_9087_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9087_cast_fp16 = reduce_max(axes = var_9087_axes_0, keep_dims = var_9086, x = var_9083_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_337_cast_fp16 = maximum(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("s_337_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_337_cast_fp16 = real_div(x = linear_189_cast_fp16, y = s_337_cast_fp16)[name = string("y_337_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9091_cast_fp16 = mul(x = y_337_cast_fp16, y = y_337_cast_fp16)[name = string("op_9091_cast_fp16")];
            tensor<int32, [1]> var_9096_axes_0 = const()[name = string("op_9096_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9096_keep_dims_0 = const()[name = string("op_9096_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9096_cast_fp16 = reduce_mean(axes = var_9096_axes_0, keep_dims = var_9096_keep_dims_0, x = var_9091_cast_fp16)[name = string("op_9096_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9097_cast_fp16 = mul(x = s_337_cast_fp16, y = s_337_cast_fp16)[name = string("op_9097_cast_fp16")];
            fp32 var_9098_epsilon_0 = const()[name = string("op_9098_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9098_cast_fp16 = inverse(epsilon = var_9098_epsilon_0, x = var_9097_cast_fp16)[name = string("op_9098_cast_fp16")];
            fp16 var_9099_to_fp16 = const()[name = string("op_9099_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9100_cast_fp16 = mul(x = var_9098_cast_fp16, y = var_9099_to_fp16)[name = string("op_9100_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9102_cast_fp16 = add(x = var_9096_cast_fp16, y = var_9100_cast_fp16)[name = string("op_9102_cast_fp16")];
            fp32 var_9103_epsilon_0 = const()[name = string("op_9103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9103_cast_fp16 = rsqrt(epsilon = var_9103_epsilon_0, x = var_9102_cast_fp16)[name = string("op_9103_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_295_cast_fp16 = mul(x = y_337_cast_fp16, y = var_9103_cast_fp16)[name = string("out_295_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120103360)))];
            tensor<fp16, [1, 32, 512]> var_9105_cast_fp16 = mul(x = out_295_cast_fp16, y = tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9105_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_41_cast_fp16 = add(x = input_249_cast_fp16, y = var_9105_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<fp16, [1]> tm_layers_20_layer_scalar_to_fp16 = const()[name = string("tm_layers_20_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ecp-1])];
            tensor<fp16, [1, 32, 512]> x_379_cast_fp16 = mul(x = h_41_cast_fp16, y = tm_layers_20_layer_scalar_to_fp16)[name = string("x_379_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9109_cast_fp16 = abs(x = x_379_cast_fp16)[name = string("op_9109_cast_fp16")];
            bool var_9112 = const()[name = string("op_9112"), val = bool(true)];
            tensor<int32, [1]> var_9113_axes_0 = const()[name = string("op_9113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9113_cast_fp16 = reduce_max(axes = var_9113_axes_0, keep_dims = var_9112, x = var_9109_cast_fp16)[name = string("op_9113_cast_fp16")];
            fp16 var_9114_to_fp16 = const()[name = string("op_9114_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_339_cast_fp16 = maximum(x = var_9113_cast_fp16, y = var_9114_to_fp16)[name = string("s_339_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_339_cast_fp16 = real_div(x = x_379_cast_fp16, y = s_339_cast_fp16)[name = string("y_339_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9117_cast_fp16 = mul(x = y_339_cast_fp16, y = y_339_cast_fp16)[name = string("op_9117_cast_fp16")];
            tensor<int32, [1]> var_9122_axes_0 = const()[name = string("op_9122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9122_keep_dims_0 = const()[name = string("op_9122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9122_cast_fp16 = reduce_mean(axes = var_9122_axes_0, keep_dims = var_9122_keep_dims_0, x = var_9117_cast_fp16)[name = string("op_9122_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9123_cast_fp16 = mul(x = s_339_cast_fp16, y = s_339_cast_fp16)[name = string("op_9123_cast_fp16")];
            fp32 var_9124_epsilon_0 = const()[name = string("op_9124_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9124_cast_fp16 = inverse(epsilon = var_9124_epsilon_0, x = var_9123_cast_fp16)[name = string("op_9124_cast_fp16")];
            fp16 var_9125_to_fp16 = const()[name = string("op_9125_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9126_cast_fp16 = mul(x = var_9124_cast_fp16, y = var_9125_to_fp16)[name = string("op_9126_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9128_cast_fp16 = add(x = var_9122_cast_fp16, y = var_9126_cast_fp16)[name = string("op_9128_cast_fp16")];
            fp32 var_9129_epsilon_0 = const()[name = string("op_9129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9129_cast_fp16 = rsqrt(epsilon = var_9129_epsilon_0, x = var_9128_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_297_cast_fp16 = mul(x = y_339_cast_fp16, y = var_9129_cast_fp16)[name = string("out_297_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120104448)))];
            tensor<fp16, [1, 32, 512]> input_253_cast_fp16 = mul(x = out_297_cast_fp16, y = tm_layers_21_input_layernorm_weight_to_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120105536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120629888))))[name = string("tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9139 = const()[name = string("op_9139"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_9140_cast_fp16 = reshape(shape = var_9139, x = linear_190_cast_fp16)[name = string("op_9140_cast_fp16")];
            tensor<int32, [4]> x_381_perm_0 = const()[name = string("x_381_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120632000))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120894208))))[name = string("tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9151 = const()[name = string("op_9151"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_9152_cast_fp16 = reshape(shape = var_9151, x = linear_191_cast_fp16)[name = string("op_9152_cast_fp16")];
            tensor<int32, [4]> x_383_perm_0 = const()[name = string("x_383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120895296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121157504))))[name = string("tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_9163 = const()[name = string("op_9163"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_9164_cast_fp16 = reshape(shape = var_9163, x = linear_192_cast_fp16)[name = string("op_9164_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_381_cast_fp16 = transpose(perm = x_381_perm_0, x = var_9140_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 4, 32, 256]> var_9168_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_9168_cast_fp16")];
            bool var_9171 = const()[name = string("op_9171"), val = bool(true)];
            tensor<int32, [1]> var_9172_axes_0 = const()[name = string("op_9172_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_9172_cast_fp16 = reduce_max(axes = var_9172_axes_0, keep_dims = var_9171, x = var_9168_cast_fp16)[name = string("op_9172_cast_fp16")];
            fp16 var_9173_to_fp16 = const()[name = string("op_9173_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_341_cast_fp16 = maximum(x = var_9172_cast_fp16, y = var_9173_to_fp16)[name = string("s_341_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_341_cast_fp16 = real_div(x = x_381_cast_fp16, y = s_341_cast_fp16)[name = string("y_341_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_9176_cast_fp16 = mul(x = y_341_cast_fp16, y = y_341_cast_fp16)[name = string("op_9176_cast_fp16")];
            tensor<int32, [1]> var_9181_axes_0 = const()[name = string("op_9181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9181_keep_dims_0 = const()[name = string("op_9181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_9181_cast_fp16 = reduce_mean(axes = var_9181_axes_0, keep_dims = var_9181_keep_dims_0, x = var_9176_cast_fp16)[name = string("op_9181_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_9182_cast_fp16 = mul(x = s_341_cast_fp16, y = s_341_cast_fp16)[name = string("op_9182_cast_fp16")];
            fp32 var_9183_epsilon_0 = const()[name = string("op_9183_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_9183_cast_fp16 = inverse(epsilon = var_9183_epsilon_0, x = var_9182_cast_fp16)[name = string("op_9183_cast_fp16")];
            fp16 var_9184_to_fp16 = const()[name = string("op_9184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_9185_cast_fp16 = mul(x = var_9183_cast_fp16, y = var_9184_to_fp16)[name = string("op_9185_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_9187_cast_fp16 = add(x = var_9181_cast_fp16, y = var_9185_cast_fp16)[name = string("op_9187_cast_fp16")];
            fp32 var_9188_epsilon_0 = const()[name = string("op_9188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_9188_cast_fp16 = rsqrt(epsilon = var_9188_epsilon_0, x = var_9187_cast_fp16)[name = string("op_9188_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_299_cast_fp16 = mul(x = y_341_cast_fp16, y = var_9188_cast_fp16)[name = string("out_299_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_85_cast_fp16 = mul(x = out_299_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_85_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_383_cast_fp16 = transpose(perm = x_383_perm_0, x = var_9152_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2, 32, 256]> var_9191_cast_fp16 = abs(x = x_383_cast_fp16)[name = string("op_9191_cast_fp16")];
            bool var_9194 = const()[name = string("op_9194"), val = bool(true)];
            tensor<int32, [1]> var_9195_axes_0 = const()[name = string("op_9195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_9195_cast_fp16 = reduce_max(axes = var_9195_axes_0, keep_dims = var_9194, x = var_9191_cast_fp16)[name = string("op_9195_cast_fp16")];
            fp16 var_9196_to_fp16 = const()[name = string("op_9196_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_343_cast_fp16 = maximum(x = var_9195_cast_fp16, y = var_9196_to_fp16)[name = string("s_343_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_343_cast_fp16 = real_div(x = x_383_cast_fp16, y = s_343_cast_fp16)[name = string("y_343_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_9199_cast_fp16 = mul(x = y_343_cast_fp16, y = y_343_cast_fp16)[name = string("op_9199_cast_fp16")];
            tensor<int32, [1]> var_9204_axes_0 = const()[name = string("op_9204_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9204_keep_dims_0 = const()[name = string("op_9204_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_9204_cast_fp16 = reduce_mean(axes = var_9204_axes_0, keep_dims = var_9204_keep_dims_0, x = var_9199_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_9205_cast_fp16 = mul(x = s_343_cast_fp16, y = s_343_cast_fp16)[name = string("op_9205_cast_fp16")];
            fp32 var_9206_epsilon_0 = const()[name = string("op_9206_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_9206_cast_fp16 = inverse(epsilon = var_9206_epsilon_0, x = var_9205_cast_fp16)[name = string("op_9206_cast_fp16")];
            fp16 var_9207_to_fp16 = const()[name = string("op_9207_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_9208_cast_fp16 = mul(x = var_9206_cast_fp16, y = var_9207_to_fp16)[name = string("op_9208_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_9210_cast_fp16 = add(x = var_9204_cast_fp16, y = var_9208_cast_fp16)[name = string("op_9210_cast_fp16")];
            fp32 var_9211_epsilon_0 = const()[name = string("op_9211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_9211_cast_fp16 = rsqrt(epsilon = var_9211_epsilon_0, x = var_9210_cast_fp16)[name = string("op_9211_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_301_cast_fp16 = mul(x = y_343_cast_fp16, y = var_9211_cast_fp16)[name = string("out_301_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_127_cast_fp16 = mul(x = out_301_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_127_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_9164_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2, 32, 256]> var_9214_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_9214_cast_fp16")];
            bool var_9217 = const()[name = string("op_9217"), val = bool(true)];
            tensor<int32, [1]> var_9218_axes_0 = const()[name = string("op_9218_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_9218_cast_fp16 = reduce_max(axes = var_9218_axes_0, keep_dims = var_9217, x = var_9214_cast_fp16)[name = string("op_9218_cast_fp16")];
            fp16 var_9219_to_fp16 = const()[name = string("op_9219_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_345_cast_fp16 = maximum(x = var_9218_cast_fp16, y = var_9219_to_fp16)[name = string("s_345_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_345_cast_fp16 = real_div(x = x_385_cast_fp16, y = s_345_cast_fp16)[name = string("y_345_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_9222_cast_fp16 = mul(x = y_345_cast_fp16, y = y_345_cast_fp16)[name = string("op_9222_cast_fp16")];
            tensor<int32, [1]> var_9227_axes_0 = const()[name = string("op_9227_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9227_keep_dims_0 = const()[name = string("op_9227_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_9227_cast_fp16 = reduce_mean(axes = var_9227_axes_0, keep_dims = var_9227_keep_dims_0, x = var_9222_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_9228_cast_fp16 = mul(x = s_345_cast_fp16, y = s_345_cast_fp16)[name = string("op_9228_cast_fp16")];
            fp32 var_9229_epsilon_0 = const()[name = string("op_9229_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_9229_cast_fp16 = inverse(epsilon = var_9229_epsilon_0, x = var_9228_cast_fp16)[name = string("op_9229_cast_fp16")];
            fp16 var_9230_to_fp16 = const()[name = string("op_9230_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_9231_cast_fp16 = mul(x = var_9229_cast_fp16, y = var_9230_to_fp16)[name = string("op_9231_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_9233_cast_fp16 = add(x = var_9227_cast_fp16, y = var_9231_cast_fp16)[name = string("op_9233_cast_fp16")];
            fp32 var_9234_epsilon_0 = const()[name = string("op_9234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_9234_cast_fp16 = rsqrt(epsilon = var_9234_epsilon_0, x = var_9233_cast_fp16)[name = string("op_9234_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_85_cast_fp16 = mul(x = y_345_cast_fp16, y = var_9234_cast_fp16)[name = string("v_85_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_9236_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9236_cast_fp16")];
            tensor<int32, [4]> var_9241_begin_0 = const()[name = string("op_9241_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9241_end_0 = const()[name = string("op_9241_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_9241_end_mask_0 = const()[name = string("op_9241_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_9241_cast_fp16 = slice_by_index(begin = var_9241_begin_0, end = var_9241_end_0, end_mask = var_9241_end_mask_0, x = q_85_cast_fp16)[name = string("op_9241_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_9242_cast_fp16 = mul(x = var_9241_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_9242_cast_fp16")];
            tensor<int32, [4]> var_9247_begin_0 = const()[name = string("op_9247_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9247_end_0 = const()[name = string("op_9247_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_9247_end_mask_0 = const()[name = string("op_9247_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_9247_cast_fp16 = slice_by_index(begin = var_9247_begin_0, end = var_9247_end_0, end_mask = var_9247_end_mask_0, x = q_85_cast_fp16)[name = string("op_9247_cast_fp16")];
            int32 var_9249 = const()[name = string("op_9249"), val = int32(-1)];
            bool var_9250_interleave_0 = const()[name = string("op_9250_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_9250_cast_fp16 = concat(axis = var_9249, interleave = var_9250_interleave_0, values = (var_9242_cast_fp16, var_9247_cast_fp16))[name = string("op_9250_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_9251_cast_fp16 = mul(x = var_9250_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9251_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_87_cast_fp16 = add(x = var_9236_cast_fp16, y = var_9251_cast_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_9254_cast_fp16 = mul(x = k_127_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9254_cast_fp16")];
            tensor<int32, [4]> var_9259_begin_0 = const()[name = string("op_9259_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9259_end_0 = const()[name = string("op_9259_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_9259_end_mask_0 = const()[name = string("op_9259_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_9259_cast_fp16 = slice_by_index(begin = var_9259_begin_0, end = var_9259_end_0, end_mask = var_9259_end_mask_0, x = k_127_cast_fp16)[name = string("op_9259_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_9260_cast_fp16 = mul(x = var_9259_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_9260_cast_fp16")];
            tensor<int32, [4]> var_9265_begin_0 = const()[name = string("op_9265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9265_end_0 = const()[name = string("op_9265_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_9265_end_mask_0 = const()[name = string("op_9265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_9265_cast_fp16 = slice_by_index(begin = var_9265_begin_0, end = var_9265_end_0, end_mask = var_9265_end_mask_0, x = k_127_cast_fp16)[name = string("op_9265_cast_fp16")];
            int32 var_9267 = const()[name = string("op_9267"), val = int32(-1)];
            bool var_9268_interleave_0 = const()[name = string("op_9268_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_9268_cast_fp16 = concat(axis = var_9267, interleave = var_9268_interleave_0, values = (var_9260_cast_fp16, var_9265_cast_fp16))[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_9269_cast_fp16 = mul(x = var_9268_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_129_cast_fp16 = add(x = var_9254_cast_fp16, y = var_9269_cast_fp16)[name = string("k_129_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_42_reps_0 = const()[name = string("tile_42_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = k_129_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 1, 32, 256]> tile_42_cast_fp16 = tile(reps = tile_42_reps_0, x = transpose_84_cast_fp16)[name = string("tile_42_cast_fp16")];
            tensor<int32, [5]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_84_cast_fp16 = reshape(shape = concat_84, x = tile_42_cast_fp16)[name = string("reshape_84_cast_fp16")];
            tensor<int32, [5]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = reshape_84_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 1, 32, 256]> reshape_85_cast_fp16 = reshape(shape = concat_85, x = transpose_85_cast_fp16)[name = string("reshape_85_cast_fp16")];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_43_reps_0 = const()[name = string("tile_43_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_86_cast_fp16 = transpose(perm = transpose_86_perm_0, x = v_85_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 1, 32, 256]> tile_43_cast_fp16 = tile(reps = tile_43_reps_0, x = transpose_86_cast_fp16)[name = string("tile_43_cast_fp16")];
            tensor<int32, [5]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_86_cast_fp16 = reshape(shape = concat_86, x = tile_43_cast_fp16)[name = string("reshape_86_cast_fp16")];
            tensor<int32, [5]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_87_cast_fp16 = transpose(perm = transpose_87_perm_0, x = reshape_86_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 1, 32, 256]> reshape_87_cast_fp16 = reshape(shape = concat_87, x = transpose_87_cast_fp16)[name = string("reshape_87_cast_fp16")];
            tensor<int32, [4]> v_87_perm_0 = const()[name = string("v_87_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9283_transpose_x_0 = const()[name = string("op_9283_transpose_x_0"), val = bool(false)];
            bool var_9283_transpose_y_0 = const()[name = string("op_9283_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_117_cast_fp16 = transpose(perm = transpose_117_perm_0, x = reshape_85_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 4, 32, 32]> var_9283_cast_fp16 = matmul(transpose_x = var_9283_transpose_x_0, transpose_y = var_9283_transpose_y_0, x = q_87_cast_fp16, y = transpose_117_cast_fp16)[name = string("op_9283_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_9285_cast_fp16 = add(x = var_9283_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9285_cast_fp16")];
            int32 var_9286 = const()[name = string("op_9286"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_345_cast_fp16 = softmax(axis = var_9286, x = var_9285_cast_fp16)[name = string("w_345_cast_fp16")];
            bool var_9289_transpose_x_0 = const()[name = string("op_9289_transpose_x_0"), val = bool(false)];
            bool var_9289_transpose_y_0 = const()[name = string("op_9289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_87_cast_fp16 = transpose(perm = v_87_perm_0, x = reshape_87_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 4, 32, 256]> var_9289_cast_fp16 = matmul(transpose_x = var_9289_transpose_x_0, transpose_y = var_9289_transpose_y_0, x = w_345_cast_fp16, y = v_87_cast_fp16)[name = string("op_9289_cast_fp16")];
            tensor<int32, [4]> var_9292_perm_0 = const()[name = string("op_9292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9296 = const()[name = string("op_9296"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_9292_cast_fp16 = transpose(perm = var_9292_perm_0, x = var_9289_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 32, 1024]> input_255_cast_fp16 = reshape(shape = var_9296, x = var_9292_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121158592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121682944))))[name = string("tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9301_cast_fp16 = abs(x = linear_193_cast_fp16)[name = string("op_9301_cast_fp16")];
            bool var_9304 = const()[name = string("op_9304"), val = bool(true)];
            tensor<int32, [1]> var_9305_axes_0 = const()[name = string("op_9305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9305_cast_fp16 = reduce_max(axes = var_9305_axes_0, keep_dims = var_9304, x = var_9301_cast_fp16)[name = string("op_9305_cast_fp16")];
            fp16 var_9306_to_fp16 = const()[name = string("op_9306_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_347_cast_fp16 = maximum(x = var_9305_cast_fp16, y = var_9306_to_fp16)[name = string("s_347_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_347_cast_fp16 = real_div(x = linear_193_cast_fp16, y = s_347_cast_fp16)[name = string("y_347_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9309_cast_fp16 = mul(x = y_347_cast_fp16, y = y_347_cast_fp16)[name = string("op_9309_cast_fp16")];
            tensor<int32, [1]> var_9314_axes_0 = const()[name = string("op_9314_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9314_keep_dims_0 = const()[name = string("op_9314_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9314_cast_fp16 = reduce_mean(axes = var_9314_axes_0, keep_dims = var_9314_keep_dims_0, x = var_9309_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9315_cast_fp16 = mul(x = s_347_cast_fp16, y = s_347_cast_fp16)[name = string("op_9315_cast_fp16")];
            fp32 var_9316_epsilon_0 = const()[name = string("op_9316_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9316_cast_fp16 = inverse(epsilon = var_9316_epsilon_0, x = var_9315_cast_fp16)[name = string("op_9316_cast_fp16")];
            fp16 var_9317_to_fp16 = const()[name = string("op_9317_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9318_cast_fp16 = mul(x = var_9316_cast_fp16, y = var_9317_to_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9320_cast_fp16 = add(x = var_9314_cast_fp16, y = var_9318_cast_fp16)[name = string("op_9320_cast_fp16")];
            fp32 var_9321_epsilon_0 = const()[name = string("op_9321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9321_cast_fp16 = rsqrt(epsilon = var_9321_epsilon_0, x = var_9320_cast_fp16)[name = string("op_9321_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_303_cast_fp16 = mul(x = y_347_cast_fp16, y = var_9321_cast_fp16)[name = string("out_303_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121684032)))];
            tensor<fp16, [1, 32, 512]> var_9323_cast_fp16 = mul(x = out_303_cast_fp16, y = tm_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("op_9323_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_389_cast_fp16 = add(x = x_379_cast_fp16, y = var_9323_cast_fp16)[name = string("x_389_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9326_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_9326_cast_fp16")];
            bool var_9329 = const()[name = string("op_9329"), val = bool(true)];
            tensor<int32, [1]> var_9330_axes_0 = const()[name = string("op_9330_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9330_cast_fp16 = reduce_max(axes = var_9330_axes_0, keep_dims = var_9329, x = var_9326_cast_fp16)[name = string("op_9330_cast_fp16")];
            fp16 var_9331_to_fp16 = const()[name = string("op_9331_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_349_cast_fp16 = maximum(x = var_9330_cast_fp16, y = var_9331_to_fp16)[name = string("s_349_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_349_cast_fp16 = real_div(x = x_389_cast_fp16, y = s_349_cast_fp16)[name = string("y_349_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9334_cast_fp16 = mul(x = y_349_cast_fp16, y = y_349_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [1]> var_9339_axes_0 = const()[name = string("op_9339_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9339_keep_dims_0 = const()[name = string("op_9339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9339_cast_fp16 = reduce_mean(axes = var_9339_axes_0, keep_dims = var_9339_keep_dims_0, x = var_9334_cast_fp16)[name = string("op_9339_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9340_cast_fp16 = mul(x = s_349_cast_fp16, y = s_349_cast_fp16)[name = string("op_9340_cast_fp16")];
            fp32 var_9341_epsilon_0 = const()[name = string("op_9341_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9341_cast_fp16 = inverse(epsilon = var_9341_epsilon_0, x = var_9340_cast_fp16)[name = string("op_9341_cast_fp16")];
            fp16 var_9342_to_fp16 = const()[name = string("op_9342_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9343_cast_fp16 = mul(x = var_9341_cast_fp16, y = var_9342_to_fp16)[name = string("op_9343_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9345_cast_fp16 = add(x = var_9339_cast_fp16, y = var_9343_cast_fp16)[name = string("op_9345_cast_fp16")];
            fp32 var_9346_epsilon_0 = const()[name = string("op_9346_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9346_cast_fp16 = rsqrt(epsilon = var_9346_epsilon_0, x = var_9345_cast_fp16)[name = string("op_9346_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_305_cast_fp16 = mul(x = y_349_cast_fp16, y = var_9346_cast_fp16)[name = string("out_305_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121685120)))];
            tensor<fp16, [1, 32, 512]> input_257_cast_fp16 = mul(x = out_305_cast_fp16, y = tm_layers_21_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121686208))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122734848))))[name = string("tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_194_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_194_cast_fp16")];
            string var_9353_mode_0 = const()[name = string("op_9353_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_9353_cast_fp16 = gelu(mode = var_9353_mode_0, x = linear_194_cast_fp16)[name = string("op_9353_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122739008))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123787648))))[name = string("tm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_195_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_259_cast_fp16 = mul(x = var_9353_cast_fp16, y = linear_195_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123791808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124840448))))[name = string("tm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9361_cast_fp16 = abs(x = linear_196_cast_fp16)[name = string("op_9361_cast_fp16")];
            bool var_9364 = const()[name = string("op_9364"), val = bool(true)];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9365_cast_fp16 = reduce_max(axes = var_9365_axes_0, keep_dims = var_9364, x = var_9361_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9366_to_fp16 = const()[name = string("op_9366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_351_cast_fp16 = maximum(x = var_9365_cast_fp16, y = var_9366_to_fp16)[name = string("s_351_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_351_cast_fp16 = real_div(x = linear_196_cast_fp16, y = s_351_cast_fp16)[name = string("y_351_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9369_cast_fp16 = mul(x = y_351_cast_fp16, y = y_351_cast_fp16)[name = string("op_9369_cast_fp16")];
            tensor<int32, [1]> var_9374_axes_0 = const()[name = string("op_9374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9374_keep_dims_0 = const()[name = string("op_9374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9374_cast_fp16 = reduce_mean(axes = var_9374_axes_0, keep_dims = var_9374_keep_dims_0, x = var_9369_cast_fp16)[name = string("op_9374_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9375_cast_fp16 = mul(x = s_351_cast_fp16, y = s_351_cast_fp16)[name = string("op_9375_cast_fp16")];
            fp32 var_9376_epsilon_0 = const()[name = string("op_9376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9376_cast_fp16 = inverse(epsilon = var_9376_epsilon_0, x = var_9375_cast_fp16)[name = string("op_9376_cast_fp16")];
            fp16 var_9377_to_fp16 = const()[name = string("op_9377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9378_cast_fp16 = mul(x = var_9376_cast_fp16, y = var_9377_to_fp16)[name = string("op_9378_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9380_cast_fp16 = add(x = var_9374_cast_fp16, y = var_9378_cast_fp16)[name = string("op_9380_cast_fp16")];
            fp32 var_9381_epsilon_0 = const()[name = string("op_9381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9381_cast_fp16 = rsqrt(epsilon = var_9381_epsilon_0, x = var_9380_cast_fp16)[name = string("op_9381_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_307_cast_fp16 = mul(x = y_351_cast_fp16, y = var_9381_cast_fp16)[name = string("out_307_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124841536)))];
            tensor<fp16, [1, 32, 512]> var_9383_cast_fp16 = mul(x = out_307_cast_fp16, y = tm_layers_21_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9383_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_261_cast_fp16 = add(x = x_389_cast_fp16, y = var_9383_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124842624))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125104832))))[name = string("tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_197_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_197_cast_fp16")];
            string x_393_mode_0 = const()[name = string("x_393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_393_cast_fp16 = gelu(mode = x_393_mode_0, x = linear_197_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<int32, [4]> var_9403_begin_0 = const()[name = string("op_9403_begin_0"), val = tensor<int32, [4]>([0, 0, 21, 0])];
            tensor<int32, [4]> var_9403_end_0 = const()[name = string("op_9403_end_0"), val = tensor<int32, [4]>([1, 32, 22, 512])];
            tensor<bool, [4]> var_9403_end_mask_0 = const()[name = string("op_9403_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9403_squeeze_mask_0 = const()[name = string("op_9403_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_9403_cast_fp16 = slice_by_index(begin = var_9403_begin_0, end = var_9403_end_0, end_mask = var_9403_end_mask_0, squeeze_mask = var_9403_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9403_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_263_cast_fp16 = mul(x = x_393_cast_fp16, y = var_9403_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125105920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125368128))))[name = string("tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9413_cast_fp16 = abs(x = linear_198_cast_fp16)[name = string("op_9413_cast_fp16")];
            bool var_9416 = const()[name = string("op_9416"), val = bool(true)];
            tensor<int32, [1]> var_9417_axes_0 = const()[name = string("op_9417_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9417_cast_fp16 = reduce_max(axes = var_9417_axes_0, keep_dims = var_9416, x = var_9413_cast_fp16)[name = string("op_9417_cast_fp16")];
            fp16 var_9418_to_fp16 = const()[name = string("op_9418_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_353_cast_fp16 = maximum(x = var_9417_cast_fp16, y = var_9418_to_fp16)[name = string("s_353_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_353_cast_fp16 = real_div(x = linear_198_cast_fp16, y = s_353_cast_fp16)[name = string("y_353_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9421_cast_fp16 = mul(x = y_353_cast_fp16, y = y_353_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<int32, [1]> var_9426_axes_0 = const()[name = string("op_9426_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9426_keep_dims_0 = const()[name = string("op_9426_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9426_cast_fp16 = reduce_mean(axes = var_9426_axes_0, keep_dims = var_9426_keep_dims_0, x = var_9421_cast_fp16)[name = string("op_9426_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9427_cast_fp16 = mul(x = s_353_cast_fp16, y = s_353_cast_fp16)[name = string("op_9427_cast_fp16")];
            fp32 var_9428_epsilon_0 = const()[name = string("op_9428_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9428_cast_fp16 = inverse(epsilon = var_9428_epsilon_0, x = var_9427_cast_fp16)[name = string("op_9428_cast_fp16")];
            fp16 var_9429_to_fp16 = const()[name = string("op_9429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9430_cast_fp16 = mul(x = var_9428_cast_fp16, y = var_9429_to_fp16)[name = string("op_9430_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9432_cast_fp16 = add(x = var_9426_cast_fp16, y = var_9430_cast_fp16)[name = string("op_9432_cast_fp16")];
            fp32 var_9433_epsilon_0 = const()[name = string("op_9433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9433_cast_fp16 = rsqrt(epsilon = var_9433_epsilon_0, x = var_9432_cast_fp16)[name = string("op_9433_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_309_cast_fp16 = mul(x = y_353_cast_fp16, y = var_9433_cast_fp16)[name = string("out_309_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125369216)))];
            tensor<fp16, [1, 32, 512]> var_9435_cast_fp16 = mul(x = out_309_cast_fp16, y = tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9435_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_43_cast_fp16 = add(x = input_261_cast_fp16, y = var_9435_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<fp16, [1]> tm_layers_21_layer_scalar_to_fp16 = const()[name = string("tm_layers_21_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e8p-1])];
            tensor<fp16, [1, 32, 512]> x_397_cast_fp16 = mul(x = h_43_cast_fp16, y = tm_layers_21_layer_scalar_to_fp16)[name = string("x_397_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9439_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_9439_cast_fp16")];
            bool var_9442 = const()[name = string("op_9442"), val = bool(true)];
            tensor<int32, [1]> var_9443_axes_0 = const()[name = string("op_9443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9443_cast_fp16 = reduce_max(axes = var_9443_axes_0, keep_dims = var_9442, x = var_9439_cast_fp16)[name = string("op_9443_cast_fp16")];
            fp16 var_9444_to_fp16 = const()[name = string("op_9444_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_355_cast_fp16 = maximum(x = var_9443_cast_fp16, y = var_9444_to_fp16)[name = string("s_355_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_355_cast_fp16 = real_div(x = x_397_cast_fp16, y = s_355_cast_fp16)[name = string("y_355_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9447_cast_fp16 = mul(x = y_355_cast_fp16, y = y_355_cast_fp16)[name = string("op_9447_cast_fp16")];
            tensor<int32, [1]> var_9452_axes_0 = const()[name = string("op_9452_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9452_keep_dims_0 = const()[name = string("op_9452_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9452_cast_fp16 = reduce_mean(axes = var_9452_axes_0, keep_dims = var_9452_keep_dims_0, x = var_9447_cast_fp16)[name = string("op_9452_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9453_cast_fp16 = mul(x = s_355_cast_fp16, y = s_355_cast_fp16)[name = string("op_9453_cast_fp16")];
            fp32 var_9454_epsilon_0 = const()[name = string("op_9454_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9454_cast_fp16 = inverse(epsilon = var_9454_epsilon_0, x = var_9453_cast_fp16)[name = string("op_9454_cast_fp16")];
            fp16 var_9455_to_fp16 = const()[name = string("op_9455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9456_cast_fp16 = mul(x = var_9454_cast_fp16, y = var_9455_to_fp16)[name = string("op_9456_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9458_cast_fp16 = add(x = var_9452_cast_fp16, y = var_9456_cast_fp16)[name = string("op_9458_cast_fp16")];
            fp32 var_9459_epsilon_0 = const()[name = string("op_9459_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9459_cast_fp16 = rsqrt(epsilon = var_9459_epsilon_0, x = var_9458_cast_fp16)[name = string("op_9459_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_311_cast_fp16 = mul(x = y_355_cast_fp16, y = var_9459_cast_fp16)[name = string("out_311_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125370304)))];
            tensor<fp16, [1, 32, 512]> input_265_cast_fp16 = mul(x = out_311_cast_fp16, y = tm_layers_22_input_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125371392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125895744))))[name = string("tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 1024]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<int32, [4]> var_9469 = const()[name = string("op_9469"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 32, 4, 256]> var_9470_cast_fp16 = reshape(shape = var_9469, x = linear_199_cast_fp16)[name = string("op_9470_cast_fp16")];
            tensor<int32, [4]> x_399_perm_0 = const()[name = string("x_399_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125897856))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126160064))))[name = string("tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<int32, [4]> var_9481 = const()[name = string("op_9481"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_9482_cast_fp16 = reshape(shape = var_9481, x = linear_200_cast_fp16)[name = string("op_9482_cast_fp16")];
            tensor<int32, [4]> x_401_perm_0 = const()[name = string("x_401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126161152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126423360))))[name = string("tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<int32, [4]> var_9493 = const()[name = string("op_9493"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 32, 2, 256]> var_9494_cast_fp16 = reshape(shape = var_9493, x = linear_201_cast_fp16)[name = string("op_9494_cast_fp16")];
            tensor<int32, [4]> x_403_perm_0 = const()[name = string("x_403_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 256]> x_399_cast_fp16 = transpose(perm = x_399_perm_0, x = var_9470_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 4, 32, 256]> var_9498_cast_fp16 = abs(x = x_399_cast_fp16)[name = string("op_9498_cast_fp16")];
            bool var_9501 = const()[name = string("op_9501"), val = bool(true)];
            tensor<int32, [1]> var_9502_axes_0 = const()[name = string("op_9502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_9502_cast_fp16 = reduce_max(axes = var_9502_axes_0, keep_dims = var_9501, x = var_9498_cast_fp16)[name = string("op_9502_cast_fp16")];
            fp16 var_9503_to_fp16 = const()[name = string("op_9503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_357_cast_fp16 = maximum(x = var_9502_cast_fp16, y = var_9503_to_fp16)[name = string("s_357_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> y_357_cast_fp16 = real_div(x = x_399_cast_fp16, y = s_357_cast_fp16)[name = string("y_357_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_9506_cast_fp16 = mul(x = y_357_cast_fp16, y = y_357_cast_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9511_axes_0 = const()[name = string("op_9511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9511_keep_dims_0 = const()[name = string("op_9511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_9511_cast_fp16 = reduce_mean(axes = var_9511_axes_0, keep_dims = var_9511_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_9512_cast_fp16 = mul(x = s_357_cast_fp16, y = s_357_cast_fp16)[name = string("op_9512_cast_fp16")];
            fp32 var_9513_epsilon_0 = const()[name = string("op_9513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_9513_cast_fp16 = inverse(epsilon = var_9513_epsilon_0, x = var_9512_cast_fp16)[name = string("op_9513_cast_fp16")];
            fp16 var_9514_to_fp16 = const()[name = string("op_9514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_9515_cast_fp16 = mul(x = var_9513_cast_fp16, y = var_9514_to_fp16)[name = string("op_9515_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_9517_cast_fp16 = add(x = var_9511_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9517_cast_fp16")];
            fp32 var_9518_epsilon_0 = const()[name = string("op_9518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_9518_cast_fp16 = rsqrt(epsilon = var_9518_epsilon_0, x = var_9517_cast_fp16)[name = string("op_9518_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> out_313_cast_fp16 = mul(x = y_357_cast_fp16, y = var_9518_cast_fp16)[name = string("out_313_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_89_cast_fp16 = mul(x = out_313_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_89_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_401_cast_fp16 = transpose(perm = x_401_perm_0, x = var_9482_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 32, 256]> var_9521_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_9521_cast_fp16")];
            bool var_9524 = const()[name = string("op_9524"), val = bool(true)];
            tensor<int32, [1]> var_9525_axes_0 = const()[name = string("op_9525_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_9525_cast_fp16 = reduce_max(axes = var_9525_axes_0, keep_dims = var_9524, x = var_9521_cast_fp16)[name = string("op_9525_cast_fp16")];
            fp16 var_9526_to_fp16 = const()[name = string("op_9526_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_359_cast_fp16 = maximum(x = var_9525_cast_fp16, y = var_9526_to_fp16)[name = string("s_359_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_359_cast_fp16 = real_div(x = x_401_cast_fp16, y = s_359_cast_fp16)[name = string("y_359_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_9529_cast_fp16 = mul(x = y_359_cast_fp16, y = y_359_cast_fp16)[name = string("op_9529_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9529_cast_fp16)[name = string("op_9534_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_9535_cast_fp16 = mul(x = s_359_cast_fp16, y = s_359_cast_fp16)[name = string("op_9535_cast_fp16")];
            fp32 var_9536_epsilon_0 = const()[name = string("op_9536_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_9536_cast_fp16 = inverse(epsilon = var_9536_epsilon_0, x = var_9535_cast_fp16)[name = string("op_9536_cast_fp16")];
            fp16 var_9537_to_fp16 = const()[name = string("op_9537_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_9538_cast_fp16 = mul(x = var_9536_cast_fp16, y = var_9537_to_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_9540_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9538_cast_fp16)[name = string("op_9540_cast_fp16")];
            fp32 var_9541_epsilon_0 = const()[name = string("op_9541_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_9541_cast_fp16 = rsqrt(epsilon = var_9541_epsilon_0, x = var_9540_cast_fp16)[name = string("op_9541_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> out_315_cast_fp16 = mul(x = y_359_cast_fp16, y = var_9541_cast_fp16)[name = string("out_315_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_133_cast_fp16 = mul(x = out_315_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_133_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> x_403_cast_fp16 = transpose(perm = x_403_perm_0, x = var_9494_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 32, 256]> var_9544_cast_fp16 = abs(x = x_403_cast_fp16)[name = string("op_9544_cast_fp16")];
            bool var_9547 = const()[name = string("op_9547"), val = bool(true)];
            tensor<int32, [1]> var_9548_axes_0 = const()[name = string("op_9548_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 32, 1]> var_9548_cast_fp16 = reduce_max(axes = var_9548_axes_0, keep_dims = var_9547, x = var_9544_cast_fp16)[name = string("op_9548_cast_fp16")];
            fp16 var_9549_to_fp16 = const()[name = string("op_9549_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 32, 1]> s_361_cast_fp16 = maximum(x = var_9548_cast_fp16, y = var_9549_to_fp16)[name = string("s_361_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> y_361_cast_fp16 = real_div(x = x_403_cast_fp16, y = s_361_cast_fp16)[name = string("y_361_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_9552_cast_fp16 = mul(x = y_361_cast_fp16, y = y_361_cast_fp16)[name = string("op_9552_cast_fp16")];
            tensor<int32, [1]> var_9557_axes_0 = const()[name = string("op_9557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9557_keep_dims_0 = const()[name = string("op_9557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 32, 1]> var_9557_cast_fp16 = reduce_mean(axes = var_9557_axes_0, keep_dims = var_9557_keep_dims_0, x = var_9552_cast_fp16)[name = string("op_9557_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_9558_cast_fp16 = mul(x = s_361_cast_fp16, y = s_361_cast_fp16)[name = string("op_9558_cast_fp16")];
            fp32 var_9559_epsilon_0 = const()[name = string("op_9559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 32, 1]> var_9559_cast_fp16 = inverse(epsilon = var_9559_epsilon_0, x = var_9558_cast_fp16)[name = string("op_9559_cast_fp16")];
            fp16 var_9560_to_fp16 = const()[name = string("op_9560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 32, 1]> var_9561_cast_fp16 = mul(x = var_9559_cast_fp16, y = var_9560_to_fp16)[name = string("op_9561_cast_fp16")];
            tensor<fp16, [1, 2, 32, 1]> var_9563_cast_fp16 = add(x = var_9557_cast_fp16, y = var_9561_cast_fp16)[name = string("op_9563_cast_fp16")];
            fp32 var_9564_epsilon_0 = const()[name = string("op_9564_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 32, 1]> var_9564_cast_fp16 = rsqrt(epsilon = var_9564_epsilon_0, x = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> v_89_cast_fp16 = mul(x = y_361_cast_fp16, y = var_9564_cast_fp16)[name = string("v_89_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_9566_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9566_cast_fp16")];
            tensor<int32, [4]> var_9571_begin_0 = const()[name = string("op_9571_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9571_end_0 = const()[name = string("op_9571_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_9571_end_mask_0 = const()[name = string("op_9571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 128]> var_9571_cast_fp16 = slice_by_index(begin = var_9571_begin_0, end = var_9571_end_0, end_mask = var_9571_end_mask_0, x = q_89_cast_fp16)[name = string("op_9571_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 128]> var_9572_cast_fp16 = mul(x = var_9571_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_9572_cast_fp16")];
            tensor<int32, [4]> var_9577_begin_0 = const()[name = string("op_9577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9577_end_0 = const()[name = string("op_9577_end_0"), val = tensor<int32, [4]>([1, 4, 32, 128])];
            tensor<bool, [4]> var_9577_end_mask_0 = const()[name = string("op_9577_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 128]> var_9577_cast_fp16 = slice_by_index(begin = var_9577_begin_0, end = var_9577_end_0, end_mask = var_9577_end_mask_0, x = q_89_cast_fp16)[name = string("op_9577_cast_fp16")];
            int32 var_9579 = const()[name = string("op_9579"), val = int32(-1)];
            bool var_9580_interleave_0 = const()[name = string("op_9580_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> var_9580_cast_fp16 = concat(axis = var_9579, interleave = var_9580_interleave_0, values = (var_9572_cast_fp16, var_9577_cast_fp16))[name = string("op_9580_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> var_9581_cast_fp16 = mul(x = var_9580_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9581_cast_fp16")];
            tensor<fp16, [1, 4, 32, 256]> q_91_cast_fp16 = add(x = var_9566_cast_fp16, y = var_9581_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_9584_cast_fp16 = mul(x = k_133_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9584_cast_fp16")];
            tensor<int32, [4]> var_9589_begin_0 = const()[name = string("op_9589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9589_end_0 = const()[name = string("op_9589_end_0"), val = tensor<int32, [4]>([1, 2, 32, 256])];
            tensor<bool, [4]> var_9589_end_mask_0 = const()[name = string("op_9589_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 128]> var_9589_cast_fp16 = slice_by_index(begin = var_9589_begin_0, end = var_9589_end_0, end_mask = var_9589_end_mask_0, x = k_133_cast_fp16)[name = string("op_9589_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 128]> var_9590_cast_fp16 = mul(x = var_9589_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_9590_cast_fp16")];
            tensor<int32, [4]> var_9595_begin_0 = const()[name = string("op_9595_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9595_end_0 = const()[name = string("op_9595_end_0"), val = tensor<int32, [4]>([1, 2, 32, 128])];
            tensor<bool, [4]> var_9595_end_mask_0 = const()[name = string("op_9595_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 128]> var_9595_cast_fp16 = slice_by_index(begin = var_9595_begin_0, end = var_9595_end_0, end_mask = var_9595_end_mask_0, x = k_133_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool var_9598_interleave_0 = const()[name = string("op_9598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 256]> var_9598_cast_fp16 = concat(axis = var_9597, interleave = var_9598_interleave_0, values = (var_9590_cast_fp16, var_9595_cast_fp16))[name = string("op_9598_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> var_9599_cast_fp16 = mul(x = var_9598_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9599_cast_fp16")];
            tensor<fp16, [1, 2, 32, 256]> k_135_cast_fp16 = add(x = var_9584_cast_fp16, y = var_9599_cast_fp16)[name = string("k_135_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_44_reps_0 = const()[name = string("tile_44_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = k_135_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 1, 32, 256]> tile_44_cast_fp16 = tile(reps = tile_44_reps_0, x = transpose_88_cast_fp16)[name = string("tile_44_cast_fp16")];
            tensor<int32, [5]> concat_88 = const()[name = string("concat_88"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_88_cast_fp16 = reshape(shape = concat_88, x = tile_44_cast_fp16)[name = string("reshape_88_cast_fp16")];
            tensor<int32, [5]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = reshape_88_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 1, 32, 256]> reshape_89_cast_fp16 = reshape(shape = concat_89, x = transpose_89_cast_fp16)[name = string("reshape_89_cast_fp16")];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_45_reps_0 = const()[name = string("tile_45_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 32, 256]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = v_89_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 1, 32, 256]> tile_45_cast_fp16 = tile(reps = tile_45_reps_0, x = transpose_90_cast_fp16)[name = string("tile_45_cast_fp16")];
            tensor<int32, [5]> concat_90 = const()[name = string("concat_90"), val = tensor<int32, [5]>([2, 2, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> reshape_90_cast_fp16 = reshape(shape = concat_90, x = tile_45_cast_fp16)[name = string("reshape_90_cast_fp16")];
            tensor<int32, [5]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_91 = const()[name = string("concat_91"), val = tensor<int32, [4]>([-1, 1, 32, 256])];
            tensor<fp16, [2, 2, 1, 32, 256]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = reshape_90_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 1, 32, 256]> reshape_91_cast_fp16 = reshape(shape = concat_91, x = transpose_91_cast_fp16)[name = string("reshape_91_cast_fp16")];
            tensor<int32, [4]> v_91_perm_0 = const()[name = string("v_91_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9613_transpose_x_0 = const()[name = string("op_9613_transpose_x_0"), val = bool(false)];
            bool var_9613_transpose_y_0 = const()[name = string("op_9613_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 32]> transpose_118_cast_fp16 = transpose(perm = transpose_118_perm_0, x = reshape_89_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 4, 32, 32]> var_9613_cast_fp16 = matmul(transpose_x = var_9613_transpose_x_0, transpose_y = var_9613_transpose_y_0, x = q_91_cast_fp16, y = transpose_118_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_9615_cast_fp16 = add(x = var_9613_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9615_cast_fp16")];
            int32 var_9616 = const()[name = string("op_9616"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_361_cast_fp16 = softmax(axis = var_9616, x = var_9615_cast_fp16)[name = string("w_361_cast_fp16")];
            bool var_9619_transpose_x_0 = const()[name = string("op_9619_transpose_x_0"), val = bool(false)];
            bool var_9619_transpose_y_0 = const()[name = string("op_9619_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 256]> v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = reshape_91_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 4, 32, 256]> var_9619_cast_fp16 = matmul(transpose_x = var_9619_transpose_x_0, transpose_y = var_9619_transpose_y_0, x = w_361_cast_fp16, y = v_91_cast_fp16)[name = string("op_9619_cast_fp16")];
            tensor<int32, [4]> var_9622_perm_0 = const()[name = string("op_9622_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9626 = const()[name = string("op_9626"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 32, 4, 256]> var_9622_cast_fp16 = transpose(perm = var_9622_perm_0, x = var_9619_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 32, 1024]> input_267_cast_fp16 = reshape(shape = var_9626, x = var_9622_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126424448))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126948800))))[name = string("tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9631_cast_fp16 = abs(x = linear_202_cast_fp16)[name = string("op_9631_cast_fp16")];
            bool var_9634 = const()[name = string("op_9634"), val = bool(true)];
            tensor<int32, [1]> var_9635_axes_0 = const()[name = string("op_9635_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9635_cast_fp16 = reduce_max(axes = var_9635_axes_0, keep_dims = var_9634, x = var_9631_cast_fp16)[name = string("op_9635_cast_fp16")];
            fp16 var_9636_to_fp16 = const()[name = string("op_9636_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_363_cast_fp16 = maximum(x = var_9635_cast_fp16, y = var_9636_to_fp16)[name = string("s_363_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_363_cast_fp16 = real_div(x = linear_202_cast_fp16, y = s_363_cast_fp16)[name = string("y_363_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9639_cast_fp16 = mul(x = y_363_cast_fp16, y = y_363_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [1]> var_9644_axes_0 = const()[name = string("op_9644_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9644_keep_dims_0 = const()[name = string("op_9644_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9644_cast_fp16 = reduce_mean(axes = var_9644_axes_0, keep_dims = var_9644_keep_dims_0, x = var_9639_cast_fp16)[name = string("op_9644_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9645_cast_fp16 = mul(x = s_363_cast_fp16, y = s_363_cast_fp16)[name = string("op_9645_cast_fp16")];
            fp32 var_9646_epsilon_0 = const()[name = string("op_9646_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9646_cast_fp16 = inverse(epsilon = var_9646_epsilon_0, x = var_9645_cast_fp16)[name = string("op_9646_cast_fp16")];
            fp16 var_9647_to_fp16 = const()[name = string("op_9647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9648_cast_fp16 = mul(x = var_9646_cast_fp16, y = var_9647_to_fp16)[name = string("op_9648_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9650_cast_fp16 = add(x = var_9644_cast_fp16, y = var_9648_cast_fp16)[name = string("op_9650_cast_fp16")];
            fp32 var_9651_epsilon_0 = const()[name = string("op_9651_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9651_cast_fp16 = rsqrt(epsilon = var_9651_epsilon_0, x = var_9650_cast_fp16)[name = string("op_9651_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_317_cast_fp16 = mul(x = y_363_cast_fp16, y = var_9651_cast_fp16)[name = string("out_317_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126949888)))];
            tensor<fp16, [1, 32, 512]> var_9653_cast_fp16 = mul(x = out_317_cast_fp16, y = tm_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_407_cast_fp16 = add(x = x_397_cast_fp16, y = var_9653_cast_fp16)[name = string("x_407_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9656_cast_fp16 = abs(x = x_407_cast_fp16)[name = string("op_9656_cast_fp16")];
            bool var_9659 = const()[name = string("op_9659"), val = bool(true)];
            tensor<int32, [1]> var_9660_axes_0 = const()[name = string("op_9660_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9660_cast_fp16 = reduce_max(axes = var_9660_axes_0, keep_dims = var_9659, x = var_9656_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp16 var_9661_to_fp16 = const()[name = string("op_9661_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_365_cast_fp16 = maximum(x = var_9660_cast_fp16, y = var_9661_to_fp16)[name = string("s_365_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_365_cast_fp16 = real_div(x = x_407_cast_fp16, y = s_365_cast_fp16)[name = string("y_365_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9664_cast_fp16 = mul(x = y_365_cast_fp16, y = y_365_cast_fp16)[name = string("op_9664_cast_fp16")];
            tensor<int32, [1]> var_9669_axes_0 = const()[name = string("op_9669_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9669_keep_dims_0 = const()[name = string("op_9669_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9669_cast_fp16 = reduce_mean(axes = var_9669_axes_0, keep_dims = var_9669_keep_dims_0, x = var_9664_cast_fp16)[name = string("op_9669_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9670_cast_fp16 = mul(x = s_365_cast_fp16, y = s_365_cast_fp16)[name = string("op_9670_cast_fp16")];
            fp32 var_9671_epsilon_0 = const()[name = string("op_9671_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9671_cast_fp16 = inverse(epsilon = var_9671_epsilon_0, x = var_9670_cast_fp16)[name = string("op_9671_cast_fp16")];
            fp16 var_9672_to_fp16 = const()[name = string("op_9672_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9673_cast_fp16 = mul(x = var_9671_cast_fp16, y = var_9672_to_fp16)[name = string("op_9673_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9675_cast_fp16 = add(x = var_9669_cast_fp16, y = var_9673_cast_fp16)[name = string("op_9675_cast_fp16")];
            fp32 var_9676_epsilon_0 = const()[name = string("op_9676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9676_cast_fp16 = rsqrt(epsilon = var_9676_epsilon_0, x = var_9675_cast_fp16)[name = string("op_9676_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_319_cast_fp16 = mul(x = y_365_cast_fp16, y = var_9676_cast_fp16)[name = string("out_319_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126950976)))];
            tensor<fp16, [1, 32, 512]> input_269_cast_fp16 = mul(x = out_319_cast_fp16, y = tm_layers_22_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126952064))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128000704))))[name = string("tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_203_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_203_cast_fp16")];
            string var_9683_mode_0 = const()[name = string("op_9683_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_9683_cast_fp16 = gelu(mode = var_9683_mode_0, x = linear_203_cast_fp16)[name = string("op_9683_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128004864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129053504))))[name = string("tm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_204_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_271_cast_fp16 = mul(x = var_9683_cast_fp16, y = linear_204_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129057664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130106304))))[name = string("tm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9691_cast_fp16 = abs(x = linear_205_cast_fp16)[name = string("op_9691_cast_fp16")];
            bool var_9694 = const()[name = string("op_9694"), val = bool(true)];
            tensor<int32, [1]> var_9695_axes_0 = const()[name = string("op_9695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9695_cast_fp16 = reduce_max(axes = var_9695_axes_0, keep_dims = var_9694, x = var_9691_cast_fp16)[name = string("op_9695_cast_fp16")];
            fp16 var_9696_to_fp16 = const()[name = string("op_9696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_367_cast_fp16 = maximum(x = var_9695_cast_fp16, y = var_9696_to_fp16)[name = string("s_367_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_367_cast_fp16 = real_div(x = linear_205_cast_fp16, y = s_367_cast_fp16)[name = string("y_367_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9699_cast_fp16 = mul(x = y_367_cast_fp16, y = y_367_cast_fp16)[name = string("op_9699_cast_fp16")];
            tensor<int32, [1]> var_9704_axes_0 = const()[name = string("op_9704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9704_keep_dims_0 = const()[name = string("op_9704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9704_cast_fp16 = reduce_mean(axes = var_9704_axes_0, keep_dims = var_9704_keep_dims_0, x = var_9699_cast_fp16)[name = string("op_9704_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9705_cast_fp16 = mul(x = s_367_cast_fp16, y = s_367_cast_fp16)[name = string("op_9705_cast_fp16")];
            fp32 var_9706_epsilon_0 = const()[name = string("op_9706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9706_cast_fp16 = inverse(epsilon = var_9706_epsilon_0, x = var_9705_cast_fp16)[name = string("op_9706_cast_fp16")];
            fp16 var_9707_to_fp16 = const()[name = string("op_9707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9708_cast_fp16 = mul(x = var_9706_cast_fp16, y = var_9707_to_fp16)[name = string("op_9708_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9710_cast_fp16 = add(x = var_9704_cast_fp16, y = var_9708_cast_fp16)[name = string("op_9710_cast_fp16")];
            fp32 var_9711_epsilon_0 = const()[name = string("op_9711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9711_cast_fp16 = rsqrt(epsilon = var_9711_epsilon_0, x = var_9710_cast_fp16)[name = string("op_9711_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_321_cast_fp16 = mul(x = y_367_cast_fp16, y = var_9711_cast_fp16)[name = string("out_321_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130107392)))];
            tensor<fp16, [1, 32, 512]> var_9713_cast_fp16 = mul(x = out_321_cast_fp16, y = tm_layers_22_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9713_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_273_cast_fp16 = add(x = x_407_cast_fp16, y = var_9713_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130108480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130370688))))[name = string("tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_206_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_206_cast_fp16")];
            string x_411_mode_0 = const()[name = string("x_411_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_411_cast_fp16 = gelu(mode = x_411_mode_0, x = linear_206_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [4]> var_9733_begin_0 = const()[name = string("op_9733_begin_0"), val = tensor<int32, [4]>([0, 0, 22, 0])];
            tensor<int32, [4]> var_9733_end_0 = const()[name = string("op_9733_end_0"), val = tensor<int32, [4]>([1, 32, 23, 512])];
            tensor<bool, [4]> var_9733_end_mask_0 = const()[name = string("op_9733_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9733_squeeze_mask_0 = const()[name = string("op_9733_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_9733_cast_fp16 = slice_by_index(begin = var_9733_begin_0, end = var_9733_end_0, end_mask = var_9733_end_mask_0, squeeze_mask = var_9733_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9733_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_275_cast_fp16 = mul(x = x_411_cast_fp16, y = var_9733_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130371776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130633984))))[name = string("tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9743_cast_fp16 = abs(x = linear_207_cast_fp16)[name = string("op_9743_cast_fp16")];
            bool var_9746 = const()[name = string("op_9746"), val = bool(true)];
            tensor<int32, [1]> var_9747_axes_0 = const()[name = string("op_9747_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9747_cast_fp16 = reduce_max(axes = var_9747_axes_0, keep_dims = var_9746, x = var_9743_cast_fp16)[name = string("op_9747_cast_fp16")];
            fp16 var_9748_to_fp16 = const()[name = string("op_9748_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_369_cast_fp16 = maximum(x = var_9747_cast_fp16, y = var_9748_to_fp16)[name = string("s_369_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_369_cast_fp16 = real_div(x = linear_207_cast_fp16, y = s_369_cast_fp16)[name = string("y_369_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9751_cast_fp16 = mul(x = y_369_cast_fp16, y = y_369_cast_fp16)[name = string("op_9751_cast_fp16")];
            tensor<int32, [1]> var_9756_axes_0 = const()[name = string("op_9756_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9756_keep_dims_0 = const()[name = string("op_9756_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9756_cast_fp16 = reduce_mean(axes = var_9756_axes_0, keep_dims = var_9756_keep_dims_0, x = var_9751_cast_fp16)[name = string("op_9756_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9757_cast_fp16 = mul(x = s_369_cast_fp16, y = s_369_cast_fp16)[name = string("op_9757_cast_fp16")];
            fp32 var_9758_epsilon_0 = const()[name = string("op_9758_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9758_cast_fp16 = inverse(epsilon = var_9758_epsilon_0, x = var_9757_cast_fp16)[name = string("op_9758_cast_fp16")];
            fp16 var_9759_to_fp16 = const()[name = string("op_9759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9760_cast_fp16 = mul(x = var_9758_cast_fp16, y = var_9759_to_fp16)[name = string("op_9760_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9762_cast_fp16 = add(x = var_9756_cast_fp16, y = var_9760_cast_fp16)[name = string("op_9762_cast_fp16")];
            fp32 var_9763_epsilon_0 = const()[name = string("op_9763_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9763_cast_fp16 = rsqrt(epsilon = var_9763_epsilon_0, x = var_9762_cast_fp16)[name = string("op_9763_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_323_cast_fp16 = mul(x = y_369_cast_fp16, y = var_9763_cast_fp16)[name = string("out_323_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130635072)))];
            tensor<fp16, [1, 32, 512]> var_9765_cast_fp16 = mul(x = out_323_cast_fp16, y = tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9765_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_45_cast_fp16 = add(x = input_273_cast_fp16, y = var_9765_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<fp16, [1]> tm_layers_22_layer_scalar_to_fp16 = const()[name = string("tm_layers_22_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 32, 512]> x_415_cast_fp16 = mul(x = h_45_cast_fp16, y = tm_layers_22_layer_scalar_to_fp16)[name = string("x_415_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9769_cast_fp16 = abs(x = x_415_cast_fp16)[name = string("op_9769_cast_fp16")];
            bool var_9772 = const()[name = string("op_9772"), val = bool(true)];
            tensor<int32, [1]> var_9773_axes_0 = const()[name = string("op_9773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9773_cast_fp16 = reduce_max(axes = var_9773_axes_0, keep_dims = var_9772, x = var_9769_cast_fp16)[name = string("op_9773_cast_fp16")];
            fp16 var_9774_to_fp16 = const()[name = string("op_9774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_371_cast_fp16 = maximum(x = var_9773_cast_fp16, y = var_9774_to_fp16)[name = string("s_371_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_371_cast_fp16 = real_div(x = x_415_cast_fp16, y = s_371_cast_fp16)[name = string("y_371_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9777_cast_fp16 = mul(x = y_371_cast_fp16, y = y_371_cast_fp16)[name = string("op_9777_cast_fp16")];
            tensor<int32, [1]> var_9782_axes_0 = const()[name = string("op_9782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9782_keep_dims_0 = const()[name = string("op_9782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9782_cast_fp16 = reduce_mean(axes = var_9782_axes_0, keep_dims = var_9782_keep_dims_0, x = var_9777_cast_fp16)[name = string("op_9782_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9783_cast_fp16 = mul(x = s_371_cast_fp16, y = s_371_cast_fp16)[name = string("op_9783_cast_fp16")];
            fp32 var_9784_epsilon_0 = const()[name = string("op_9784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9784_cast_fp16 = inverse(epsilon = var_9784_epsilon_0, x = var_9783_cast_fp16)[name = string("op_9784_cast_fp16")];
            fp16 var_9785_to_fp16 = const()[name = string("op_9785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9786_cast_fp16 = mul(x = var_9784_cast_fp16, y = var_9785_to_fp16)[name = string("op_9786_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9788_cast_fp16 = add(x = var_9782_cast_fp16, y = var_9786_cast_fp16)[name = string("op_9788_cast_fp16")];
            fp32 var_9789_epsilon_0 = const()[name = string("op_9789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9789_cast_fp16 = rsqrt(epsilon = var_9789_epsilon_0, x = var_9788_cast_fp16)[name = string("op_9789_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_325_cast_fp16 = mul(x = y_371_cast_fp16, y = var_9789_cast_fp16)[name = string("out_325_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130636160)))];
            tensor<fp16, [1, 32, 512]> input_277_cast_fp16 = mul(x = out_325_cast_fp16, y = tm_layers_23_input_layernorm_weight_to_fp16)[name = string("input_277_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130637248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131685888))))[name = string("tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_208_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<int32, [4]> var_9799 = const()[name = string("op_9799"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 32, 4, 512]> var_9800_cast_fp16 = reshape(shape = var_9799, x = linear_208_cast_fp16)[name = string("op_9800_cast_fp16")];
            tensor<int32, [4]> x_417_perm_0 = const()[name = string("x_417_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131690048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131952256))))[name = string("tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<int32, [4]> var_9811 = const()[name = string("op_9811"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 32, 1, 512]> var_9812_cast_fp16 = reshape(shape = var_9811, x = linear_209_cast_fp16)[name = string("op_9812_cast_fp16")];
            tensor<int32, [4]> x_419_perm_0 = const()[name = string("x_419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131953344))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132215552))))[name = string("tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<int32, [4]> var_9823 = const()[name = string("op_9823"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 32, 1, 512]> var_9824_cast_fp16 = reshape(shape = var_9823, x = linear_210_cast_fp16)[name = string("op_9824_cast_fp16")];
            tensor<int32, [4]> x_421_perm_0 = const()[name = string("x_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 32, 512]> x_417_cast_fp16 = transpose(perm = x_417_perm_0, x = var_9800_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 4, 32, 512]> var_9828_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_9828_cast_fp16")];
            bool var_9831 = const()[name = string("op_9831"), val = bool(true)];
            tensor<int32, [1]> var_9832_axes_0 = const()[name = string("op_9832_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 32, 1]> var_9832_cast_fp16 = reduce_max(axes = var_9832_axes_0, keep_dims = var_9831, x = var_9828_cast_fp16)[name = string("op_9832_cast_fp16")];
            fp16 var_9833_to_fp16 = const()[name = string("op_9833_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 32, 1]> s_373_cast_fp16 = maximum(x = var_9832_cast_fp16, y = var_9833_to_fp16)[name = string("s_373_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> y_373_cast_fp16 = real_div(x = x_417_cast_fp16, y = s_373_cast_fp16)[name = string("y_373_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_9836_cast_fp16 = mul(x = y_373_cast_fp16, y = y_373_cast_fp16)[name = string("op_9836_cast_fp16")];
            tensor<int32, [1]> var_9841_axes_0 = const()[name = string("op_9841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9841_keep_dims_0 = const()[name = string("op_9841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 32, 1]> var_9841_cast_fp16 = reduce_mean(axes = var_9841_axes_0, keep_dims = var_9841_keep_dims_0, x = var_9836_cast_fp16)[name = string("op_9841_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_9842_cast_fp16 = mul(x = s_373_cast_fp16, y = s_373_cast_fp16)[name = string("op_9842_cast_fp16")];
            fp32 var_9843_epsilon_0 = const()[name = string("op_9843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 32, 1]> var_9843_cast_fp16 = inverse(epsilon = var_9843_epsilon_0, x = var_9842_cast_fp16)[name = string("op_9843_cast_fp16")];
            fp16 var_9844_to_fp16 = const()[name = string("op_9844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 32, 1]> var_9845_cast_fp16 = mul(x = var_9843_cast_fp16, y = var_9844_to_fp16)[name = string("op_9845_cast_fp16")];
            tensor<fp16, [1, 4, 32, 1]> var_9847_cast_fp16 = add(x = var_9841_cast_fp16, y = var_9845_cast_fp16)[name = string("op_9847_cast_fp16")];
            fp32 var_9848_epsilon_0 = const()[name = string("op_9848_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 32, 1]> var_9848_cast_fp16 = rsqrt(epsilon = var_9848_epsilon_0, x = var_9847_cast_fp16)[name = string("op_9848_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> out_327_cast_fp16 = mul(x = y_373_cast_fp16, y = var_9848_cast_fp16)[name = string("out_327_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132216640)))];
            tensor<fp16, [1, 4, 32, 512]> q_93_cast_fp16 = mul(x = out_327_cast_fp16, y = tm_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("q_93_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> x_419_cast_fp16 = transpose(perm = x_419_perm_0, x = var_9812_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1, 32, 512]> var_9851_cast_fp16 = abs(x = x_419_cast_fp16)[name = string("op_9851_cast_fp16")];
            bool var_9854 = const()[name = string("op_9854"), val = bool(true)];
            tensor<int32, [1]> var_9855_axes_0 = const()[name = string("op_9855_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 32, 1]> var_9855_cast_fp16 = reduce_max(axes = var_9855_axes_0, keep_dims = var_9854, x = var_9851_cast_fp16)[name = string("op_9855_cast_fp16")];
            fp16 var_9856_to_fp16 = const()[name = string("op_9856_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 32, 1]> s_375_cast_fp16 = maximum(x = var_9855_cast_fp16, y = var_9856_to_fp16)[name = string("s_375_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> y_375_cast_fp16 = real_div(x = x_419_cast_fp16, y = s_375_cast_fp16)[name = string("y_375_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_9859_cast_fp16 = mul(x = y_375_cast_fp16, y = y_375_cast_fp16)[name = string("op_9859_cast_fp16")];
            tensor<int32, [1]> var_9864_axes_0 = const()[name = string("op_9864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9864_keep_dims_0 = const()[name = string("op_9864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 32, 1]> var_9864_cast_fp16 = reduce_mean(axes = var_9864_axes_0, keep_dims = var_9864_keep_dims_0, x = var_9859_cast_fp16)[name = string("op_9864_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_9865_cast_fp16 = mul(x = s_375_cast_fp16, y = s_375_cast_fp16)[name = string("op_9865_cast_fp16")];
            fp32 var_9866_epsilon_0 = const()[name = string("op_9866_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 32, 1]> var_9866_cast_fp16 = inverse(epsilon = var_9866_epsilon_0, x = var_9865_cast_fp16)[name = string("op_9866_cast_fp16")];
            fp16 var_9867_to_fp16 = const()[name = string("op_9867_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 32, 1]> var_9868_cast_fp16 = mul(x = var_9866_cast_fp16, y = var_9867_to_fp16)[name = string("op_9868_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_9870_cast_fp16 = add(x = var_9864_cast_fp16, y = var_9868_cast_fp16)[name = string("op_9870_cast_fp16")];
            fp32 var_9871_epsilon_0 = const()[name = string("op_9871_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 32, 1]> var_9871_cast_fp16 = rsqrt(epsilon = var_9871_epsilon_0, x = var_9870_cast_fp16)[name = string("op_9871_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> out_329_cast_fp16 = mul(x = y_375_cast_fp16, y = var_9871_cast_fp16)[name = string("out_329_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132217728)))];
            tensor<fp16, [1, 1, 32, 512]> k_139_cast_fp16 = mul(x = out_329_cast_fp16, y = tm_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("k_139_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> x_421_cast_fp16 = transpose(perm = x_421_perm_0, x = var_9824_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 1, 32, 512]> var_9874_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_9874_cast_fp16")];
            bool var_9877 = const()[name = string("op_9877"), val = bool(true)];
            tensor<int32, [1]> var_9878_axes_0 = const()[name = string("op_9878_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 32, 1]> var_9878_cast_fp16 = reduce_max(axes = var_9878_axes_0, keep_dims = var_9877, x = var_9874_cast_fp16)[name = string("op_9878_cast_fp16")];
            fp16 var_9879_to_fp16 = const()[name = string("op_9879_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 32, 1]> s_377_cast_fp16 = maximum(x = var_9878_cast_fp16, y = var_9879_to_fp16)[name = string("s_377_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> y_377_cast_fp16 = real_div(x = x_421_cast_fp16, y = s_377_cast_fp16)[name = string("y_377_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_9882_cast_fp16 = mul(x = y_377_cast_fp16, y = y_377_cast_fp16)[name = string("op_9882_cast_fp16")];
            tensor<int32, [1]> var_9887_axes_0 = const()[name = string("op_9887_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9887_keep_dims_0 = const()[name = string("op_9887_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 32, 1]> var_9887_cast_fp16 = reduce_mean(axes = var_9887_axes_0, keep_dims = var_9887_keep_dims_0, x = var_9882_cast_fp16)[name = string("op_9887_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_9888_cast_fp16 = mul(x = s_377_cast_fp16, y = s_377_cast_fp16)[name = string("op_9888_cast_fp16")];
            fp32 var_9889_epsilon_0 = const()[name = string("op_9889_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 32, 1]> var_9889_cast_fp16 = inverse(epsilon = var_9889_epsilon_0, x = var_9888_cast_fp16)[name = string("op_9889_cast_fp16")];
            fp16 var_9890_to_fp16 = const()[name = string("op_9890_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 32, 1]> var_9891_cast_fp16 = mul(x = var_9889_cast_fp16, y = var_9890_to_fp16)[name = string("op_9891_cast_fp16")];
            tensor<fp16, [1, 1, 32, 1]> var_9893_cast_fp16 = add(x = var_9887_cast_fp16, y = var_9891_cast_fp16)[name = string("op_9893_cast_fp16")];
            fp32 var_9894_epsilon_0 = const()[name = string("op_9894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 32, 1]> var_9894_cast_fp16 = rsqrt(epsilon = var_9894_epsilon_0, x = var_9893_cast_fp16)[name = string("op_9894_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> v_93_cast_fp16 = mul(x = y_377_cast_fp16, y = var_9894_cast_fp16)[name = string("v_93_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_9896_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9896_cast_fp16")];
            tensor<int32, [4]> var_9901_begin_0 = const()[name = string("op_9901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9901_end_0 = const()[name = string("op_9901_end_0"), val = tensor<int32, [4]>([1, 4, 32, 512])];
            tensor<bool, [4]> var_9901_end_mask_0 = const()[name = string("op_9901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 32, 256]> var_9901_cast_fp16 = slice_by_index(begin = var_9901_begin_0, end = var_9901_end_0, end_mask = var_9901_end_mask_0, x = q_93_cast_fp16)[name = string("op_9901_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 32, 256]> var_9902_cast_fp16 = mul(x = var_9901_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_9902_cast_fp16")];
            tensor<int32, [4]> var_9907_begin_0 = const()[name = string("op_9907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9907_end_0 = const()[name = string("op_9907_end_0"), val = tensor<int32, [4]>([1, 4, 32, 256])];
            tensor<bool, [4]> var_9907_end_mask_0 = const()[name = string("op_9907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 32, 256]> var_9907_cast_fp16 = slice_by_index(begin = var_9907_begin_0, end = var_9907_end_0, end_mask = var_9907_end_mask_0, x = q_93_cast_fp16)[name = string("op_9907_cast_fp16")];
            int32 var_9909 = const()[name = string("op_9909"), val = int32(-1)];
            bool var_9910_interleave_0 = const()[name = string("op_9910_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 512]> var_9910_cast_fp16 = concat(axis = var_9909, interleave = var_9910_interleave_0, values = (var_9902_cast_fp16, var_9907_cast_fp16))[name = string("op_9910_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> var_9911_cast_fp16 = mul(x = var_9910_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9911_cast_fp16")];
            tensor<fp16, [1, 4, 32, 512]> q_cast_fp16 = add(x = var_9896_cast_fp16, y = var_9911_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_9914_cast_fp16 = mul(x = k_139_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9914_cast_fp16")];
            tensor<int32, [4]> var_9919_begin_0 = const()[name = string("op_9919_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9919_end_0 = const()[name = string("op_9919_end_0"), val = tensor<int32, [4]>([1, 1, 32, 512])];
            tensor<bool, [4]> var_9919_end_mask_0 = const()[name = string("op_9919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 32, 256]> var_9919_cast_fp16 = slice_by_index(begin = var_9919_begin_0, end = var_9919_end_0, end_mask = var_9919_end_mask_0, x = k_139_cast_fp16)[name = string("op_9919_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 32, 256]> var_9920_cast_fp16 = mul(x = var_9919_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_9920_cast_fp16")];
            tensor<int32, [4]> var_9925_begin_0 = const()[name = string("op_9925_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9925_end_0 = const()[name = string("op_9925_end_0"), val = tensor<int32, [4]>([1, 1, 32, 256])];
            tensor<bool, [4]> var_9925_end_mask_0 = const()[name = string("op_9925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 32, 256]> var_9925_cast_fp16 = slice_by_index(begin = var_9925_begin_0, end = var_9925_end_0, end_mask = var_9925_end_mask_0, x = k_139_cast_fp16)[name = string("op_9925_cast_fp16")];
            int32 var_9927 = const()[name = string("op_9927"), val = int32(-1)];
            bool var_9928_interleave_0 = const()[name = string("op_9928_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 32, 512]> var_9928_cast_fp16 = concat(axis = var_9927, interleave = var_9928_interleave_0, values = (var_9920_cast_fp16, var_9925_cast_fp16))[name = string("op_9928_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> var_9929_cast_fp16 = mul(x = var_9928_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9929_cast_fp16")];
            tensor<fp16, [1, 1, 32, 512]> k_141_cast_fp16 = add(x = var_9914_cast_fp16, y = var_9929_cast_fp16)[name = string("k_141_cast_fp16")];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_46_reps_0 = const()[name = string("tile_46_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 32, 512]> transpose_92_cast_fp16 = transpose(perm = transpose_92_perm_0, x = k_141_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 1, 32, 512]> tile_46_cast_fp16 = tile(reps = tile_46_reps_0, x = transpose_92_cast_fp16)[name = string("tile_46_cast_fp16")];
            tensor<int32, [5]> concat_92 = const()[name = string("concat_92"), val = tensor<int32, [5]>([4, 1, 1, 32, 512])];
            tensor<fp16, [4, 1, 1, 32, 512]> reshape_92_cast_fp16 = reshape(shape = concat_92, x = tile_46_cast_fp16)[name = string("reshape_92_cast_fp16")];
            tensor<int32, [5]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_93 = const()[name = string("concat_93"), val = tensor<int32, [4]>([-1, 1, 32, 512])];
            tensor<fp16, [1, 4, 1, 32, 512]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = reshape_92_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 1, 32, 512]> reshape_93_cast_fp16 = reshape(shape = concat_93, x = transpose_93_cast_fp16)[name = string("reshape_93_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_47_reps_0 = const()[name = string("tile_47_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 32, 512]> transpose_94_cast_fp16 = transpose(perm = transpose_94_perm_0, x = v_93_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 1, 32, 512]> tile_47_cast_fp16 = tile(reps = tile_47_reps_0, x = transpose_94_cast_fp16)[name = string("tile_47_cast_fp16")];
            tensor<int32, [5]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [5]>([4, 1, 1, 32, 512])];
            tensor<fp16, [4, 1, 1, 32, 512]> reshape_94_cast_fp16 = reshape(shape = concat_94, x = tile_47_cast_fp16)[name = string("reshape_94_cast_fp16")];
            tensor<int32, [5]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [4]>([-1, 1, 32, 512])];
            tensor<fp16, [1, 4, 1, 32, 512]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = reshape_94_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 1, 32, 512]> reshape_95_cast_fp16 = reshape(shape = concat_95, x = transpose_95_cast_fp16)[name = string("reshape_95_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9943_transpose_x_0 = const()[name = string("op_9943_transpose_x_0"), val = bool(false)];
            bool var_9943_transpose_y_0 = const()[name = string("op_9943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 32]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = reshape_93_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 4, 32, 32]> var_9943_cast_fp16 = matmul(transpose_x = var_9943_transpose_x_0, transpose_y = var_9943_transpose_y_0, x = q_cast_fp16, y = transpose_119_cast_fp16)[name = string("op_9943_cast_fp16")];
            tensor<fp16, [1, 4, 32, 32]> var_9945_cast_fp16 = add(x = var_9943_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9945_cast_fp16")];
            int32 var_9946 = const()[name = string("op_9946"), val = int32(-1)];
            tensor<fp16, [1, 4, 32, 32]> w_377_cast_fp16 = softmax(axis = var_9946, x = var_9945_cast_fp16)[name = string("w_377_cast_fp16")];
            bool var_9949_transpose_x_0 = const()[name = string("op_9949_transpose_x_0"), val = bool(false)];
            bool var_9949_transpose_y_0 = const()[name = string("op_9949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 32, 512]> v_cast_fp16 = transpose(perm = v_perm_0, x = reshape_95_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 4, 32, 512]> var_9949_cast_fp16 = matmul(transpose_x = var_9949_transpose_x_0, transpose_y = var_9949_transpose_y_0, x = w_377_cast_fp16, y = v_cast_fp16)[name = string("op_9949_cast_fp16")];
            tensor<int32, [4]> var_9952_perm_0 = const()[name = string("op_9952_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9956 = const()[name = string("op_9956"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 32, 4, 512]> var_9952_cast_fp16 = transpose(perm = var_9952_perm_0, x = var_9949_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 32, 2048]> input_279_cast_fp16 = reshape(shape = var_9956, x = var_9952_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132218816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133267456))))[name = string("tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9961_cast_fp16 = abs(x = linear_211_cast_fp16)[name = string("op_9961_cast_fp16")];
            bool var_9964 = const()[name = string("op_9964"), val = bool(true)];
            tensor<int32, [1]> var_9965_axes_0 = const()[name = string("op_9965_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9965_cast_fp16 = reduce_max(axes = var_9965_axes_0, keep_dims = var_9964, x = var_9961_cast_fp16)[name = string("op_9965_cast_fp16")];
            fp16 var_9966_to_fp16 = const()[name = string("op_9966_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_379_cast_fp16 = maximum(x = var_9965_cast_fp16, y = var_9966_to_fp16)[name = string("s_379_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_379_cast_fp16 = real_div(x = linear_211_cast_fp16, y = s_379_cast_fp16)[name = string("y_379_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9969_cast_fp16 = mul(x = y_379_cast_fp16, y = y_379_cast_fp16)[name = string("op_9969_cast_fp16")];
            tensor<int32, [1]> var_9974_axes_0 = const()[name = string("op_9974_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9974_keep_dims_0 = const()[name = string("op_9974_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9974_cast_fp16 = reduce_mean(axes = var_9974_axes_0, keep_dims = var_9974_keep_dims_0, x = var_9969_cast_fp16)[name = string("op_9974_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9975_cast_fp16 = mul(x = s_379_cast_fp16, y = s_379_cast_fp16)[name = string("op_9975_cast_fp16")];
            fp32 var_9976_epsilon_0 = const()[name = string("op_9976_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_9976_cast_fp16 = inverse(epsilon = var_9976_epsilon_0, x = var_9975_cast_fp16)[name = string("op_9976_cast_fp16")];
            fp16 var_9977_to_fp16 = const()[name = string("op_9977_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_9978_cast_fp16 = mul(x = var_9976_cast_fp16, y = var_9977_to_fp16)[name = string("op_9978_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_9980_cast_fp16 = add(x = var_9974_cast_fp16, y = var_9978_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp32 var_9981_epsilon_0 = const()[name = string("op_9981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_9981_cast_fp16 = rsqrt(epsilon = var_9981_epsilon_0, x = var_9980_cast_fp16)[name = string("op_9981_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_331_cast_fp16 = mul(x = y_379_cast_fp16, y = var_9981_cast_fp16)[name = string("out_331_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133268544)))];
            tensor<fp16, [1, 32, 512]> var_9983_cast_fp16 = mul(x = out_331_cast_fp16, y = tm_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("op_9983_cast_fp16")];
            tensor<fp16, [1, 32, 512]> x_425_cast_fp16 = add(x = x_415_cast_fp16, y = var_9983_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9986_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_9986_cast_fp16")];
            bool var_9989 = const()[name = string("op_9989"), val = bool(true)];
            tensor<int32, [1]> var_9990_axes_0 = const()[name = string("op_9990_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_9990_cast_fp16 = reduce_max(axes = var_9990_axes_0, keep_dims = var_9989, x = var_9986_cast_fp16)[name = string("op_9990_cast_fp16")];
            fp16 var_9991_to_fp16 = const()[name = string("op_9991_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_381_cast_fp16 = maximum(x = var_9990_cast_fp16, y = var_9991_to_fp16)[name = string("s_381_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_381_cast_fp16 = real_div(x = x_425_cast_fp16, y = s_381_cast_fp16)[name = string("y_381_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_9994_cast_fp16 = mul(x = y_381_cast_fp16, y = y_381_cast_fp16)[name = string("op_9994_cast_fp16")];
            tensor<int32, [1]> var_9999_axes_0 = const()[name = string("op_9999_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9999_keep_dims_0 = const()[name = string("op_9999_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9999_cast_fp16 = reduce_mean(axes = var_9999_axes_0, keep_dims = var_9999_keep_dims_0, x = var_9994_cast_fp16)[name = string("op_9999_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_10000_cast_fp16 = mul(x = s_381_cast_fp16, y = s_381_cast_fp16)[name = string("op_10000_cast_fp16")];
            fp32 var_10001_epsilon_0 = const()[name = string("op_10001_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_10001_cast_fp16 = inverse(epsilon = var_10001_epsilon_0, x = var_10000_cast_fp16)[name = string("op_10001_cast_fp16")];
            fp16 var_10002_to_fp16 = const()[name = string("op_10002_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_10003_cast_fp16 = mul(x = var_10001_cast_fp16, y = var_10002_to_fp16)[name = string("op_10003_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_10005_cast_fp16 = add(x = var_9999_cast_fp16, y = var_10003_cast_fp16)[name = string("op_10005_cast_fp16")];
            fp32 var_10006_epsilon_0 = const()[name = string("op_10006_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_10006_cast_fp16 = rsqrt(epsilon = var_10006_epsilon_0, x = var_10005_cast_fp16)[name = string("op_10006_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_333_cast_fp16 = mul(x = y_381_cast_fp16, y = var_10006_cast_fp16)[name = string("out_333_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133269632)))];
            tensor<fp16, [1, 32, 512]> input_281_cast_fp16 = mul(x = out_333_cast_fp16, y = tm_layers_23_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133270720))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134319360))))[name = string("tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_212_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_212_cast_fp16")];
            string var_10013_mode_0 = const()[name = string("op_10013_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 2048]> var_10013_cast_fp16 = gelu(mode = var_10013_mode_0, x = linear_212_cast_fp16)[name = string("op_10013_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134323520))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135372160))))[name = string("tm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 2048]> linear_213_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_213_cast_fp16")];
            tensor<fp16, [1, 32, 2048]> input_283_cast_fp16 = mul(x = var_10013_cast_fp16, y = linear_213_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135376320))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136424960))))[name = string("tm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_283_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_10021_cast_fp16 = abs(x = linear_214_cast_fp16)[name = string("op_10021_cast_fp16")];
            bool var_10024 = const()[name = string("op_10024"), val = bool(true)];
            tensor<int32, [1]> var_10025_axes_0 = const()[name = string("op_10025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_10025_cast_fp16 = reduce_max(axes = var_10025_axes_0, keep_dims = var_10024, x = var_10021_cast_fp16)[name = string("op_10025_cast_fp16")];
            fp16 var_10026_to_fp16 = const()[name = string("op_10026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_383_cast_fp16 = maximum(x = var_10025_cast_fp16, y = var_10026_to_fp16)[name = string("s_383_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_383_cast_fp16 = real_div(x = linear_214_cast_fp16, y = s_383_cast_fp16)[name = string("y_383_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_10029_cast_fp16 = mul(x = y_383_cast_fp16, y = y_383_cast_fp16)[name = string("op_10029_cast_fp16")];
            tensor<int32, [1]> var_10034_axes_0 = const()[name = string("op_10034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10034_keep_dims_0 = const()[name = string("op_10034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_10034_cast_fp16 = reduce_mean(axes = var_10034_axes_0, keep_dims = var_10034_keep_dims_0, x = var_10029_cast_fp16)[name = string("op_10034_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_10035_cast_fp16 = mul(x = s_383_cast_fp16, y = s_383_cast_fp16)[name = string("op_10035_cast_fp16")];
            fp32 var_10036_epsilon_0 = const()[name = string("op_10036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_10036_cast_fp16 = inverse(epsilon = var_10036_epsilon_0, x = var_10035_cast_fp16)[name = string("op_10036_cast_fp16")];
            fp16 var_10037_to_fp16 = const()[name = string("op_10037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_10038_cast_fp16 = mul(x = var_10036_cast_fp16, y = var_10037_to_fp16)[name = string("op_10038_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_10040_cast_fp16 = add(x = var_10034_cast_fp16, y = var_10038_cast_fp16)[name = string("op_10040_cast_fp16")];
            fp32 var_10041_epsilon_0 = const()[name = string("op_10041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_10041_cast_fp16 = rsqrt(epsilon = var_10041_epsilon_0, x = var_10040_cast_fp16)[name = string("op_10041_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_335_cast_fp16 = mul(x = y_383_cast_fp16, y = var_10041_cast_fp16)[name = string("out_335_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136426048)))];
            tensor<fp16, [1, 32, 512]> var_10043_cast_fp16 = mul(x = out_335_cast_fp16, y = tm_layers_23_post_feedforward_layernorm_weight_to_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_285_cast_fp16 = add(x = x_425_cast_fp16, y = var_10043_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136427136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136689344))))[name = string("tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_215_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = string("linear_215_cast_fp16")];
            string x_429_mode_0 = const()[name = string("x_429_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 32, 512]> x_429_cast_fp16 = gelu(mode = x_429_mode_0, x = linear_215_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [4]> var_10063_begin_0 = const()[name = string("op_10063_begin_0"), val = tensor<int32, [4]>([0, 0, 23, 0])];
            tensor<int32, [4]> var_10063_end_0 = const()[name = string("op_10063_end_0"), val = tensor<int32, [4]>([1, 32, 24, 512])];
            tensor<bool, [4]> var_10063_end_mask_0 = const()[name = string("op_10063_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_10063_squeeze_mask_0 = const()[name = string("op_10063_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 32, 512]> var_10063_cast_fp16 = slice_by_index(begin = var_10063_begin_0, end = var_10063_end_0, end_mask = var_10063_end_mask_0, squeeze_mask = var_10063_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 32, 512]> input_287_cast_fp16 = mul(x = x_429_cast_fp16, y = var_10063_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136690432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136952640))))[name = string("tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 512]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_10073_cast_fp16 = abs(x = linear_216_cast_fp16)[name = string("op_10073_cast_fp16")];
            bool var_10076 = const()[name = string("op_10076"), val = bool(true)];
            tensor<int32, [1]> var_10077_axes_0 = const()[name = string("op_10077_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_10077_cast_fp16 = reduce_max(axes = var_10077_axes_0, keep_dims = var_10076, x = var_10073_cast_fp16)[name = string("op_10077_cast_fp16")];
            fp16 var_10078_to_fp16 = const()[name = string("op_10078_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_385_cast_fp16 = maximum(x = var_10077_cast_fp16, y = var_10078_to_fp16)[name = string("s_385_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_385_cast_fp16 = real_div(x = linear_216_cast_fp16, y = s_385_cast_fp16)[name = string("y_385_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_10081_cast_fp16 = mul(x = y_385_cast_fp16, y = y_385_cast_fp16)[name = string("op_10081_cast_fp16")];
            tensor<int32, [1]> var_10086_axes_0 = const()[name = string("op_10086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10086_keep_dims_0 = const()[name = string("op_10086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_10086_cast_fp16 = reduce_mean(axes = var_10086_axes_0, keep_dims = var_10086_keep_dims_0, x = var_10081_cast_fp16)[name = string("op_10086_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_10087_cast_fp16 = mul(x = s_385_cast_fp16, y = s_385_cast_fp16)[name = string("op_10087_cast_fp16")];
            fp32 var_10088_epsilon_0 = const()[name = string("op_10088_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_10088_cast_fp16 = inverse(epsilon = var_10088_epsilon_0, x = var_10087_cast_fp16)[name = string("op_10088_cast_fp16")];
            fp16 var_10089_to_fp16 = const()[name = string("op_10089_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_10090_cast_fp16 = mul(x = var_10088_cast_fp16, y = var_10089_to_fp16)[name = string("op_10090_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_10092_cast_fp16 = add(x = var_10086_cast_fp16, y = var_10090_cast_fp16)[name = string("op_10092_cast_fp16")];
            fp32 var_10093_epsilon_0 = const()[name = string("op_10093_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_10093_cast_fp16 = rsqrt(epsilon = var_10093_epsilon_0, x = var_10092_cast_fp16)[name = string("op_10093_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_337_cast_fp16 = mul(x = y_385_cast_fp16, y = var_10093_cast_fp16)[name = string("out_337_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136953728)))];
            tensor<fp16, [1, 32, 512]> var_10095_cast_fp16 = mul(x = out_337_cast_fp16, y = tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_10095_cast_fp16")];
            tensor<fp16, [1, 32, 512]> h_47_cast_fp16 = add(x = input_285_cast_fp16, y = var_10095_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<fp16, [1]> tm_layers_23_layer_scalar_to_fp16 = const()[name = string("tm_layers_23_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d6p-4])];
            tensor<fp16, [1, 32, 512]> x_cast_fp16 = mul(x = h_47_cast_fp16, y = tm_layers_23_layer_scalar_to_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_10099_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_10099_cast_fp16")];
            bool var_10102 = const()[name = string("op_10102"), val = bool(true)];
            tensor<int32, [1]> var_10103_axes_0 = const()[name = string("op_10103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_10103_cast_fp16 = reduce_max(axes = var_10103_axes_0, keep_dims = var_10102, x = var_10099_cast_fp16)[name = string("op_10103_cast_fp16")];
            fp16 var_10104_to_fp16 = const()[name = string("op_10104_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 32, 1]> s_387_cast_fp16 = maximum(x = var_10103_cast_fp16, y = var_10104_to_fp16)[name = string("s_387_cast_fp16")];
            tensor<fp16, [1, 32, 512]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_387_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 32, 512]> var_10107_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_10107_cast_fp16")];
            tensor<int32, [1]> var_10112_axes_0 = const()[name = string("op_10112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10112_keep_dims_0 = const()[name = string("op_10112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_10112_cast_fp16 = reduce_mean(axes = var_10112_axes_0, keep_dims = var_10112_keep_dims_0, x = var_10107_cast_fp16)[name = string("op_10112_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_10113_cast_fp16 = mul(x = s_387_cast_fp16, y = s_387_cast_fp16)[name = string("op_10113_cast_fp16")];
            fp32 var_10114_epsilon_0 = const()[name = string("op_10114_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_10114_cast_fp16 = inverse(epsilon = var_10114_epsilon_0, x = var_10113_cast_fp16)[name = string("op_10114_cast_fp16")];
            fp16 var_10115_to_fp16 = const()[name = string("op_10115_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_10116_cast_fp16 = mul(x = var_10114_cast_fp16, y = var_10115_to_fp16)[name = string("op_10116_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_10118_cast_fp16 = add(x = var_10112_cast_fp16, y = var_10116_cast_fp16)[name = string("op_10118_cast_fp16")];
            fp32 var_10119_epsilon_0 = const()[name = string("op_10119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_10119_cast_fp16 = rsqrt(epsilon = var_10119_epsilon_0, x = var_10118_cast_fp16)[name = string("op_10119_cast_fp16")];
            tensor<fp16, [1, 32, 512]> out_cast_fp16 = mul(x = y_cast_fp16, y = var_10119_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [512]> tm_norm_weight_to_fp16 = const()[name = string("tm_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136954816)))];
            tensor<fp16, [1, 32, 512]> h_cast_fp16 = mul(x = out_cast_fp16, y = tm_norm_weight_to_fp16)[name = string("h_cast_fp16")];
            tensor<int32, [1]> var_10126_axes_0 = const()[name = string("op_10126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10126_keep_dims_0 = const()[name = string("op_10126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10126_cast_fp16 = reduce_sum(axes = var_10126_axes_0, keep_dims = var_10126_keep_dims_0, x = attention_mask)[name = string("op_10126_cast_fp16")];
            tensor<fp16, [1, 32]> var_10127_cast_fp16 = real_div(x = attention_mask, y = var_10126_cast_fp16)[name = string("op_10127_cast_fp16")];
            tensor<int32, [1]> var_10134_axes_0 = const()[name = string("op_10134_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 32]> var_10134_cast_fp16 = expand_dims(axes = var_10134_axes_0, x = var_10127_cast_fp16)[name = string("op_10134_cast_fp16")];
            bool input_transpose_x_0 = const()[name = string("input_transpose_x_0"), val = bool(false)];
            bool input_transpose_y_0 = const()[name = string("input_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512]> input_cast_fp16 = matmul(transpose_x = input_transpose_x_0, transpose_y = input_transpose_y_0, x = var_10134_cast_fp16, y = h_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [768, 512]> tm_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136955904))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137349184))))[name = string("tm_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_217_bias_0_to_fp16 = const()[name = string("linear_217_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137350784)))];
            tensor<fp16, [1, 1, 768]> linear_217_cast_fp16 = linear(bias = linear_217_bias_0_to_fp16, weight = tm_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<int32, [3]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_10151_squeeze_mask_0 = const()[name = string("op_10151_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, squeeze_mask = var_10151_squeeze_mask_0, x = linear_217_cast_fp16)[name = string("op_10151_cast_fp16")];
            tensor<fp16, [1, 768]> var_10157_cast_fp16 = abs(x = var_10151_cast_fp16)[name = string("op_10157_cast_fp16")];
            bool var_10160 = const()[name = string("op_10160"), val = bool(true)];
            tensor<int32, [1]> s_axes_0 = const()[name = string("s_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> s_cast_fp16 = reduce_max(axes = s_axes_0, keep_dims = var_10160, x = var_10157_cast_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [1, 768]> pooled_cast_fp16 = real_div(x = var_10151_cast_fp16, y = s_cast_fp16)[name = string("pooled_cast_fp16")];
            tensor<fp16, [1, 768]> var_10163_cast_fp16 = mul(x = pooled_cast_fp16, y = pooled_cast_fp16)[name = string("op_10163_cast_fp16")];
            tensor<int32, [1]> var_10168_axes_0 = const()[name = string("op_10168_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10168_keep_dims_0 = const()[name = string("op_10168_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10168_cast_fp16 = reduce_sum(axes = var_10168_axes_0, keep_dims = var_10168_keep_dims_0, x = var_10163_cast_fp16)[name = string("op_10168_cast_fp16")];
            fp32 var_10169_epsilon_0 = const()[name = string("op_10169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_10169_cast_fp16 = rsqrt(epsilon = var_10169_epsilon_0, x = var_10168_cast_fp16)[name = string("op_10169_cast_fp16")];
            tensor<fp16, [1, 768]> var_10170_cast_fp16 = mul(x = pooled_cast_fp16, y = var_10169_cast_fp16)[name = string("op_10170_cast_fp16")];
            string var_10170_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_10170_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 768]> embedding = cast(dtype = var_10170_cast_fp16_to_fp32_dtype_0, x = var_10170_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func embed_48<ios18>(tensor<fp16, [1, 48]> attention_mask, tensor<fp16, [1, 48, 512]> inputs_embeds) {
            tensor<fp16, [12288, 512]> tm_ple_per_layer_model_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [12288, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1728))), scale = tensor<fp16, [12288, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6293248))))[name = string("tm_ple_per_layer_model_projection_weight_to_fp16_quantized")];
            tensor<fp16, [12288]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [12288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6317888)))];
            tensor<fp16, [1, 48, 12288]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = tm_ple_per_layer_model_projection_weight_to_fp16_quantized, x = inputs_embeds)[name = string("linear_0_cast_fp16")];
            fp16 var_2129_to_fp16 = const()[name = string("op_2129_to_fp16"), val = fp16(0x1.6ap-5)];
            tensor<fp16, [1, 48, 12288]> ple_1_cast_fp16 = mul(x = linear_0_cast_fp16, y = var_2129_to_fp16)[name = string("ple_1_cast_fp16")];
            tensor<int32, [4]> var_2135 = const()[name = string("op_2135"), val = tensor<int32, [4]>([1, -1, 24, 512])];
            tensor<fp16, [1, 48, 24, 512]> x_1_cast_fp16 = reshape(shape = var_2135, x = ple_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 48, 24, 512]> var_2137_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_2137_cast_fp16")];
            bool var_2140 = const()[name = string("op_2140"), val = bool(true)];
            tensor<int32, [1]> var_2141_axes_0 = const()[name = string("op_2141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 24, 1]> var_2141_cast_fp16 = reduce_max(axes = var_2141_axes_0, keep_dims = var_2140, x = var_2137_cast_fp16)[name = string("op_2141_cast_fp16")];
            fp16 var_2142_to_fp16 = const()[name = string("op_2142_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 24, 1]> s_1_cast_fp16 = maximum(x = var_2141_cast_fp16, y = var_2142_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 48, 24, 512]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 48, 24, 512]> var_2145_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [1]> var_2150_axes_0 = const()[name = string("op_2150_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2150_keep_dims_0 = const()[name = string("op_2150_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 24, 1]> var_2150_cast_fp16 = reduce_mean(axes = var_2150_axes_0, keep_dims = var_2150_keep_dims_0, x = var_2145_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 48, 24, 1]> var_2151_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 24, 1]> var_2152_cast_fp16 = inverse(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            fp16 var_2153_to_fp16 = const()[name = string("op_2153_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 24, 1]> var_2154_cast_fp16 = mul(x = var_2152_cast_fp16, y = var_2153_to_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 48, 24, 1]> var_2156_cast_fp16 = add(x = var_2150_cast_fp16, y = var_2154_cast_fp16)[name = string("op_2156_cast_fp16")];
            fp32 var_2157_epsilon_0 = const()[name = string("op_2157_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 24, 1]> var_2157_cast_fp16 = rsqrt(epsilon = var_2157_epsilon_0, x = var_2156_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<fp16, [1, 48, 24, 512]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_2157_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [512]> tm_ple_per_layer_projection_norm_weight_to_fp16 = const()[name = string("tm_ple_per_layer_projection_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6342528)))];
            tensor<fp16, [1, 48, 24, 512]> ple_3_cast_fp16 = mul(x = out_1_cast_fp16, y = tm_ple_per_layer_projection_norm_weight_to_fp16)[name = string("ple_3_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 48]> var_2162_cast_fp16 = sub(x = var_2160_to_fp16, y = attention_mask)[name = string("op_2162_cast_fp16")];
            fp16 var_2163_to_fp16 = const()[name = string("op_2163_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 48]> var_2164_cast_fp16 = mul(x = var_2162_cast_fp16, y = var_2163_to_fp16)[name = string("op_2164_cast_fp16")];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 48]> var_2171_cast_fp16 = expand_dims(axes = var_2171_axes_0, x = var_2164_cast_fp16)[name = string("op_2171_cast_fp16")];
            tensor<int32, [1]> var_2173_axes_0 = const()[name = string("op_2173_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 48]> var_2173_cast_fp16 = expand_dims(axes = var_2173_axes_0, x = var_2171_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2179_cast_fp16 = abs(x = inputs_embeds)[name = string("op_2179_cast_fp16")];
            bool var_2182 = const()[name = string("op_2182"), val = bool(true)];
            tensor<int32, [1]> var_2183_axes_0 = const()[name = string("op_2183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2183_cast_fp16 = reduce_max(axes = var_2183_axes_0, keep_dims = var_2182, x = var_2179_cast_fp16)[name = string("op_2183_cast_fp16")];
            fp16 var_2184_to_fp16 = const()[name = string("op_2184_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_3_cast_fp16 = maximum(x = var_2183_cast_fp16, y = var_2184_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_3_cast_fp16 = real_div(x = inputs_embeds, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2187_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> var_2192_axes_0 = const()[name = string("op_2192_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2192_keep_dims_0 = const()[name = string("op_2192_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2192_cast_fp16 = reduce_mean(axes = var_2192_axes_0, keep_dims = var_2192_keep_dims_0, x = var_2187_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2193_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp32 var_2194_epsilon_0 = const()[name = string("op_2194_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2194_cast_fp16 = inverse(epsilon = var_2194_epsilon_0, x = var_2193_cast_fp16)[name = string("op_2194_cast_fp16")];
            fp16 var_2195_to_fp16 = const()[name = string("op_2195_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2196_cast_fp16 = mul(x = var_2194_cast_fp16, y = var_2195_to_fp16)[name = string("op_2196_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2198_cast_fp16 = add(x = var_2192_cast_fp16, y = var_2196_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_2199_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6343616)))];
            tensor<fp16, [1, 48, 512]> input_1_cast_fp16 = mul(x = out_3_cast_fp16, y = tm_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6344704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6869056))))[name = string("tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6871168)))];
            tensor<fp16, [1, 48, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_1_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6873280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7135488))))[name = string("tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7136576)))];
            tensor<fp16, [1, 48, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_2221 = const()[name = string("op_2221"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_2222_cast_fp16 = reshape(shape = var_2221, x = linear_2_cast_fp16)[name = string("op_2222_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7137664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7399872))))[name = string("tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_2234_cast_fp16 = reshape(shape = var_2233, x = linear_3_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<int32, [4]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_2210_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 4, 48, 256]> var_2238_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_2238_cast_fp16")];
            bool var_2241 = const()[name = string("op_2241"), val = bool(true)];
            tensor<int32, [1]> var_2242_axes_0 = const()[name = string("op_2242_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_2242_cast_fp16 = reduce_max(axes = var_2242_axes_0, keep_dims = var_2241, x = var_2238_cast_fp16)[name = string("op_2242_cast_fp16")];
            fp16 var_2243_to_fp16 = const()[name = string("op_2243_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_5_cast_fp16 = maximum(x = var_2242_cast_fp16, y = var_2243_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_2246_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<int32, [1]> var_2251_axes_0 = const()[name = string("op_2251_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2251_keep_dims_0 = const()[name = string("op_2251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_2251_cast_fp16 = reduce_mean(axes = var_2251_axes_0, keep_dims = var_2251_keep_dims_0, x = var_2246_cast_fp16)[name = string("op_2251_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_2252_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_2252_cast_fp16")];
            fp32 var_2253_epsilon_0 = const()[name = string("op_2253_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_2253_cast_fp16 = inverse(epsilon = var_2253_epsilon_0, x = var_2252_cast_fp16)[name = string("op_2253_cast_fp16")];
            fp16 var_2254_to_fp16 = const()[name = string("op_2254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_2255_cast_fp16 = mul(x = var_2253_cast_fp16, y = var_2254_to_fp16)[name = string("op_2255_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_2257_cast_fp16 = add(x = var_2251_cast_fp16, y = var_2255_cast_fp16)[name = string("op_2257_cast_fp16")];
            fp32 var_2258_epsilon_0 = const()[name = string("op_2258_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_2258_cast_fp16 = rsqrt(epsilon = var_2258_epsilon_0, x = var_2257_cast_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_2258_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7400960)))];
            tensor<fp16, [1, 4, 48, 256]> q_1_cast_fp16 = mul(x = out_5_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_2222_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 2, 48, 256]> var_2261_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_2261_cast_fp16")];
            bool var_2264 = const()[name = string("op_2264"), val = bool(true)];
            tensor<int32, [1]> var_2265_axes_0 = const()[name = string("op_2265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_2265_cast_fp16 = reduce_max(axes = var_2265_axes_0, keep_dims = var_2264, x = var_2261_cast_fp16)[name = string("op_2265_cast_fp16")];
            fp16 var_2266_to_fp16 = const()[name = string("op_2266_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_7_cast_fp16 = maximum(x = var_2265_cast_fp16, y = var_2266_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2269_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> var_2274_axes_0 = const()[name = string("op_2274_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2274_keep_dims_0 = const()[name = string("op_2274_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_2274_cast_fp16 = reduce_mean(axes = var_2274_axes_0, keep_dims = var_2274_keep_dims_0, x = var_2269_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2275_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_2276_cast_fp16 = inverse(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            fp16 var_2277_to_fp16 = const()[name = string("op_2277_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_2278_cast_fp16 = mul(x = var_2276_cast_fp16, y = var_2277_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2280_cast_fp16 = add(x = var_2274_cast_fp16, y = var_2278_cast_fp16)[name = string("op_2280_cast_fp16")];
            fp32 var_2281_epsilon_0 = const()[name = string("op_2281_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_2281_cast_fp16 = rsqrt(epsilon = var_2281_epsilon_0, x = var_2280_cast_fp16)[name = string("op_2281_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_7_cast_fp16 = mul(x = y_7_cast_fp16, y = var_2281_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7401536)))];
            tensor<fp16, [1, 2, 48, 256]> k_1_cast_fp16 = mul(x = out_7_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = var_2234_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2, 48, 256]> var_2284_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_2284_cast_fp16")];
            bool var_2287 = const()[name = string("op_2287"), val = bool(true)];
            tensor<int32, [1]> var_2288_axes_0 = const()[name = string("op_2288_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_2288_cast_fp16 = reduce_max(axes = var_2288_axes_0, keep_dims = var_2287, x = var_2284_cast_fp16)[name = string("op_2288_cast_fp16")];
            fp16 var_2289_to_fp16 = const()[name = string("op_2289_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_9_cast_fp16 = maximum(x = var_2288_cast_fp16, y = var_2289_to_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_9_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2292_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_2292_cast_fp16")];
            tensor<int32, [1]> var_2297_axes_0 = const()[name = string("op_2297_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2297_keep_dims_0 = const()[name = string("op_2297_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_2297_cast_fp16 = reduce_mean(axes = var_2297_axes_0, keep_dims = var_2297_keep_dims_0, x = var_2292_cast_fp16)[name = string("op_2297_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2298_cast_fp16 = mul(x = s_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_2298_cast_fp16")];
            fp32 var_2299_epsilon_0 = const()[name = string("op_2299_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_2299_cast_fp16 = inverse(epsilon = var_2299_epsilon_0, x = var_2298_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 var_2300_to_fp16 = const()[name = string("op_2300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_2301_cast_fp16 = mul(x = var_2299_cast_fp16, y = var_2300_to_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2303_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2301_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp32 var_2304_epsilon_0 = const()[name = string("op_2304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_2304_cast_fp16 = rsqrt(epsilon = var_2304_epsilon_0, x = var_2303_cast_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_1_cast_fp16 = mul(x = y_9_cast_fp16, y = var_2304_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<fp16, [48, 256]> cos_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [48, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137352384))), scale = tensor<fp16, [48, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137364736))))[name = string("cos_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 48, 256]> var_2306_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2306_cast_fp16")];
            tensor<int32, [4]> var_2311_begin_0 = const()[name = string("op_2311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2311_end_0 = const()[name = string("op_2311_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_2311_end_mask_0 = const()[name = string("op_2311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_2311_cast_fp16 = slice_by_index(begin = var_2311_begin_0, end = var_2311_end_0, end_mask = var_2311_end_mask_0, x = q_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = q_1_cast_fp16)[name = string("op_2317_cast_fp16")];
            int32 var_2319 = const()[name = string("op_2319"), val = int32(-1)];
            bool var_2320_interleave_0 = const()[name = string("op_2320_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_2320_cast_fp16 = concat(axis = var_2319, interleave = var_2320_interleave_0, values = (var_2312_cast_fp16, var_2317_cast_fp16))[name = string("op_2320_cast_fp16")];
            tensor<fp16, [48, 256]> sin_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [48, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137364928))), scale = tensor<fp16, [48, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137377280))))[name = string("sin_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 48, 256]> var_2321_cast_fp16 = mul(x = var_2320_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_3_cast_fp16 = add(x = var_2306_cast_fp16, y = var_2321_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2324_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2329_begin_0 = const()[name = string("op_2329_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2329_end_0 = const()[name = string("op_2329_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_2329_end_mask_0 = const()[name = string("op_2329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, x = k_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_2330_cast_fp16 = mul(x = var_2329_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_2330_cast_fp16")];
            tensor<int32, [4]> var_2335_begin_0 = const()[name = string("op_2335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2335_end_0 = const()[name = string("op_2335_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_2335_end_mask_0 = const()[name = string("op_2335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_2335_cast_fp16 = slice_by_index(begin = var_2335_begin_0, end = var_2335_end_0, end_mask = var_2335_end_mask_0, x = k_1_cast_fp16)[name = string("op_2335_cast_fp16")];
            int32 var_2337 = const()[name = string("op_2337"), val = int32(-1)];
            bool var_2338_interleave_0 = const()[name = string("op_2338_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_2338_cast_fp16 = concat(axis = var_2337, interleave = var_2338_interleave_0, values = (var_2330_cast_fp16, var_2335_cast_fp16))[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2339_cast_fp16 = mul(x = var_2338_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2339_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_3_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2339_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = k_3_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [4, 1, 48, 256]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
            tensor<int32, [5]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_0_cast_fp16 = reshape(shape = concat_0, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [4, 1, 48, 256]> reshape_1_cast_fp16 = reshape(shape = concat_1, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = v_1_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [4, 1, 48, 256]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
            tensor<int32, [5]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_2_cast_fp16 = reshape(shape = concat_2, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [4, 1, 48, 256]> reshape_3_cast_fp16 = reshape(shape = concat_3, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2353_transpose_x_0 = const()[name = string("op_2353_transpose_x_0"), val = bool(false)];
            bool var_2353_transpose_y_0 = const()[name = string("op_2353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 4, 48, 48]> var_2353_cast_fp16 = matmul(transpose_x = var_2353_transpose_x_0, transpose_y = var_2353_transpose_y_0, x = q_3_cast_fp16, y = transpose_96_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_2355_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2356 = const()[name = string("op_2356"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_9_cast_fp16 = softmax(axis = var_2356, x = var_2355_cast_fp16)[name = string("w_9_cast_fp16")];
            bool var_2359_transpose_x_0 = const()[name = string("op_2359_transpose_x_0"), val = bool(false)];
            bool var_2359_transpose_y_0 = const()[name = string("op_2359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 4, 48, 256]> var_2359_cast_fp16 = matmul(transpose_x = var_2359_transpose_x_0, transpose_y = var_2359_transpose_y_0, x = w_9_cast_fp16, y = v_3_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<int32, [4]> var_2362_perm_0 = const()[name = string("op_2362_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2366 = const()[name = string("op_2366"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_2362_cast_fp16 = transpose(perm = var_2362_perm_0, x = var_2359_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 48, 1024]> input_3_cast_fp16 = reshape(shape = var_2366, x = var_2362_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7402112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7926464))))[name = string("tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2371_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_2371_cast_fp16")];
            bool var_2374 = const()[name = string("op_2374"), val = bool(true)];
            tensor<int32, [1]> var_2375_axes_0 = const()[name = string("op_2375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2375_cast_fp16 = reduce_max(axes = var_2375_axes_0, keep_dims = var_2374, x = var_2371_cast_fp16)[name = string("op_2375_cast_fp16")];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_11_cast_fp16 = maximum(x = var_2375_cast_fp16, y = var_2376_to_fp16)[name = string("s_11_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_11_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_11_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2379_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_2379_cast_fp16")];
            tensor<int32, [1]> var_2384_axes_0 = const()[name = string("op_2384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2384_keep_dims_0 = const()[name = string("op_2384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2384_cast_fp16 = reduce_mean(axes = var_2384_axes_0, keep_dims = var_2384_keep_dims_0, x = var_2379_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2385_cast_fp16 = mul(x = s_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2385_cast_fp16")];
            fp32 var_2386_epsilon_0 = const()[name = string("op_2386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2386_cast_fp16 = inverse(epsilon = var_2386_epsilon_0, x = var_2385_cast_fp16)[name = string("op_2386_cast_fp16")];
            fp16 var_2387_to_fp16 = const()[name = string("op_2387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2388_cast_fp16 = mul(x = var_2386_cast_fp16, y = var_2387_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2390_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2388_cast_fp16)[name = string("op_2390_cast_fp16")];
            fp32 var_2391_epsilon_0 = const()[name = string("op_2391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2391_cast_fp16 = rsqrt(epsilon = var_2391_epsilon_0, x = var_2390_cast_fp16)[name = string("op_2391_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_2391_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7927552)))];
            tensor<fp16, [1, 48, 512]> var_2393_cast_fp16 = mul(x = out_9_cast_fp16, y = tm_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_11_cast_fp16 = add(x = inputs_embeds, y = var_2393_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2396_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_2396_cast_fp16")];
            bool var_2399 = const()[name = string("op_2399"), val = bool(true)];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_2399, x = var_2396_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 var_2401_to_fp16 = const()[name = string("op_2401_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_13_cast_fp16 = maximum(x = var_2400_cast_fp16, y = var_2401_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = s_13_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2404_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [1]> var_2409_axes_0 = const()[name = string("op_2409_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2409_keep_dims_0 = const()[name = string("op_2409_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2409_cast_fp16 = reduce_mean(axes = var_2409_axes_0, keep_dims = var_2409_keep_dims_0, x = var_2404_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2410_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2411_cast_fp16 = inverse(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2413_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("op_2413_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2415_cast_fp16 = add(x = var_2409_cast_fp16, y = var_2413_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp32 var_2416_epsilon_0 = const()[name = string("op_2416_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2416_cast_fp16 = rsqrt(epsilon = var_2416_epsilon_0, x = var_2415_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_2416_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7928640)))];
            tensor<fp16, [1, 48, 512]> input_5_cast_fp16 = mul(x = out_11_cast_fp16, y = tm_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7929728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8978368))))[name = string("tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8982528)))];
            tensor<fp16, [1, 48, 2048]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_2423_mode_0 = const()[name = string("op_2423_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_2423_cast_fp16 = gelu(mode = var_2423_mode_0, x = linear_5_cast_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8986688))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10035328))))[name = string("tm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_7_cast_fp16 = mul(x = var_2423_cast_fp16, y = linear_6_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10039488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11088128))))[name = string("tm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2431_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_2431_cast_fp16")];
            bool var_2434 = const()[name = string("op_2434"), val = bool(true)];
            tensor<int32, [1]> var_2435_axes_0 = const()[name = string("op_2435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2435_cast_fp16 = reduce_max(axes = var_2435_axes_0, keep_dims = var_2434, x = var_2431_cast_fp16)[name = string("op_2435_cast_fp16")];
            fp16 var_2436_to_fp16 = const()[name = string("op_2436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_15_cast_fp16 = maximum(x = var_2435_cast_fp16, y = var_2436_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_15_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_15_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2439_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_2439_cast_fp16")];
            tensor<int32, [1]> var_2444_axes_0 = const()[name = string("op_2444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2444_keep_dims_0 = const()[name = string("op_2444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2444_cast_fp16 = reduce_mean(axes = var_2444_axes_0, keep_dims = var_2444_keep_dims_0, x = var_2439_cast_fp16)[name = string("op_2444_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2445_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_2445_cast_fp16")];
            fp32 var_2446_epsilon_0 = const()[name = string("op_2446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2446_cast_fp16 = inverse(epsilon = var_2446_epsilon_0, x = var_2445_cast_fp16)[name = string("op_2446_cast_fp16")];
            fp16 var_2447_to_fp16 = const()[name = string("op_2447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2448_cast_fp16 = mul(x = var_2446_cast_fp16, y = var_2447_to_fp16)[name = string("op_2448_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2450_cast_fp16 = add(x = var_2444_cast_fp16, y = var_2448_cast_fp16)[name = string("op_2450_cast_fp16")];
            fp32 var_2451_epsilon_0 = const()[name = string("op_2451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2451_cast_fp16 = rsqrt(epsilon = var_2451_epsilon_0, x = var_2450_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_2451_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11089216)))];
            tensor<fp16, [1, 48, 512]> var_2453_cast_fp16 = mul(x = out_13_cast_fp16, y = tm_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_9_cast_fp16 = add(x = x_11_cast_fp16, y = var_2453_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11090304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11352512))))[name = string("tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_8_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_8_cast_fp16")];
            string x_15_mode_0 = const()[name = string("x_15_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_15_cast_fp16 = gelu(mode = x_15_mode_0, x = linear_8_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2473_squeeze_mask_0 = const()[name = string("op_2473_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, squeeze_mask = var_2473_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2473_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_11_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2473_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11353600))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11615808))))[name = string("tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2483_cast_fp16 = abs(x = linear_9_cast_fp16)[name = string("op_2483_cast_fp16")];
            bool var_2486 = const()[name = string("op_2486"), val = bool(true)];
            tensor<int32, [1]> var_2487_axes_0 = const()[name = string("op_2487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2487_cast_fp16 = reduce_max(axes = var_2487_axes_0, keep_dims = var_2486, x = var_2483_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp16 var_2488_to_fp16 = const()[name = string("op_2488_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_17_cast_fp16 = maximum(x = var_2487_cast_fp16, y = var_2488_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_17_cast_fp16 = real_div(x = linear_9_cast_fp16, y = s_17_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2491_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_2491_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2496_keep_dims_0 = const()[name = string("op_2496_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2496_cast_fp16 = reduce_mean(axes = var_2496_axes_0, keep_dims = var_2496_keep_dims_0, x = var_2491_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2497_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_2497_cast_fp16")];
            fp32 var_2498_epsilon_0 = const()[name = string("op_2498_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2498_cast_fp16 = inverse(epsilon = var_2498_epsilon_0, x = var_2497_cast_fp16)[name = string("op_2498_cast_fp16")];
            fp16 var_2499_to_fp16 = const()[name = string("op_2499_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2500_cast_fp16 = mul(x = var_2498_cast_fp16, y = var_2499_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2502_cast_fp16 = add(x = var_2496_cast_fp16, y = var_2500_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp32 var_2503_epsilon_0 = const()[name = string("op_2503_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2503_cast_fp16 = rsqrt(epsilon = var_2503_epsilon_0, x = var_2502_cast_fp16)[name = string("op_2503_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_2503_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11616896)))];
            tensor<fp16, [1, 48, 512]> var_2505_cast_fp16 = mul(x = out_15_cast_fp16, y = tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_1_cast_fp16 = add(x = input_9_cast_fp16, y = var_2505_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<fp16, [1]> tm_layers_0_layer_scalar_to_fp16 = const()[name = string("tm_layers_0_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9p-2])];
            tensor<fp16, [1, 48, 512]> x_19_cast_fp16 = mul(x = h_1_cast_fp16, y = tm_layers_0_layer_scalar_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2509_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool var_2512 = const()[name = string("op_2512"), val = bool(true)];
            tensor<int32, [1]> var_2513_axes_0 = const()[name = string("op_2513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2513_cast_fp16 = reduce_max(axes = var_2513_axes_0, keep_dims = var_2512, x = var_2509_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_19_cast_fp16 = maximum(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_19_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2517_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [1]> var_2522_axes_0 = const()[name = string("op_2522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2522_keep_dims_0 = const()[name = string("op_2522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2522_cast_fp16 = reduce_mean(axes = var_2522_axes_0, keep_dims = var_2522_keep_dims_0, x = var_2517_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2523_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_2523_cast_fp16")];
            fp32 var_2524_epsilon_0 = const()[name = string("op_2524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2524_cast_fp16 = inverse(epsilon = var_2524_epsilon_0, x = var_2523_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp16 var_2525_to_fp16 = const()[name = string("op_2525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2526_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_2525_to_fp16)[name = string("op_2526_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2528_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2526_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_2529_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11617984)))];
            tensor<fp16, [1, 48, 512]> input_13_cast_fp16 = mul(x = out_17_cast_fp16, y = tm_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11619072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12143424))))[name = string("tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_2540_cast_fp16 = reshape(shape = var_2539, x = linear_10_cast_fp16)[name = string("op_2540_cast_fp16")];
            tensor<int32, [4]> x_21_perm_0 = const()[name = string("x_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12145536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12407744))))[name = string("tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<int32, [4]> var_2551 = const()[name = string("op_2551"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_2552_cast_fp16 = reshape(shape = var_2551, x = linear_11_cast_fp16)[name = string("op_2552_cast_fp16")];
            tensor<int32, [4]> x_23_perm_0 = const()[name = string("x_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12408832))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671040))))[name = string("tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [4]> var_2563 = const()[name = string("op_2563"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_2564_cast_fp16 = reshape(shape = var_2563, x = linear_12_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = var_2540_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 4, 48, 256]> var_2568_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_2568_cast_fp16")];
            bool var_2571 = const()[name = string("op_2571"), val = bool(true)];
            tensor<int32, [1]> var_2572_axes_0 = const()[name = string("op_2572_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_2572_cast_fp16 = reduce_max(axes = var_2572_axes_0, keep_dims = var_2571, x = var_2568_cast_fp16)[name = string("op_2572_cast_fp16")];
            fp16 var_2573_to_fp16 = const()[name = string("op_2573_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_21_cast_fp16 = maximum(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_21_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_2576_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> var_2581_axes_0 = const()[name = string("op_2581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2581_keep_dims_0 = const()[name = string("op_2581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_2581_cast_fp16 = reduce_mean(axes = var_2581_axes_0, keep_dims = var_2581_keep_dims_0, x = var_2576_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_2582_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp32 var_2583_epsilon_0 = const()[name = string("op_2583_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_2583_cast_fp16 = inverse(epsilon = var_2583_epsilon_0, x = var_2582_cast_fp16)[name = string("op_2583_cast_fp16")];
            fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_2585_cast_fp16 = mul(x = var_2583_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_2587_cast_fp16 = add(x = var_2581_cast_fp16, y = var_2585_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_19_cast_fp16 = mul(x = y_21_cast_fp16, y = var_2588_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672128)))];
            tensor<fp16, [1, 4, 48, 256]> q_5_cast_fp16 = mul(x = out_19_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_23_cast_fp16 = transpose(perm = x_23_perm_0, x = var_2552_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2, 48, 256]> var_2591_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2594 = const()[name = string("op_2594"), val = bool(true)];
            tensor<int32, [1]> var_2595_axes_0 = const()[name = string("op_2595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_2595_cast_fp16 = reduce_max(axes = var_2595_axes_0, keep_dims = var_2594, x = var_2591_cast_fp16)[name = string("op_2595_cast_fp16")];
            fp16 var_2596_to_fp16 = const()[name = string("op_2596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_23_cast_fp16 = maximum(x = var_2595_cast_fp16, y = var_2596_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_23_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2599_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2599_cast_fp16")];
            tensor<int32, [1]> var_2604_axes_0 = const()[name = string("op_2604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2604_keep_dims_0 = const()[name = string("op_2604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_2604_cast_fp16 = reduce_mean(axes = var_2604_axes_0, keep_dims = var_2604_keep_dims_0, x = var_2599_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2605_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_2605_cast_fp16")];
            fp32 var_2606_epsilon_0 = const()[name = string("op_2606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_2606_cast_fp16 = inverse(epsilon = var_2606_epsilon_0, x = var_2605_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp16 var_2607_to_fp16 = const()[name = string("op_2607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_2608_cast_fp16 = mul(x = var_2606_cast_fp16, y = var_2607_to_fp16)[name = string("op_2608_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2610_cast_fp16 = add(x = var_2604_cast_fp16, y = var_2608_cast_fp16)[name = string("op_2610_cast_fp16")];
            fp32 var_2611_epsilon_0 = const()[name = string("op_2611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_2611_cast_fp16 = rsqrt(epsilon = var_2611_epsilon_0, x = var_2610_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_21_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2611_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672704)))];
            tensor<fp16, [1, 2, 48, 256]> k_7_cast_fp16 = mul(x = out_21_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_2564_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 2, 48, 256]> var_2614_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_2614_cast_fp16")];
            bool var_2617 = const()[name = string("op_2617"), val = bool(true)];
            tensor<int32, [1]> var_2618_axes_0 = const()[name = string("op_2618_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_2618_cast_fp16 = reduce_max(axes = var_2618_axes_0, keep_dims = var_2617, x = var_2614_cast_fp16)[name = string("op_2618_cast_fp16")];
            fp16 var_2619_to_fp16 = const()[name = string("op_2619_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_25_cast_fp16 = maximum(x = var_2618_cast_fp16, y = var_2619_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_25_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2622_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2622_cast_fp16")];
            tensor<int32, [1]> var_2627_axes_0 = const()[name = string("op_2627_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2627_keep_dims_0 = const()[name = string("op_2627_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_2627_cast_fp16 = reduce_mean(axes = var_2627_axes_0, keep_dims = var_2627_keep_dims_0, x = var_2622_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2628_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_2628_cast_fp16")];
            fp32 var_2629_epsilon_0 = const()[name = string("op_2629_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_2629_cast_fp16 = inverse(epsilon = var_2629_epsilon_0, x = var_2628_cast_fp16)[name = string("op_2629_cast_fp16")];
            fp16 var_2630_to_fp16 = const()[name = string("op_2630_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_2631_cast_fp16 = mul(x = var_2629_cast_fp16, y = var_2630_to_fp16)[name = string("op_2631_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2633_cast_fp16 = add(x = var_2627_cast_fp16, y = var_2631_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_2634_cast_fp16 = rsqrt(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_5_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2634_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_2636_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2636_cast_fp16")];
            tensor<int32, [4]> var_2641_begin_0 = const()[name = string("op_2641_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2641_end_0 = const()[name = string("op_2641_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_2641_end_mask_0 = const()[name = string("op_2641_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_2641_cast_fp16 = slice_by_index(begin = var_2641_begin_0, end = var_2641_end_0, end_mask = var_2641_end_mask_0, x = q_5_cast_fp16)[name = string("op_2641_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_2642_cast_fp16 = mul(x = var_2641_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_2642_cast_fp16")];
            tensor<int32, [4]> var_2647_begin_0 = const()[name = string("op_2647_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2647_end_0 = const()[name = string("op_2647_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_2647_end_mask_0 = const()[name = string("op_2647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_2647_cast_fp16 = slice_by_index(begin = var_2647_begin_0, end = var_2647_end_0, end_mask = var_2647_end_mask_0, x = q_5_cast_fp16)[name = string("op_2647_cast_fp16")];
            int32 var_2649 = const()[name = string("op_2649"), val = int32(-1)];
            bool var_2650_interleave_0 = const()[name = string("op_2650_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_2650_cast_fp16 = concat(axis = var_2649, interleave = var_2650_interleave_0, values = (var_2642_cast_fp16, var_2647_cast_fp16))[name = string("op_2650_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2651_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_7_cast_fp16 = add(x = var_2636_cast_fp16, y = var_2651_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2654_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = k_7_cast_fp16)[name = string("op_2659_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_2660_cast_fp16 = mul(x = var_2659_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2660_cast_fp16")];
            tensor<int32, [4]> var_2665_begin_0 = const()[name = string("op_2665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2665_end_0 = const()[name = string("op_2665_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_2665_end_mask_0 = const()[name = string("op_2665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_2665_cast_fp16 = slice_by_index(begin = var_2665_begin_0, end = var_2665_end_0, end_mask = var_2665_end_mask_0, x = k_7_cast_fp16)[name = string("op_2665_cast_fp16")];
            int32 var_2667 = const()[name = string("op_2667"), val = int32(-1)];
            bool var_2668_interleave_0 = const()[name = string("op_2668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_2668_cast_fp16 = concat(axis = var_2667, interleave = var_2668_interleave_0, values = (var_2660_cast_fp16, var_2665_cast_fp16))[name = string("op_2668_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2669_cast_fp16 = mul(x = var_2668_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_9_cast_fp16 = add(x = var_2654_cast_fp16, y = var_2669_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = k_9_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [4, 1, 48, 256]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_4_cast_fp16 = reshape(shape = concat_4, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [4, 1, 48, 256]> reshape_5_cast_fp16 = reshape(shape = concat_5, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = v_5_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [4, 1, 48, 256]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_6_cast_fp16 = reshape(shape = concat_6, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [4, 1, 48, 256]> reshape_7_cast_fp16 = reshape(shape = concat_7, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2683_transpose_x_0 = const()[name = string("op_2683_transpose_x_0"), val = bool(false)];
            bool var_2683_transpose_y_0 = const()[name = string("op_2683_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_97_cast_fp16 = transpose(perm = transpose_97_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 4, 48, 48]> var_2683_cast_fp16 = matmul(transpose_x = var_2683_transpose_x_0, transpose_y = var_2683_transpose_y_0, x = q_7_cast_fp16, y = transpose_97_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_2685_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2685_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_25_cast_fp16 = softmax(axis = var_2686, x = var_2685_cast_fp16)[name = string("w_25_cast_fp16")];
            bool var_2689_transpose_x_0 = const()[name = string("op_2689_transpose_x_0"), val = bool(false)];
            bool var_2689_transpose_y_0 = const()[name = string("op_2689_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 4, 48, 256]> var_2689_cast_fp16 = matmul(transpose_x = var_2689_transpose_x_0, transpose_y = var_2689_transpose_y_0, x = w_25_cast_fp16, y = v_7_cast_fp16)[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> var_2692_perm_0 = const()[name = string("op_2692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2696 = const()[name = string("op_2696"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_2692_cast_fp16 = transpose(perm = var_2692_perm_0, x = var_2689_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 48, 1024]> input_15_cast_fp16 = reshape(shape = var_2696, x = var_2692_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13197632))))[name = string("tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2701_cast_fp16 = abs(x = linear_13_cast_fp16)[name = string("op_2701_cast_fp16")];
            bool var_2704 = const()[name = string("op_2704"), val = bool(true)];
            tensor<int32, [1]> var_2705_axes_0 = const()[name = string("op_2705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2705_cast_fp16 = reduce_max(axes = var_2705_axes_0, keep_dims = var_2704, x = var_2701_cast_fp16)[name = string("op_2705_cast_fp16")];
            fp16 var_2706_to_fp16 = const()[name = string("op_2706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_27_cast_fp16 = maximum(x = var_2705_cast_fp16, y = var_2706_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_27_cast_fp16 = real_div(x = linear_13_cast_fp16, y = s_27_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2709_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2709_cast_fp16")];
            tensor<int32, [1]> var_2714_axes_0 = const()[name = string("op_2714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2714_keep_dims_0 = const()[name = string("op_2714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2714_cast_fp16 = reduce_mean(axes = var_2714_axes_0, keep_dims = var_2714_keep_dims_0, x = var_2709_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2715_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2715_cast_fp16")];
            fp32 var_2716_epsilon_0 = const()[name = string("op_2716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2716_cast_fp16 = inverse(epsilon = var_2716_epsilon_0, x = var_2715_cast_fp16)[name = string("op_2716_cast_fp16")];
            fp16 var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2718_cast_fp16 = mul(x = var_2716_cast_fp16, y = var_2717_to_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2720_cast_fp16 = add(x = var_2714_cast_fp16, y = var_2718_cast_fp16)[name = string("op_2720_cast_fp16")];
            fp32 var_2721_epsilon_0 = const()[name = string("op_2721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2721_cast_fp16 = rsqrt(epsilon = var_2721_epsilon_0, x = var_2720_cast_fp16)[name = string("op_2721_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2721_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13198720)))];
            tensor<fp16, [1, 48, 512]> var_2723_cast_fp16 = mul(x = out_23_cast_fp16, y = tm_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2723_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_29_cast_fp16 = add(x = x_19_cast_fp16, y = var_2723_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2726_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_2726_cast_fp16")];
            bool var_2729 = const()[name = string("op_2729"), val = bool(true)];
            tensor<int32, [1]> var_2730_axes_0 = const()[name = string("op_2730_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2730_cast_fp16 = reduce_max(axes = var_2730_axes_0, keep_dims = var_2729, x = var_2726_cast_fp16)[name = string("op_2730_cast_fp16")];
            fp16 var_2731_to_fp16 = const()[name = string("op_2731_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_29_cast_fp16 = maximum(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_29_cast_fp16 = real_div(x = x_29_cast_fp16, y = s_29_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2734_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2734_cast_fp16")];
            tensor<int32, [1]> var_2739_axes_0 = const()[name = string("op_2739_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2739_keep_dims_0 = const()[name = string("op_2739_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2739_cast_fp16 = reduce_mean(axes = var_2739_axes_0, keep_dims = var_2739_keep_dims_0, x = var_2734_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2740_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2740_cast_fp16")];
            fp32 var_2741_epsilon_0 = const()[name = string("op_2741_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2741_cast_fp16 = inverse(epsilon = var_2741_epsilon_0, x = var_2740_cast_fp16)[name = string("op_2741_cast_fp16")];
            fp16 var_2742_to_fp16 = const()[name = string("op_2742_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2743_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_2742_to_fp16)[name = string("op_2743_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2745_cast_fp16 = add(x = var_2739_cast_fp16, y = var_2743_cast_fp16)[name = string("op_2745_cast_fp16")];
            fp32 var_2746_epsilon_0 = const()[name = string("op_2746_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2746_cast_fp16 = rsqrt(epsilon = var_2746_epsilon_0, x = var_2745_cast_fp16)[name = string("op_2746_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2746_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13199808)))];
            tensor<fp16, [1, 48, 512]> input_17_cast_fp16 = mul(x = out_25_cast_fp16, y = tm_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13200896))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14249536))))[name = string("tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_14_cast_fp16")];
            string var_2753_mode_0 = const()[name = string("op_2753_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_2753_cast_fp16 = gelu(mode = var_2753_mode_0, x = linear_14_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14253696))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15302336))))[name = string("tm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_15_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_19_cast_fp16 = mul(x = var_2753_cast_fp16, y = linear_15_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15306496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16355136))))[name = string("tm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2761_cast_fp16 = abs(x = linear_16_cast_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2764 = const()[name = string("op_2764"), val = bool(true)];
            tensor<int32, [1]> var_2765_axes_0 = const()[name = string("op_2765_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2765_cast_fp16 = reduce_max(axes = var_2765_axes_0, keep_dims = var_2764, x = var_2761_cast_fp16)[name = string("op_2765_cast_fp16")];
            fp16 var_2766_to_fp16 = const()[name = string("op_2766_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_31_cast_fp16 = maximum(x = var_2765_cast_fp16, y = var_2766_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_31_cast_fp16 = real_div(x = linear_16_cast_fp16, y = s_31_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2769_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [1]> var_2774_axes_0 = const()[name = string("op_2774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2774_keep_dims_0 = const()[name = string("op_2774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2774_cast_fp16 = reduce_mean(axes = var_2774_axes_0, keep_dims = var_2774_keep_dims_0, x = var_2769_cast_fp16)[name = string("op_2774_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2775_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp32 var_2776_epsilon_0 = const()[name = string("op_2776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2776_cast_fp16 = inverse(epsilon = var_2776_epsilon_0, x = var_2775_cast_fp16)[name = string("op_2776_cast_fp16")];
            fp16 var_2777_to_fp16 = const()[name = string("op_2777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2778_cast_fp16 = mul(x = var_2776_cast_fp16, y = var_2777_to_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2780_cast_fp16 = add(x = var_2774_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2780_cast_fp16")];
            fp32 var_2781_epsilon_0 = const()[name = string("op_2781_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2781_cast_fp16 = rsqrt(epsilon = var_2781_epsilon_0, x = var_2780_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2781_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16356224)))];
            tensor<fp16, [1, 48, 512]> var_2783_cast_fp16 = mul(x = out_27_cast_fp16, y = tm_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2783_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_21_cast_fp16 = add(x = x_29_cast_fp16, y = var_2783_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16357312))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16619520))))[name = string("tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_17_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            string x_33_mode_0 = const()[name = string("x_33_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_33_cast_fp16 = gelu(mode = x_33_mode_0, x = linear_17_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> var_2803_begin_0 = const()[name = string("op_2803_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2803_end_0 = const()[name = string("op_2803_end_0"), val = tensor<int32, [4]>([1, 48, 2, 512])];
            tensor<bool, [4]> var_2803_end_mask_0 = const()[name = string("op_2803_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2803_squeeze_mask_0 = const()[name = string("op_2803_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_2803_cast_fp16 = slice_by_index(begin = var_2803_begin_0, end = var_2803_end_0, end_mask = var_2803_end_mask_0, squeeze_mask = var_2803_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_23_cast_fp16 = mul(x = x_33_cast_fp16, y = var_2803_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16620608))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16882816))))[name = string("tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2813_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2813_cast_fp16")];
            bool var_2816 = const()[name = string("op_2816"), val = bool(true)];
            tensor<int32, [1]> var_2817_axes_0 = const()[name = string("op_2817_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2817_cast_fp16 = reduce_max(axes = var_2817_axes_0, keep_dims = var_2816, x = var_2813_cast_fp16)[name = string("op_2817_cast_fp16")];
            fp16 var_2818_to_fp16 = const()[name = string("op_2818_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_33_cast_fp16 = maximum(x = var_2817_cast_fp16, y = var_2818_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_33_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_33_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2821_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<int32, [1]> var_2826_axes_0 = const()[name = string("op_2826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2826_keep_dims_0 = const()[name = string("op_2826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2826_cast_fp16 = reduce_mean(axes = var_2826_axes_0, keep_dims = var_2826_keep_dims_0, x = var_2821_cast_fp16)[name = string("op_2826_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2827_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp32 var_2828_epsilon_0 = const()[name = string("op_2828_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2828_cast_fp16 = inverse(epsilon = var_2828_epsilon_0, x = var_2827_cast_fp16)[name = string("op_2828_cast_fp16")];
            fp16 var_2829_to_fp16 = const()[name = string("op_2829_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2830_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_2829_to_fp16)[name = string("op_2830_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2832_cast_fp16 = add(x = var_2826_cast_fp16, y = var_2830_cast_fp16)[name = string("op_2832_cast_fp16")];
            fp32 var_2833_epsilon_0 = const()[name = string("op_2833_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2833_cast_fp16 = rsqrt(epsilon = var_2833_epsilon_0, x = var_2832_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2833_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16883904)))];
            tensor<fp16, [1, 48, 512]> var_2835_cast_fp16 = mul(x = out_29_cast_fp16, y = tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_3_cast_fp16 = add(x = input_21_cast_fp16, y = var_2835_cast_fp16)[name = string("h_3_cast_fp16")];
            tensor<fp16, [1]> tm_layers_1_layer_scalar_to_fp16 = const()[name = string("tm_layers_1_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.5ap-3])];
            tensor<fp16, [1, 48, 512]> x_37_cast_fp16 = mul(x = h_3_cast_fp16, y = tm_layers_1_layer_scalar_to_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2839_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2839_cast_fp16")];
            bool var_2842 = const()[name = string("op_2842"), val = bool(true)];
            tensor<int32, [1]> var_2843_axes_0 = const()[name = string("op_2843_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_2843_cast_fp16 = reduce_max(axes = var_2843_axes_0, keep_dims = var_2842, x = var_2839_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp16 var_2844_to_fp16 = const()[name = string("op_2844_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_35_cast_fp16 = maximum(x = var_2843_cast_fp16, y = var_2844_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_35_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_35_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_2847_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2847_cast_fp16")];
            tensor<int32, [1]> var_2852_axes_0 = const()[name = string("op_2852_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2852_keep_dims_0 = const()[name = string("op_2852_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_2852_cast_fp16 = reduce_mean(axes = var_2852_axes_0, keep_dims = var_2852_keep_dims_0, x = var_2847_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2853_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp32 var_2854_epsilon_0 = const()[name = string("op_2854_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_2854_cast_fp16 = inverse(epsilon = var_2854_epsilon_0, x = var_2853_cast_fp16)[name = string("op_2854_cast_fp16")];
            fp16 var_2855_to_fp16 = const()[name = string("op_2855_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_2856_cast_fp16 = mul(x = var_2854_cast_fp16, y = var_2855_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_2858_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2856_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp32 var_2859_epsilon_0 = const()[name = string("op_2859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_2859_cast_fp16 = rsqrt(epsilon = var_2859_epsilon_0, x = var_2858_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_31_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2859_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16884992)))];
            tensor<fp16, [1, 48, 512]> input_25_cast_fp16 = mul(x = out_31_cast_fp16, y = tm_layers_2_input_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16886080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17410432))))[name = string("tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [4]> var_2869 = const()[name = string("op_2869"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_2870_cast_fp16 = reshape(shape = var_2869, x = linear_19_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17412544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17674752))))[name = string("tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [4]> var_2881 = const()[name = string("op_2881"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_2882_cast_fp16 = reshape(shape = var_2881, x = linear_20_cast_fp16)[name = string("op_2882_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17675840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17938048))))[name = string("tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_2894_cast_fp16 = reshape(shape = var_2893, x = linear_21_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [4]> x_43_perm_0 = const()[name = string("x_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2870_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 4, 48, 256]> var_2898_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2898_cast_fp16")];
            bool var_2901 = const()[name = string("op_2901"), val = bool(true)];
            tensor<int32, [1]> var_2902_axes_0 = const()[name = string("op_2902_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_2902_cast_fp16 = reduce_max(axes = var_2902_axes_0, keep_dims = var_2901, x = var_2898_cast_fp16)[name = string("op_2902_cast_fp16")];
            fp16 var_2903_to_fp16 = const()[name = string("op_2903_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_37_cast_fp16 = maximum(x = var_2902_cast_fp16, y = var_2903_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_37_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_37_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_2906_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<int32, [1]> var_2911_axes_0 = const()[name = string("op_2911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2911_keep_dims_0 = const()[name = string("op_2911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_2911_cast_fp16 = reduce_mean(axes = var_2911_axes_0, keep_dims = var_2911_keep_dims_0, x = var_2906_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_2912_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2912_cast_fp16")];
            fp32 var_2913_epsilon_0 = const()[name = string("op_2913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_2913_cast_fp16 = inverse(epsilon = var_2913_epsilon_0, x = var_2912_cast_fp16)[name = string("op_2913_cast_fp16")];
            fp16 var_2914_to_fp16 = const()[name = string("op_2914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_2915_cast_fp16 = mul(x = var_2913_cast_fp16, y = var_2914_to_fp16)[name = string("op_2915_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_2917_cast_fp16 = add(x = var_2911_cast_fp16, y = var_2915_cast_fp16)[name = string("op_2917_cast_fp16")];
            fp32 var_2918_epsilon_0 = const()[name = string("op_2918_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_2918_cast_fp16 = rsqrt(epsilon = var_2918_epsilon_0, x = var_2917_cast_fp16)[name = string("op_2918_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_33_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2918_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_9_cast_fp16 = mul(x = out_33_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2882_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 2, 48, 256]> var_2921_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2921_cast_fp16")];
            bool var_2924 = const()[name = string("op_2924"), val = bool(true)];
            tensor<int32, [1]> var_2925_axes_0 = const()[name = string("op_2925_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_2925_cast_fp16 = reduce_max(axes = var_2925_axes_0, keep_dims = var_2924, x = var_2921_cast_fp16)[name = string("op_2925_cast_fp16")];
            fp16 var_2926_to_fp16 = const()[name = string("op_2926_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_39_cast_fp16 = maximum(x = var_2925_cast_fp16, y = var_2926_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_39_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_39_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2929_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2929_cast_fp16")];
            tensor<int32, [1]> var_2934_axes_0 = const()[name = string("op_2934_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2934_keep_dims_0 = const()[name = string("op_2934_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_2934_cast_fp16 = reduce_mean(axes = var_2934_axes_0, keep_dims = var_2934_keep_dims_0, x = var_2929_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2935_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2935_cast_fp16")];
            fp32 var_2936_epsilon_0 = const()[name = string("op_2936_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_2936_cast_fp16 = inverse(epsilon = var_2936_epsilon_0, x = var_2935_cast_fp16)[name = string("op_2936_cast_fp16")];
            fp16 var_2937_to_fp16 = const()[name = string("op_2937_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_2938_cast_fp16 = mul(x = var_2936_cast_fp16, y = var_2937_to_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2940_cast_fp16 = add(x = var_2934_cast_fp16, y = var_2938_cast_fp16)[name = string("op_2940_cast_fp16")];
            fp32 var_2941_epsilon_0 = const()[name = string("op_2941_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_2941_cast_fp16 = rsqrt(epsilon = var_2941_epsilon_0, x = var_2940_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_35_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2941_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_13_cast_fp16 = mul(x = out_35_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_43_cast_fp16 = transpose(perm = x_43_perm_0, x = var_2894_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 2, 48, 256]> var_2944_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2944_cast_fp16")];
            bool var_2947 = const()[name = string("op_2947"), val = bool(true)];
            tensor<int32, [1]> var_2948_axes_0 = const()[name = string("op_2948_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_2948_cast_fp16 = reduce_max(axes = var_2948_axes_0, keep_dims = var_2947, x = var_2944_cast_fp16)[name = string("op_2948_cast_fp16")];
            fp16 var_2949_to_fp16 = const()[name = string("op_2949_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_41_cast_fp16 = maximum(x = var_2948_cast_fp16, y = var_2949_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_41_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_41_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2952_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2952_cast_fp16")];
            tensor<int32, [1]> var_2957_axes_0 = const()[name = string("op_2957_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2957_keep_dims_0 = const()[name = string("op_2957_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_2957_cast_fp16 = reduce_mean(axes = var_2957_axes_0, keep_dims = var_2957_keep_dims_0, x = var_2952_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2958_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2958_cast_fp16")];
            fp32 var_2959_epsilon_0 = const()[name = string("op_2959_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_2959_cast_fp16 = inverse(epsilon = var_2959_epsilon_0, x = var_2958_cast_fp16)[name = string("op_2959_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_2961_cast_fp16 = mul(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_2963_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2961_cast_fp16)[name = string("op_2963_cast_fp16")];
            fp32 var_2964_epsilon_0 = const()[name = string("op_2964_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_2964_cast_fp16 = rsqrt(epsilon = var_2964_epsilon_0, x = var_2963_cast_fp16)[name = string("op_2964_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_9_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2964_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_2966_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2966_cast_fp16")];
            tensor<int32, [4]> var_2971_begin_0 = const()[name = string("op_2971_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2971_end_0 = const()[name = string("op_2971_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_2971_end_mask_0 = const()[name = string("op_2971_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_2971_cast_fp16 = slice_by_index(begin = var_2971_begin_0, end = var_2971_end_0, end_mask = var_2971_end_mask_0, x = q_9_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_2972_cast_fp16 = mul(x = var_2971_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2972_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            int32 var_2979 = const()[name = string("op_2979"), val = int32(-1)];
            bool var_2980_interleave_0 = const()[name = string("op_2980_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_2980_cast_fp16 = concat(axis = var_2979, interleave = var_2980_interleave_0, values = (var_2972_cast_fp16, var_2977_cast_fp16))[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_2981_cast_fp16 = mul(x = var_2980_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2981_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_11_cast_fp16 = add(x = var_2966_cast_fp16, y = var_2981_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2984_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2989_begin_0 = const()[name = string("op_2989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2989_end_0 = const()[name = string("op_2989_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_2989_end_mask_0 = const()[name = string("op_2989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_2989_cast_fp16 = slice_by_index(begin = var_2989_begin_0, end = var_2989_end_0, end_mask = var_2989_end_mask_0, x = k_13_cast_fp16)[name = string("op_2989_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_2990_cast_fp16 = mul(x = var_2989_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2990_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = k_13_cast_fp16)[name = string("op_2995_cast_fp16")];
            int32 var_2997 = const()[name = string("op_2997"), val = int32(-1)];
            bool var_2998_interleave_0 = const()[name = string("op_2998_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_2998_cast_fp16 = concat(axis = var_2997, interleave = var_2998_interleave_0, values = (var_2990_cast_fp16, var_2995_cast_fp16))[name = string("op_2998_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_2999_cast_fp16 = mul(x = var_2998_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2999_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_15_cast_fp16 = add(x = var_2984_cast_fp16, y = var_2999_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = k_15_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [4, 1, 48, 256]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
            tensor<int32, [5]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_8_cast_fp16 = reshape(shape = concat_8, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [4, 1, 48, 256]> reshape_9_cast_fp16 = reshape(shape = concat_9, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = v_9_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [4, 1, 48, 256]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
            tensor<int32, [5]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_10_cast_fp16 = reshape(shape = concat_10, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [4, 1, 48, 256]> reshape_11_cast_fp16 = reshape(shape = concat_11, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3013_transpose_x_0 = const()[name = string("op_3013_transpose_x_0"), val = bool(false)];
            bool var_3013_transpose_y_0 = const()[name = string("op_3013_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 4, 48, 48]> var_3013_cast_fp16 = matmul(transpose_x = var_3013_transpose_x_0, transpose_y = var_3013_transpose_y_0, x = q_11_cast_fp16, y = transpose_98_cast_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_3015_cast_fp16 = add(x = var_3013_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3015_cast_fp16")];
            int32 var_3016 = const()[name = string("op_3016"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_41_cast_fp16 = softmax(axis = var_3016, x = var_3015_cast_fp16)[name = string("w_41_cast_fp16")];
            bool var_3019_transpose_x_0 = const()[name = string("op_3019_transpose_x_0"), val = bool(false)];
            bool var_3019_transpose_y_0 = const()[name = string("op_3019_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 4, 48, 256]> var_3019_cast_fp16 = matmul(transpose_x = var_3019_transpose_x_0, transpose_y = var_3019_transpose_y_0, x = w_41_cast_fp16, y = v_11_cast_fp16)[name = string("op_3019_cast_fp16")];
            tensor<int32, [4]> var_3022_perm_0 = const()[name = string("op_3022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3026 = const()[name = string("op_3026"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_3022_cast_fp16 = transpose(perm = var_3022_perm_0, x = var_3019_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 48, 1024]> input_27_cast_fp16 = reshape(shape = var_3026, x = var_3022_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17939136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18463488))))[name = string("tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3031_cast_fp16 = abs(x = linear_22_cast_fp16)[name = string("op_3031_cast_fp16")];
            bool var_3034 = const()[name = string("op_3034"), val = bool(true)];
            tensor<int32, [1]> var_3035_axes_0 = const()[name = string("op_3035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3035_cast_fp16 = reduce_max(axes = var_3035_axes_0, keep_dims = var_3034, x = var_3031_cast_fp16)[name = string("op_3035_cast_fp16")];
            fp16 var_3036_to_fp16 = const()[name = string("op_3036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_43_cast_fp16 = maximum(x = var_3035_cast_fp16, y = var_3036_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_43_cast_fp16 = real_div(x = linear_22_cast_fp16, y = s_43_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3039_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3039_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3045_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_3045_cast_fp16")];
            fp32 var_3046_epsilon_0 = const()[name = string("op_3046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3046_cast_fp16 = inverse(epsilon = var_3046_epsilon_0, x = var_3045_cast_fp16)[name = string("op_3046_cast_fp16")];
            fp16 var_3047_to_fp16 = const()[name = string("op_3047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3048_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3047_to_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3050_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3048_cast_fp16)[name = string("op_3050_cast_fp16")];
            fp32 var_3051_epsilon_0 = const()[name = string("op_3051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3051_cast_fp16 = rsqrt(epsilon = var_3051_epsilon_0, x = var_3050_cast_fp16)[name = string("op_3051_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_3051_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18464576)))];
            tensor<fp16, [1, 48, 512]> var_3053_cast_fp16 = mul(x = out_37_cast_fp16, y = tm_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = var_3053_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3056_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_3056_cast_fp16")];
            bool var_3059 = const()[name = string("op_3059"), val = bool(true)];
            tensor<int32, [1]> var_3060_axes_0 = const()[name = string("op_3060_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3060_cast_fp16 = reduce_max(axes = var_3060_axes_0, keep_dims = var_3059, x = var_3056_cast_fp16)[name = string("op_3060_cast_fp16")];
            fp16 var_3061_to_fp16 = const()[name = string("op_3061_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_45_cast_fp16 = maximum(x = var_3060_cast_fp16, y = var_3061_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_45_cast_fp16 = real_div(x = x_47_cast_fp16, y = s_45_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3064_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<int32, [1]> var_3069_axes_0 = const()[name = string("op_3069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3069_keep_dims_0 = const()[name = string("op_3069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3069_cast_fp16 = reduce_mean(axes = var_3069_axes_0, keep_dims = var_3069_keep_dims_0, x = var_3064_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3070_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp32 var_3071_epsilon_0 = const()[name = string("op_3071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3071_cast_fp16 = inverse(epsilon = var_3071_epsilon_0, x = var_3070_cast_fp16)[name = string("op_3071_cast_fp16")];
            fp16 var_3072_to_fp16 = const()[name = string("op_3072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3073_cast_fp16 = mul(x = var_3071_cast_fp16, y = var_3072_to_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3075_cast_fp16 = add(x = var_3069_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3075_cast_fp16")];
            fp32 var_3076_epsilon_0 = const()[name = string("op_3076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3076_cast_fp16 = rsqrt(epsilon = var_3076_epsilon_0, x = var_3075_cast_fp16)[name = string("op_3076_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_3076_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18465664)))];
            tensor<fp16, [1, 48, 512]> input_29_cast_fp16 = mul(x = out_39_cast_fp16, y = tm_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18466752))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19515392))))[name = string("tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_23_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_23_cast_fp16")];
            string var_3083_mode_0 = const()[name = string("op_3083_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_3083_cast_fp16 = gelu(mode = var_3083_mode_0, x = linear_23_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19519552))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20568192))))[name = string("tm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_24_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_31_cast_fp16 = mul(x = var_3083_cast_fp16, y = linear_24_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20572352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21620992))))[name = string("tm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3091_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_3091_cast_fp16")];
            bool var_3094 = const()[name = string("op_3094"), val = bool(true)];
            tensor<int32, [1]> var_3095_axes_0 = const()[name = string("op_3095_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3095_cast_fp16 = reduce_max(axes = var_3095_axes_0, keep_dims = var_3094, x = var_3091_cast_fp16)[name = string("op_3095_cast_fp16")];
            fp16 var_3096_to_fp16 = const()[name = string("op_3096_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_47_cast_fp16 = maximum(x = var_3095_cast_fp16, y = var_3096_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_47_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_47_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3099_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3104_keep_dims_0 = const()[name = string("op_3104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3104_cast_fp16 = reduce_mean(axes = var_3104_axes_0, keep_dims = var_3104_keep_dims_0, x = var_3099_cast_fp16)[name = string("op_3104_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3105_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_3105_cast_fp16")];
            fp32 var_3106_epsilon_0 = const()[name = string("op_3106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3106_cast_fp16 = inverse(epsilon = var_3106_epsilon_0, x = var_3105_cast_fp16)[name = string("op_3106_cast_fp16")];
            fp16 var_3107_to_fp16 = const()[name = string("op_3107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3108_cast_fp16 = mul(x = var_3106_cast_fp16, y = var_3107_to_fp16)[name = string("op_3108_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3110_cast_fp16 = add(x = var_3104_cast_fp16, y = var_3108_cast_fp16)[name = string("op_3110_cast_fp16")];
            fp32 var_3111_epsilon_0 = const()[name = string("op_3111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3111_cast_fp16 = rsqrt(epsilon = var_3111_epsilon_0, x = var_3110_cast_fp16)[name = string("op_3111_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_3111_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21622080)))];
            tensor<fp16, [1, 48, 512]> var_3113_cast_fp16 = mul(x = out_41_cast_fp16, y = tm_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_33_cast_fp16 = add(x = x_47_cast_fp16, y = var_3113_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21623168))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21885376))))[name = string("tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_26_cast_fp16")];
            string x_51_mode_0 = const()[name = string("x_51_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_51_cast_fp16 = gelu(mode = x_51_mode_0, x = linear_26_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([1, 48, 3, 512])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3133_squeeze_mask_0 = const()[name = string("op_3133_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, squeeze_mask = var_3133_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_35_cast_fp16 = mul(x = x_51_cast_fp16, y = var_3133_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21886464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22148672))))[name = string("tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3143_cast_fp16 = abs(x = linear_27_cast_fp16)[name = string("op_3143_cast_fp16")];
            bool var_3146 = const()[name = string("op_3146"), val = bool(true)];
            tensor<int32, [1]> var_3147_axes_0 = const()[name = string("op_3147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3147_cast_fp16 = reduce_max(axes = var_3147_axes_0, keep_dims = var_3146, x = var_3143_cast_fp16)[name = string("op_3147_cast_fp16")];
            fp16 var_3148_to_fp16 = const()[name = string("op_3148_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_49_cast_fp16 = maximum(x = var_3147_cast_fp16, y = var_3148_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_49_cast_fp16 = real_div(x = linear_27_cast_fp16, y = s_49_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3151_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_3151_cast_fp16")];
            tensor<int32, [1]> var_3156_axes_0 = const()[name = string("op_3156_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3156_keep_dims_0 = const()[name = string("op_3156_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3156_cast_fp16 = reduce_mean(axes = var_3156_axes_0, keep_dims = var_3156_keep_dims_0, x = var_3151_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3157_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_3157_cast_fp16")];
            fp32 var_3158_epsilon_0 = const()[name = string("op_3158_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3158_cast_fp16 = inverse(epsilon = var_3158_epsilon_0, x = var_3157_cast_fp16)[name = string("op_3158_cast_fp16")];
            fp16 var_3159_to_fp16 = const()[name = string("op_3159_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3160_cast_fp16 = mul(x = var_3158_cast_fp16, y = var_3159_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3162_cast_fp16 = add(x = var_3156_cast_fp16, y = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp32 var_3163_epsilon_0 = const()[name = string("op_3163_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3163_cast_fp16 = rsqrt(epsilon = var_3163_epsilon_0, x = var_3162_cast_fp16)[name = string("op_3163_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_43_cast_fp16 = mul(x = y_49_cast_fp16, y = var_3163_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22149760)))];
            tensor<fp16, [1, 48, 512]> var_3165_cast_fp16 = mul(x = out_43_cast_fp16, y = tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_5_cast_fp16 = add(x = input_33_cast_fp16, y = var_3165_cast_fp16)[name = string("h_5_cast_fp16")];
            tensor<fp16, [1]> tm_layers_2_layer_scalar_to_fp16 = const()[name = string("tm_layers_2_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ep-1])];
            tensor<fp16, [1, 48, 512]> x_55_cast_fp16 = mul(x = h_5_cast_fp16, y = tm_layers_2_layer_scalar_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3169_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_3169_cast_fp16")];
            bool var_3172 = const()[name = string("op_3172"), val = bool(true)];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3173_cast_fp16 = reduce_max(axes = var_3173_axes_0, keep_dims = var_3172, x = var_3169_cast_fp16)[name = string("op_3173_cast_fp16")];
            fp16 var_3174_to_fp16 = const()[name = string("op_3174_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_51_cast_fp16 = maximum(x = var_3173_cast_fp16, y = var_3174_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_51_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_51_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3177_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_3177_cast_fp16")];
            tensor<int32, [1]> var_3182_axes_0 = const()[name = string("op_3182_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3182_keep_dims_0 = const()[name = string("op_3182_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3182_cast_fp16 = reduce_mean(axes = var_3182_axes_0, keep_dims = var_3182_keep_dims_0, x = var_3177_cast_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3183_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_3183_cast_fp16")];
            fp32 var_3184_epsilon_0 = const()[name = string("op_3184_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3184_cast_fp16 = inverse(epsilon = var_3184_epsilon_0, x = var_3183_cast_fp16)[name = string("op_3184_cast_fp16")];
            fp16 var_3185_to_fp16 = const()[name = string("op_3185_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3186_cast_fp16 = mul(x = var_3184_cast_fp16, y = var_3185_to_fp16)[name = string("op_3186_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3188_cast_fp16 = add(x = var_3182_cast_fp16, y = var_3186_cast_fp16)[name = string("op_3188_cast_fp16")];
            fp32 var_3189_epsilon_0 = const()[name = string("op_3189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3189_cast_fp16 = rsqrt(epsilon = var_3189_epsilon_0, x = var_3188_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_45_cast_fp16 = mul(x = y_51_cast_fp16, y = var_3189_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22150848)))];
            tensor<fp16, [1, 48, 512]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = tm_layers_3_input_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22151936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22676288))))[name = string("tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_3200_cast_fp16 = reshape(shape = var_3199, x = linear_28_cast_fp16)[name = string("op_3200_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22678400))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22940608))))[name = string("tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_29_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_59_perm_0 = const()[name = string("x_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22941696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23203904))))[name = string("tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_3223 = const()[name = string("op_3223"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_3224_cast_fp16 = reshape(shape = var_3223, x = linear_30_cast_fp16)[name = string("op_3224_cast_fp16")];
            tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_3200_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 4, 48, 256]> var_3228_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3231 = const()[name = string("op_3231"), val = bool(true)];
            tensor<int32, [1]> var_3232_axes_0 = const()[name = string("op_3232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_3232_cast_fp16 = reduce_max(axes = var_3232_axes_0, keep_dims = var_3231, x = var_3228_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_53_cast_fp16 = maximum(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_53_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_53_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_3236_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_3236_cast_fp16")];
            tensor<int32, [1]> var_3241_axes_0 = const()[name = string("op_3241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3241_keep_dims_0 = const()[name = string("op_3241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_3241_cast_fp16 = reduce_mean(axes = var_3241_axes_0, keep_dims = var_3241_keep_dims_0, x = var_3236_cast_fp16)[name = string("op_3241_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_3242_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_3242_cast_fp16")];
            fp32 var_3243_epsilon_0 = const()[name = string("op_3243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_3243_cast_fp16 = inverse(epsilon = var_3243_epsilon_0, x = var_3242_cast_fp16)[name = string("op_3243_cast_fp16")];
            fp16 var_3244_to_fp16 = const()[name = string("op_3244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_3245_cast_fp16 = mul(x = var_3243_cast_fp16, y = var_3244_to_fp16)[name = string("op_3245_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_3247_cast_fp16 = add(x = var_3241_cast_fp16, y = var_3245_cast_fp16)[name = string("op_3247_cast_fp16")];
            fp32 var_3248_epsilon_0 = const()[name = string("op_3248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_3248_cast_fp16 = rsqrt(epsilon = var_3248_epsilon_0, x = var_3247_cast_fp16)[name = string("op_3248_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_47_cast_fp16 = mul(x = y_53_cast_fp16, y = var_3248_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23204992)))];
            tensor<fp16, [1, 4, 48, 256]> q_13_cast_fp16 = mul(x = out_47_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_59_cast_fp16 = transpose(perm = x_59_perm_0, x = var_3212_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 2, 48, 256]> var_3251_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_3251_cast_fp16")];
            bool var_3254 = const()[name = string("op_3254"), val = bool(true)];
            tensor<int32, [1]> var_3255_axes_0 = const()[name = string("op_3255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_3255_cast_fp16 = reduce_max(axes = var_3255_axes_0, keep_dims = var_3254, x = var_3251_cast_fp16)[name = string("op_3255_cast_fp16")];
            fp16 var_3256_to_fp16 = const()[name = string("op_3256_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_55_cast_fp16 = maximum(x = var_3255_cast_fp16, y = var_3256_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_55_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_55_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_3259_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<int32, [1]> var_3264_axes_0 = const()[name = string("op_3264_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3264_keep_dims_0 = const()[name = string("op_3264_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_3264_cast_fp16 = reduce_mean(axes = var_3264_axes_0, keep_dims = var_3264_keep_dims_0, x = var_3259_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_3265_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_3265_cast_fp16")];
            fp32 var_3266_epsilon_0 = const()[name = string("op_3266_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_3266_cast_fp16 = inverse(epsilon = var_3266_epsilon_0, x = var_3265_cast_fp16)[name = string("op_3266_cast_fp16")];
            fp16 var_3267_to_fp16 = const()[name = string("op_3267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_3268_cast_fp16 = mul(x = var_3266_cast_fp16, y = var_3267_to_fp16)[name = string("op_3268_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_3270_cast_fp16 = add(x = var_3264_cast_fp16, y = var_3268_cast_fp16)[name = string("op_3270_cast_fp16")];
            fp32 var_3271_epsilon_0 = const()[name = string("op_3271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_3271_cast_fp16 = rsqrt(epsilon = var_3271_epsilon_0, x = var_3270_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_49_cast_fp16 = mul(x = y_55_cast_fp16, y = var_3271_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23205568)))];
            tensor<fp16, [1, 2, 48, 256]> k_19_cast_fp16 = mul(x = out_49_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_3224_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2, 48, 256]> var_3274_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_3274_cast_fp16")];
            bool var_3277 = const()[name = string("op_3277"), val = bool(true)];
            tensor<int32, [1]> var_3278_axes_0 = const()[name = string("op_3278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_3278_cast_fp16 = reduce_max(axes = var_3278_axes_0, keep_dims = var_3277, x = var_3274_cast_fp16)[name = string("op_3278_cast_fp16")];
            fp16 var_3279_to_fp16 = const()[name = string("op_3279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_57_cast_fp16 = maximum(x = var_3278_cast_fp16, y = var_3279_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_57_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_57_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_3282_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_3282_cast_fp16")];
            tensor<int32, [1]> var_3287_axes_0 = const()[name = string("op_3287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3287_keep_dims_0 = const()[name = string("op_3287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_3287_cast_fp16 = reduce_mean(axes = var_3287_axes_0, keep_dims = var_3287_keep_dims_0, x = var_3282_cast_fp16)[name = string("op_3287_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_3288_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_3289_cast_fp16 = inverse(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            fp16 var_3290_to_fp16 = const()[name = string("op_3290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_3291_cast_fp16 = mul(x = var_3289_cast_fp16, y = var_3290_to_fp16)[name = string("op_3291_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_3293_cast_fp16 = add(x = var_3287_cast_fp16, y = var_3291_cast_fp16)[name = string("op_3293_cast_fp16")];
            fp32 var_3294_epsilon_0 = const()[name = string("op_3294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_3294_cast_fp16 = rsqrt(epsilon = var_3294_epsilon_0, x = var_3293_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_13_cast_fp16 = mul(x = y_57_cast_fp16, y = var_3294_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_3296_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3296_cast_fp16")];
            tensor<int32, [4]> var_3301_begin_0 = const()[name = string("op_3301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3301_end_0 = const()[name = string("op_3301_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_3301_end_mask_0 = const()[name = string("op_3301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_3301_cast_fp16 = slice_by_index(begin = var_3301_begin_0, end = var_3301_end_0, end_mask = var_3301_end_mask_0, x = q_13_cast_fp16)[name = string("op_3301_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_3302_cast_fp16 = mul(x = var_3301_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3307_begin_0 = const()[name = string("op_3307_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3307_end_0 = const()[name = string("op_3307_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_3307_end_mask_0 = const()[name = string("op_3307_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_3307_cast_fp16 = slice_by_index(begin = var_3307_begin_0, end = var_3307_end_0, end_mask = var_3307_end_mask_0, x = q_13_cast_fp16)[name = string("op_3307_cast_fp16")];
            int32 var_3309 = const()[name = string("op_3309"), val = int32(-1)];
            bool var_3310_interleave_0 = const()[name = string("op_3310_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_3310_cast_fp16 = concat(axis = var_3309, interleave = var_3310_interleave_0, values = (var_3302_cast_fp16, var_3307_cast_fp16))[name = string("op_3310_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_3311_cast_fp16 = mul(x = var_3310_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3311_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_15_cast_fp16 = add(x = var_3296_cast_fp16, y = var_3311_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_3314_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3314_cast_fp16")];
            tensor<int32, [4]> var_3319_begin_0 = const()[name = string("op_3319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3319_end_0 = const()[name = string("op_3319_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_3319_end_mask_0 = const()[name = string("op_3319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_3319_cast_fp16 = slice_by_index(begin = var_3319_begin_0, end = var_3319_end_0, end_mask = var_3319_end_mask_0, x = k_19_cast_fp16)[name = string("op_3319_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_3320_cast_fp16 = mul(x = var_3319_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_3320_cast_fp16")];
            tensor<int32, [4]> var_3325_begin_0 = const()[name = string("op_3325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3325_end_0 = const()[name = string("op_3325_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_3325_end_mask_0 = const()[name = string("op_3325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_3325_cast_fp16 = slice_by_index(begin = var_3325_begin_0, end = var_3325_end_0, end_mask = var_3325_end_mask_0, x = k_19_cast_fp16)[name = string("op_3325_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), val = int32(-1)];
            bool var_3328_interleave_0 = const()[name = string("op_3328_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_3328_cast_fp16 = concat(axis = var_3327, interleave = var_3328_interleave_0, values = (var_3320_cast_fp16, var_3325_cast_fp16))[name = string("op_3328_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_3329_cast_fp16 = mul(x = var_3328_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3329_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_21_cast_fp16 = add(x = var_3314_cast_fp16, y = var_3329_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = k_21_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [4, 1, 48, 256]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_12_cast_fp16 = reshape(shape = concat_12, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [4, 1, 48, 256]> reshape_13_cast_fp16 = reshape(shape = concat_13, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = v_13_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [4, 1, 48, 256]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_14_cast_fp16 = reshape(shape = concat_14, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [4, 1, 48, 256]> reshape_15_cast_fp16 = reshape(shape = concat_15, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3343_transpose_x_0 = const()[name = string("op_3343_transpose_x_0"), val = bool(false)];
            bool var_3343_transpose_y_0 = const()[name = string("op_3343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 4, 48, 48]> var_3343_cast_fp16 = matmul(transpose_x = var_3343_transpose_x_0, transpose_y = var_3343_transpose_y_0, x = q_15_cast_fp16, y = transpose_99_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_3345_cast_fp16 = add(x = var_3343_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3345_cast_fp16")];
            int32 var_3346 = const()[name = string("op_3346"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_57_cast_fp16 = softmax(axis = var_3346, x = var_3345_cast_fp16)[name = string("w_57_cast_fp16")];
            bool var_3349_transpose_x_0 = const()[name = string("op_3349_transpose_x_0"), val = bool(false)];
            bool var_3349_transpose_y_0 = const()[name = string("op_3349_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 4, 48, 256]> var_3349_cast_fp16 = matmul(transpose_x = var_3349_transpose_x_0, transpose_y = var_3349_transpose_y_0, x = w_57_cast_fp16, y = v_15_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<int32, [4]> var_3352_perm_0 = const()[name = string("op_3352_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3356 = const()[name = string("op_3356"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_3352_cast_fp16 = transpose(perm = var_3352_perm_0, x = var_3349_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 48, 1024]> input_39_cast_fp16 = reshape(shape = var_3356, x = var_3352_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23206144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23730496))))[name = string("tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3361_cast_fp16 = abs(x = linear_31_cast_fp16)[name = string("op_3361_cast_fp16")];
            bool var_3364 = const()[name = string("op_3364"), val = bool(true)];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3365_cast_fp16 = reduce_max(axes = var_3365_axes_0, keep_dims = var_3364, x = var_3361_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_59_cast_fp16 = maximum(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_59_cast_fp16 = real_div(x = linear_31_cast_fp16, y = s_59_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3369_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<int32, [1]> var_3374_axes_0 = const()[name = string("op_3374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3374_keep_dims_0 = const()[name = string("op_3374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3374_cast_fp16 = reduce_mean(axes = var_3374_axes_0, keep_dims = var_3374_keep_dims_0, x = var_3369_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3375_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_3375_cast_fp16")];
            fp32 var_3376_epsilon_0 = const()[name = string("op_3376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3376_cast_fp16 = inverse(epsilon = var_3376_epsilon_0, x = var_3375_cast_fp16)[name = string("op_3376_cast_fp16")];
            fp16 var_3377_to_fp16 = const()[name = string("op_3377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3378_cast_fp16 = mul(x = var_3376_cast_fp16, y = var_3377_to_fp16)[name = string("op_3378_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3380_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3380_cast_fp16")];
            fp32 var_3381_epsilon_0 = const()[name = string("op_3381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3381_cast_fp16 = rsqrt(epsilon = var_3381_epsilon_0, x = var_3380_cast_fp16)[name = string("op_3381_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_3381_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23731584)))];
            tensor<fp16, [1, 48, 512]> var_3383_cast_fp16 = mul(x = out_51_cast_fp16, y = tm_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_65_cast_fp16 = add(x = x_55_cast_fp16, y = var_3383_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3386_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_3386_cast_fp16")];
            bool var_3389 = const()[name = string("op_3389"), val = bool(true)];
            tensor<int32, [1]> var_3390_axes_0 = const()[name = string("op_3390_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3390_cast_fp16 = reduce_max(axes = var_3390_axes_0, keep_dims = var_3389, x = var_3386_cast_fp16)[name = string("op_3390_cast_fp16")];
            fp16 var_3391_to_fp16 = const()[name = string("op_3391_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_61_cast_fp16 = maximum(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_61_cast_fp16 = real_div(x = x_65_cast_fp16, y = s_61_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3394_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<int32, [1]> var_3399_axes_0 = const()[name = string("op_3399_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3399_keep_dims_0 = const()[name = string("op_3399_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3399_cast_fp16 = reduce_mean(axes = var_3399_axes_0, keep_dims = var_3399_keep_dims_0, x = var_3394_cast_fp16)[name = string("op_3399_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3400_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_3400_cast_fp16")];
            fp32 var_3401_epsilon_0 = const()[name = string("op_3401_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3401_cast_fp16 = inverse(epsilon = var_3401_epsilon_0, x = var_3400_cast_fp16)[name = string("op_3401_cast_fp16")];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3403_cast_fp16 = mul(x = var_3401_cast_fp16, y = var_3402_to_fp16)[name = string("op_3403_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3405_cast_fp16 = add(x = var_3399_cast_fp16, y = var_3403_cast_fp16)[name = string("op_3405_cast_fp16")];
            fp32 var_3406_epsilon_0 = const()[name = string("op_3406_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3406_cast_fp16 = rsqrt(epsilon = var_3406_epsilon_0, x = var_3405_cast_fp16)[name = string("op_3406_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_3406_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23732672)))];
            tensor<fp16, [1, 48, 512]> input_41_cast_fp16 = mul(x = out_53_cast_fp16, y = tm_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23733760))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24782400))))[name = string("tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_32_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_32_cast_fp16")];
            string var_3413_mode_0 = const()[name = string("op_3413_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_3413_cast_fp16 = gelu(mode = var_3413_mode_0, x = linear_32_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24786560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25835200))))[name = string("tm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_43_cast_fp16 = mul(x = var_3413_cast_fp16, y = linear_33_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25839360))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26888000))))[name = string("tm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3421_cast_fp16 = abs(x = linear_34_cast_fp16)[name = string("op_3421_cast_fp16")];
            bool var_3424 = const()[name = string("op_3424"), val = bool(true)];
            tensor<int32, [1]> var_3425_axes_0 = const()[name = string("op_3425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3425_cast_fp16 = reduce_max(axes = var_3425_axes_0, keep_dims = var_3424, x = var_3421_cast_fp16)[name = string("op_3425_cast_fp16")];
            fp16 var_3426_to_fp16 = const()[name = string("op_3426_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_63_cast_fp16 = maximum(x = var_3425_cast_fp16, y = var_3426_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_63_cast_fp16 = real_div(x = linear_34_cast_fp16, y = s_63_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3429_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_3429_cast_fp16")];
            tensor<int32, [1]> var_3434_axes_0 = const()[name = string("op_3434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3434_keep_dims_0 = const()[name = string("op_3434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3434_cast_fp16 = reduce_mean(axes = var_3434_axes_0, keep_dims = var_3434_keep_dims_0, x = var_3429_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3435_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp32 var_3436_epsilon_0 = const()[name = string("op_3436_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3436_cast_fp16 = inverse(epsilon = var_3436_epsilon_0, x = var_3435_cast_fp16)[name = string("op_3436_cast_fp16")];
            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3438_cast_fp16 = mul(x = var_3436_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3440_cast_fp16 = add(x = var_3434_cast_fp16, y = var_3438_cast_fp16)[name = string("op_3440_cast_fp16")];
            fp32 var_3441_epsilon_0 = const()[name = string("op_3441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3441_cast_fp16 = rsqrt(epsilon = var_3441_epsilon_0, x = var_3440_cast_fp16)[name = string("op_3441_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_55_cast_fp16 = mul(x = y_63_cast_fp16, y = var_3441_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26889088)))];
            tensor<fp16, [1, 48, 512]> var_3443_cast_fp16 = mul(x = out_55_cast_fp16, y = tm_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_45_cast_fp16 = add(x = x_65_cast_fp16, y = var_3443_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26890176))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27152384))))[name = string("tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_35_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_35_cast_fp16")];
            string x_69_mode_0 = const()[name = string("x_69_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_69_cast_fp16 = gelu(mode = x_69_mode_0, x = linear_35_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [4]> var_3463_begin_0 = const()[name = string("op_3463_begin_0"), val = tensor<int32, [4]>([0, 0, 3, 0])];
            tensor<int32, [4]> var_3463_end_0 = const()[name = string("op_3463_end_0"), val = tensor<int32, [4]>([1, 48, 4, 512])];
            tensor<bool, [4]> var_3463_end_mask_0 = const()[name = string("op_3463_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3463_squeeze_mask_0 = const()[name = string("op_3463_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_3463_cast_fp16 = slice_by_index(begin = var_3463_begin_0, end = var_3463_end_0, end_mask = var_3463_end_mask_0, squeeze_mask = var_3463_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3463_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_47_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3463_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27153472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27415680))))[name = string("tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3473_cast_fp16 = abs(x = linear_36_cast_fp16)[name = string("op_3473_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<int32, [1]> var_3477_axes_0 = const()[name = string("op_3477_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3477_cast_fp16 = reduce_max(axes = var_3477_axes_0, keep_dims = var_3476, x = var_3473_cast_fp16)[name = string("op_3477_cast_fp16")];
            fp16 var_3478_to_fp16 = const()[name = string("op_3478_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_65_cast_fp16 = maximum(x = var_3477_cast_fp16, y = var_3478_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_65_cast_fp16 = real_div(x = linear_36_cast_fp16, y = s_65_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3481_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [1]> var_3486_axes_0 = const()[name = string("op_3486_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3486_keep_dims_0 = const()[name = string("op_3486_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3486_cast_fp16 = reduce_mean(axes = var_3486_axes_0, keep_dims = var_3486_keep_dims_0, x = var_3481_cast_fp16)[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3487_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp32 var_3488_epsilon_0 = const()[name = string("op_3488_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3488_cast_fp16 = inverse(epsilon = var_3488_epsilon_0, x = var_3487_cast_fp16)[name = string("op_3488_cast_fp16")];
            fp16 var_3489_to_fp16 = const()[name = string("op_3489_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3490_cast_fp16 = mul(x = var_3488_cast_fp16, y = var_3489_to_fp16)[name = string("op_3490_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3492_cast_fp16 = add(x = var_3486_cast_fp16, y = var_3490_cast_fp16)[name = string("op_3492_cast_fp16")];
            fp32 var_3493_epsilon_0 = const()[name = string("op_3493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3493_cast_fp16 = rsqrt(epsilon = var_3493_epsilon_0, x = var_3492_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_57_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3493_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27416768)))];
            tensor<fp16, [1, 48, 512]> var_3495_cast_fp16 = mul(x = out_57_cast_fp16, y = tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3495_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_7_cast_fp16 = add(x = input_45_cast_fp16, y = var_3495_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<fp16, [1]> tm_layers_3_layer_scalar_to_fp16 = const()[name = string("tm_layers_3_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 48, 512]> x_73_cast_fp16 = mul(x = h_7_cast_fp16, y = tm_layers_3_layer_scalar_to_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3499_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_3499_cast_fp16")];
            bool var_3502 = const()[name = string("op_3502"), val = bool(true)];
            tensor<int32, [1]> var_3503_axes_0 = const()[name = string("op_3503_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3503_cast_fp16 = reduce_max(axes = var_3503_axes_0, keep_dims = var_3502, x = var_3499_cast_fp16)[name = string("op_3503_cast_fp16")];
            fp16 var_3504_to_fp16 = const()[name = string("op_3504_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_67_cast_fp16 = maximum(x = var_3503_cast_fp16, y = var_3504_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_67_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_67_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3507_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<int32, [1]> var_3512_axes_0 = const()[name = string("op_3512_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3512_keep_dims_0 = const()[name = string("op_3512_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3512_cast_fp16 = reduce_mean(axes = var_3512_axes_0, keep_dims = var_3512_keep_dims_0, x = var_3507_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3513_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_3513_cast_fp16")];
            fp32 var_3514_epsilon_0 = const()[name = string("op_3514_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3514_cast_fp16 = inverse(epsilon = var_3514_epsilon_0, x = var_3513_cast_fp16)[name = string("op_3514_cast_fp16")];
            fp16 var_3515_to_fp16 = const()[name = string("op_3515_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3516_cast_fp16 = mul(x = var_3514_cast_fp16, y = var_3515_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3518_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp32 var_3519_epsilon_0 = const()[name = string("op_3519_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3519_cast_fp16 = rsqrt(epsilon = var_3519_epsilon_0, x = var_3518_cast_fp16)[name = string("op_3519_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_59_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3519_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27417856)))];
            tensor<fp16, [1, 48, 512]> input_49_cast_fp16 = mul(x = out_59_cast_fp16, y = tm_layers_4_input_layernorm_weight_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27418944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27943296))))[name = string("tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3529 = const()[name = string("op_3529"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_3530_cast_fp16 = reshape(shape = var_3529, x = linear_37_cast_fp16)[name = string("op_3530_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27945408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28207616))))[name = string("tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3541 = const()[name = string("op_3541"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_3542_cast_fp16 = reshape(shape = var_3541, x = linear_38_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28208704))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28470912))))[name = string("tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_3553 = const()[name = string("op_3553"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_3554_cast_fp16 = reshape(shape = var_3553, x = linear_39_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<int32, [4]> x_79_perm_0 = const()[name = string("x_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3530_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 4, 48, 256]> var_3558_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3558_cast_fp16")];
            bool var_3561 = const()[name = string("op_3561"), val = bool(true)];
            tensor<int32, [1]> var_3562_axes_0 = const()[name = string("op_3562_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_3562_cast_fp16 = reduce_max(axes = var_3562_axes_0, keep_dims = var_3561, x = var_3558_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp16 var_3563_to_fp16 = const()[name = string("op_3563_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_69_cast_fp16 = maximum(x = var_3562_cast_fp16, y = var_3563_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_69_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_69_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_3566_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3566_cast_fp16")];
            tensor<int32, [1]> var_3571_axes_0 = const()[name = string("op_3571_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3571_keep_dims_0 = const()[name = string("op_3571_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_3571_cast_fp16 = reduce_mean(axes = var_3571_axes_0, keep_dims = var_3571_keep_dims_0, x = var_3566_cast_fp16)[name = string("op_3571_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_3572_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3572_cast_fp16")];
            fp32 var_3573_epsilon_0 = const()[name = string("op_3573_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_3573_cast_fp16 = inverse(epsilon = var_3573_epsilon_0, x = var_3572_cast_fp16)[name = string("op_3573_cast_fp16")];
            fp16 var_3574_to_fp16 = const()[name = string("op_3574_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_3575_cast_fp16 = mul(x = var_3573_cast_fp16, y = var_3574_to_fp16)[name = string("op_3575_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_3577_cast_fp16 = add(x = var_3571_cast_fp16, y = var_3575_cast_fp16)[name = string("op_3577_cast_fp16")];
            fp32 var_3578_epsilon_0 = const()[name = string("op_3578_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_3578_cast_fp16 = rsqrt(epsilon = var_3578_epsilon_0, x = var_3577_cast_fp16)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_61_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3578_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [256]> tm_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472000)))];
            tensor<fp16, [1, 4, 48, 256]> q_17_cast_fp16 = mul(x = out_61_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3542_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2, 48, 256]> var_3581_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3581_cast_fp16")];
            bool var_3584 = const()[name = string("op_3584"), val = bool(true)];
            tensor<int32, [1]> var_3585_axes_0 = const()[name = string("op_3585_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_3585_cast_fp16 = reduce_max(axes = var_3585_axes_0, keep_dims = var_3584, x = var_3581_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp16 var_3586_to_fp16 = const()[name = string("op_3586_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_71_cast_fp16 = maximum(x = var_3585_cast_fp16, y = var_3586_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_71_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_71_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_3589_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3589_cast_fp16")];
            tensor<int32, [1]> var_3594_axes_0 = const()[name = string("op_3594_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3594_keep_dims_0 = const()[name = string("op_3594_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_3594_cast_fp16 = reduce_mean(axes = var_3594_axes_0, keep_dims = var_3594_keep_dims_0, x = var_3589_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_3595_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp32 var_3596_epsilon_0 = const()[name = string("op_3596_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_3596_cast_fp16 = inverse(epsilon = var_3596_epsilon_0, x = var_3595_cast_fp16)[name = string("op_3596_cast_fp16")];
            fp16 var_3597_to_fp16 = const()[name = string("op_3597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_3598_cast_fp16 = mul(x = var_3596_cast_fp16, y = var_3597_to_fp16)[name = string("op_3598_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_3600_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3598_cast_fp16)[name = string("op_3600_cast_fp16")];
            fp32 var_3601_epsilon_0 = const()[name = string("op_3601_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_3601_cast_fp16 = rsqrt(epsilon = var_3601_epsilon_0, x = var_3600_cast_fp16)[name = string("op_3601_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_63_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3601_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472576)))];
            tensor<fp16, [1, 2, 48, 256]> k_25_cast_fp16 = mul(x = out_63_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_79_cast_fp16 = transpose(perm = x_79_perm_0, x = var_3554_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 2, 48, 256]> var_3604_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3604_cast_fp16")];
            bool var_3607 = const()[name = string("op_3607"), val = bool(true)];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_3607, x = var_3604_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_73_cast_fp16 = maximum(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_73_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_73_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_3612_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<int32, [1]> var_3617_axes_0 = const()[name = string("op_3617_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3617_keep_dims_0 = const()[name = string("op_3617_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_3617_cast_fp16 = reduce_mean(axes = var_3617_axes_0, keep_dims = var_3617_keep_dims_0, x = var_3612_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_3618_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_3619_cast_fp16 = inverse(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            fp16 var_3620_to_fp16 = const()[name = string("op_3620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_3621_cast_fp16 = mul(x = var_3619_cast_fp16, y = var_3620_to_fp16)[name = string("op_3621_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_3623_cast_fp16 = add(x = var_3617_cast_fp16, y = var_3621_cast_fp16)[name = string("op_3623_cast_fp16")];
            fp32 var_3624_epsilon_0 = const()[name = string("op_3624_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_3624_cast_fp16 = rsqrt(epsilon = var_3624_epsilon_0, x = var_3623_cast_fp16)[name = string("op_3624_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_17_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3624_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_3626_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3626_cast_fp16")];
            tensor<int32, [4]> var_3631_begin_0 = const()[name = string("op_3631_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3631_end_0 = const()[name = string("op_3631_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_3631_end_mask_0 = const()[name = string("op_3631_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_3631_cast_fp16 = slice_by_index(begin = var_3631_begin_0, end = var_3631_end_0, end_mask = var_3631_end_mask_0, x = q_17_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_3632_cast_fp16 = mul(x = var_3631_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> var_3637_begin_0 = const()[name = string("op_3637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3637_end_0 = const()[name = string("op_3637_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_3637_end_mask_0 = const()[name = string("op_3637_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_3637_cast_fp16 = slice_by_index(begin = var_3637_begin_0, end = var_3637_end_0, end_mask = var_3637_end_mask_0, x = q_17_cast_fp16)[name = string("op_3637_cast_fp16")];
            int32 var_3639 = const()[name = string("op_3639"), val = int32(-1)];
            bool var_3640_interleave_0 = const()[name = string("op_3640_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_3640_cast_fp16 = concat(axis = var_3639, interleave = var_3640_interleave_0, values = (var_3632_cast_fp16, var_3637_cast_fp16))[name = string("op_3640_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_3641_cast_fp16 = mul(x = var_3640_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3641_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_19_cast_fp16 = add(x = var_3626_cast_fp16, y = var_3641_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_3644_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = k_25_cast_fp16)[name = string("op_3649_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_3650_cast_fp16 = mul(x = var_3649_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            tensor<int32, [4]> var_3655_begin_0 = const()[name = string("op_3655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3655_end_0 = const()[name = string("op_3655_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_3655_end_mask_0 = const()[name = string("op_3655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_3655_cast_fp16 = slice_by_index(begin = var_3655_begin_0, end = var_3655_end_0, end_mask = var_3655_end_mask_0, x = k_25_cast_fp16)[name = string("op_3655_cast_fp16")];
            int32 var_3657 = const()[name = string("op_3657"), val = int32(-1)];
            bool var_3658_interleave_0 = const()[name = string("op_3658_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_3658_cast_fp16 = concat(axis = var_3657, interleave = var_3658_interleave_0, values = (var_3650_cast_fp16, var_3655_cast_fp16))[name = string("op_3658_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_3659_cast_fp16 = mul(x = var_3658_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3659_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_27_cast_fp16 = add(x = var_3644_cast_fp16, y = var_3659_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = k_27_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [4, 1, 48, 256]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
            tensor<int32, [5]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_16_cast_fp16 = reshape(shape = concat_16, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [4, 1, 48, 256]> reshape_17_cast_fp16 = reshape(shape = concat_17, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = v_17_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [4, 1, 48, 256]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
            tensor<int32, [5]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_18_cast_fp16 = reshape(shape = concat_18, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [4, 1, 48, 256]> reshape_19_cast_fp16 = reshape(shape = concat_19, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3673_transpose_x_0 = const()[name = string("op_3673_transpose_x_0"), val = bool(false)];
            bool var_3673_transpose_y_0 = const()[name = string("op_3673_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 4, 48, 48]> var_3673_cast_fp16 = matmul(transpose_x = var_3673_transpose_x_0, transpose_y = var_3673_transpose_y_0, x = q_19_cast_fp16, y = transpose_100_cast_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_3675_cast_fp16 = add(x = var_3673_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3675_cast_fp16")];
            int32 var_3676 = const()[name = string("op_3676"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_73_cast_fp16 = softmax(axis = var_3676, x = var_3675_cast_fp16)[name = string("w_73_cast_fp16")];
            bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)];
            bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 4, 48, 256]> var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = w_73_cast_fp16, y = v_19_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<int32, [4]> var_3682_perm_0 = const()[name = string("op_3682_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3686 = const()[name = string("op_3686"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_3682_cast_fp16 = transpose(perm = var_3682_perm_0, x = var_3679_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 48, 1024]> input_51_cast_fp16 = reshape(shape = var_3686, x = var_3682_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28473152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28997504))))[name = string("tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3691_cast_fp16 = abs(x = linear_40_cast_fp16)[name = string("op_3691_cast_fp16")];
            bool var_3694 = const()[name = string("op_3694"), val = bool(true)];
            tensor<int32, [1]> var_3695_axes_0 = const()[name = string("op_3695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3695_cast_fp16 = reduce_max(axes = var_3695_axes_0, keep_dims = var_3694, x = var_3691_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 var_3696_to_fp16 = const()[name = string("op_3696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_75_cast_fp16 = maximum(x = var_3695_cast_fp16, y = var_3696_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_75_cast_fp16 = real_div(x = linear_40_cast_fp16, y = s_75_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3699_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3699_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3704_keep_dims_0 = const()[name = string("op_3704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3704_cast_fp16 = reduce_mean(axes = var_3704_axes_0, keep_dims = var_3704_keep_dims_0, x = var_3699_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3705_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3705_cast_fp16")];
            fp32 var_3706_epsilon_0 = const()[name = string("op_3706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3706_cast_fp16 = inverse(epsilon = var_3706_epsilon_0, x = var_3705_cast_fp16)[name = string("op_3706_cast_fp16")];
            fp16 var_3707_to_fp16 = const()[name = string("op_3707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3708_cast_fp16 = mul(x = var_3706_cast_fp16, y = var_3707_to_fp16)[name = string("op_3708_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3710_cast_fp16 = add(x = var_3704_cast_fp16, y = var_3708_cast_fp16)[name = string("op_3710_cast_fp16")];
            fp32 var_3711_epsilon_0 = const()[name = string("op_3711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3711_cast_fp16 = rsqrt(epsilon = var_3711_epsilon_0, x = var_3710_cast_fp16)[name = string("op_3711_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3711_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28998592)))];
            tensor<fp16, [1, 48, 512]> var_3713_cast_fp16 = mul(x = out_65_cast_fp16, y = tm_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = var_3713_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3716_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3716_cast_fp16")];
            bool var_3719 = const()[name = string("op_3719"), val = bool(true)];
            tensor<int32, [1]> var_3720_axes_0 = const()[name = string("op_3720_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3720_cast_fp16 = reduce_max(axes = var_3720_axes_0, keep_dims = var_3719, x = var_3716_cast_fp16)[name = string("op_3720_cast_fp16")];
            fp16 var_3721_to_fp16 = const()[name = string("op_3721_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_77_cast_fp16 = maximum(x = var_3720_cast_fp16, y = var_3721_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_77_cast_fp16 = real_div(x = x_83_cast_fp16, y = s_77_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3724_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3729_axes_0 = const()[name = string("op_3729_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3729_keep_dims_0 = const()[name = string("op_3729_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3729_cast_fp16 = reduce_mean(axes = var_3729_axes_0, keep_dims = var_3729_keep_dims_0, x = var_3724_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3730_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3730_cast_fp16")];
            fp32 var_3731_epsilon_0 = const()[name = string("op_3731_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3731_cast_fp16 = inverse(epsilon = var_3731_epsilon_0, x = var_3730_cast_fp16)[name = string("op_3731_cast_fp16")];
            fp16 var_3732_to_fp16 = const()[name = string("op_3732_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3733_cast_fp16 = mul(x = var_3731_cast_fp16, y = var_3732_to_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3735_cast_fp16 = add(x = var_3729_cast_fp16, y = var_3733_cast_fp16)[name = string("op_3735_cast_fp16")];
            fp32 var_3736_epsilon_0 = const()[name = string("op_3736_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3736_cast_fp16 = rsqrt(epsilon = var_3736_epsilon_0, x = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_67_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3736_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28999680)))];
            tensor<fp16, [1, 48, 512]> input_53_cast_fp16 = mul(x = out_67_cast_fp16, y = tm_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29000768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30049408))))[name = string("tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_41_cast_fp16")];
            string var_3743_mode_0 = const()[name = string("op_3743_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_3743_cast_fp16 = gelu(mode = var_3743_mode_0, x = linear_41_cast_fp16)[name = string("op_3743_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30053568))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31102208))))[name = string("tm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_42_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_55_cast_fp16 = mul(x = var_3743_cast_fp16, y = linear_42_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31106368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32155008))))[name = string("tm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3751_cast_fp16 = abs(x = linear_43_cast_fp16)[name = string("op_3751_cast_fp16")];
            bool var_3754 = const()[name = string("op_3754"), val = bool(true)];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_3754, x = var_3751_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 var_3756_to_fp16 = const()[name = string("op_3756_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_79_cast_fp16 = maximum(x = var_3755_cast_fp16, y = var_3756_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_79_cast_fp16 = real_div(x = linear_43_cast_fp16, y = s_79_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3759_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<int32, [1]> var_3764_axes_0 = const()[name = string("op_3764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3764_keep_dims_0 = const()[name = string("op_3764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3764_cast_fp16 = reduce_mean(axes = var_3764_axes_0, keep_dims = var_3764_keep_dims_0, x = var_3759_cast_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3765_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3766_cast_fp16 = inverse(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3768_cast_fp16 = mul(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("op_3768_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3770_cast_fp16 = add(x = var_3764_cast_fp16, y = var_3768_cast_fp16)[name = string("op_3770_cast_fp16")];
            fp32 var_3771_epsilon_0 = const()[name = string("op_3771_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3771_cast_fp16 = rsqrt(epsilon = var_3771_epsilon_0, x = var_3770_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_69_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3771_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32156096)))];
            tensor<fp16, [1, 48, 512]> var_3773_cast_fp16 = mul(x = out_69_cast_fp16, y = tm_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3773_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_57_cast_fp16 = add(x = x_83_cast_fp16, y = var_3773_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32157184))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32419392))))[name = string("tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_44_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_44_cast_fp16")];
            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_87_cast_fp16 = gelu(mode = x_87_mode_0, x = linear_44_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [4]> var_3793_begin_0 = const()[name = string("op_3793_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_3793_end_0 = const()[name = string("op_3793_end_0"), val = tensor<int32, [4]>([1, 48, 5, 512])];
            tensor<bool, [4]> var_3793_end_mask_0 = const()[name = string("op_3793_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3793_squeeze_mask_0 = const()[name = string("op_3793_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_3793_cast_fp16 = slice_by_index(begin = var_3793_begin_0, end = var_3793_end_0, end_mask = var_3793_end_mask_0, squeeze_mask = var_3793_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_59_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3793_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32420480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32682688))))[name = string("tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3803_cast_fp16 = abs(x = linear_45_cast_fp16)[name = string("op_3803_cast_fp16")];
            bool var_3806 = const()[name = string("op_3806"), val = bool(true)];
            tensor<int32, [1]> var_3807_axes_0 = const()[name = string("op_3807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3807_cast_fp16 = reduce_max(axes = var_3807_axes_0, keep_dims = var_3806, x = var_3803_cast_fp16)[name = string("op_3807_cast_fp16")];
            fp16 var_3808_to_fp16 = const()[name = string("op_3808_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_81_cast_fp16 = maximum(x = var_3807_cast_fp16, y = var_3808_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_81_cast_fp16 = real_div(x = linear_45_cast_fp16, y = s_81_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3811_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [1]> var_3816_axes_0 = const()[name = string("op_3816_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3816_keep_dims_0 = const()[name = string("op_3816_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3816_cast_fp16 = reduce_mean(axes = var_3816_axes_0, keep_dims = var_3816_keep_dims_0, x = var_3811_cast_fp16)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3817_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3817_cast_fp16")];
            fp32 var_3818_epsilon_0 = const()[name = string("op_3818_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3818_cast_fp16 = inverse(epsilon = var_3818_epsilon_0, x = var_3817_cast_fp16)[name = string("op_3818_cast_fp16")];
            fp16 var_3819_to_fp16 = const()[name = string("op_3819_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3820_cast_fp16 = mul(x = var_3818_cast_fp16, y = var_3819_to_fp16)[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3822_cast_fp16 = add(x = var_3816_cast_fp16, y = var_3820_cast_fp16)[name = string("op_3822_cast_fp16")];
            fp32 var_3823_epsilon_0 = const()[name = string("op_3823_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3823_cast_fp16 = rsqrt(epsilon = var_3823_epsilon_0, x = var_3822_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_71_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3823_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32683776)))];
            tensor<fp16, [1, 48, 512]> var_3825_cast_fp16 = mul(x = out_71_cast_fp16, y = tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3825_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_9_cast_fp16 = add(x = input_57_cast_fp16, y = var_3825_cast_fp16)[name = string("h_9_cast_fp16")];
            tensor<fp16, [1]> tm_layers_4_layer_scalar_to_fp16 = const()[name = string("tm_layers_4_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.62p-1])];
            tensor<fp16, [1, 48, 512]> x_91_cast_fp16 = mul(x = h_9_cast_fp16, y = tm_layers_4_layer_scalar_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3829_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3829_cast_fp16")];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<int32, [1]> var_3833_axes_0 = const()[name = string("op_3833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_3833_cast_fp16 = reduce_max(axes = var_3833_axes_0, keep_dims = var_3832, x = var_3829_cast_fp16)[name = string("op_3833_cast_fp16")];
            fp16 var_3834_to_fp16 = const()[name = string("op_3834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_83_cast_fp16 = maximum(x = var_3833_cast_fp16, y = var_3834_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_83_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_83_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_3837_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3837_cast_fp16")];
            tensor<int32, [1]> var_3842_axes_0 = const()[name = string("op_3842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3842_keep_dims_0 = const()[name = string("op_3842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_3842_cast_fp16 = reduce_mean(axes = var_3842_axes_0, keep_dims = var_3842_keep_dims_0, x = var_3837_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3843_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp32 var_3844_epsilon_0 = const()[name = string("op_3844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_3844_cast_fp16 = inverse(epsilon = var_3844_epsilon_0, x = var_3843_cast_fp16)[name = string("op_3844_cast_fp16")];
            fp16 var_3845_to_fp16 = const()[name = string("op_3845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_3846_cast_fp16 = mul(x = var_3844_cast_fp16, y = var_3845_to_fp16)[name = string("op_3846_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_3848_cast_fp16 = add(x = var_3842_cast_fp16, y = var_3846_cast_fp16)[name = string("op_3848_cast_fp16")];
            fp32 var_3849_epsilon_0 = const()[name = string("op_3849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_3849_cast_fp16 = rsqrt(epsilon = var_3849_epsilon_0, x = var_3848_cast_fp16)[name = string("op_3849_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_73_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3849_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32684864)))];
            tensor<fp16, [1, 48, 512]> input_61_cast_fp16 = mul(x = out_73_cast_fp16, y = tm_layers_5_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32685952))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33734592))))[name = string("tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_46_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [4]> var_3859 = const()[name = string("op_3859"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 48, 4, 512]> var_3860_cast_fp16 = reshape(shape = var_3859, x = linear_46_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<int32, [4]> x_93_perm_0 = const()[name = string("x_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33738752))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34000960))))[name = string("tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<int32, [4]> var_3871 = const()[name = string("op_3871"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 48, 1, 512]> var_3872_cast_fp16 = reshape(shape = var_3871, x = linear_47_cast_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [4]> x_95_perm_0 = const()[name = string("x_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34002048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34264256))))[name = string("tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [4]> var_3883 = const()[name = string("op_3883"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 48, 1, 512]> var_3884_cast_fp16 = reshape(shape = var_3883, x = linear_48_cast_fp16)[name = string("op_3884_cast_fp16")];
            tensor<int32, [4]> x_97_perm_0 = const()[name = string("x_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 512]> x_93_cast_fp16 = transpose(perm = x_93_perm_0, x = var_3860_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 4, 48, 512]> var_3888_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3888_cast_fp16")];
            bool var_3891 = const()[name = string("op_3891"), val = bool(true)];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_3892_cast_fp16 = reduce_max(axes = var_3892_axes_0, keep_dims = var_3891, x = var_3888_cast_fp16)[name = string("op_3892_cast_fp16")];
            fp16 var_3893_to_fp16 = const()[name = string("op_3893_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_85_cast_fp16 = maximum(x = var_3892_cast_fp16, y = var_3893_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> y_85_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_85_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_3896_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3896_cast_fp16")];
            tensor<int32, [1]> var_3901_axes_0 = const()[name = string("op_3901_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3901_keep_dims_0 = const()[name = string("op_3901_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_3901_cast_fp16 = reduce_mean(axes = var_3901_axes_0, keep_dims = var_3901_keep_dims_0, x = var_3896_cast_fp16)[name = string("op_3901_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_3902_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3902_cast_fp16")];
            fp32 var_3903_epsilon_0 = const()[name = string("op_3903_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_3903_cast_fp16 = inverse(epsilon = var_3903_epsilon_0, x = var_3902_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_3905_cast_fp16 = mul(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_3907_cast_fp16 = add(x = var_3901_cast_fp16, y = var_3905_cast_fp16)[name = string("op_3907_cast_fp16")];
            fp32 var_3908_epsilon_0 = const()[name = string("op_3908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_3908_cast_fp16 = rsqrt(epsilon = var_3908_epsilon_0, x = var_3907_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> out_75_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3908_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34265344)))];
            tensor<fp16, [1, 4, 48, 512]> q_21_cast_fp16 = mul(x = out_75_cast_fp16, y = tm_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> x_95_cast_fp16 = transpose(perm = x_95_perm_0, x = var_3872_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1, 48, 512]> var_3911_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 48, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 48, 1]> s_87_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> y_87_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_87_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_3919_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 48, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_3925_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 48, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 48, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 48, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> out_77_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3931_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34266432)))];
            tensor<fp16, [1, 1, 48, 512]> k_31_cast_fp16 = mul(x = out_77_cast_fp16, y = tm_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = var_3884_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 1, 48, 512]> var_3934_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3934_cast_fp16")];
            bool var_3937 = const()[name = string("op_3937"), val = bool(true)];
            tensor<int32, [1]> var_3938_axes_0 = const()[name = string("op_3938_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 48, 1]> var_3938_cast_fp16 = reduce_max(axes = var_3938_axes_0, keep_dims = var_3937, x = var_3934_cast_fp16)[name = string("op_3938_cast_fp16")];
            fp16 var_3939_to_fp16 = const()[name = string("op_3939_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 48, 1]> s_89_cast_fp16 = maximum(x = var_3938_cast_fp16, y = var_3939_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> y_89_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_89_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_3942_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3942_cast_fp16")];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3947_keep_dims_0 = const()[name = string("op_3947_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 48, 1]> var_3947_cast_fp16 = reduce_mean(axes = var_3947_axes_0, keep_dims = var_3947_keep_dims_0, x = var_3942_cast_fp16)[name = string("op_3947_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_3948_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3948_cast_fp16")];
            fp32 var_3949_epsilon_0 = const()[name = string("op_3949_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 48, 1]> var_3949_cast_fp16 = inverse(epsilon = var_3949_epsilon_0, x = var_3948_cast_fp16)[name = string("op_3949_cast_fp16")];
            fp16 var_3950_to_fp16 = const()[name = string("op_3950_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 48, 1]> var_3951_cast_fp16 = mul(x = var_3949_cast_fp16, y = var_3950_to_fp16)[name = string("op_3951_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_3953_cast_fp16 = add(x = var_3947_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3953_cast_fp16")];
            fp32 var_3954_epsilon_0 = const()[name = string("op_3954_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 48, 1]> var_3954_cast_fp16 = rsqrt(epsilon = var_3954_epsilon_0, x = var_3953_cast_fp16)[name = string("op_3954_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> v_21_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3954_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<fp16, [48, 512]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [48, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137377472))), scale = tensor<fp16, [48, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137364736))))[name = string("cos_to_fp16_quantized")];
            tensor<fp16, [1, 4, 48, 512]> var_3956_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3956_cast_fp16")];
            tensor<int32, [4]> var_3961_begin_0 = const()[name = string("op_3961_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3961_end_0 = const()[name = string("op_3961_end_0"), val = tensor<int32, [4]>([1, 4, 48, 512])];
            tensor<bool, [4]> var_3961_end_mask_0 = const()[name = string("op_3961_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 256]> var_3961_cast_fp16 = slice_by_index(begin = var_3961_begin_0, end = var_3961_end_0, end_mask = var_3961_end_mask_0, x = q_21_cast_fp16)[name = string("op_3961_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 256]> var_3962_cast_fp16 = mul(x = var_3961_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_3962_cast_fp16")];
            tensor<int32, [4]> var_3967_begin_0 = const()[name = string("op_3967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3967_end_0 = const()[name = string("op_3967_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_3967_end_mask_0 = const()[name = string("op_3967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 256]> var_3967_cast_fp16 = slice_by_index(begin = var_3967_begin_0, end = var_3967_end_0, end_mask = var_3967_end_mask_0, x = q_21_cast_fp16)[name = string("op_3967_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(-1)];
            bool var_3970_interleave_0 = const()[name = string("op_3970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 512]> var_3970_cast_fp16 = concat(axis = var_3969, interleave = var_3970_interleave_0, values = (var_3962_cast_fp16, var_3967_cast_fp16))[name = string("op_3970_cast_fp16")];
            tensor<fp16, [48, 512]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [48, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137402112))), scale = tensor<fp16, [48, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137426752))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [1, 4, 48, 512]> var_3971_cast_fp16 = mul(x = var_3970_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3971_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> q_23_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3971_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_3974_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3974_cast_fp16")];
            tensor<int32, [4]> var_3979_begin_0 = const()[name = string("op_3979_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3979_end_0 = const()[name = string("op_3979_end_0"), val = tensor<int32, [4]>([1, 1, 48, 512])];
            tensor<bool, [4]> var_3979_end_mask_0 = const()[name = string("op_3979_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 48, 256]> var_3979_cast_fp16 = slice_by_index(begin = var_3979_begin_0, end = var_3979_end_0, end_mask = var_3979_end_mask_0, x = k_31_cast_fp16)[name = string("op_3979_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 48, 256]> var_3980_cast_fp16 = mul(x = var_3979_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_3980_cast_fp16")];
            tensor<int32, [4]> var_3985_begin_0 = const()[name = string("op_3985_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3985_end_0 = const()[name = string("op_3985_end_0"), val = tensor<int32, [4]>([1, 1, 48, 256])];
            tensor<bool, [4]> var_3985_end_mask_0 = const()[name = string("op_3985_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 48, 256]> var_3985_cast_fp16 = slice_by_index(begin = var_3985_begin_0, end = var_3985_end_0, end_mask = var_3985_end_mask_0, x = k_31_cast_fp16)[name = string("op_3985_cast_fp16")];
            int32 var_3987 = const()[name = string("op_3987"), val = int32(-1)];
            bool var_3988_interleave_0 = const()[name = string("op_3988_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 48, 512]> var_3988_cast_fp16 = concat(axis = var_3987, interleave = var_3988_interleave_0, values = (var_3980_cast_fp16, var_3985_cast_fp16))[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_3989_cast_fp16 = mul(x = var_3988_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> k_33_cast_fp16 = add(x = var_3974_cast_fp16, y = var_3989_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 48, 512]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = k_33_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [4, 1, 48, 512]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([4, 1, 1, 48, 512])];
            tensor<fp16, [4, 1, 1, 48, 512]> reshape_20_cast_fp16 = reshape(shape = concat_20, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 48, 512])];
            tensor<fp16, [1, 4, 1, 48, 512]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [4, 1, 48, 512]> reshape_21_cast_fp16 = reshape(shape = concat_21, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 48, 512]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = v_21_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [4, 1, 48, 512]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([4, 1, 1, 48, 512])];
            tensor<fp16, [4, 1, 1, 48, 512]> reshape_22_cast_fp16 = reshape(shape = concat_22, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 48, 512])];
            tensor<fp16, [1, 4, 1, 48, 512]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [4, 1, 48, 512]> reshape_23_cast_fp16 = reshape(shape = concat_23, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4003_transpose_x_0 = const()[name = string("op_4003_transpose_x_0"), val = bool(false)];
            bool var_4003_transpose_y_0 = const()[name = string("op_4003_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 48]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 4, 48, 48]> var_4003_cast_fp16 = matmul(transpose_x = var_4003_transpose_x_0, transpose_y = var_4003_transpose_y_0, x = q_23_cast_fp16, y = transpose_101_cast_fp16)[name = string("op_4003_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_4005_cast_fp16 = add(x = var_4003_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4005_cast_fp16")];
            int32 var_4006 = const()[name = string("op_4006"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_89_cast_fp16 = softmax(axis = var_4006, x = var_4005_cast_fp16)[name = string("w_89_cast_fp16")];
            bool var_4009_transpose_x_0 = const()[name = string("op_4009_transpose_x_0"), val = bool(false)];
            bool var_4009_transpose_y_0 = const()[name = string("op_4009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 512]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 4, 48, 512]> var_4009_cast_fp16 = matmul(transpose_x = var_4009_transpose_x_0, transpose_y = var_4009_transpose_y_0, x = w_89_cast_fp16, y = v_23_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4012_perm_0 = const()[name = string("op_4012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4016 = const()[name = string("op_4016"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 48, 4, 512]> var_4012_cast_fp16 = transpose(perm = var_4012_perm_0, x = var_4009_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 48, 2048]> input_63_cast_fp16 = reshape(shape = var_4016, x = var_4012_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34267520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35316160))))[name = string("tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4021_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_4021_cast_fp16")];
            bool var_4024 = const()[name = string("op_4024"), val = bool(true)];
            tensor<int32, [1]> var_4025_axes_0 = const()[name = string("op_4025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4025_cast_fp16 = reduce_max(axes = var_4025_axes_0, keep_dims = var_4024, x = var_4021_cast_fp16)[name = string("op_4025_cast_fp16")];
            fp16 var_4026_to_fp16 = const()[name = string("op_4026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_91_cast_fp16 = maximum(x = var_4025_cast_fp16, y = var_4026_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_91_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_91_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4029_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<int32, [1]> var_4034_axes_0 = const()[name = string("op_4034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4034_keep_dims_0 = const()[name = string("op_4034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4034_cast_fp16 = reduce_mean(axes = var_4034_axes_0, keep_dims = var_4034_keep_dims_0, x = var_4029_cast_fp16)[name = string("op_4034_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4035_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_4035_cast_fp16")];
            fp32 var_4036_epsilon_0 = const()[name = string("op_4036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4036_cast_fp16 = inverse(epsilon = var_4036_epsilon_0, x = var_4035_cast_fp16)[name = string("op_4036_cast_fp16")];
            fp16 var_4037_to_fp16 = const()[name = string("op_4037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4038_cast_fp16 = mul(x = var_4036_cast_fp16, y = var_4037_to_fp16)[name = string("op_4038_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4040_cast_fp16 = add(x = var_4034_cast_fp16, y = var_4038_cast_fp16)[name = string("op_4040_cast_fp16")];
            fp32 var_4041_epsilon_0 = const()[name = string("op_4041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4041_cast_fp16 = rsqrt(epsilon = var_4041_epsilon_0, x = var_4040_cast_fp16)[name = string("op_4041_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_79_cast_fp16 = mul(x = y_91_cast_fp16, y = var_4041_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35317248)))];
            tensor<fp16, [1, 48, 512]> var_4043_cast_fp16 = mul(x = out_79_cast_fp16, y = tm_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_4043_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_101_cast_fp16 = add(x = x_91_cast_fp16, y = var_4043_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4046_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_4046_cast_fp16")];
            bool var_4049 = const()[name = string("op_4049"), val = bool(true)];
            tensor<int32, [1]> var_4050_axes_0 = const()[name = string("op_4050_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4050_cast_fp16 = reduce_max(axes = var_4050_axes_0, keep_dims = var_4049, x = var_4046_cast_fp16)[name = string("op_4050_cast_fp16")];
            fp16 var_4051_to_fp16 = const()[name = string("op_4051_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_93_cast_fp16 = maximum(x = var_4050_cast_fp16, y = var_4051_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_93_cast_fp16 = real_div(x = x_101_cast_fp16, y = s_93_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4054_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [1]> var_4059_axes_0 = const()[name = string("op_4059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4059_keep_dims_0 = const()[name = string("op_4059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4059_cast_fp16 = reduce_mean(axes = var_4059_axes_0, keep_dims = var_4059_keep_dims_0, x = var_4054_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4060_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_4060_cast_fp16")];
            fp32 var_4061_epsilon_0 = const()[name = string("op_4061_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4061_cast_fp16 = inverse(epsilon = var_4061_epsilon_0, x = var_4060_cast_fp16)[name = string("op_4061_cast_fp16")];
            fp16 var_4062_to_fp16 = const()[name = string("op_4062_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4063_cast_fp16 = mul(x = var_4061_cast_fp16, y = var_4062_to_fp16)[name = string("op_4063_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4065_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4063_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp32 var_4066_epsilon_0 = const()[name = string("op_4066_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4066_cast_fp16 = rsqrt(epsilon = var_4066_epsilon_0, x = var_4065_cast_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_81_cast_fp16 = mul(x = y_93_cast_fp16, y = var_4066_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35318336)))];
            tensor<fp16, [1, 48, 512]> input_65_cast_fp16 = mul(x = out_81_cast_fp16, y = tm_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35319424))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36368064))))[name = string("tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_50_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_50_cast_fp16")];
            string var_4073_mode_0 = const()[name = string("op_4073_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_4073_cast_fp16 = gelu(mode = var_4073_mode_0, x = linear_50_cast_fp16)[name = string("op_4073_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36372224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37420864))))[name = string("tm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_51_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_67_cast_fp16 = mul(x = var_4073_cast_fp16, y = linear_51_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37425024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38473664))))[name = string("tm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4081_cast_fp16 = abs(x = linear_52_cast_fp16)[name = string("op_4081_cast_fp16")];
            bool var_4084 = const()[name = string("op_4084"), val = bool(true)];
            tensor<int32, [1]> var_4085_axes_0 = const()[name = string("op_4085_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4085_cast_fp16 = reduce_max(axes = var_4085_axes_0, keep_dims = var_4084, x = var_4081_cast_fp16)[name = string("op_4085_cast_fp16")];
            fp16 var_4086_to_fp16 = const()[name = string("op_4086_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_95_cast_fp16 = maximum(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_95_cast_fp16 = real_div(x = linear_52_cast_fp16, y = s_95_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4089_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_4089_cast_fp16")];
            tensor<int32, [1]> var_4094_axes_0 = const()[name = string("op_4094_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4094_keep_dims_0 = const()[name = string("op_4094_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4094_cast_fp16 = reduce_mean(axes = var_4094_axes_0, keep_dims = var_4094_keep_dims_0, x = var_4089_cast_fp16)[name = string("op_4094_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4095_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_4095_cast_fp16")];
            fp32 var_4096_epsilon_0 = const()[name = string("op_4096_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4096_cast_fp16 = inverse(epsilon = var_4096_epsilon_0, x = var_4095_cast_fp16)[name = string("op_4096_cast_fp16")];
            fp16 var_4097_to_fp16 = const()[name = string("op_4097_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4098_cast_fp16 = mul(x = var_4096_cast_fp16, y = var_4097_to_fp16)[name = string("op_4098_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4100_cast_fp16 = add(x = var_4094_cast_fp16, y = var_4098_cast_fp16)[name = string("op_4100_cast_fp16")];
            fp32 var_4101_epsilon_0 = const()[name = string("op_4101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4101_cast_fp16 = rsqrt(epsilon = var_4101_epsilon_0, x = var_4100_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_83_cast_fp16 = mul(x = y_95_cast_fp16, y = var_4101_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38474752)))];
            tensor<fp16, [1, 48, 512]> var_4103_cast_fp16 = mul(x = out_83_cast_fp16, y = tm_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_69_cast_fp16 = add(x = x_101_cast_fp16, y = var_4103_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38475840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38738048))))[name = string("tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_53_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_53_cast_fp16")];
            string x_105_mode_0 = const()[name = string("x_105_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_105_cast_fp16 = gelu(mode = x_105_mode_0, x = linear_53_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [4]> var_4123_begin_0 = const()[name = string("op_4123_begin_0"), val = tensor<int32, [4]>([0, 0, 5, 0])];
            tensor<int32, [4]> var_4123_end_0 = const()[name = string("op_4123_end_0"), val = tensor<int32, [4]>([1, 48, 6, 512])];
            tensor<bool, [4]> var_4123_end_mask_0 = const()[name = string("op_4123_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4123_squeeze_mask_0 = const()[name = string("op_4123_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_4123_cast_fp16 = slice_by_index(begin = var_4123_begin_0, end = var_4123_end_0, end_mask = var_4123_end_mask_0, squeeze_mask = var_4123_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_71_cast_fp16 = mul(x = x_105_cast_fp16, y = var_4123_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38739136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39001344))))[name = string("tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4133_cast_fp16 = abs(x = linear_54_cast_fp16)[name = string("op_4133_cast_fp16")];
            bool var_4136 = const()[name = string("op_4136"), val = bool(true)];
            tensor<int32, [1]> var_4137_axes_0 = const()[name = string("op_4137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4137_cast_fp16 = reduce_max(axes = var_4137_axes_0, keep_dims = var_4136, x = var_4133_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 var_4138_to_fp16 = const()[name = string("op_4138_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_97_cast_fp16 = maximum(x = var_4137_cast_fp16, y = var_4138_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_97_cast_fp16 = real_div(x = linear_54_cast_fp16, y = s_97_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4141_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_4141_cast_fp16")];
            tensor<int32, [1]> var_4146_axes_0 = const()[name = string("op_4146_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4146_keep_dims_0 = const()[name = string("op_4146_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4146_cast_fp16 = reduce_mean(axes = var_4146_axes_0, keep_dims = var_4146_keep_dims_0, x = var_4141_cast_fp16)[name = string("op_4146_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4147_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_4147_cast_fp16")];
            fp32 var_4148_epsilon_0 = const()[name = string("op_4148_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4148_cast_fp16 = inverse(epsilon = var_4148_epsilon_0, x = var_4147_cast_fp16)[name = string("op_4148_cast_fp16")];
            fp16 var_4149_to_fp16 = const()[name = string("op_4149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4150_cast_fp16 = mul(x = var_4148_cast_fp16, y = var_4149_to_fp16)[name = string("op_4150_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4152_cast_fp16 = add(x = var_4146_cast_fp16, y = var_4150_cast_fp16)[name = string("op_4152_cast_fp16")];
            fp32 var_4153_epsilon_0 = const()[name = string("op_4153_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4153_cast_fp16 = rsqrt(epsilon = var_4153_epsilon_0, x = var_4152_cast_fp16)[name = string("op_4153_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_85_cast_fp16 = mul(x = y_97_cast_fp16, y = var_4153_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39002432)))];
            tensor<fp16, [1, 48, 512]> var_4155_cast_fp16 = mul(x = out_85_cast_fp16, y = tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4155_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_11_cast_fp16 = add(x = input_69_cast_fp16, y = var_4155_cast_fp16)[name = string("h_11_cast_fp16")];
            tensor<fp16, [1]> tm_layers_5_layer_scalar_to_fp16 = const()[name = string("tm_layers_5_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-2])];
            tensor<fp16, [1, 48, 512]> x_109_cast_fp16 = mul(x = h_11_cast_fp16, y = tm_layers_5_layer_scalar_to_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4159_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_4159_cast_fp16")];
            bool var_4162 = const()[name = string("op_4162"), val = bool(true)];
            tensor<int32, [1]> var_4163_axes_0 = const()[name = string("op_4163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4163_cast_fp16 = reduce_max(axes = var_4163_axes_0, keep_dims = var_4162, x = var_4159_cast_fp16)[name = string("op_4163_cast_fp16")];
            fp16 var_4164_to_fp16 = const()[name = string("op_4164_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_99_cast_fp16 = maximum(x = var_4163_cast_fp16, y = var_4164_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_99_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_99_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4167_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_4167_cast_fp16")];
            tensor<int32, [1]> var_4172_axes_0 = const()[name = string("op_4172_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4172_keep_dims_0 = const()[name = string("op_4172_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4172_cast_fp16 = reduce_mean(axes = var_4172_axes_0, keep_dims = var_4172_keep_dims_0, x = var_4167_cast_fp16)[name = string("op_4172_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4173_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4174_cast_fp16 = inverse(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            fp16 var_4175_to_fp16 = const()[name = string("op_4175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4176_cast_fp16 = mul(x = var_4174_cast_fp16, y = var_4175_to_fp16)[name = string("op_4176_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4178_cast_fp16 = add(x = var_4172_cast_fp16, y = var_4176_cast_fp16)[name = string("op_4178_cast_fp16")];
            fp32 var_4179_epsilon_0 = const()[name = string("op_4179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4179_cast_fp16 = rsqrt(epsilon = var_4179_epsilon_0, x = var_4178_cast_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_87_cast_fp16 = mul(x = y_99_cast_fp16, y = var_4179_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39003520)))];
            tensor<fp16, [1, 48, 512]> input_73_cast_fp16 = mul(x = out_87_cast_fp16, y = tm_layers_6_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39004608))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39528960))))[name = string("tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [4]> var_4189 = const()[name = string("op_4189"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_4190_cast_fp16 = reshape(shape = var_4189, x = linear_55_cast_fp16)[name = string("op_4190_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39531072))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39793280))))[name = string("tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_4201 = const()[name = string("op_4201"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_4202_cast_fp16 = reshape(shape = var_4201, x = linear_56_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39794368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40056576))))[name = string("tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_57_cast_fp16)[name = string("op_4214_cast_fp16")];
            tensor<int32, [4]> x_115_perm_0 = const()[name = string("x_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4190_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 4, 48, 256]> var_4218_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_4218_cast_fp16")];
            bool var_4221 = const()[name = string("op_4221"), val = bool(true)];
            tensor<int32, [1]> var_4222_axes_0 = const()[name = string("op_4222_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_4222_cast_fp16 = reduce_max(axes = var_4222_axes_0, keep_dims = var_4221, x = var_4218_cast_fp16)[name = string("op_4222_cast_fp16")];
            fp16 var_4223_to_fp16 = const()[name = string("op_4223_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_101_cast_fp16 = maximum(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_101_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_101_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4226_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [1]> var_4231_axes_0 = const()[name = string("op_4231_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4231_keep_dims_0 = const()[name = string("op_4231_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_4231_cast_fp16 = reduce_mean(axes = var_4231_axes_0, keep_dims = var_4231_keep_dims_0, x = var_4226_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_4232_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_4232_cast_fp16")];
            fp32 var_4233_epsilon_0 = const()[name = string("op_4233_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_4233_cast_fp16 = inverse(epsilon = var_4233_epsilon_0, x = var_4232_cast_fp16)[name = string("op_4233_cast_fp16")];
            fp16 var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_4235_cast_fp16 = mul(x = var_4233_cast_fp16, y = var_4234_to_fp16)[name = string("op_4235_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_4237_cast_fp16 = add(x = var_4231_cast_fp16, y = var_4235_cast_fp16)[name = string("op_4237_cast_fp16")];
            fp32 var_4238_epsilon_0 = const()[name = string("op_4238_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_4238_cast_fp16 = rsqrt(epsilon = var_4238_epsilon_0, x = var_4237_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_89_cast_fp16 = mul(x = y_101_cast_fp16, y = var_4238_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40057664)))];
            tensor<fp16, [1, 4, 48, 256]> q_25_cast_fp16 = mul(x = out_89_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4202_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 2, 48, 256]> var_4241_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_4241_cast_fp16")];
            bool var_4244 = const()[name = string("op_4244"), val = bool(true)];
            tensor<int32, [1]> var_4245_axes_0 = const()[name = string("op_4245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_4245_cast_fp16 = reduce_max(axes = var_4245_axes_0, keep_dims = var_4244, x = var_4241_cast_fp16)[name = string("op_4245_cast_fp16")];
            fp16 var_4246_to_fp16 = const()[name = string("op_4246_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_103_cast_fp16 = maximum(x = var_4245_cast_fp16, y = var_4246_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_103_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_103_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4249_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<int32, [1]> var_4254_axes_0 = const()[name = string("op_4254_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4254_keep_dims_0 = const()[name = string("op_4254_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_4254_cast_fp16 = reduce_mean(axes = var_4254_axes_0, keep_dims = var_4254_keep_dims_0, x = var_4249_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4255_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_4255_cast_fp16")];
            fp32 var_4256_epsilon_0 = const()[name = string("op_4256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_4256_cast_fp16 = inverse(epsilon = var_4256_epsilon_0, x = var_4255_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp16 var_4257_to_fp16 = const()[name = string("op_4257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_4258_cast_fp16 = mul(x = var_4256_cast_fp16, y = var_4257_to_fp16)[name = string("op_4258_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4260_cast_fp16 = add(x = var_4254_cast_fp16, y = var_4258_cast_fp16)[name = string("op_4260_cast_fp16")];
            fp32 var_4261_epsilon_0 = const()[name = string("op_4261_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_4261_cast_fp16 = rsqrt(epsilon = var_4261_epsilon_0, x = var_4260_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_91_cast_fp16 = mul(x = y_103_cast_fp16, y = var_4261_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058240)))];
            tensor<fp16, [1, 2, 48, 256]> k_37_cast_fp16 = mul(x = out_91_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_37_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_115_cast_fp16 = transpose(perm = x_115_perm_0, x = var_4214_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 2, 48, 256]> var_4264_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_4264_cast_fp16")];
            bool var_4267 = const()[name = string("op_4267"), val = bool(true)];
            tensor<int32, [1]> var_4268_axes_0 = const()[name = string("op_4268_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_4268_cast_fp16 = reduce_max(axes = var_4268_axes_0, keep_dims = var_4267, x = var_4264_cast_fp16)[name = string("op_4268_cast_fp16")];
            fp16 var_4269_to_fp16 = const()[name = string("op_4269_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_105_cast_fp16 = maximum(x = var_4268_cast_fp16, y = var_4269_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_105_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_105_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4272_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_4272_cast_fp16")];
            tensor<int32, [1]> var_4277_axes_0 = const()[name = string("op_4277_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4277_keep_dims_0 = const()[name = string("op_4277_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_4277_cast_fp16 = reduce_mean(axes = var_4277_axes_0, keep_dims = var_4277_keep_dims_0, x = var_4272_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4278_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp32 var_4279_epsilon_0 = const()[name = string("op_4279_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_4279_cast_fp16 = inverse(epsilon = var_4279_epsilon_0, x = var_4278_cast_fp16)[name = string("op_4279_cast_fp16")];
            fp16 var_4280_to_fp16 = const()[name = string("op_4280_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_4281_cast_fp16 = mul(x = var_4279_cast_fp16, y = var_4280_to_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4283_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4281_cast_fp16)[name = string("op_4283_cast_fp16")];
            fp32 var_4284_epsilon_0 = const()[name = string("op_4284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_4284_cast_fp16 = rsqrt(epsilon = var_4284_epsilon_0, x = var_4283_cast_fp16)[name = string("op_4284_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_25_cast_fp16 = mul(x = y_105_cast_fp16, y = var_4284_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4286_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4286_cast_fp16")];
            tensor<int32, [4]> var_4291_begin_0 = const()[name = string("op_4291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4291_end_0 = const()[name = string("op_4291_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_4291_end_mask_0 = const()[name = string("op_4291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_4291_cast_fp16 = slice_by_index(begin = var_4291_begin_0, end = var_4291_end_0, end_mask = var_4291_end_mask_0, x = q_25_cast_fp16)[name = string("op_4291_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_4292_cast_fp16 = mul(x = var_4291_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = q_25_cast_fp16)[name = string("op_4297_cast_fp16")];
            int32 var_4299 = const()[name = string("op_4299"), val = int32(-1)];
            bool var_4300_interleave_0 = const()[name = string("op_4300_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_4300_cast_fp16 = concat(axis = var_4299, interleave = var_4300_interleave_0, values = (var_4292_cast_fp16, var_4297_cast_fp16))[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4301_cast_fp16 = mul(x = var_4300_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4301_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_27_cast_fp16 = add(x = var_4286_cast_fp16, y = var_4301_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4304_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4304_cast_fp16")];
            tensor<int32, [4]> var_4309_begin_0 = const()[name = string("op_4309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4309_end_0 = const()[name = string("op_4309_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_4309_end_mask_0 = const()[name = string("op_4309_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_4309_cast_fp16 = slice_by_index(begin = var_4309_begin_0, end = var_4309_end_0, end_mask = var_4309_end_mask_0, x = k_37_cast_fp16)[name = string("op_4309_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_4310_cast_fp16 = mul(x = var_4309_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_4310_cast_fp16")];
            tensor<int32, [4]> var_4315_begin_0 = const()[name = string("op_4315_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4315_end_0 = const()[name = string("op_4315_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_4315_end_mask_0 = const()[name = string("op_4315_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_4315_cast_fp16 = slice_by_index(begin = var_4315_begin_0, end = var_4315_end_0, end_mask = var_4315_end_mask_0, x = k_37_cast_fp16)[name = string("op_4315_cast_fp16")];
            int32 var_4317 = const()[name = string("op_4317"), val = int32(-1)];
            bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_4318_cast_fp16 = concat(axis = var_4317, interleave = var_4318_interleave_0, values = (var_4310_cast_fp16, var_4315_cast_fp16))[name = string("op_4318_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4319_cast_fp16 = mul(x = var_4318_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_39_cast_fp16 = add(x = var_4304_cast_fp16, y = var_4319_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = k_39_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [4, 1, 48, 256]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
            tensor<int32, [5]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_24_cast_fp16 = reshape(shape = concat_24, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [4, 1, 48, 256]> reshape_25_cast_fp16 = reshape(shape = concat_25, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = v_25_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [4, 1, 48, 256]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
            tensor<int32, [5]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_26_cast_fp16 = reshape(shape = concat_26, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [4, 1, 48, 256]> reshape_27_cast_fp16 = reshape(shape = concat_27, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4333_transpose_x_0 = const()[name = string("op_4333_transpose_x_0"), val = bool(false)];
            bool var_4333_transpose_y_0 = const()[name = string("op_4333_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_102_cast_fp16 = transpose(perm = transpose_102_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 4, 48, 48]> var_4333_cast_fp16 = matmul(transpose_x = var_4333_transpose_x_0, transpose_y = var_4333_transpose_y_0, x = q_27_cast_fp16, y = transpose_102_cast_fp16)[name = string("op_4333_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_4335_cast_fp16 = add(x = var_4333_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4335_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_105_cast_fp16 = softmax(axis = var_4336, x = var_4335_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4339_transpose_x_0 = const()[name = string("op_4339_transpose_x_0"), val = bool(false)];
            bool var_4339_transpose_y_0 = const()[name = string("op_4339_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 4, 48, 256]> var_4339_cast_fp16 = matmul(transpose_x = var_4339_transpose_x_0, transpose_y = var_4339_transpose_y_0, x = w_105_cast_fp16, y = v_27_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<int32, [4]> var_4342_perm_0 = const()[name = string("op_4342_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4346 = const()[name = string("op_4346"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_4342_cast_fp16 = transpose(perm = var_4342_perm_0, x = var_4339_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 48, 1024]> input_75_cast_fp16 = reshape(shape = var_4346, x = var_4342_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40583168))))[name = string("tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4351_cast_fp16 = abs(x = linear_58_cast_fp16)[name = string("op_4351_cast_fp16")];
            bool var_4354 = const()[name = string("op_4354"), val = bool(true)];
            tensor<int32, [1]> var_4355_axes_0 = const()[name = string("op_4355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4355_cast_fp16 = reduce_max(axes = var_4355_axes_0, keep_dims = var_4354, x = var_4351_cast_fp16)[name = string("op_4355_cast_fp16")];
            fp16 var_4356_to_fp16 = const()[name = string("op_4356_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_107_cast_fp16 = maximum(x = var_4355_cast_fp16, y = var_4356_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_107_cast_fp16 = real_div(x = linear_58_cast_fp16, y = s_107_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4359_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<int32, [1]> var_4364_axes_0 = const()[name = string("op_4364_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4364_keep_dims_0 = const()[name = string("op_4364_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4364_cast_fp16 = reduce_mean(axes = var_4364_axes_0, keep_dims = var_4364_keep_dims_0, x = var_4359_cast_fp16)[name = string("op_4364_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4365_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_4365_cast_fp16")];
            fp32 var_4366_epsilon_0 = const()[name = string("op_4366_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4366_cast_fp16 = inverse(epsilon = var_4366_epsilon_0, x = var_4365_cast_fp16)[name = string("op_4366_cast_fp16")];
            fp16 var_4367_to_fp16 = const()[name = string("op_4367_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4368_cast_fp16 = mul(x = var_4366_cast_fp16, y = var_4367_to_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4370_cast_fp16 = add(x = var_4364_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            fp32 var_4371_epsilon_0 = const()[name = string("op_4371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4371_cast_fp16 = rsqrt(epsilon = var_4371_epsilon_0, x = var_4370_cast_fp16)[name = string("op_4371_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_93_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4371_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40584256)))];
            tensor<fp16, [1, 48, 512]> var_4373_cast_fp16 = mul(x = out_93_cast_fp16, y = tm_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_4373_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = var_4373_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4376_cast_fp16 = abs(x = x_119_cast_fp16)[name = string("op_4376_cast_fp16")];
            bool var_4379 = const()[name = string("op_4379"), val = bool(true)];
            tensor<int32, [1]> var_4380_axes_0 = const()[name = string("op_4380_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4380_cast_fp16 = reduce_max(axes = var_4380_axes_0, keep_dims = var_4379, x = var_4376_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 var_4381_to_fp16 = const()[name = string("op_4381_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_109_cast_fp16 = maximum(x = var_4380_cast_fp16, y = var_4381_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_109_cast_fp16 = real_div(x = x_119_cast_fp16, y = s_109_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4384_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4384_cast_fp16")];
            tensor<int32, [1]> var_4389_axes_0 = const()[name = string("op_4389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4389_keep_dims_0 = const()[name = string("op_4389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4389_cast_fp16 = reduce_mean(axes = var_4389_axes_0, keep_dims = var_4389_keep_dims_0, x = var_4384_cast_fp16)[name = string("op_4389_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4390_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_4390_cast_fp16")];
            fp32 var_4391_epsilon_0 = const()[name = string("op_4391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4391_cast_fp16 = inverse(epsilon = var_4391_epsilon_0, x = var_4390_cast_fp16)[name = string("op_4391_cast_fp16")];
            fp16 var_4392_to_fp16 = const()[name = string("op_4392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = string("op_4393_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4395_cast_fp16 = add(x = var_4389_cast_fp16, y = var_4393_cast_fp16)[name = string("op_4395_cast_fp16")];
            fp32 var_4396_epsilon_0 = const()[name = string("op_4396_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4396_cast_fp16 = rsqrt(epsilon = var_4396_epsilon_0, x = var_4395_cast_fp16)[name = string("op_4396_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_95_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4396_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40585344)))];
            tensor<fp16, [1, 48, 512]> input_77_cast_fp16 = mul(x = out_95_cast_fp16, y = tm_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40586432))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41635072))))[name = string("tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_59_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_59_cast_fp16")];
            string var_4403_mode_0 = const()[name = string("op_4403_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_4403_cast_fp16 = gelu(mode = var_4403_mode_0, x = linear_59_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41639232))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42687872))))[name = string("tm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_60_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_79_cast_fp16 = mul(x = var_4403_cast_fp16, y = linear_60_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42692032))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43740672))))[name = string("tm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4411_cast_fp16 = abs(x = linear_61_cast_fp16)[name = string("op_4411_cast_fp16")];
            bool var_4414 = const()[name = string("op_4414"), val = bool(true)];
            tensor<int32, [1]> var_4415_axes_0 = const()[name = string("op_4415_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4415_cast_fp16 = reduce_max(axes = var_4415_axes_0, keep_dims = var_4414, x = var_4411_cast_fp16)[name = string("op_4415_cast_fp16")];
            fp16 var_4416_to_fp16 = const()[name = string("op_4416_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_111_cast_fp16 = maximum(x = var_4415_cast_fp16, y = var_4416_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_111_cast_fp16 = real_div(x = linear_61_cast_fp16, y = s_111_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4419_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4419_cast_fp16")];
            tensor<int32, [1]> var_4424_axes_0 = const()[name = string("op_4424_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4424_keep_dims_0 = const()[name = string("op_4424_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4424_cast_fp16 = reduce_mean(axes = var_4424_axes_0, keep_dims = var_4424_keep_dims_0, x = var_4419_cast_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4425_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4426_cast_fp16 = inverse(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            fp16 var_4427_to_fp16 = const()[name = string("op_4427_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4428_cast_fp16 = mul(x = var_4426_cast_fp16, y = var_4427_to_fp16)[name = string("op_4428_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4430_cast_fp16 = add(x = var_4424_cast_fp16, y = var_4428_cast_fp16)[name = string("op_4430_cast_fp16")];
            fp32 var_4431_epsilon_0 = const()[name = string("op_4431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4431_cast_fp16 = rsqrt(epsilon = var_4431_epsilon_0, x = var_4430_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_97_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4431_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43741760)))];
            tensor<fp16, [1, 48, 512]> var_4433_cast_fp16 = mul(x = out_97_cast_fp16, y = tm_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4433_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_81_cast_fp16 = add(x = x_119_cast_fp16, y = var_4433_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43742848))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44005056))))[name = string("tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_62_cast_fp16")];
            string x_123_mode_0 = const()[name = string("x_123_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_123_cast_fp16 = gelu(mode = x_123_mode_0, x = linear_62_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [4]> var_4453_begin_0 = const()[name = string("op_4453_begin_0"), val = tensor<int32, [4]>([0, 0, 6, 0])];
            tensor<int32, [4]> var_4453_end_0 = const()[name = string("op_4453_end_0"), val = tensor<int32, [4]>([1, 48, 7, 512])];
            tensor<bool, [4]> var_4453_end_mask_0 = const()[name = string("op_4453_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4453_squeeze_mask_0 = const()[name = string("op_4453_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_4453_cast_fp16 = slice_by_index(begin = var_4453_begin_0, end = var_4453_end_0, end_mask = var_4453_end_mask_0, squeeze_mask = var_4453_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_83_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4453_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44006144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44268352))))[name = string("tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4463_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4463_cast_fp16")];
            bool var_4466 = const()[name = string("op_4466"), val = bool(true)];
            tensor<int32, [1]> var_4467_axes_0 = const()[name = string("op_4467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4467_cast_fp16 = reduce_max(axes = var_4467_axes_0, keep_dims = var_4466, x = var_4463_cast_fp16)[name = string("op_4467_cast_fp16")];
            fp16 var_4468_to_fp16 = const()[name = string("op_4468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_113_cast_fp16 = maximum(x = var_4467_cast_fp16, y = var_4468_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_113_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_113_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4471_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [1]> var_4476_axes_0 = const()[name = string("op_4476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4476_keep_dims_0 = const()[name = string("op_4476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4476_cast_fp16 = reduce_mean(axes = var_4476_axes_0, keep_dims = var_4476_keep_dims_0, x = var_4471_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4477_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_fp16)[name = string("op_4477_cast_fp16")];
            fp32 var_4478_epsilon_0 = const()[name = string("op_4478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4478_cast_fp16 = inverse(epsilon = var_4478_epsilon_0, x = var_4477_cast_fp16)[name = string("op_4478_cast_fp16")];
            fp16 var_4479_to_fp16 = const()[name = string("op_4479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4480_cast_fp16 = mul(x = var_4478_cast_fp16, y = var_4479_to_fp16)[name = string("op_4480_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4482_cast_fp16 = add(x = var_4476_cast_fp16, y = var_4480_cast_fp16)[name = string("op_4482_cast_fp16")];
            fp32 var_4483_epsilon_0 = const()[name = string("op_4483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4483_cast_fp16 = rsqrt(epsilon = var_4483_epsilon_0, x = var_4482_cast_fp16)[name = string("op_4483_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_99_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4483_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44269440)))];
            tensor<fp16, [1, 48, 512]> var_4485_cast_fp16 = mul(x = out_99_cast_fp16, y = tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_13_cast_fp16 = add(x = input_81_cast_fp16, y = var_4485_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<fp16, [1]> tm_layers_6_layer_scalar_to_fp16 = const()[name = string("tm_layers_6_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.cep-3])];
            tensor<fp16, [1, 48, 512]> x_127_cast_fp16 = mul(x = h_13_cast_fp16, y = tm_layers_6_layer_scalar_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4489_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_4489_cast_fp16")];
            bool var_4492 = const()[name = string("op_4492"), val = bool(true)];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4493_cast_fp16 = reduce_max(axes = var_4493_axes_0, keep_dims = var_4492, x = var_4489_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_115_cast_fp16 = maximum(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_115_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_115_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4497_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<int32, [1]> var_4502_axes_0 = const()[name = string("op_4502_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4502_keep_dims_0 = const()[name = string("op_4502_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4502_cast_fp16 = reduce_mean(axes = var_4502_axes_0, keep_dims = var_4502_keep_dims_0, x = var_4497_cast_fp16)[name = string("op_4502_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4503_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4504_cast_fp16 = inverse(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            fp16 var_4505_to_fp16 = const()[name = string("op_4505_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4506_cast_fp16 = mul(x = var_4504_cast_fp16, y = var_4505_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4508_cast_fp16 = add(x = var_4502_cast_fp16, y = var_4506_cast_fp16)[name = string("op_4508_cast_fp16")];
            fp32 var_4509_epsilon_0 = const()[name = string("op_4509_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4509_cast_fp16 = rsqrt(epsilon = var_4509_epsilon_0, x = var_4508_cast_fp16)[name = string("op_4509_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_101_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4509_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44270528)))];
            tensor<fp16, [1, 48, 512]> input_85_cast_fp16 = mul(x = out_101_cast_fp16, y = tm_layers_7_input_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44271616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44795968))))[name = string("tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4519 = const()[name = string("op_4519"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_4520_cast_fp16 = reshape(shape = var_4519, x = linear_64_cast_fp16)[name = string("op_4520_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44798080))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45060288))))[name = string("tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4531 = const()[name = string("op_4531"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_4532_cast_fp16 = reshape(shape = var_4531, x = linear_65_cast_fp16)[name = string("op_4532_cast_fp16")];
            tensor<int32, [4]> x_131_perm_0 = const()[name = string("x_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45061376))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45323584))))[name = string("tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4543 = const()[name = string("op_4543"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_4544_cast_fp16 = reshape(shape = var_4543, x = linear_66_cast_fp16)[name = string("op_4544_cast_fp16")];
            tensor<int32, [4]> x_133_perm_0 = const()[name = string("x_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_4520_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 4, 48, 256]> var_4548_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_4548_cast_fp16")];
            bool var_4551 = const()[name = string("op_4551"), val = bool(true)];
            tensor<int32, [1]> var_4552_axes_0 = const()[name = string("op_4552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_4552_cast_fp16 = reduce_max(axes = var_4552_axes_0, keep_dims = var_4551, x = var_4548_cast_fp16)[name = string("op_4552_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_117_cast_fp16 = maximum(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_117_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_117_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4556_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<int32, [1]> var_4561_axes_0 = const()[name = string("op_4561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4561_keep_dims_0 = const()[name = string("op_4561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_4561_cast_fp16 = reduce_mean(axes = var_4561_axes_0, keep_dims = var_4561_keep_dims_0, x = var_4556_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_4562_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4562_cast_fp16")];
            fp32 var_4563_epsilon_0 = const()[name = string("op_4563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_4563_cast_fp16 = inverse(epsilon = var_4563_epsilon_0, x = var_4562_cast_fp16)[name = string("op_4563_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_4565_cast_fp16 = mul(x = var_4563_cast_fp16, y = var_4564_to_fp16)[name = string("op_4565_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_4567_cast_fp16 = add(x = var_4561_cast_fp16, y = var_4565_cast_fp16)[name = string("op_4567_cast_fp16")];
            fp32 var_4568_epsilon_0 = const()[name = string("op_4568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_4568_cast_fp16 = rsqrt(epsilon = var_4568_epsilon_0, x = var_4567_cast_fp16)[name = string("op_4568_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_103_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4568_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_29_cast_fp16 = mul(x = out_103_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_131_cast_fp16 = transpose(perm = x_131_perm_0, x = var_4532_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2, 48, 256]> var_4571_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_4571_cast_fp16")];
            bool var_4574 = const()[name = string("op_4574"), val = bool(true)];
            tensor<int32, [1]> var_4575_axes_0 = const()[name = string("op_4575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_4575_cast_fp16 = reduce_max(axes = var_4575_axes_0, keep_dims = var_4574, x = var_4571_cast_fp16)[name = string("op_4575_cast_fp16")];
            fp16 var_4576_to_fp16 = const()[name = string("op_4576_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_119_cast_fp16 = maximum(x = var_4575_cast_fp16, y = var_4576_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_119_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_119_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4579_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4579_cast_fp16")];
            tensor<int32, [1]> var_4584_axes_0 = const()[name = string("op_4584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4584_keep_dims_0 = const()[name = string("op_4584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_4584_cast_fp16 = reduce_mean(axes = var_4584_axes_0, keep_dims = var_4584_keep_dims_0, x = var_4579_cast_fp16)[name = string("op_4584_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4585_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4585_cast_fp16")];
            fp32 var_4586_epsilon_0 = const()[name = string("op_4586_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_4586_cast_fp16 = inverse(epsilon = var_4586_epsilon_0, x = var_4585_cast_fp16)[name = string("op_4586_cast_fp16")];
            fp16 var_4587_to_fp16 = const()[name = string("op_4587_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_4588_cast_fp16 = mul(x = var_4586_cast_fp16, y = var_4587_to_fp16)[name = string("op_4588_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4590_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4588_cast_fp16)[name = string("op_4590_cast_fp16")];
            fp32 var_4591_epsilon_0 = const()[name = string("op_4591_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_4591_cast_fp16 = rsqrt(epsilon = var_4591_epsilon_0, x = var_4590_cast_fp16)[name = string("op_4591_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_105_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4591_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [256]> tm_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45324672)))];
            tensor<fp16, [1, 2, 48, 256]> k_43_cast_fp16 = mul(x = out_105_cast_fp16, y = tm_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_133_cast_fp16 = transpose(perm = x_133_perm_0, x = var_4544_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 2, 48, 256]> var_4594_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_4594_cast_fp16")];
            bool var_4597 = const()[name = string("op_4597"), val = bool(true)];
            tensor<int32, [1]> var_4598_axes_0 = const()[name = string("op_4598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_4598_cast_fp16 = reduce_max(axes = var_4598_axes_0, keep_dims = var_4597, x = var_4594_cast_fp16)[name = string("op_4598_cast_fp16")];
            fp16 var_4599_to_fp16 = const()[name = string("op_4599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_121_cast_fp16 = maximum(x = var_4598_cast_fp16, y = var_4599_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_121_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_121_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4602_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4602_cast_fp16")];
            tensor<int32, [1]> var_4607_axes_0 = const()[name = string("op_4607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4607_keep_dims_0 = const()[name = string("op_4607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_4607_cast_fp16 = reduce_mean(axes = var_4607_axes_0, keep_dims = var_4607_keep_dims_0, x = var_4602_cast_fp16)[name = string("op_4607_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4608_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4608_cast_fp16")];
            fp32 var_4609_epsilon_0 = const()[name = string("op_4609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_4609_cast_fp16 = inverse(epsilon = var_4609_epsilon_0, x = var_4608_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_4611_cast_fp16 = mul(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("op_4611_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4613_cast_fp16 = add(x = var_4607_cast_fp16, y = var_4611_cast_fp16)[name = string("op_4613_cast_fp16")];
            fp32 var_4614_epsilon_0 = const()[name = string("op_4614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_4614_cast_fp16 = rsqrt(epsilon = var_4614_epsilon_0, x = var_4613_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_29_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4614_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4616_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4616_cast_fp16")];
            tensor<int32, [4]> var_4621_begin_0 = const()[name = string("op_4621_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4621_end_0 = const()[name = string("op_4621_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_4621_end_mask_0 = const()[name = string("op_4621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_4621_cast_fp16 = slice_by_index(begin = var_4621_begin_0, end = var_4621_end_0, end_mask = var_4621_end_mask_0, x = q_29_cast_fp16)[name = string("op_4621_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_4622_cast_fp16 = mul(x = var_4621_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<int32, [4]> var_4627_begin_0 = const()[name = string("op_4627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4627_end_0 = const()[name = string("op_4627_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_4627_end_mask_0 = const()[name = string("op_4627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_4627_cast_fp16 = slice_by_index(begin = var_4627_begin_0, end = var_4627_end_0, end_mask = var_4627_end_mask_0, x = q_29_cast_fp16)[name = string("op_4627_cast_fp16")];
            int32 var_4629 = const()[name = string("op_4629"), val = int32(-1)];
            bool var_4630_interleave_0 = const()[name = string("op_4630_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_4630_cast_fp16 = concat(axis = var_4629, interleave = var_4630_interleave_0, values = (var_4622_cast_fp16, var_4627_cast_fp16))[name = string("op_4630_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4631_cast_fp16 = mul(x = var_4630_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4631_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_31_cast_fp16 = add(x = var_4616_cast_fp16, y = var_4631_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4634_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4634_cast_fp16")];
            tensor<int32, [4]> var_4639_begin_0 = const()[name = string("op_4639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4639_end_0 = const()[name = string("op_4639_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_4639_end_mask_0 = const()[name = string("op_4639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_4639_cast_fp16 = slice_by_index(begin = var_4639_begin_0, end = var_4639_end_0, end_mask = var_4639_end_mask_0, x = k_43_cast_fp16)[name = string("op_4639_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_4640_cast_fp16 = mul(x = var_4639_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4640_cast_fp16")];
            tensor<int32, [4]> var_4645_begin_0 = const()[name = string("op_4645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4645_end_0 = const()[name = string("op_4645_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_4645_end_mask_0 = const()[name = string("op_4645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_4645_cast_fp16 = slice_by_index(begin = var_4645_begin_0, end = var_4645_end_0, end_mask = var_4645_end_mask_0, x = k_43_cast_fp16)[name = string("op_4645_cast_fp16")];
            int32 var_4647 = const()[name = string("op_4647"), val = int32(-1)];
            bool var_4648_interleave_0 = const()[name = string("op_4648_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_4648_cast_fp16 = concat(axis = var_4647, interleave = var_4648_interleave_0, values = (var_4640_cast_fp16, var_4645_cast_fp16))[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4649_cast_fp16 = mul(x = var_4648_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4649_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_45_cast_fp16 = add(x = var_4634_cast_fp16, y = var_4649_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_14_reps_0 = const()[name = string("tile_14_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = k_45_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [4, 1, 48, 256]> tile_14_cast_fp16 = tile(reps = tile_14_reps_0, x = transpose_28_cast_fp16)[name = string("tile_14_cast_fp16")];
            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_28_cast_fp16 = reshape(shape = concat_28, x = tile_14_cast_fp16)[name = string("reshape_28_cast_fp16")];
            tensor<int32, [5]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_28_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [4, 1, 48, 256]> reshape_29_cast_fp16 = reshape(shape = concat_29, x = transpose_29_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_15_reps_0 = const()[name = string("tile_15_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = v_29_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [4, 1, 48, 256]> tile_15_cast_fp16 = tile(reps = tile_15_reps_0, x = transpose_30_cast_fp16)[name = string("tile_15_cast_fp16")];
            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_30_cast_fp16 = reshape(shape = concat_30, x = tile_15_cast_fp16)[name = string("reshape_30_cast_fp16")];
            tensor<int32, [5]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_30_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [4, 1, 48, 256]> reshape_31_cast_fp16 = reshape(shape = concat_31, x = transpose_31_cast_fp16)[name = string("reshape_31_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4663_transpose_x_0 = const()[name = string("op_4663_transpose_x_0"), val = bool(false)];
            bool var_4663_transpose_y_0 = const()[name = string("op_4663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 4, 48, 48]> var_4663_cast_fp16 = matmul(transpose_x = var_4663_transpose_x_0, transpose_y = var_4663_transpose_y_0, x = q_31_cast_fp16, y = transpose_103_cast_fp16)[name = string("op_4663_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_4665_cast_fp16 = add(x = var_4663_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4666 = const()[name = string("op_4666"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_121_cast_fp16 = softmax(axis = var_4666, x = var_4665_cast_fp16)[name = string("w_121_cast_fp16")];
            bool var_4669_transpose_x_0 = const()[name = string("op_4669_transpose_x_0"), val = bool(false)];
            bool var_4669_transpose_y_0 = const()[name = string("op_4669_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = reshape_31_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 4, 48, 256]> var_4669_cast_fp16 = matmul(transpose_x = var_4669_transpose_x_0, transpose_y = var_4669_transpose_y_0, x = w_121_cast_fp16, y = v_31_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<int32, [4]> var_4672_perm_0 = const()[name = string("op_4672_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4676 = const()[name = string("op_4676"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_4672_cast_fp16 = transpose(perm = var_4672_perm_0, x = var_4669_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 48, 1024]> input_87_cast_fp16 = reshape(shape = var_4676, x = var_4672_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45325248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45849600))))[name = string("tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4681_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4684 = const()[name = string("op_4684"), val = bool(true)];
            tensor<int32, [1]> var_4685_axes_0 = const()[name = string("op_4685_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4685_cast_fp16 = reduce_max(axes = var_4685_axes_0, keep_dims = var_4684, x = var_4681_cast_fp16)[name = string("op_4685_cast_fp16")];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_123_cast_fp16 = maximum(x = var_4685_cast_fp16, y = var_4686_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_123_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_123_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4689_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [1]> var_4694_axes_0 = const()[name = string("op_4694_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4694_keep_dims_0 = const()[name = string("op_4694_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4694_cast_fp16 = reduce_mean(axes = var_4694_axes_0, keep_dims = var_4694_keep_dims_0, x = var_4689_cast_fp16)[name = string("op_4694_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4695_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4695_cast_fp16")];
            fp32 var_4696_epsilon_0 = const()[name = string("op_4696_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4696_cast_fp16 = inverse(epsilon = var_4696_epsilon_0, x = var_4695_cast_fp16)[name = string("op_4696_cast_fp16")];
            fp16 var_4697_to_fp16 = const()[name = string("op_4697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4698_cast_fp16 = mul(x = var_4696_cast_fp16, y = var_4697_to_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4700_cast_fp16 = add(x = var_4694_cast_fp16, y = var_4698_cast_fp16)[name = string("op_4700_cast_fp16")];
            fp32 var_4701_epsilon_0 = const()[name = string("op_4701_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4701_cast_fp16 = rsqrt(epsilon = var_4701_epsilon_0, x = var_4700_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_107_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4701_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45850688)))];
            tensor<fp16, [1, 48, 512]> var_4703_cast_fp16 = mul(x = out_107_cast_fp16, y = tm_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4703_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_137_cast_fp16 = add(x = x_127_cast_fp16, y = var_4703_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4706_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_4706_cast_fp16")];
            bool var_4709 = const()[name = string("op_4709"), val = bool(true)];
            tensor<int32, [1]> var_4710_axes_0 = const()[name = string("op_4710_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4710_cast_fp16 = reduce_max(axes = var_4710_axes_0, keep_dims = var_4709, x = var_4706_cast_fp16)[name = string("op_4710_cast_fp16")];
            fp16 var_4711_to_fp16 = const()[name = string("op_4711_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_125_cast_fp16 = maximum(x = var_4710_cast_fp16, y = var_4711_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_125_cast_fp16 = real_div(x = x_137_cast_fp16, y = s_125_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4714_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4714_cast_fp16")];
            tensor<int32, [1]> var_4719_axes_0 = const()[name = string("op_4719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4719_keep_dims_0 = const()[name = string("op_4719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4719_cast_fp16 = reduce_mean(axes = var_4719_axes_0, keep_dims = var_4719_keep_dims_0, x = var_4714_cast_fp16)[name = string("op_4719_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4720_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4720_cast_fp16")];
            fp32 var_4721_epsilon_0 = const()[name = string("op_4721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4721_cast_fp16 = inverse(epsilon = var_4721_epsilon_0, x = var_4720_cast_fp16)[name = string("op_4721_cast_fp16")];
            fp16 var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4723_cast_fp16 = mul(x = var_4721_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4725_cast_fp16 = add(x = var_4719_cast_fp16, y = var_4723_cast_fp16)[name = string("op_4725_cast_fp16")];
            fp32 var_4726_epsilon_0 = const()[name = string("op_4726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4726_cast_fp16 = rsqrt(epsilon = var_4726_epsilon_0, x = var_4725_cast_fp16)[name = string("op_4726_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_109_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4726_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45851776)))];
            tensor<fp16, [1, 48, 512]> input_89_cast_fp16 = mul(x = out_109_cast_fp16, y = tm_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45852864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46901504))))[name = string("tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4733_mode_0 = const()[name = string("op_4733_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_4733_cast_fp16 = gelu(mode = var_4733_mode_0, x = linear_68_cast_fp16)[name = string("op_4733_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46905664))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47954304))))[name = string("tm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_91_cast_fp16 = mul(x = var_4733_cast_fp16, y = linear_69_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47958464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49007104))))[name = string("tm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4741_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4741_cast_fp16")];
            bool var_4744 = const()[name = string("op_4744"), val = bool(true)];
            tensor<int32, [1]> var_4745_axes_0 = const()[name = string("op_4745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4745_cast_fp16 = reduce_max(axes = var_4745_axes_0, keep_dims = var_4744, x = var_4741_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_127_cast_fp16 = maximum(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_127_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_127_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4749_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<int32, [1]> var_4754_axes_0 = const()[name = string("op_4754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4754_keep_dims_0 = const()[name = string("op_4754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4754_cast_fp16 = reduce_mean(axes = var_4754_axes_0, keep_dims = var_4754_keep_dims_0, x = var_4749_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4755_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4755_cast_fp16")];
            fp32 var_4756_epsilon_0 = const()[name = string("op_4756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4756_cast_fp16 = inverse(epsilon = var_4756_epsilon_0, x = var_4755_cast_fp16)[name = string("op_4756_cast_fp16")];
            fp16 var_4757_to_fp16 = const()[name = string("op_4757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4758_cast_fp16 = mul(x = var_4756_cast_fp16, y = var_4757_to_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4760_cast_fp16 = add(x = var_4754_cast_fp16, y = var_4758_cast_fp16)[name = string("op_4760_cast_fp16")];
            fp32 var_4761_epsilon_0 = const()[name = string("op_4761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4761_cast_fp16 = rsqrt(epsilon = var_4761_epsilon_0, x = var_4760_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_111_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4761_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49008192)))];
            tensor<fp16, [1, 48, 512]> var_4763_cast_fp16 = mul(x = out_111_cast_fp16, y = tm_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4763_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_93_cast_fp16 = add(x = x_137_cast_fp16, y = var_4763_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49009280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49271488))))[name = string("tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_71_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_71_cast_fp16")];
            string x_141_mode_0 = const()[name = string("x_141_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_141_cast_fp16 = gelu(mode = x_141_mode_0, x = linear_71_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_4783_begin_0 = const()[name = string("op_4783_begin_0"), val = tensor<int32, [4]>([0, 0, 7, 0])];
            tensor<int32, [4]> var_4783_end_0 = const()[name = string("op_4783_end_0"), val = tensor<int32, [4]>([1, 48, 8, 512])];
            tensor<bool, [4]> var_4783_end_mask_0 = const()[name = string("op_4783_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4783_squeeze_mask_0 = const()[name = string("op_4783_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_4783_cast_fp16 = slice_by_index(begin = var_4783_begin_0, end = var_4783_end_0, end_mask = var_4783_end_mask_0, squeeze_mask = var_4783_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_95_cast_fp16 = mul(x = x_141_cast_fp16, y = var_4783_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49272576))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49534784))))[name = string("tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4793_cast_fp16 = abs(x = linear_72_cast_fp16)[name = string("op_4793_cast_fp16")];
            bool var_4796 = const()[name = string("op_4796"), val = bool(true)];
            tensor<int32, [1]> var_4797_axes_0 = const()[name = string("op_4797_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4797_cast_fp16 = reduce_max(axes = var_4797_axes_0, keep_dims = var_4796, x = var_4793_cast_fp16)[name = string("op_4797_cast_fp16")];
            fp16 var_4798_to_fp16 = const()[name = string("op_4798_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_129_cast_fp16 = maximum(x = var_4797_cast_fp16, y = var_4798_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_129_cast_fp16 = real_div(x = linear_72_cast_fp16, y = s_129_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4801_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4801_cast_fp16")];
            tensor<int32, [1]> var_4806_axes_0 = const()[name = string("op_4806_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4806_keep_dims_0 = const()[name = string("op_4806_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4806_cast_fp16 = reduce_mean(axes = var_4806_axes_0, keep_dims = var_4806_keep_dims_0, x = var_4801_cast_fp16)[name = string("op_4806_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4807_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4807_cast_fp16")];
            fp32 var_4808_epsilon_0 = const()[name = string("op_4808_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4808_cast_fp16 = inverse(epsilon = var_4808_epsilon_0, x = var_4807_cast_fp16)[name = string("op_4808_cast_fp16")];
            fp16 var_4809_to_fp16 = const()[name = string("op_4809_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4810_cast_fp16 = mul(x = var_4808_cast_fp16, y = var_4809_to_fp16)[name = string("op_4810_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4812_cast_fp16 = add(x = var_4806_cast_fp16, y = var_4810_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp32 var_4813_epsilon_0 = const()[name = string("op_4813_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4813_cast_fp16 = rsqrt(epsilon = var_4813_epsilon_0, x = var_4812_cast_fp16)[name = string("op_4813_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_113_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4813_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49535872)))];
            tensor<fp16, [1, 48, 512]> var_4815_cast_fp16 = mul(x = out_113_cast_fp16, y = tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4815_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_15_cast_fp16 = add(x = input_93_cast_fp16, y = var_4815_cast_fp16)[name = string("h_15_cast_fp16")];
            tensor<fp16, [1]> tm_layers_7_layer_scalar_to_fp16 = const()[name = string("tm_layers_7_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-1])];
            tensor<fp16, [1, 48, 512]> x_145_cast_fp16 = mul(x = h_15_cast_fp16, y = tm_layers_7_layer_scalar_to_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4819_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4819_cast_fp16")];
            bool var_4822 = const()[name = string("op_4822"), val = bool(true)];
            tensor<int32, [1]> var_4823_axes_0 = const()[name = string("op_4823_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_4823_cast_fp16 = reduce_max(axes = var_4823_axes_0, keep_dims = var_4822, x = var_4819_cast_fp16)[name = string("op_4823_cast_fp16")];
            fp16 var_4824_to_fp16 = const()[name = string("op_4824_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_131_cast_fp16 = maximum(x = var_4823_cast_fp16, y = var_4824_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_131_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_131_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_4827_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [1]> var_4832_axes_0 = const()[name = string("op_4832_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4832_keep_dims_0 = const()[name = string("op_4832_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_4832_cast_fp16 = reduce_mean(axes = var_4832_axes_0, keep_dims = var_4832_keep_dims_0, x = var_4827_cast_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4833_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4833_cast_fp16")];
            fp32 var_4834_epsilon_0 = const()[name = string("op_4834_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_4834_cast_fp16 = inverse(epsilon = var_4834_epsilon_0, x = var_4833_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp16 var_4835_to_fp16 = const()[name = string("op_4835_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_4836_cast_fp16 = mul(x = var_4834_cast_fp16, y = var_4835_to_fp16)[name = string("op_4836_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_4838_cast_fp16 = add(x = var_4832_cast_fp16, y = var_4836_cast_fp16)[name = string("op_4838_cast_fp16")];
            fp32 var_4839_epsilon_0 = const()[name = string("op_4839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_4839_cast_fp16 = rsqrt(epsilon = var_4839_epsilon_0, x = var_4838_cast_fp16)[name = string("op_4839_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_115_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4839_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49536960)))];
            tensor<fp16, [1, 48, 512]> input_97_cast_fp16 = mul(x = out_115_cast_fp16, y = tm_layers_8_input_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49538048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50062400))))[name = string("tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4849 = const()[name = string("op_4849"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_4850_cast_fp16 = reshape(shape = var_4849, x = linear_73_cast_fp16)[name = string("op_4850_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50064512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50326720))))[name = string("tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [4]> var_4861 = const()[name = string("op_4861"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_4862_cast_fp16 = reshape(shape = var_4861, x = linear_74_cast_fp16)[name = string("op_4862_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50327808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50590016))))[name = string("tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [4]> var_4873 = const()[name = string("op_4873"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_4874_cast_fp16 = reshape(shape = var_4873, x = linear_75_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<int32, [4]> x_151_perm_0 = const()[name = string("x_151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_4850_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 4, 48, 256]> var_4878_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4878_cast_fp16")];
            bool var_4881 = const()[name = string("op_4881"), val = bool(true)];
            tensor<int32, [1]> var_4882_axes_0 = const()[name = string("op_4882_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_4882_cast_fp16 = reduce_max(axes = var_4882_axes_0, keep_dims = var_4881, x = var_4878_cast_fp16)[name = string("op_4882_cast_fp16")];
            fp16 var_4883_to_fp16 = const()[name = string("op_4883_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_133_cast_fp16 = maximum(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_133_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_133_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4886_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4886_cast_fp16")];
            tensor<int32, [1]> var_4891_axes_0 = const()[name = string("op_4891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4891_keep_dims_0 = const()[name = string("op_4891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_4891_cast_fp16 = reduce_mean(axes = var_4891_axes_0, keep_dims = var_4891_keep_dims_0, x = var_4886_cast_fp16)[name = string("op_4891_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_4892_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4892_cast_fp16")];
            fp32 var_4893_epsilon_0 = const()[name = string("op_4893_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_4893_cast_fp16 = inverse(epsilon = var_4893_epsilon_0, x = var_4892_cast_fp16)[name = string("op_4893_cast_fp16")];
            fp16 var_4894_to_fp16 = const()[name = string("op_4894_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_4895_cast_fp16 = mul(x = var_4893_cast_fp16, y = var_4894_to_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_4897_cast_fp16 = add(x = var_4891_cast_fp16, y = var_4895_cast_fp16)[name = string("op_4897_cast_fp16")];
            fp32 var_4898_epsilon_0 = const()[name = string("op_4898_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_4898_cast_fp16 = rsqrt(epsilon = var_4898_epsilon_0, x = var_4897_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_117_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4898_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591104)))];
            tensor<fp16, [1, 4, 48, 256]> q_33_cast_fp16 = mul(x = out_117_cast_fp16, y = tm_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_4862_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 2, 48, 256]> var_4901_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4901_cast_fp16")];
            bool var_4904 = const()[name = string("op_4904"), val = bool(true)];
            tensor<int32, [1]> var_4905_axes_0 = const()[name = string("op_4905_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_4905_cast_fp16 = reduce_max(axes = var_4905_axes_0, keep_dims = var_4904, x = var_4901_cast_fp16)[name = string("op_4905_cast_fp16")];
            fp16 var_4906_to_fp16 = const()[name = string("op_4906_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_135_cast_fp16 = maximum(x = var_4905_cast_fp16, y = var_4906_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_135_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_135_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4909_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [1]> var_4914_axes_0 = const()[name = string("op_4914_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4914_keep_dims_0 = const()[name = string("op_4914_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_4914_cast_fp16 = reduce_mean(axes = var_4914_axes_0, keep_dims = var_4914_keep_dims_0, x = var_4909_cast_fp16)[name = string("op_4914_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4915_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4915_cast_fp16")];
            fp32 var_4916_epsilon_0 = const()[name = string("op_4916_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_4916_cast_fp16 = inverse(epsilon = var_4916_epsilon_0, x = var_4915_cast_fp16)[name = string("op_4916_cast_fp16")];
            fp16 var_4917_to_fp16 = const()[name = string("op_4917_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_4918_cast_fp16 = mul(x = var_4916_cast_fp16, y = var_4917_to_fp16)[name = string("op_4918_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4920_cast_fp16 = add(x = var_4914_cast_fp16, y = var_4918_cast_fp16)[name = string("op_4920_cast_fp16")];
            fp32 var_4921_epsilon_0 = const()[name = string("op_4921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_4921_cast_fp16 = rsqrt(epsilon = var_4921_epsilon_0, x = var_4920_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_119_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4921_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591680)))];
            tensor<fp16, [1, 2, 48, 256]> k_49_cast_fp16 = mul(x = out_119_cast_fp16, y = tm_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_151_cast_fp16 = transpose(perm = x_151_perm_0, x = var_4874_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 2, 48, 256]> var_4924_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4924_cast_fp16")];
            bool var_4927 = const()[name = string("op_4927"), val = bool(true)];
            tensor<int32, [1]> var_4928_axes_0 = const()[name = string("op_4928_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_4928_cast_fp16 = reduce_max(axes = var_4928_axes_0, keep_dims = var_4927, x = var_4924_cast_fp16)[name = string("op_4928_cast_fp16")];
            fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_137_cast_fp16 = maximum(x = var_4928_cast_fp16, y = var_4929_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_137_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_137_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4932_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4932_cast_fp16")];
            tensor<int32, [1]> var_4937_axes_0 = const()[name = string("op_4937_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4937_keep_dims_0 = const()[name = string("op_4937_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_4937_cast_fp16 = reduce_mean(axes = var_4937_axes_0, keep_dims = var_4937_keep_dims_0, x = var_4932_cast_fp16)[name = string("op_4937_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4938_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4938_cast_fp16")];
            fp32 var_4939_epsilon_0 = const()[name = string("op_4939_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_4939_cast_fp16 = inverse(epsilon = var_4939_epsilon_0, x = var_4938_cast_fp16)[name = string("op_4939_cast_fp16")];
            fp16 var_4940_to_fp16 = const()[name = string("op_4940_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_4941_cast_fp16 = mul(x = var_4939_cast_fp16, y = var_4940_to_fp16)[name = string("op_4941_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_4943_cast_fp16 = add(x = var_4937_cast_fp16, y = var_4941_cast_fp16)[name = string("op_4943_cast_fp16")];
            fp32 var_4944_epsilon_0 = const()[name = string("op_4944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_4944_cast_fp16 = rsqrt(epsilon = var_4944_epsilon_0, x = var_4943_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_33_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4944_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4946_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = q_33_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_4952_cast_fp16 = mul(x = var_4951_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4957_begin_0 = const()[name = string("op_4957_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4957_end_0 = const()[name = string("op_4957_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_4957_end_mask_0 = const()[name = string("op_4957_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_4957_cast_fp16 = slice_by_index(begin = var_4957_begin_0, end = var_4957_end_0, end_mask = var_4957_end_mask_0, x = q_33_cast_fp16)[name = string("op_4957_cast_fp16")];
            int32 var_4959 = const()[name = string("op_4959"), val = int32(-1)];
            bool var_4960_interleave_0 = const()[name = string("op_4960_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_4960_cast_fp16 = concat(axis = var_4959, interleave = var_4960_interleave_0, values = (var_4952_cast_fp16, var_4957_cast_fp16))[name = string("op_4960_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_4961_cast_fp16 = mul(x = var_4960_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4961_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_35_cast_fp16 = add(x = var_4946_cast_fp16, y = var_4961_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4964_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4964_cast_fp16")];
            tensor<int32, [4]> var_4969_begin_0 = const()[name = string("op_4969_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4969_end_0 = const()[name = string("op_4969_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_4969_end_mask_0 = const()[name = string("op_4969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_4969_cast_fp16 = slice_by_index(begin = var_4969_begin_0, end = var_4969_end_0, end_mask = var_4969_end_mask_0, x = k_49_cast_fp16)[name = string("op_4969_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_4970_cast_fp16 = mul(x = var_4969_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_4970_cast_fp16")];
            tensor<int32, [4]> var_4975_begin_0 = const()[name = string("op_4975_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4975_end_0 = const()[name = string("op_4975_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_4975_end_mask_0 = const()[name = string("op_4975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_4975_cast_fp16 = slice_by_index(begin = var_4975_begin_0, end = var_4975_end_0, end_mask = var_4975_end_mask_0, x = k_49_cast_fp16)[name = string("op_4975_cast_fp16")];
            int32 var_4977 = const()[name = string("op_4977"), val = int32(-1)];
            bool var_4978_interleave_0 = const()[name = string("op_4978_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_4978_cast_fp16 = concat(axis = var_4977, interleave = var_4978_interleave_0, values = (var_4970_cast_fp16, var_4975_cast_fp16))[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_4979_cast_fp16 = mul(x = var_4978_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_51_cast_fp16 = add(x = var_4964_cast_fp16, y = var_4979_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_16_reps_0 = const()[name = string("tile_16_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = k_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [4, 1, 48, 256]> tile_16_cast_fp16 = tile(reps = tile_16_reps_0, x = transpose_32_cast_fp16)[name = string("tile_16_cast_fp16")];
            tensor<int32, [5]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_32_cast_fp16 = reshape(shape = concat_32, x = tile_16_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [5]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [4, 1, 48, 256]> reshape_33_cast_fp16 = reshape(shape = concat_33, x = transpose_33_cast_fp16)[name = string("reshape_33_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_17_reps_0 = const()[name = string("tile_17_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = v_33_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [4, 1, 48, 256]> tile_17_cast_fp16 = tile(reps = tile_17_reps_0, x = transpose_34_cast_fp16)[name = string("tile_17_cast_fp16")];
            tensor<int32, [5]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_34_cast_fp16 = reshape(shape = concat_34, x = tile_17_cast_fp16)[name = string("reshape_34_cast_fp16")];
            tensor<int32, [5]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = reshape_34_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [4, 1, 48, 256]> reshape_35_cast_fp16 = reshape(shape = concat_35, x = transpose_35_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = reshape_33_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 4, 48, 48]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = q_35_cast_fp16, y = transpose_104_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_4995_cast_fp16 = add(x = var_4993_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4995_cast_fp16")];
            int32 var_4996 = const()[name = string("op_4996"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_137_cast_fp16 = softmax(axis = var_4996, x = var_4995_cast_fp16)[name = string("w_137_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 4, 48, 256]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = w_137_cast_fp16, y = v_35_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5002_perm_0 = const()[name = string("op_5002_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5006 = const()[name = string("op_5006"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_5002_cast_fp16 = transpose(perm = var_5002_perm_0, x = var_4999_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 48, 1024]> input_99_cast_fp16 = reshape(shape = var_5006, x = var_5002_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50592256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51116608))))[name = string("tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5011_cast_fp16 = abs(x = linear_76_cast_fp16)[name = string("op_5011_cast_fp16")];
            bool var_5014 = const()[name = string("op_5014"), val = bool(true)];
            tensor<int32, [1]> var_5015_axes_0 = const()[name = string("op_5015_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5015_cast_fp16 = reduce_max(axes = var_5015_axes_0, keep_dims = var_5014, x = var_5011_cast_fp16)[name = string("op_5015_cast_fp16")];
            fp16 var_5016_to_fp16 = const()[name = string("op_5016_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_139_cast_fp16 = maximum(x = var_5015_cast_fp16, y = var_5016_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_139_cast_fp16 = real_div(x = linear_76_cast_fp16, y = s_139_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5019_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_5019_cast_fp16")];
            tensor<int32, [1]> var_5024_axes_0 = const()[name = string("op_5024_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5024_keep_dims_0 = const()[name = string("op_5024_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5024_cast_fp16 = reduce_mean(axes = var_5024_axes_0, keep_dims = var_5024_keep_dims_0, x = var_5019_cast_fp16)[name = string("op_5024_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5025_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_5025_cast_fp16")];
            fp32 var_5026_epsilon_0 = const()[name = string("op_5026_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5026_cast_fp16 = inverse(epsilon = var_5026_epsilon_0, x = var_5025_cast_fp16)[name = string("op_5026_cast_fp16")];
            fp16 var_5027_to_fp16 = const()[name = string("op_5027_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5028_cast_fp16 = mul(x = var_5026_cast_fp16, y = var_5027_to_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5030_cast_fp16 = add(x = var_5024_cast_fp16, y = var_5028_cast_fp16)[name = string("op_5030_cast_fp16")];
            fp32 var_5031_epsilon_0 = const()[name = string("op_5031_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5031_cast_fp16 = rsqrt(epsilon = var_5031_epsilon_0, x = var_5030_cast_fp16)[name = string("op_5031_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_121_cast_fp16 = mul(x = y_139_cast_fp16, y = var_5031_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51117696)))];
            tensor<fp16, [1, 48, 512]> var_5033_cast_fp16 = mul(x = out_121_cast_fp16, y = tm_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_5033_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = var_5033_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5036_cast_fp16 = abs(x = x_155_cast_fp16)[name = string("op_5036_cast_fp16")];
            bool var_5039 = const()[name = string("op_5039"), val = bool(true)];
            tensor<int32, [1]> var_5040_axes_0 = const()[name = string("op_5040_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5040_cast_fp16 = reduce_max(axes = var_5040_axes_0, keep_dims = var_5039, x = var_5036_cast_fp16)[name = string("op_5040_cast_fp16")];
            fp16 var_5041_to_fp16 = const()[name = string("op_5041_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_141_cast_fp16 = maximum(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_141_cast_fp16 = real_div(x = x_155_cast_fp16, y = s_141_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5044_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<int32, [1]> var_5049_axes_0 = const()[name = string("op_5049_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5049_keep_dims_0 = const()[name = string("op_5049_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5049_cast_fp16 = reduce_mean(axes = var_5049_axes_0, keep_dims = var_5049_keep_dims_0, x = var_5044_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5050_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_5050_cast_fp16")];
            fp32 var_5051_epsilon_0 = const()[name = string("op_5051_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5051_cast_fp16 = inverse(epsilon = var_5051_epsilon_0, x = var_5050_cast_fp16)[name = string("op_5051_cast_fp16")];
            fp16 var_5052_to_fp16 = const()[name = string("op_5052_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5053_cast_fp16 = mul(x = var_5051_cast_fp16, y = var_5052_to_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5055_cast_fp16 = add(x = var_5049_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            fp32 var_5056_epsilon_0 = const()[name = string("op_5056_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5056_cast_fp16 = rsqrt(epsilon = var_5056_epsilon_0, x = var_5055_cast_fp16)[name = string("op_5056_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_123_cast_fp16 = mul(x = y_141_cast_fp16, y = var_5056_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51118784)))];
            tensor<fp16, [1, 48, 512]> input_101_cast_fp16 = mul(x = out_123_cast_fp16, y = tm_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51119872))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52168512))))[name = string("tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_77_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_77_cast_fp16")];
            string var_5063_mode_0 = const()[name = string("op_5063_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_5063_cast_fp16 = gelu(mode = var_5063_mode_0, x = linear_77_cast_fp16)[name = string("op_5063_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52172672))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53221312))))[name = string("tm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_78_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_103_cast_fp16 = mul(x = var_5063_cast_fp16, y = linear_78_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53225472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54274112))))[name = string("tm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5071_cast_fp16 = abs(x = linear_79_cast_fp16)[name = string("op_5071_cast_fp16")];
            bool var_5074 = const()[name = string("op_5074"), val = bool(true)];
            tensor<int32, [1]> var_5075_axes_0 = const()[name = string("op_5075_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5075_cast_fp16 = reduce_max(axes = var_5075_axes_0, keep_dims = var_5074, x = var_5071_cast_fp16)[name = string("op_5075_cast_fp16")];
            fp16 var_5076_to_fp16 = const()[name = string("op_5076_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_143_cast_fp16 = maximum(x = var_5075_cast_fp16, y = var_5076_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_143_cast_fp16 = real_div(x = linear_79_cast_fp16, y = s_143_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5079_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_5079_cast_fp16")];
            tensor<int32, [1]> var_5084_axes_0 = const()[name = string("op_5084_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5084_keep_dims_0 = const()[name = string("op_5084_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5084_cast_fp16 = reduce_mean(axes = var_5084_axes_0, keep_dims = var_5084_keep_dims_0, x = var_5079_cast_fp16)[name = string("op_5084_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5085_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5086_cast_fp16 = inverse(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            fp16 var_5087_to_fp16 = const()[name = string("op_5087_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5088_cast_fp16 = mul(x = var_5086_cast_fp16, y = var_5087_to_fp16)[name = string("op_5088_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5090_cast_fp16 = add(x = var_5084_cast_fp16, y = var_5088_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp32 var_5091_epsilon_0 = const()[name = string("op_5091_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5091_cast_fp16 = rsqrt(epsilon = var_5091_epsilon_0, x = var_5090_cast_fp16)[name = string("op_5091_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_125_cast_fp16 = mul(x = y_143_cast_fp16, y = var_5091_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54275200)))];
            tensor<fp16, [1, 48, 512]> var_5093_cast_fp16 = mul(x = out_125_cast_fp16, y = tm_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5093_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_105_cast_fp16 = add(x = x_155_cast_fp16, y = var_5093_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54276288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54538496))))[name = string("tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_80_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_80_cast_fp16")];
            string x_159_mode_0 = const()[name = string("x_159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_159_cast_fp16 = gelu(mode = x_159_mode_0, x = linear_80_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [4]> var_5113_begin_0 = const()[name = string("op_5113_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_5113_end_0 = const()[name = string("op_5113_end_0"), val = tensor<int32, [4]>([1, 48, 9, 512])];
            tensor<bool, [4]> var_5113_end_mask_0 = const()[name = string("op_5113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5113_squeeze_mask_0 = const()[name = string("op_5113_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_5113_cast_fp16 = slice_by_index(begin = var_5113_begin_0, end = var_5113_end_0, end_mask = var_5113_end_mask_0, squeeze_mask = var_5113_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_107_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5113_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54539584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54801792))))[name = string("tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5123_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5123_cast_fp16")];
            bool var_5126 = const()[name = string("op_5126"), val = bool(true)];
            tensor<int32, [1]> var_5127_axes_0 = const()[name = string("op_5127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5127_cast_fp16 = reduce_max(axes = var_5127_axes_0, keep_dims = var_5126, x = var_5123_cast_fp16)[name = string("op_5127_cast_fp16")];
            fp16 var_5128_to_fp16 = const()[name = string("op_5128_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_145_cast_fp16 = maximum(x = var_5127_cast_fp16, y = var_5128_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_145_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_145_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5131_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5136_axes_0 = const()[name = string("op_5136_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5136_keep_dims_0 = const()[name = string("op_5136_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5136_cast_fp16 = reduce_mean(axes = var_5136_axes_0, keep_dims = var_5136_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5136_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5137_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5138_cast_fp16 = inverse(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp16 var_5139_to_fp16 = const()[name = string("op_5139_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5140_cast_fp16 = mul(x = var_5138_cast_fp16, y = var_5139_to_fp16)[name = string("op_5140_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5142_cast_fp16 = add(x = var_5136_cast_fp16, y = var_5140_cast_fp16)[name = string("op_5142_cast_fp16")];
            fp32 var_5143_epsilon_0 = const()[name = string("op_5143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5143_cast_fp16 = rsqrt(epsilon = var_5143_epsilon_0, x = var_5142_cast_fp16)[name = string("op_5143_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_127_cast_fp16 = mul(x = y_145_cast_fp16, y = var_5143_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54802880)))];
            tensor<fp16, [1, 48, 512]> var_5145_cast_fp16 = mul(x = out_127_cast_fp16, y = tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5145_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_17_cast_fp16 = add(x = input_105_cast_fp16, y = var_5145_cast_fp16)[name = string("h_17_cast_fp16")];
            tensor<fp16, [1]> tm_layers_8_layer_scalar_to_fp16 = const()[name = string("tm_layers_8_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bcp-1])];
            tensor<fp16, [1, 48, 512]> x_163_cast_fp16 = mul(x = h_17_cast_fp16, y = tm_layers_8_layer_scalar_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5149_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_5149_cast_fp16")];
            bool var_5152 = const()[name = string("op_5152"), val = bool(true)];
            tensor<int32, [1]> var_5153_axes_0 = const()[name = string("op_5153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5153_cast_fp16 = reduce_max(axes = var_5153_axes_0, keep_dims = var_5152, x = var_5149_cast_fp16)[name = string("op_5153_cast_fp16")];
            fp16 var_5154_to_fp16 = const()[name = string("op_5154_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_147_cast_fp16 = maximum(x = var_5153_cast_fp16, y = var_5154_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_147_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_147_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5157_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_5157_cast_fp16")];
            tensor<int32, [1]> var_5162_axes_0 = const()[name = string("op_5162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5162_keep_dims_0 = const()[name = string("op_5162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5162_cast_fp16 = reduce_mean(axes = var_5162_axes_0, keep_dims = var_5162_keep_dims_0, x = var_5157_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5163_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_5163_cast_fp16")];
            fp32 var_5164_epsilon_0 = const()[name = string("op_5164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5164_cast_fp16 = inverse(epsilon = var_5164_epsilon_0, x = var_5163_cast_fp16)[name = string("op_5164_cast_fp16")];
            fp16 var_5165_to_fp16 = const()[name = string("op_5165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5166_cast_fp16 = mul(x = var_5164_cast_fp16, y = var_5165_to_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5168_cast_fp16 = add(x = var_5162_cast_fp16, y = var_5166_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp32 var_5169_epsilon_0 = const()[name = string("op_5169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5169_cast_fp16 = rsqrt(epsilon = var_5169_epsilon_0, x = var_5168_cast_fp16)[name = string("op_5169_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_129_cast_fp16 = mul(x = y_147_cast_fp16, y = var_5169_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54803968)))];
            tensor<fp16, [1, 48, 512]> input_109_cast_fp16 = mul(x = out_129_cast_fp16, y = tm_layers_9_input_layernorm_weight_to_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54805056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55329408))))[name = string("tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [4]> var_5179 = const()[name = string("op_5179"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_5180_cast_fp16 = reshape(shape = var_5179, x = linear_82_cast_fp16)[name = string("op_5180_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55331520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55593728))))[name = string("tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [4]> var_5191 = const()[name = string("op_5191"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_5192_cast_fp16 = reshape(shape = var_5191, x = linear_83_cast_fp16)[name = string("op_5192_cast_fp16")];
            tensor<int32, [4]> x_167_perm_0 = const()[name = string("x_167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55594816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55857024))))[name = string("tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5203 = const()[name = string("op_5203"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_5204_cast_fp16 = reshape(shape = var_5203, x = linear_84_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_5180_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 4, 48, 256]> var_5208_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_5208_cast_fp16")];
            bool var_5211 = const()[name = string("op_5211"), val = bool(true)];
            tensor<int32, [1]> var_5212_axes_0 = const()[name = string("op_5212_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_5212_cast_fp16 = reduce_max(axes = var_5212_axes_0, keep_dims = var_5211, x = var_5208_cast_fp16)[name = string("op_5212_cast_fp16")];
            fp16 var_5213_to_fp16 = const()[name = string("op_5213_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_149_cast_fp16 = maximum(x = var_5212_cast_fp16, y = var_5213_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_149_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_149_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_5216_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<int32, [1]> var_5221_axes_0 = const()[name = string("op_5221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5221_keep_dims_0 = const()[name = string("op_5221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_5221_cast_fp16 = reduce_mean(axes = var_5221_axes_0, keep_dims = var_5221_keep_dims_0, x = var_5216_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_5222_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_5223_cast_fp16 = inverse(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            fp16 var_5224_to_fp16 = const()[name = string("op_5224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_5225_cast_fp16 = mul(x = var_5223_cast_fp16, y = var_5224_to_fp16)[name = string("op_5225_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_5227_cast_fp16 = add(x = var_5221_cast_fp16, y = var_5225_cast_fp16)[name = string("op_5227_cast_fp16")];
            fp32 var_5228_epsilon_0 = const()[name = string("op_5228_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_5228_cast_fp16 = rsqrt(epsilon = var_5228_epsilon_0, x = var_5227_cast_fp16)[name = string("op_5228_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_131_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5228_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = var_5192_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 2, 48, 256]> var_5231_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_5231_cast_fp16")];
            bool var_5234 = const()[name = string("op_5234"), val = bool(true)];
            tensor<int32, [1]> var_5235_axes_0 = const()[name = string("op_5235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_5235_cast_fp16 = reduce_max(axes = var_5235_axes_0, keep_dims = var_5234, x = var_5231_cast_fp16)[name = string("op_5235_cast_fp16")];
            fp16 var_5236_to_fp16 = const()[name = string("op_5236_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_151_cast_fp16 = maximum(x = var_5235_cast_fp16, y = var_5236_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_151_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_151_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_5239_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5239_cast_fp16")];
            tensor<int32, [1]> var_5244_axes_0 = const()[name = string("op_5244_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5244_keep_dims_0 = const()[name = string("op_5244_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_5244_cast_fp16 = reduce_mean(axes = var_5244_axes_0, keep_dims = var_5244_keep_dims_0, x = var_5239_cast_fp16)[name = string("op_5244_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_5245_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_5245_cast_fp16")];
            fp32 var_5246_epsilon_0 = const()[name = string("op_5246_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_5246_cast_fp16 = inverse(epsilon = var_5246_epsilon_0, x = var_5245_cast_fp16)[name = string("op_5246_cast_fp16")];
            fp16 var_5247_to_fp16 = const()[name = string("op_5247_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_5248_cast_fp16 = mul(x = var_5246_cast_fp16, y = var_5247_to_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_5250_cast_fp16 = add(x = var_5244_cast_fp16, y = var_5248_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp32 var_5251_epsilon_0 = const()[name = string("op_5251_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_5251_cast_fp16 = rsqrt(epsilon = var_5251_epsilon_0, x = var_5250_cast_fp16)[name = string("op_5251_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_133_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5251_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [256]> tm_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858112)))];
            tensor<fp16, [1, 2, 48, 256]> k_55_cast_fp16 = mul(x = out_133_cast_fp16, y = tm_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_5204_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2, 48, 256]> var_5254_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_5254_cast_fp16")];
            bool var_5257 = const()[name = string("op_5257"), val = bool(true)];
            tensor<int32, [1]> var_5258_axes_0 = const()[name = string("op_5258_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_5258_cast_fp16 = reduce_max(axes = var_5258_axes_0, keep_dims = var_5257, x = var_5254_cast_fp16)[name = string("op_5258_cast_fp16")];
            fp16 var_5259_to_fp16 = const()[name = string("op_5259_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_153_cast_fp16 = maximum(x = var_5258_cast_fp16, y = var_5259_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_153_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_153_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_5262_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5262_cast_fp16")];
            tensor<int32, [1]> var_5267_axes_0 = const()[name = string("op_5267_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5267_keep_dims_0 = const()[name = string("op_5267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_5267_cast_fp16 = reduce_mean(axes = var_5267_axes_0, keep_dims = var_5267_keep_dims_0, x = var_5262_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_5268_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5268_cast_fp16")];
            fp32 var_5269_epsilon_0 = const()[name = string("op_5269_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_5269_cast_fp16 = inverse(epsilon = var_5269_epsilon_0, x = var_5268_cast_fp16)[name = string("op_5269_cast_fp16")];
            fp16 var_5270_to_fp16 = const()[name = string("op_5270_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_5271_cast_fp16 = mul(x = var_5269_cast_fp16, y = var_5270_to_fp16)[name = string("op_5271_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_5273_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5271_cast_fp16)[name = string("op_5273_cast_fp16")];
            fp32 var_5274_epsilon_0 = const()[name = string("op_5274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_5274_cast_fp16 = rsqrt(epsilon = var_5274_epsilon_0, x = var_5273_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_37_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5274_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_5276_cast_fp16 = mul(x = out_131_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5276_cast_fp16")];
            tensor<int32, [4]> var_5281_begin_0 = const()[name = string("op_5281_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5281_end_0 = const()[name = string("op_5281_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_5281_end_mask_0 = const()[name = string("op_5281_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_5281_cast_fp16 = slice_by_index(begin = var_5281_begin_0, end = var_5281_end_0, end_mask = var_5281_end_mask_0, x = out_131_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_5282_cast_fp16 = mul(x = var_5281_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_5282_cast_fp16")];
            tensor<int32, [4]> var_5287_begin_0 = const()[name = string("op_5287_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5287_end_0 = const()[name = string("op_5287_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_5287_end_mask_0 = const()[name = string("op_5287_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_5287_cast_fp16 = slice_by_index(begin = var_5287_begin_0, end = var_5287_end_0, end_mask = var_5287_end_mask_0, x = out_131_cast_fp16)[name = string("op_5287_cast_fp16")];
            int32 var_5289 = const()[name = string("op_5289"), val = int32(-1)];
            bool var_5290_interleave_0 = const()[name = string("op_5290_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_5290_cast_fp16 = concat(axis = var_5289, interleave = var_5290_interleave_0, values = (var_5282_cast_fp16, var_5287_cast_fp16))[name = string("op_5290_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_5291_cast_fp16 = mul(x = var_5290_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5291_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_39_cast_fp16 = add(x = var_5276_cast_fp16, y = var_5291_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_5294_cast_fp16 = mul(x = k_55_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5294_cast_fp16")];
            tensor<int32, [4]> var_5299_begin_0 = const()[name = string("op_5299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5299_end_0 = const()[name = string("op_5299_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_5299_end_mask_0 = const()[name = string("op_5299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_5299_cast_fp16 = slice_by_index(begin = var_5299_begin_0, end = var_5299_end_0, end_mask = var_5299_end_mask_0, x = k_55_cast_fp16)[name = string("op_5299_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_5300_cast_fp16 = mul(x = var_5299_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_5300_cast_fp16")];
            tensor<int32, [4]> var_5305_begin_0 = const()[name = string("op_5305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5305_end_0 = const()[name = string("op_5305_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_5305_end_mask_0 = const()[name = string("op_5305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_5305_cast_fp16 = slice_by_index(begin = var_5305_begin_0, end = var_5305_end_0, end_mask = var_5305_end_mask_0, x = k_55_cast_fp16)[name = string("op_5305_cast_fp16")];
            int32 var_5307 = const()[name = string("op_5307"), val = int32(-1)];
            bool var_5308_interleave_0 = const()[name = string("op_5308_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_5308_cast_fp16 = concat(axis = var_5307, interleave = var_5308_interleave_0, values = (var_5300_cast_fp16, var_5305_cast_fp16))[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_5309_cast_fp16 = mul(x = var_5308_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_57_cast_fp16 = add(x = var_5294_cast_fp16, y = var_5309_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_18_reps_0 = const()[name = string("tile_18_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = k_57_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [4, 1, 48, 256]> tile_18_cast_fp16 = tile(reps = tile_18_reps_0, x = transpose_36_cast_fp16)[name = string("tile_18_cast_fp16")];
            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_36_cast_fp16 = reshape(shape = concat_36, x = tile_18_cast_fp16)[name = string("reshape_36_cast_fp16")];
            tensor<int32, [5]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = reshape_36_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [4, 1, 48, 256]> reshape_37_cast_fp16 = reshape(shape = concat_37, x = transpose_37_cast_fp16)[name = string("reshape_37_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_19_reps_0 = const()[name = string("tile_19_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = v_37_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [4, 1, 48, 256]> tile_19_cast_fp16 = tile(reps = tile_19_reps_0, x = transpose_38_cast_fp16)[name = string("tile_19_cast_fp16")];
            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_38_cast_fp16 = reshape(shape = concat_38, x = tile_19_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [5]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [4, 1, 48, 256]> reshape_39_cast_fp16 = reshape(shape = concat_39, x = transpose_39_cast_fp16)[name = string("reshape_39_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5323_transpose_x_0 = const()[name = string("op_5323_transpose_x_0"), val = bool(false)];
            bool var_5323_transpose_y_0 = const()[name = string("op_5323_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = reshape_37_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 4, 48, 48]> var_5323_cast_fp16 = matmul(transpose_x = var_5323_transpose_x_0, transpose_y = var_5323_transpose_y_0, x = q_39_cast_fp16, y = transpose_105_cast_fp16)[name = string("op_5323_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_5325_cast_fp16 = add(x = var_5323_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5325_cast_fp16")];
            int32 var_5326 = const()[name = string("op_5326"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_153_cast_fp16 = softmax(axis = var_5326, x = var_5325_cast_fp16)[name = string("w_153_cast_fp16")];
            bool var_5329_transpose_x_0 = const()[name = string("op_5329_transpose_x_0"), val = bool(false)];
            bool var_5329_transpose_y_0 = const()[name = string("op_5329_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = reshape_39_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 4, 48, 256]> var_5329_cast_fp16 = matmul(transpose_x = var_5329_transpose_x_0, transpose_y = var_5329_transpose_y_0, x = w_153_cast_fp16, y = v_39_cast_fp16)[name = string("op_5329_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = var_5329_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 48, 1024]> input_111_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56383040))))[name = string("tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5341_cast_fp16 = abs(x = linear_85_cast_fp16)[name = string("op_5341_cast_fp16")];
            bool var_5344 = const()[name = string("op_5344"), val = bool(true)];
            tensor<int32, [1]> var_5345_axes_0 = const()[name = string("op_5345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5345_cast_fp16 = reduce_max(axes = var_5345_axes_0, keep_dims = var_5344, x = var_5341_cast_fp16)[name = string("op_5345_cast_fp16")];
            fp16 var_5346_to_fp16 = const()[name = string("op_5346_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_155_cast_fp16 = maximum(x = var_5345_cast_fp16, y = var_5346_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_155_cast_fp16 = real_div(x = linear_85_cast_fp16, y = s_155_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5349_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5354_axes_0 = const()[name = string("op_5354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5354_keep_dims_0 = const()[name = string("op_5354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5354_cast_fp16 = reduce_mean(axes = var_5354_axes_0, keep_dims = var_5354_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5355_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5355_cast_fp16")];
            fp32 var_5356_epsilon_0 = const()[name = string("op_5356_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5356_cast_fp16 = inverse(epsilon = var_5356_epsilon_0, x = var_5355_cast_fp16)[name = string("op_5356_cast_fp16")];
            fp16 var_5357_to_fp16 = const()[name = string("op_5357_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5358_cast_fp16 = mul(x = var_5356_cast_fp16, y = var_5357_to_fp16)[name = string("op_5358_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5360_cast_fp16 = add(x = var_5354_cast_fp16, y = var_5358_cast_fp16)[name = string("op_5360_cast_fp16")];
            fp32 var_5361_epsilon_0 = const()[name = string("op_5361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5361_cast_fp16 = rsqrt(epsilon = var_5361_epsilon_0, x = var_5360_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_135_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5361_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56384128)))];
            tensor<fp16, [1, 48, 512]> var_5363_cast_fp16 = mul(x = out_135_cast_fp16, y = tm_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_5363_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_173_cast_fp16 = add(x = x_163_cast_fp16, y = var_5363_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5366_cast_fp16 = abs(x = x_173_cast_fp16)[name = string("op_5366_cast_fp16")];
            bool var_5369 = const()[name = string("op_5369"), val = bool(true)];
            tensor<int32, [1]> var_5370_axes_0 = const()[name = string("op_5370_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5370_cast_fp16 = reduce_max(axes = var_5370_axes_0, keep_dims = var_5369, x = var_5366_cast_fp16)[name = string("op_5370_cast_fp16")];
            fp16 var_5371_to_fp16 = const()[name = string("op_5371_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_157_cast_fp16 = maximum(x = var_5370_cast_fp16, y = var_5371_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_157_cast_fp16 = real_div(x = x_173_cast_fp16, y = s_157_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5374_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [1]> var_5379_axes_0 = const()[name = string("op_5379_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5379_keep_dims_0 = const()[name = string("op_5379_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5379_cast_fp16 = reduce_mean(axes = var_5379_axes_0, keep_dims = var_5379_keep_dims_0, x = var_5374_cast_fp16)[name = string("op_5379_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5380_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5381_cast_fp16 = inverse(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            fp16 var_5382_to_fp16 = const()[name = string("op_5382_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5383_cast_fp16 = mul(x = var_5381_cast_fp16, y = var_5382_to_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5385_cast_fp16 = add(x = var_5379_cast_fp16, y = var_5383_cast_fp16)[name = string("op_5385_cast_fp16")];
            fp32 var_5386_epsilon_0 = const()[name = string("op_5386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5386_cast_fp16 = rsqrt(epsilon = var_5386_epsilon_0, x = var_5385_cast_fp16)[name = string("op_5386_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_137_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5386_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56385216)))];
            tensor<fp16, [1, 48, 512]> input_113_cast_fp16 = mul(x = out_137_cast_fp16, y = tm_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56386304))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57434944))))[name = string("tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_86_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_86_cast_fp16")];
            string var_5393_mode_0 = const()[name = string("op_5393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_5393_cast_fp16 = gelu(mode = var_5393_mode_0, x = linear_86_cast_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57439104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58487744))))[name = string("tm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_87_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_115_cast_fp16 = mul(x = var_5393_cast_fp16, y = linear_87_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58491904))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59540544))))[name = string("tm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5401_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5401_cast_fp16")];
            bool var_5404 = const()[name = string("op_5404"), val = bool(true)];
            tensor<int32, [1]> var_5405_axes_0 = const()[name = string("op_5405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5405_cast_fp16 = reduce_max(axes = var_5405_axes_0, keep_dims = var_5404, x = var_5401_cast_fp16)[name = string("op_5405_cast_fp16")];
            fp16 var_5406_to_fp16 = const()[name = string("op_5406_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_159_cast_fp16 = maximum(x = var_5405_cast_fp16, y = var_5406_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_159_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_159_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5409_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5409_cast_fp16")];
            tensor<int32, [1]> var_5414_axes_0 = const()[name = string("op_5414_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5414_keep_dims_0 = const()[name = string("op_5414_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5414_cast_fp16 = reduce_mean(axes = var_5414_axes_0, keep_dims = var_5414_keep_dims_0, x = var_5409_cast_fp16)[name = string("op_5414_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5415_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5415_cast_fp16")];
            fp32 var_5416_epsilon_0 = const()[name = string("op_5416_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5416_cast_fp16 = inverse(epsilon = var_5416_epsilon_0, x = var_5415_cast_fp16)[name = string("op_5416_cast_fp16")];
            fp16 var_5417_to_fp16 = const()[name = string("op_5417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5418_cast_fp16 = mul(x = var_5416_cast_fp16, y = var_5417_to_fp16)[name = string("op_5418_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5420_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5418_cast_fp16)[name = string("op_5420_cast_fp16")];
            fp32 var_5421_epsilon_0 = const()[name = string("op_5421_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5421_cast_fp16 = rsqrt(epsilon = var_5421_epsilon_0, x = var_5420_cast_fp16)[name = string("op_5421_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_139_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5421_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59541632)))];
            tensor<fp16, [1, 48, 512]> var_5423_cast_fp16 = mul(x = out_139_cast_fp16, y = tm_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5423_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_117_cast_fp16 = add(x = x_173_cast_fp16, y = var_5423_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59542720))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59804928))))[name = string("tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_89_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_89_cast_fp16")];
            string x_177_mode_0 = const()[name = string("x_177_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_177_cast_fp16 = gelu(mode = x_177_mode_0, x = linear_89_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [4]> var_5443_begin_0 = const()[name = string("op_5443_begin_0"), val = tensor<int32, [4]>([0, 0, 9, 0])];
            tensor<int32, [4]> var_5443_end_0 = const()[name = string("op_5443_end_0"), val = tensor<int32, [4]>([1, 48, 10, 512])];
            tensor<bool, [4]> var_5443_end_mask_0 = const()[name = string("op_5443_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5443_squeeze_mask_0 = const()[name = string("op_5443_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_5443_cast_fp16 = slice_by_index(begin = var_5443_begin_0, end = var_5443_end_0, end_mask = var_5443_end_mask_0, squeeze_mask = var_5443_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5443_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_119_cast_fp16 = mul(x = x_177_cast_fp16, y = var_5443_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59806016))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60068224))))[name = string("tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5453_cast_fp16 = abs(x = linear_90_cast_fp16)[name = string("op_5453_cast_fp16")];
            bool var_5456 = const()[name = string("op_5456"), val = bool(true)];
            tensor<int32, [1]> var_5457_axes_0 = const()[name = string("op_5457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5457_cast_fp16 = reduce_max(axes = var_5457_axes_0, keep_dims = var_5456, x = var_5453_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_161_cast_fp16 = maximum(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_161_cast_fp16 = real_div(x = linear_90_cast_fp16, y = s_161_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5461_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<int32, [1]> var_5466_axes_0 = const()[name = string("op_5466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5466_keep_dims_0 = const()[name = string("op_5466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5466_cast_fp16 = reduce_mean(axes = var_5466_axes_0, keep_dims = var_5466_keep_dims_0, x = var_5461_cast_fp16)[name = string("op_5466_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5467_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5467_cast_fp16")];
            fp32 var_5468_epsilon_0 = const()[name = string("op_5468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5468_cast_fp16 = inverse(epsilon = var_5468_epsilon_0, x = var_5467_cast_fp16)[name = string("op_5468_cast_fp16")];
            fp16 var_5469_to_fp16 = const()[name = string("op_5469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5470_cast_fp16 = mul(x = var_5468_cast_fp16, y = var_5469_to_fp16)[name = string("op_5470_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5472_cast_fp16 = add(x = var_5466_cast_fp16, y = var_5470_cast_fp16)[name = string("op_5472_cast_fp16")];
            fp32 var_5473_epsilon_0 = const()[name = string("op_5473_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5473_cast_fp16 = rsqrt(epsilon = var_5473_epsilon_0, x = var_5472_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_141_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5473_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60069312)))];
            tensor<fp16, [1, 48, 512]> var_5475_cast_fp16 = mul(x = out_141_cast_fp16, y = tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5475_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_19_cast_fp16 = add(x = input_117_cast_fp16, y = var_5475_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<fp16, [1]> tm_layers_9_layer_scalar_to_fp16 = const()[name = string("tm_layers_9_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9ep-1])];
            tensor<fp16, [1, 48, 512]> x_181_cast_fp16 = mul(x = h_19_cast_fp16, y = tm_layers_9_layer_scalar_to_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5479_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_5479_cast_fp16")];
            bool var_5482 = const()[name = string("op_5482"), val = bool(true)];
            tensor<int32, [1]> var_5483_axes_0 = const()[name = string("op_5483_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5483_cast_fp16 = reduce_max(axes = var_5483_axes_0, keep_dims = var_5482, x = var_5479_cast_fp16)[name = string("op_5483_cast_fp16")];
            fp16 var_5484_to_fp16 = const()[name = string("op_5484_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_163_cast_fp16 = maximum(x = var_5483_cast_fp16, y = var_5484_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_163_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_163_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5487_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5492_axes_0 = const()[name = string("op_5492_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5492_keep_dims_0 = const()[name = string("op_5492_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5492_cast_fp16 = reduce_mean(axes = var_5492_axes_0, keep_dims = var_5492_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5493_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5493_cast_fp16")];
            fp32 var_5494_epsilon_0 = const()[name = string("op_5494_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5494_cast_fp16 = inverse(epsilon = var_5494_epsilon_0, x = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp16 var_5495_to_fp16 = const()[name = string("op_5495_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5496_cast_fp16 = mul(x = var_5494_cast_fp16, y = var_5495_to_fp16)[name = string("op_5496_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5498_cast_fp16 = add(x = var_5492_cast_fp16, y = var_5496_cast_fp16)[name = string("op_5498_cast_fp16")];
            fp32 var_5499_epsilon_0 = const()[name = string("op_5499_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5499_cast_fp16 = rsqrt(epsilon = var_5499_epsilon_0, x = var_5498_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_143_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5499_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60070400)))];
            tensor<fp16, [1, 48, 512]> input_121_cast_fp16 = mul(x = out_143_cast_fp16, y = tm_layers_10_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60071488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60595840))))[name = string("tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5509 = const()[name = string("op_5509"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_5510_cast_fp16 = reshape(shape = var_5509, x = linear_91_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60597952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60860160))))[name = string("tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5521 = const()[name = string("op_5521"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_5522_cast_fp16 = reshape(shape = var_5521, x = linear_92_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60861248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61123456))))[name = string("tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5533 = const()[name = string("op_5533"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_5534_cast_fp16 = reshape(shape = var_5533, x = linear_93_cast_fp16)[name = string("op_5534_cast_fp16")];
            tensor<int32, [4]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5510_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 4, 48, 256]> var_5538_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_5538_cast_fp16")];
            bool var_5541 = const()[name = string("op_5541"), val = bool(true)];
            tensor<int32, [1]> var_5542_axes_0 = const()[name = string("op_5542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_5542_cast_fp16 = reduce_max(axes = var_5542_axes_0, keep_dims = var_5541, x = var_5538_cast_fp16)[name = string("op_5542_cast_fp16")];
            fp16 var_5543_to_fp16 = const()[name = string("op_5543_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_165_cast_fp16 = maximum(x = var_5542_cast_fp16, y = var_5543_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_165_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_165_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_5546_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<int32, [1]> var_5551_axes_0 = const()[name = string("op_5551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5551_keep_dims_0 = const()[name = string("op_5551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_5551_cast_fp16 = reduce_mean(axes = var_5551_axes_0, keep_dims = var_5551_keep_dims_0, x = var_5546_cast_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_5552_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5552_cast_fp16")];
            fp32 var_5553_epsilon_0 = const()[name = string("op_5553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_5553_cast_fp16 = inverse(epsilon = var_5553_epsilon_0, x = var_5552_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp16 var_5554_to_fp16 = const()[name = string("op_5554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_5555_cast_fp16 = mul(x = var_5553_cast_fp16, y = var_5554_to_fp16)[name = string("op_5555_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_5557_cast_fp16 = add(x = var_5551_cast_fp16, y = var_5555_cast_fp16)[name = string("op_5557_cast_fp16")];
            fp32 var_5558_epsilon_0 = const()[name = string("op_5558_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_5558_cast_fp16 = rsqrt(epsilon = var_5558_epsilon_0, x = var_5557_cast_fp16)[name = string("op_5558_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_145_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5558_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61124544)))];
            tensor<fp16, [1, 4, 48, 256]> q_41_cast_fp16 = mul(x = out_145_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_41_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_5522_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2, 48, 256]> var_5561_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_5561_cast_fp16")];
            bool var_5564 = const()[name = string("op_5564"), val = bool(true)];
            tensor<int32, [1]> var_5565_axes_0 = const()[name = string("op_5565_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_5565_cast_fp16 = reduce_max(axes = var_5565_axes_0, keep_dims = var_5564, x = var_5561_cast_fp16)[name = string("op_5565_cast_fp16")];
            fp16 var_5566_to_fp16 = const()[name = string("op_5566_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_167_cast_fp16 = maximum(x = var_5565_cast_fp16, y = var_5566_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_167_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_167_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_5569_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5569_cast_fp16")];
            tensor<int32, [1]> var_5574_axes_0 = const()[name = string("op_5574_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5574_keep_dims_0 = const()[name = string("op_5574_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_5574_cast_fp16 = reduce_mean(axes = var_5574_axes_0, keep_dims = var_5574_keep_dims_0, x = var_5569_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_5575_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5575_cast_fp16")];
            fp32 var_5576_epsilon_0 = const()[name = string("op_5576_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_5576_cast_fp16 = inverse(epsilon = var_5576_epsilon_0, x = var_5575_cast_fp16)[name = string("op_5576_cast_fp16")];
            fp16 var_5577_to_fp16 = const()[name = string("op_5577_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_5578_cast_fp16 = mul(x = var_5576_cast_fp16, y = var_5577_to_fp16)[name = string("op_5578_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_5580_cast_fp16 = add(x = var_5574_cast_fp16, y = var_5578_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_5581_cast_fp16 = rsqrt(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_147_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5581_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [256]> tm_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125120)))];
            tensor<fp16, [1, 2, 48, 256]> k_61_cast_fp16 = mul(x = out_147_cast_fp16, y = tm_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = var_5534_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 2, 48, 256]> var_5584_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_5584_cast_fp16")];
            bool var_5587 = const()[name = string("op_5587"), val = bool(true)];
            tensor<int32, [1]> var_5588_axes_0 = const()[name = string("op_5588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_5588_cast_fp16 = reduce_max(axes = var_5588_axes_0, keep_dims = var_5587, x = var_5584_cast_fp16)[name = string("op_5588_cast_fp16")];
            fp16 var_5589_to_fp16 = const()[name = string("op_5589_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_169_cast_fp16 = maximum(x = var_5588_cast_fp16, y = var_5589_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_169_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_169_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_5592_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5592_cast_fp16")];
            tensor<int32, [1]> var_5597_axes_0 = const()[name = string("op_5597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5597_keep_dims_0 = const()[name = string("op_5597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_5597_cast_fp16 = reduce_mean(axes = var_5597_axes_0, keep_dims = var_5597_keep_dims_0, x = var_5592_cast_fp16)[name = string("op_5597_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_5598_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5598_cast_fp16")];
            fp32 var_5599_epsilon_0 = const()[name = string("op_5599_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_5599_cast_fp16 = inverse(epsilon = var_5599_epsilon_0, x = var_5598_cast_fp16)[name = string("op_5599_cast_fp16")];
            fp16 var_5600_to_fp16 = const()[name = string("op_5600_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_5601_cast_fp16 = mul(x = var_5599_cast_fp16, y = var_5600_to_fp16)[name = string("op_5601_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_5603_cast_fp16 = add(x = var_5597_cast_fp16, y = var_5601_cast_fp16)[name = string("op_5603_cast_fp16")];
            fp32 var_5604_epsilon_0 = const()[name = string("op_5604_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_5604_cast_fp16 = rsqrt(epsilon = var_5604_epsilon_0, x = var_5603_cast_fp16)[name = string("op_5604_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_41_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5604_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_5606_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5611_begin_0 = const()[name = string("op_5611_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5611_end_0 = const()[name = string("op_5611_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_5611_end_mask_0 = const()[name = string("op_5611_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_5611_cast_fp16 = slice_by_index(begin = var_5611_begin_0, end = var_5611_end_0, end_mask = var_5611_end_mask_0, x = q_41_cast_fp16)[name = string("op_5611_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_5612_cast_fp16 = mul(x = var_5611_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_5612_cast_fp16")];
            tensor<int32, [4]> var_5617_begin_0 = const()[name = string("op_5617_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5617_end_0 = const()[name = string("op_5617_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_5617_end_mask_0 = const()[name = string("op_5617_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_5617_cast_fp16 = slice_by_index(begin = var_5617_begin_0, end = var_5617_end_0, end_mask = var_5617_end_mask_0, x = q_41_cast_fp16)[name = string("op_5617_cast_fp16")];
            int32 var_5619 = const()[name = string("op_5619"), val = int32(-1)];
            bool var_5620_interleave_0 = const()[name = string("op_5620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_5620_cast_fp16 = concat(axis = var_5619, interleave = var_5620_interleave_0, values = (var_5612_cast_fp16, var_5617_cast_fp16))[name = string("op_5620_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_5621_cast_fp16 = mul(x = var_5620_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5621_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_43_cast_fp16 = add(x = var_5606_cast_fp16, y = var_5621_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_5624_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = k_61_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = k_61_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool var_5638_interleave_0 = const()[name = string("op_5638_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_5638_cast_fp16 = concat(axis = var_5637, interleave = var_5638_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("op_5638_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_5639_cast_fp16 = mul(x = var_5638_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_63_cast_fp16 = add(x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_20_reps_0 = const()[name = string("tile_20_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = k_63_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [4, 1, 48, 256]> tile_20_cast_fp16 = tile(reps = tile_20_reps_0, x = transpose_40_cast_fp16)[name = string("tile_20_cast_fp16")];
            tensor<int32, [5]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_40_cast_fp16 = reshape(shape = concat_40, x = tile_20_cast_fp16)[name = string("reshape_40_cast_fp16")];
            tensor<int32, [5]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = reshape_40_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [4, 1, 48, 256]> reshape_41_cast_fp16 = reshape(shape = concat_41, x = transpose_41_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_21_reps_0 = const()[name = string("tile_21_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = v_41_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [4, 1, 48, 256]> tile_21_cast_fp16 = tile(reps = tile_21_reps_0, x = transpose_42_cast_fp16)[name = string("tile_21_cast_fp16")];
            tensor<int32, [5]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_42_cast_fp16 = reshape(shape = concat_42, x = tile_21_cast_fp16)[name = string("reshape_42_cast_fp16")];
            tensor<int32, [5]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = reshape_42_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [4, 1, 48, 256]> reshape_43_cast_fp16 = reshape(shape = concat_43, x = transpose_43_cast_fp16)[name = string("reshape_43_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5653_transpose_x_0 = const()[name = string("op_5653_transpose_x_0"), val = bool(false)];
            bool var_5653_transpose_y_0 = const()[name = string("op_5653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 4, 48, 48]> var_5653_cast_fp16 = matmul(transpose_x = var_5653_transpose_x_0, transpose_y = var_5653_transpose_y_0, x = q_43_cast_fp16, y = transpose_106_cast_fp16)[name = string("op_5653_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_5655_cast_fp16 = add(x = var_5653_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5655_cast_fp16")];
            int32 var_5656 = const()[name = string("op_5656"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_169_cast_fp16 = softmax(axis = var_5656, x = var_5655_cast_fp16)[name = string("w_169_cast_fp16")];
            bool var_5659_transpose_x_0 = const()[name = string("op_5659_transpose_x_0"), val = bool(false)];
            bool var_5659_transpose_y_0 = const()[name = string("op_5659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = reshape_43_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 4, 48, 256]> var_5659_cast_fp16 = matmul(transpose_x = var_5659_transpose_x_0, transpose_y = var_5659_transpose_y_0, x = w_169_cast_fp16, y = v_43_cast_fp16)[name = string("op_5659_cast_fp16")];
            tensor<int32, [4]> var_5662_perm_0 = const()[name = string("op_5662_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5666 = const()[name = string("op_5666"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_5662_cast_fp16 = transpose(perm = var_5662_perm_0, x = var_5659_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 48, 1024]> input_123_cast_fp16 = reshape(shape = var_5666, x = var_5662_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61650048))))[name = string("tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5671_cast_fp16 = abs(x = linear_94_cast_fp16)[name = string("op_5671_cast_fp16")];
            bool var_5674 = const()[name = string("op_5674"), val = bool(true)];
            tensor<int32, [1]> var_5675_axes_0 = const()[name = string("op_5675_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5675_cast_fp16 = reduce_max(axes = var_5675_axes_0, keep_dims = var_5674, x = var_5671_cast_fp16)[name = string("op_5675_cast_fp16")];
            fp16 var_5676_to_fp16 = const()[name = string("op_5676_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_171_cast_fp16 = maximum(x = var_5675_cast_fp16, y = var_5676_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_171_cast_fp16 = real_div(x = linear_94_cast_fp16, y = s_171_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5679_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5679_cast_fp16")];
            tensor<int32, [1]> var_5684_axes_0 = const()[name = string("op_5684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5684_keep_dims_0 = const()[name = string("op_5684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5684_cast_fp16 = reduce_mean(axes = var_5684_axes_0, keep_dims = var_5684_keep_dims_0, x = var_5679_cast_fp16)[name = string("op_5684_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5685_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5685_cast_fp16")];
            fp32 var_5686_epsilon_0 = const()[name = string("op_5686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5686_cast_fp16 = inverse(epsilon = var_5686_epsilon_0, x = var_5685_cast_fp16)[name = string("op_5686_cast_fp16")];
            fp16 var_5687_to_fp16 = const()[name = string("op_5687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5688_cast_fp16 = mul(x = var_5686_cast_fp16, y = var_5687_to_fp16)[name = string("op_5688_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5690_cast_fp16 = add(x = var_5684_cast_fp16, y = var_5688_cast_fp16)[name = string("op_5690_cast_fp16")];
            fp32 var_5691_epsilon_0 = const()[name = string("op_5691_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5691_cast_fp16 = rsqrt(epsilon = var_5691_epsilon_0, x = var_5690_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_149_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5691_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61651136)))];
            tensor<fp16, [1, 48, 512]> var_5693_cast_fp16 = mul(x = out_149_cast_fp16, y = tm_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5693_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = var_5693_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5696_cast_fp16 = abs(x = x_191_cast_fp16)[name = string("op_5696_cast_fp16")];
            bool var_5699 = const()[name = string("op_5699"), val = bool(true)];
            tensor<int32, [1]> var_5700_axes_0 = const()[name = string("op_5700_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5700_cast_fp16 = reduce_max(axes = var_5700_axes_0, keep_dims = var_5699, x = var_5696_cast_fp16)[name = string("op_5700_cast_fp16")];
            fp16 var_5701_to_fp16 = const()[name = string("op_5701_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_173_cast_fp16 = maximum(x = var_5700_cast_fp16, y = var_5701_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_173_cast_fp16 = real_div(x = x_191_cast_fp16, y = s_173_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5704_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5704_cast_fp16")];
            tensor<int32, [1]> var_5709_axes_0 = const()[name = string("op_5709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5709_keep_dims_0 = const()[name = string("op_5709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5709_cast_fp16 = reduce_mean(axes = var_5709_axes_0, keep_dims = var_5709_keep_dims_0, x = var_5704_cast_fp16)[name = string("op_5709_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5710_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5710_cast_fp16")];
            fp32 var_5711_epsilon_0 = const()[name = string("op_5711_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5711_cast_fp16 = inverse(epsilon = var_5711_epsilon_0, x = var_5710_cast_fp16)[name = string("op_5711_cast_fp16")];
            fp16 var_5712_to_fp16 = const()[name = string("op_5712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = string("op_5713_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5715_cast_fp16 = add(x = var_5709_cast_fp16, y = var_5713_cast_fp16)[name = string("op_5715_cast_fp16")];
            fp32 var_5716_epsilon_0 = const()[name = string("op_5716_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5716_cast_fp16 = rsqrt(epsilon = var_5716_epsilon_0, x = var_5715_cast_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_151_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5716_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61652224)))];
            tensor<fp16, [1, 48, 512]> input_125_cast_fp16 = mul(x = out_151_cast_fp16, y = tm_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61653312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62701952))))[name = string("tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_95_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_95_cast_fp16")];
            string var_5723_mode_0 = const()[name = string("op_5723_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_5723_cast_fp16 = gelu(mode = var_5723_mode_0, x = linear_95_cast_fp16)[name = string("op_5723_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62706112))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63754752))))[name = string("tm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_127_cast_fp16 = mul(x = var_5723_cast_fp16, y = linear_96_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63758912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64807552))))[name = string("tm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5731_cast_fp16 = abs(x = linear_97_cast_fp16)[name = string("op_5731_cast_fp16")];
            bool var_5734 = const()[name = string("op_5734"), val = bool(true)];
            tensor<int32, [1]> var_5735_axes_0 = const()[name = string("op_5735_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5735_cast_fp16 = reduce_max(axes = var_5735_axes_0, keep_dims = var_5734, x = var_5731_cast_fp16)[name = string("op_5735_cast_fp16")];
            fp16 var_5736_to_fp16 = const()[name = string("op_5736_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_175_cast_fp16 = maximum(x = var_5735_cast_fp16, y = var_5736_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_175_cast_fp16 = real_div(x = linear_97_cast_fp16, y = s_175_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5739_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [1]> var_5744_axes_0 = const()[name = string("op_5744_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5744_keep_dims_0 = const()[name = string("op_5744_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5744_cast_fp16 = reduce_mean(axes = var_5744_axes_0, keep_dims = var_5744_keep_dims_0, x = var_5739_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5745_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5745_cast_fp16")];
            fp32 var_5746_epsilon_0 = const()[name = string("op_5746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5746_cast_fp16 = inverse(epsilon = var_5746_epsilon_0, x = var_5745_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5747_to_fp16 = const()[name = string("op_5747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5748_cast_fp16 = mul(x = var_5746_cast_fp16, y = var_5747_to_fp16)[name = string("op_5748_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5750_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5748_cast_fp16)[name = string("op_5750_cast_fp16")];
            fp32 var_5751_epsilon_0 = const()[name = string("op_5751_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5751_cast_fp16 = rsqrt(epsilon = var_5751_epsilon_0, x = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_153_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5751_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64808640)))];
            tensor<fp16, [1, 48, 512]> var_5753_cast_fp16 = mul(x = out_153_cast_fp16, y = tm_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_129_cast_fp16 = add(x = x_191_cast_fp16, y = var_5753_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64809728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65071936))))[name = string("tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_98_cast_fp16")];
            string x_195_mode_0 = const()[name = string("x_195_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_195_cast_fp16 = gelu(mode = x_195_mode_0, x = linear_98_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [4]> var_5773_begin_0 = const()[name = string("op_5773_begin_0"), val = tensor<int32, [4]>([0, 0, 10, 0])];
            tensor<int32, [4]> var_5773_end_0 = const()[name = string("op_5773_end_0"), val = tensor<int32, [4]>([1, 48, 11, 512])];
            tensor<bool, [4]> var_5773_end_mask_0 = const()[name = string("op_5773_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5773_squeeze_mask_0 = const()[name = string("op_5773_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_5773_cast_fp16 = slice_by_index(begin = var_5773_begin_0, end = var_5773_end_0, end_mask = var_5773_end_mask_0, squeeze_mask = var_5773_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_131_cast_fp16 = mul(x = x_195_cast_fp16, y = var_5773_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65073024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65335232))))[name = string("tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5783_cast_fp16 = abs(x = linear_99_cast_fp16)[name = string("op_5783_cast_fp16")];
            bool var_5786 = const()[name = string("op_5786"), val = bool(true)];
            tensor<int32, [1]> var_5787_axes_0 = const()[name = string("op_5787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5787_cast_fp16 = reduce_max(axes = var_5787_axes_0, keep_dims = var_5786, x = var_5783_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp16 var_5788_to_fp16 = const()[name = string("op_5788_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_177_cast_fp16 = maximum(x = var_5787_cast_fp16, y = var_5788_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_177_cast_fp16 = real_div(x = linear_99_cast_fp16, y = s_177_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5791_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5791_cast_fp16")];
            tensor<int32, [1]> var_5796_axes_0 = const()[name = string("op_5796_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5796_keep_dims_0 = const()[name = string("op_5796_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5796_cast_fp16 = reduce_mean(axes = var_5796_axes_0, keep_dims = var_5796_keep_dims_0, x = var_5791_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5797_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5797_cast_fp16")];
            fp32 var_5798_epsilon_0 = const()[name = string("op_5798_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5798_cast_fp16 = inverse(epsilon = var_5798_epsilon_0, x = var_5797_cast_fp16)[name = string("op_5798_cast_fp16")];
            fp16 var_5799_to_fp16 = const()[name = string("op_5799_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5800_cast_fp16 = mul(x = var_5798_cast_fp16, y = var_5799_to_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5802_cast_fp16 = add(x = var_5796_cast_fp16, y = var_5800_cast_fp16)[name = string("op_5802_cast_fp16")];
            fp32 var_5803_epsilon_0 = const()[name = string("op_5803_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5803_cast_fp16 = rsqrt(epsilon = var_5803_epsilon_0, x = var_5802_cast_fp16)[name = string("op_5803_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_155_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5803_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65336320)))];
            tensor<fp16, [1, 48, 512]> var_5805_cast_fp16 = mul(x = out_155_cast_fp16, y = tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5805_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_21_cast_fp16 = add(x = input_129_cast_fp16, y = var_5805_cast_fp16)[name = string("h_21_cast_fp16")];
            tensor<fp16, [1]> tm_layers_10_layer_scalar_to_fp16 = const()[name = string("tm_layers_10_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.1cp-1])];
            tensor<fp16, [1, 48, 512]> x_199_cast_fp16 = mul(x = h_21_cast_fp16, y = tm_layers_10_layer_scalar_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5809_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5809_cast_fp16")];
            bool var_5812 = const()[name = string("op_5812"), val = bool(true)];
            tensor<int32, [1]> var_5813_axes_0 = const()[name = string("op_5813_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_5813_cast_fp16 = reduce_max(axes = var_5813_axes_0, keep_dims = var_5812, x = var_5809_cast_fp16)[name = string("op_5813_cast_fp16")];
            fp16 var_5814_to_fp16 = const()[name = string("op_5814_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_179_cast_fp16 = maximum(x = var_5813_cast_fp16, y = var_5814_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_179_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_179_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_5817_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5817_cast_fp16")];
            tensor<int32, [1]> var_5822_axes_0 = const()[name = string("op_5822_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5822_keep_dims_0 = const()[name = string("op_5822_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_5822_cast_fp16 = reduce_mean(axes = var_5822_axes_0, keep_dims = var_5822_keep_dims_0, x = var_5817_cast_fp16)[name = string("op_5822_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5823_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp32 var_5824_epsilon_0 = const()[name = string("op_5824_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_5824_cast_fp16 = inverse(epsilon = var_5824_epsilon_0, x = var_5823_cast_fp16)[name = string("op_5824_cast_fp16")];
            fp16 var_5825_to_fp16 = const()[name = string("op_5825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_5826_cast_fp16 = mul(x = var_5824_cast_fp16, y = var_5825_to_fp16)[name = string("op_5826_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_5828_cast_fp16 = add(x = var_5822_cast_fp16, y = var_5826_cast_fp16)[name = string("op_5828_cast_fp16")];
            fp32 var_5829_epsilon_0 = const()[name = string("op_5829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_5829_cast_fp16 = rsqrt(epsilon = var_5829_epsilon_0, x = var_5828_cast_fp16)[name = string("op_5829_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_157_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5829_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65337408)))];
            tensor<fp16, [1, 48, 512]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = tm_layers_11_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65338496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66387136))))[name = string("tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_100_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5839 = const()[name = string("op_5839"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 48, 4, 512]> var_5840_cast_fp16 = reshape(shape = var_5839, x = linear_100_cast_fp16)[name = string("op_5840_cast_fp16")];
            tensor<int32, [4]> x_201_perm_0 = const()[name = string("x_201_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66391296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66653504))))[name = string("tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_5851 = const()[name = string("op_5851"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 48, 1, 512]> var_5852_cast_fp16 = reshape(shape = var_5851, x = linear_101_cast_fp16)[name = string("op_5852_cast_fp16")];
            tensor<int32, [4]> x_203_perm_0 = const()[name = string("x_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66654592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66916800))))[name = string("tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 48, 1, 512]> var_5864_cast_fp16 = reshape(shape = var_5863, x = linear_102_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 512]> x_201_cast_fp16 = transpose(perm = x_201_perm_0, x = var_5840_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 4, 48, 512]> var_5868_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5868_cast_fp16")];
            bool var_5871 = const()[name = string("op_5871"), val = bool(true)];
            tensor<int32, [1]> var_5872_axes_0 = const()[name = string("op_5872_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_5872_cast_fp16 = reduce_max(axes = var_5872_axes_0, keep_dims = var_5871, x = var_5868_cast_fp16)[name = string("op_5872_cast_fp16")];
            fp16 var_5873_to_fp16 = const()[name = string("op_5873_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_181_cast_fp16 = maximum(x = var_5872_cast_fp16, y = var_5873_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> y_181_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_181_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_5876_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5876_cast_fp16")];
            tensor<int32, [1]> var_5881_axes_0 = const()[name = string("op_5881_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5881_keep_dims_0 = const()[name = string("op_5881_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_5881_cast_fp16 = reduce_mean(axes = var_5881_axes_0, keep_dims = var_5881_keep_dims_0, x = var_5876_cast_fp16)[name = string("op_5881_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_5882_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5882_cast_fp16")];
            fp32 var_5883_epsilon_0 = const()[name = string("op_5883_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_5883_cast_fp16 = inverse(epsilon = var_5883_epsilon_0, x = var_5882_cast_fp16)[name = string("op_5883_cast_fp16")];
            fp16 var_5884_to_fp16 = const()[name = string("op_5884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_5885_cast_fp16 = mul(x = var_5883_cast_fp16, y = var_5884_to_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_5887_cast_fp16 = add(x = var_5881_cast_fp16, y = var_5885_cast_fp16)[name = string("op_5887_cast_fp16")];
            fp32 var_5888_epsilon_0 = const()[name = string("op_5888_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_5888_cast_fp16 = rsqrt(epsilon = var_5888_epsilon_0, x = var_5887_cast_fp16)[name = string("op_5888_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> out_159_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5888_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66917888)))];
            tensor<fp16, [1, 4, 48, 512]> q_45_cast_fp16 = mul(x = out_159_cast_fp16, y = tm_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> x_203_cast_fp16 = transpose(perm = x_203_perm_0, x = var_5852_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1, 48, 512]> var_5891_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5891_cast_fp16")];
            bool var_5894 = const()[name = string("op_5894"), val = bool(true)];
            tensor<int32, [1]> var_5895_axes_0 = const()[name = string("op_5895_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 48, 1]> var_5895_cast_fp16 = reduce_max(axes = var_5895_axes_0, keep_dims = var_5894, x = var_5891_cast_fp16)[name = string("op_5895_cast_fp16")];
            fp16 var_5896_to_fp16 = const()[name = string("op_5896_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 48, 1]> s_183_cast_fp16 = maximum(x = var_5895_cast_fp16, y = var_5896_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> y_183_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_183_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_5899_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5899_cast_fp16")];
            tensor<int32, [1]> var_5904_axes_0 = const()[name = string("op_5904_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5904_keep_dims_0 = const()[name = string("op_5904_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 48, 1]> var_5904_cast_fp16 = reduce_mean(axes = var_5904_axes_0, keep_dims = var_5904_keep_dims_0, x = var_5899_cast_fp16)[name = string("op_5904_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_5905_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5905_cast_fp16")];
            fp32 var_5906_epsilon_0 = const()[name = string("op_5906_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 48, 1]> var_5906_cast_fp16 = inverse(epsilon = var_5906_epsilon_0, x = var_5905_cast_fp16)[name = string("op_5906_cast_fp16")];
            fp16 var_5907_to_fp16 = const()[name = string("op_5907_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 48, 1]> var_5908_cast_fp16 = mul(x = var_5906_cast_fp16, y = var_5907_to_fp16)[name = string("op_5908_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_5910_cast_fp16 = add(x = var_5904_cast_fp16, y = var_5908_cast_fp16)[name = string("op_5910_cast_fp16")];
            fp32 var_5911_epsilon_0 = const()[name = string("op_5911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 48, 1]> var_5911_cast_fp16 = rsqrt(epsilon = var_5911_epsilon_0, x = var_5910_cast_fp16)[name = string("op_5911_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> out_161_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5911_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66918976)))];
            tensor<fp16, [1, 1, 48, 512]> k_67_cast_fp16 = mul(x = out_161_cast_fp16, y = tm_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_5864_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1, 48, 512]> var_5914_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5914_cast_fp16")];
            bool var_5917 = const()[name = string("op_5917"), val = bool(true)];
            tensor<int32, [1]> var_5918_axes_0 = const()[name = string("op_5918_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 48, 1]> var_5918_cast_fp16 = reduce_max(axes = var_5918_axes_0, keep_dims = var_5917, x = var_5914_cast_fp16)[name = string("op_5918_cast_fp16")];
            fp16 var_5919_to_fp16 = const()[name = string("op_5919_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 48, 1]> s_185_cast_fp16 = maximum(x = var_5918_cast_fp16, y = var_5919_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> y_185_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_185_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_5922_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5922_cast_fp16")];
            tensor<int32, [1]> var_5927_axes_0 = const()[name = string("op_5927_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5927_keep_dims_0 = const()[name = string("op_5927_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 48, 1]> var_5927_cast_fp16 = reduce_mean(axes = var_5927_axes_0, keep_dims = var_5927_keep_dims_0, x = var_5922_cast_fp16)[name = string("op_5927_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_5928_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5928_cast_fp16")];
            fp32 var_5929_epsilon_0 = const()[name = string("op_5929_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 48, 1]> var_5929_cast_fp16 = inverse(epsilon = var_5929_epsilon_0, x = var_5928_cast_fp16)[name = string("op_5929_cast_fp16")];
            fp16 var_5930_to_fp16 = const()[name = string("op_5930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 48, 1]> var_5931_cast_fp16 = mul(x = var_5929_cast_fp16, y = var_5930_to_fp16)[name = string("op_5931_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_5933_cast_fp16 = add(x = var_5927_cast_fp16, y = var_5931_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp32 var_5934_epsilon_0 = const()[name = string("op_5934_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 48, 1]> var_5934_cast_fp16 = rsqrt(epsilon = var_5934_epsilon_0, x = var_5933_cast_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> v_45_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5934_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_5936_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5936_cast_fp16")];
            tensor<int32, [4]> var_5941_begin_0 = const()[name = string("op_5941_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5941_end_0 = const()[name = string("op_5941_end_0"), val = tensor<int32, [4]>([1, 4, 48, 512])];
            tensor<bool, [4]> var_5941_end_mask_0 = const()[name = string("op_5941_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 256]> var_5941_cast_fp16 = slice_by_index(begin = var_5941_begin_0, end = var_5941_end_0, end_mask = var_5941_end_mask_0, x = q_45_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 256]> var_5942_cast_fp16 = mul(x = var_5941_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_5942_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 256]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = q_45_cast_fp16)[name = string("op_5947_cast_fp16")];
            int32 var_5949 = const()[name = string("op_5949"), val = int32(-1)];
            bool var_5950_interleave_0 = const()[name = string("op_5950_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 512]> var_5950_cast_fp16 = concat(axis = var_5949, interleave = var_5950_interleave_0, values = (var_5942_cast_fp16, var_5947_cast_fp16))[name = string("op_5950_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_5951_cast_fp16 = mul(x = var_5950_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5951_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> q_47_cast_fp16 = add(x = var_5936_cast_fp16, y = var_5951_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_5954_cast_fp16 = mul(x = k_67_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5954_cast_fp16")];
            tensor<int32, [4]> var_5959_begin_0 = const()[name = string("op_5959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5959_end_0 = const()[name = string("op_5959_end_0"), val = tensor<int32, [4]>([1, 1, 48, 512])];
            tensor<bool, [4]> var_5959_end_mask_0 = const()[name = string("op_5959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 48, 256]> var_5959_cast_fp16 = slice_by_index(begin = var_5959_begin_0, end = var_5959_end_0, end_mask = var_5959_end_mask_0, x = k_67_cast_fp16)[name = string("op_5959_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 48, 256]> var_5960_cast_fp16 = mul(x = var_5959_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_5960_cast_fp16")];
            tensor<int32, [4]> var_5965_begin_0 = const()[name = string("op_5965_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5965_end_0 = const()[name = string("op_5965_end_0"), val = tensor<int32, [4]>([1, 1, 48, 256])];
            tensor<bool, [4]> var_5965_end_mask_0 = const()[name = string("op_5965_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 48, 256]> var_5965_cast_fp16 = slice_by_index(begin = var_5965_begin_0, end = var_5965_end_0, end_mask = var_5965_end_mask_0, x = k_67_cast_fp16)[name = string("op_5965_cast_fp16")];
            int32 var_5967 = const()[name = string("op_5967"), val = int32(-1)];
            bool var_5968_interleave_0 = const()[name = string("op_5968_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 48, 512]> var_5968_cast_fp16 = concat(axis = var_5967, interleave = var_5968_interleave_0, values = (var_5960_cast_fp16, var_5965_cast_fp16))[name = string("op_5968_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_5969_cast_fp16 = mul(x = var_5968_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5969_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> k_69_cast_fp16 = add(x = var_5954_cast_fp16, y = var_5969_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_22_reps_0 = const()[name = string("tile_22_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 48, 512]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = k_69_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [4, 1, 48, 512]> tile_22_cast_fp16 = tile(reps = tile_22_reps_0, x = transpose_44_cast_fp16)[name = string("tile_22_cast_fp16")];
            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([4, 1, 1, 48, 512])];
            tensor<fp16, [4, 1, 1, 48, 512]> reshape_44_cast_fp16 = reshape(shape = concat_44, x = tile_22_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [5]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 48, 512])];
            tensor<fp16, [1, 4, 1, 48, 512]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [4, 1, 48, 512]> reshape_45_cast_fp16 = reshape(shape = concat_45, x = transpose_45_cast_fp16)[name = string("reshape_45_cast_fp16")];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_23_reps_0 = const()[name = string("tile_23_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 48, 512]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = v_45_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [4, 1, 48, 512]> tile_23_cast_fp16 = tile(reps = tile_23_reps_0, x = transpose_46_cast_fp16)[name = string("tile_23_cast_fp16")];
            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([4, 1, 1, 48, 512])];
            tensor<fp16, [4, 1, 1, 48, 512]> reshape_46_cast_fp16 = reshape(shape = concat_46, x = tile_23_cast_fp16)[name = string("reshape_46_cast_fp16")];
            tensor<int32, [5]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 48, 512])];
            tensor<fp16, [1, 4, 1, 48, 512]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = reshape_46_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [4, 1, 48, 512]> reshape_47_cast_fp16 = reshape(shape = concat_47, x = transpose_47_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5983_transpose_x_0 = const()[name = string("op_5983_transpose_x_0"), val = bool(false)];
            bool var_5983_transpose_y_0 = const()[name = string("op_5983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 48]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = reshape_45_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 4, 48, 48]> var_5983_cast_fp16 = matmul(transpose_x = var_5983_transpose_x_0, transpose_y = var_5983_transpose_y_0, x = q_47_cast_fp16, y = transpose_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_5985_cast_fp16 = add(x = var_5983_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5985_cast_fp16")];
            int32 var_5986 = const()[name = string("op_5986"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_185_cast_fp16 = softmax(axis = var_5986, x = var_5985_cast_fp16)[name = string("w_185_cast_fp16")];
            bool var_5989_transpose_x_0 = const()[name = string("op_5989_transpose_x_0"), val = bool(false)];
            bool var_5989_transpose_y_0 = const()[name = string("op_5989_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 512]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 4, 48, 512]> var_5989_cast_fp16 = matmul(transpose_x = var_5989_transpose_x_0, transpose_y = var_5989_transpose_y_0, x = w_185_cast_fp16, y = v_47_cast_fp16)[name = string("op_5989_cast_fp16")];
            tensor<int32, [4]> var_5992_perm_0 = const()[name = string("op_5992_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5996 = const()[name = string("op_5996"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 48, 4, 512]> var_5992_cast_fp16 = transpose(perm = var_5992_perm_0, x = var_5989_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 48, 2048]> input_135_cast_fp16 = reshape(shape = var_5996, x = var_5992_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66920064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67968704))))[name = string("tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6001_cast_fp16 = abs(x = linear_103_cast_fp16)[name = string("op_6001_cast_fp16")];
            bool var_6004 = const()[name = string("op_6004"), val = bool(true)];
            tensor<int32, [1]> var_6005_axes_0 = const()[name = string("op_6005_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6005_cast_fp16 = reduce_max(axes = var_6005_axes_0, keep_dims = var_6004, x = var_6001_cast_fp16)[name = string("op_6005_cast_fp16")];
            fp16 var_6006_to_fp16 = const()[name = string("op_6006_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_187_cast_fp16 = maximum(x = var_6005_cast_fp16, y = var_6006_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_187_cast_fp16 = real_div(x = linear_103_cast_fp16, y = s_187_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6009_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_6009_cast_fp16")];
            tensor<int32, [1]> var_6014_axes_0 = const()[name = string("op_6014_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6014_keep_dims_0 = const()[name = string("op_6014_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6014_cast_fp16 = reduce_mean(axes = var_6014_axes_0, keep_dims = var_6014_keep_dims_0, x = var_6009_cast_fp16)[name = string("op_6014_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6015_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_6015_cast_fp16")];
            fp32 var_6016_epsilon_0 = const()[name = string("op_6016_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6016_cast_fp16 = inverse(epsilon = var_6016_epsilon_0, x = var_6015_cast_fp16)[name = string("op_6016_cast_fp16")];
            fp16 var_6017_to_fp16 = const()[name = string("op_6017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6018_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6020_cast_fp16 = add(x = var_6014_cast_fp16, y = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp32 var_6021_epsilon_0 = const()[name = string("op_6021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6021_cast_fp16 = rsqrt(epsilon = var_6021_epsilon_0, x = var_6020_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_163_cast_fp16 = mul(x = y_187_cast_fp16, y = var_6021_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969792)))];
            tensor<fp16, [1, 48, 512]> var_6023_cast_fp16 = mul(x = out_163_cast_fp16, y = tm_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_209_cast_fp16 = add(x = x_199_cast_fp16, y = var_6023_cast_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6026_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_6026_cast_fp16")];
            bool var_6029 = const()[name = string("op_6029"), val = bool(true)];
            tensor<int32, [1]> var_6030_axes_0 = const()[name = string("op_6030_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6030_cast_fp16 = reduce_max(axes = var_6030_axes_0, keep_dims = var_6029, x = var_6026_cast_fp16)[name = string("op_6030_cast_fp16")];
            fp16 var_6031_to_fp16 = const()[name = string("op_6031_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_189_cast_fp16 = maximum(x = var_6030_cast_fp16, y = var_6031_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_189_cast_fp16 = real_div(x = x_209_cast_fp16, y = s_189_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6034_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [1]> var_6039_axes_0 = const()[name = string("op_6039_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6039_keep_dims_0 = const()[name = string("op_6039_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6039_cast_fp16 = reduce_mean(axes = var_6039_axes_0, keep_dims = var_6039_keep_dims_0, x = var_6034_cast_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6040_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6041_cast_fp16 = inverse(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            fp16 var_6042_to_fp16 = const()[name = string("op_6042_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6043_cast_fp16 = mul(x = var_6041_cast_fp16, y = var_6042_to_fp16)[name = string("op_6043_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6045_cast_fp16 = add(x = var_6039_cast_fp16, y = var_6043_cast_fp16)[name = string("op_6045_cast_fp16")];
            fp32 var_6046_epsilon_0 = const()[name = string("op_6046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6046_cast_fp16 = rsqrt(epsilon = var_6046_epsilon_0, x = var_6045_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_165_cast_fp16 = mul(x = y_189_cast_fp16, y = var_6046_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67970880)))];
            tensor<fp16, [1, 48, 512]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = tm_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971968))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69020608))))[name = string("tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_104_cast_fp16")];
            string var_6053_mode_0 = const()[name = string("op_6053_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_6053_cast_fp16 = gelu(mode = var_6053_mode_0, x = linear_104_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70073408))))[name = string("tm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_105_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_139_cast_fp16 = mul(x = var_6053_cast_fp16, y = linear_105_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70077568))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71126208))))[name = string("tm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6061_cast_fp16 = abs(x = linear_106_cast_fp16)[name = string("op_6061_cast_fp16")];
            bool var_6064 = const()[name = string("op_6064"), val = bool(true)];
            tensor<int32, [1]> var_6065_axes_0 = const()[name = string("op_6065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6065_cast_fp16 = reduce_max(axes = var_6065_axes_0, keep_dims = var_6064, x = var_6061_cast_fp16)[name = string("op_6065_cast_fp16")];
            fp16 var_6066_to_fp16 = const()[name = string("op_6066_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_191_cast_fp16 = maximum(x = var_6065_cast_fp16, y = var_6066_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_191_cast_fp16 = real_div(x = linear_106_cast_fp16, y = s_191_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6069_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<int32, [1]> var_6074_axes_0 = const()[name = string("op_6074_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6074_keep_dims_0 = const()[name = string("op_6074_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6074_cast_fp16 = reduce_mean(axes = var_6074_axes_0, keep_dims = var_6074_keep_dims_0, x = var_6069_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6075_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_6075_cast_fp16")];
            fp32 var_6076_epsilon_0 = const()[name = string("op_6076_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6076_cast_fp16 = inverse(epsilon = var_6076_epsilon_0, x = var_6075_cast_fp16)[name = string("op_6076_cast_fp16")];
            fp16 var_6077_to_fp16 = const()[name = string("op_6077_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6078_cast_fp16 = mul(x = var_6076_cast_fp16, y = var_6077_to_fp16)[name = string("op_6078_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6080_cast_fp16 = add(x = var_6074_cast_fp16, y = var_6078_cast_fp16)[name = string("op_6080_cast_fp16")];
            fp32 var_6081_epsilon_0 = const()[name = string("op_6081_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6081_cast_fp16 = rsqrt(epsilon = var_6081_epsilon_0, x = var_6080_cast_fp16)[name = string("op_6081_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_167_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6081_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71127296)))];
            tensor<fp16, [1, 48, 512]> var_6083_cast_fp16 = mul(x = out_167_cast_fp16, y = tm_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6083_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_141_cast_fp16 = add(x = x_209_cast_fp16, y = var_6083_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71128384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71390592))))[name = string("tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_107_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_107_cast_fp16")];
            string x_213_mode_0 = const()[name = string("x_213_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_213_cast_fp16 = gelu(mode = x_213_mode_0, x = linear_107_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [4]> var_6103_begin_0 = const()[name = string("op_6103_begin_0"), val = tensor<int32, [4]>([0, 0, 11, 0])];
            tensor<int32, [4]> var_6103_end_0 = const()[name = string("op_6103_end_0"), val = tensor<int32, [4]>([1, 48, 12, 512])];
            tensor<bool, [4]> var_6103_end_mask_0 = const()[name = string("op_6103_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6103_squeeze_mask_0 = const()[name = string("op_6103_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_6103_cast_fp16 = slice_by_index(begin = var_6103_begin_0, end = var_6103_end_0, end_mask = var_6103_end_mask_0, squeeze_mask = var_6103_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6103_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_143_cast_fp16 = mul(x = x_213_cast_fp16, y = var_6103_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71391680))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71653888))))[name = string("tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6113_cast_fp16 = abs(x = linear_108_cast_fp16)[name = string("op_6113_cast_fp16")];
            bool var_6116 = const()[name = string("op_6116"), val = bool(true)];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6117_cast_fp16 = reduce_max(axes = var_6117_axes_0, keep_dims = var_6116, x = var_6113_cast_fp16)[name = string("op_6117_cast_fp16")];
            fp16 var_6118_to_fp16 = const()[name = string("op_6118_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_193_cast_fp16 = maximum(x = var_6117_cast_fp16, y = var_6118_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_193_cast_fp16 = real_div(x = linear_108_cast_fp16, y = s_193_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6121_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6121_cast_fp16")];
            tensor<int32, [1]> var_6126_axes_0 = const()[name = string("op_6126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6126_keep_dims_0 = const()[name = string("op_6126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6126_cast_fp16 = reduce_mean(axes = var_6126_axes_0, keep_dims = var_6126_keep_dims_0, x = var_6121_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6127_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_6127_cast_fp16")];
            fp32 var_6128_epsilon_0 = const()[name = string("op_6128_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6128_cast_fp16 = inverse(epsilon = var_6128_epsilon_0, x = var_6127_cast_fp16)[name = string("op_6128_cast_fp16")];
            fp16 var_6129_to_fp16 = const()[name = string("op_6129_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6130_cast_fp16 = mul(x = var_6128_cast_fp16, y = var_6129_to_fp16)[name = string("op_6130_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6132_cast_fp16 = add(x = var_6126_cast_fp16, y = var_6130_cast_fp16)[name = string("op_6132_cast_fp16")];
            fp32 var_6133_epsilon_0 = const()[name = string("op_6133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6133_cast_fp16 = rsqrt(epsilon = var_6133_epsilon_0, x = var_6132_cast_fp16)[name = string("op_6133_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_169_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6133_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71654976)))];
            tensor<fp16, [1, 48, 512]> var_6135_cast_fp16 = mul(x = out_169_cast_fp16, y = tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6135_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_23_cast_fp16 = add(x = input_141_cast_fp16, y = var_6135_cast_fp16)[name = string("h_23_cast_fp16")];
            tensor<fp16, [1]> tm_layers_11_layer_scalar_to_fp16 = const()[name = string("tm_layers_11_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.86p-2])];
            tensor<fp16, [1, 48, 512]> x_217_cast_fp16 = mul(x = h_23_cast_fp16, y = tm_layers_11_layer_scalar_to_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6139_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_195_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_195_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_195_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6153_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_171_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71656064)))];
            tensor<fp16, [1, 48, 512]> input_145_cast_fp16 = mul(x = out_171_cast_fp16, y = tm_layers_12_input_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71657152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72181504))))[name = string("tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<int32, [4]> var_6169 = const()[name = string("op_6169"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_6170_cast_fp16 = reshape(shape = var_6169, x = linear_109_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72183616))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72445824))))[name = string("tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [4]> var_6181 = const()[name = string("op_6181"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_6182_cast_fp16 = reshape(shape = var_6181, x = linear_110_cast_fp16)[name = string("op_6182_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72446912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72709120))))[name = string("tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<int32, [4]> var_6193 = const()[name = string("op_6193"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_6194_cast_fp16 = reshape(shape = var_6193, x = linear_111_cast_fp16)[name = string("op_6194_cast_fp16")];
            tensor<int32, [4]> x_223_perm_0 = const()[name = string("x_223_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6170_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 4, 48, 256]> var_6198_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_6198_cast_fp16")];
            bool var_6201 = const()[name = string("op_6201"), val = bool(true)];
            tensor<int32, [1]> var_6202_axes_0 = const()[name = string("op_6202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_6202_cast_fp16 = reduce_max(axes = var_6202_axes_0, keep_dims = var_6201, x = var_6198_cast_fp16)[name = string("op_6202_cast_fp16")];
            fp16 var_6203_to_fp16 = const()[name = string("op_6203_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_197_cast_fp16 = maximum(x = var_6202_cast_fp16, y = var_6203_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_197_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_197_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6206_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6206_cast_fp16")];
            tensor<int32, [1]> var_6211_axes_0 = const()[name = string("op_6211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6211_keep_dims_0 = const()[name = string("op_6211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_6211_cast_fp16 = reduce_mean(axes = var_6211_axes_0, keep_dims = var_6211_keep_dims_0, x = var_6206_cast_fp16)[name = string("op_6211_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_6212_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6212_cast_fp16")];
            fp32 var_6213_epsilon_0 = const()[name = string("op_6213_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_6213_cast_fp16 = inverse(epsilon = var_6213_epsilon_0, x = var_6212_cast_fp16)[name = string("op_6213_cast_fp16")];
            fp16 var_6214_to_fp16 = const()[name = string("op_6214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_6215_cast_fp16 = mul(x = var_6213_cast_fp16, y = var_6214_to_fp16)[name = string("op_6215_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_6217_cast_fp16 = add(x = var_6211_cast_fp16, y = var_6215_cast_fp16)[name = string("op_6217_cast_fp16")];
            fp32 var_6218_epsilon_0 = const()[name = string("op_6218_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_6218_cast_fp16 = rsqrt(epsilon = var_6218_epsilon_0, x = var_6217_cast_fp16)[name = string("op_6218_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_173_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6218_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_49_cast_fp16 = mul(x = out_173_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_49_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6182_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 2, 48, 256]> var_6221_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_6221_cast_fp16")];
            bool var_6224 = const()[name = string("op_6224"), val = bool(true)];
            tensor<int32, [1]> var_6225_axes_0 = const()[name = string("op_6225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_6225_cast_fp16 = reduce_max(axes = var_6225_axes_0, keep_dims = var_6224, x = var_6221_cast_fp16)[name = string("op_6225_cast_fp16")];
            fp16 var_6226_to_fp16 = const()[name = string("op_6226_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_199_cast_fp16 = maximum(x = var_6225_cast_fp16, y = var_6226_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_199_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_199_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6229_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6229_cast_fp16")];
            tensor<int32, [1]> var_6234_axes_0 = const()[name = string("op_6234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6234_keep_dims_0 = const()[name = string("op_6234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_6234_cast_fp16 = reduce_mean(axes = var_6234_axes_0, keep_dims = var_6234_keep_dims_0, x = var_6229_cast_fp16)[name = string("op_6234_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6235_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6235_cast_fp16")];
            fp32 var_6236_epsilon_0 = const()[name = string("op_6236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_6236_cast_fp16 = inverse(epsilon = var_6236_epsilon_0, x = var_6235_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_6238_cast_fp16 = mul(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6240_cast_fp16 = add(x = var_6234_cast_fp16, y = var_6238_cast_fp16)[name = string("op_6240_cast_fp16")];
            fp32 var_6241_epsilon_0 = const()[name = string("op_6241_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_6241_cast_fp16 = rsqrt(epsilon = var_6241_epsilon_0, x = var_6240_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_175_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6241_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_73_cast_fp16 = mul(x = out_175_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_223_cast_fp16 = transpose(perm = x_223_perm_0, x = var_6194_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2, 48, 256]> var_6244_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_6244_cast_fp16")];
            bool var_6247 = const()[name = string("op_6247"), val = bool(true)];
            tensor<int32, [1]> var_6248_axes_0 = const()[name = string("op_6248_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_6248_cast_fp16 = reduce_max(axes = var_6248_axes_0, keep_dims = var_6247, x = var_6244_cast_fp16)[name = string("op_6248_cast_fp16")];
            fp16 var_6249_to_fp16 = const()[name = string("op_6249_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_201_cast_fp16 = maximum(x = var_6248_cast_fp16, y = var_6249_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_201_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_201_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6252_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6252_cast_fp16")];
            tensor<int32, [1]> var_6257_axes_0 = const()[name = string("op_6257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6257_keep_dims_0 = const()[name = string("op_6257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_6257_cast_fp16 = reduce_mean(axes = var_6257_axes_0, keep_dims = var_6257_keep_dims_0, x = var_6252_cast_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6258_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6258_cast_fp16")];
            fp32 var_6259_epsilon_0 = const()[name = string("op_6259_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_6259_cast_fp16 = inverse(epsilon = var_6259_epsilon_0, x = var_6258_cast_fp16)[name = string("op_6259_cast_fp16")];
            fp16 var_6260_to_fp16 = const()[name = string("op_6260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_6261_cast_fp16 = mul(x = var_6259_cast_fp16, y = var_6260_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6263_cast_fp16 = add(x = var_6257_cast_fp16, y = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp32 var_6264_epsilon_0 = const()[name = string("op_6264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_6264_cast_fp16 = rsqrt(epsilon = var_6264_epsilon_0, x = var_6263_cast_fp16)[name = string("op_6264_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_49_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6264_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6266_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6266_cast_fp16")];
            tensor<int32, [4]> var_6271_begin_0 = const()[name = string("op_6271_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6271_end_0 = const()[name = string("op_6271_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_6271_end_mask_0 = const()[name = string("op_6271_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_6271_cast_fp16 = slice_by_index(begin = var_6271_begin_0, end = var_6271_end_0, end_mask = var_6271_end_mask_0, x = q_49_cast_fp16)[name = string("op_6271_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_6272_cast_fp16 = mul(x = var_6271_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_6272_cast_fp16")];
            tensor<int32, [4]> var_6277_begin_0 = const()[name = string("op_6277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6277_end_0 = const()[name = string("op_6277_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_6277_end_mask_0 = const()[name = string("op_6277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_6277_cast_fp16 = slice_by_index(begin = var_6277_begin_0, end = var_6277_end_0, end_mask = var_6277_end_mask_0, x = q_49_cast_fp16)[name = string("op_6277_cast_fp16")];
            int32 var_6279 = const()[name = string("op_6279"), val = int32(-1)];
            bool var_6280_interleave_0 = const()[name = string("op_6280_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_6280_cast_fp16 = concat(axis = var_6279, interleave = var_6280_interleave_0, values = (var_6272_cast_fp16, var_6277_cast_fp16))[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6281_cast_fp16 = mul(x = var_6280_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6281_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_51_cast_fp16 = add(x = var_6266_cast_fp16, y = var_6281_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6284_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6284_cast_fp16")];
            tensor<int32, [4]> var_6289_begin_0 = const()[name = string("op_6289_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6289_end_0 = const()[name = string("op_6289_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_6289_end_mask_0 = const()[name = string("op_6289_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_6289_cast_fp16 = slice_by_index(begin = var_6289_begin_0, end = var_6289_end_0, end_mask = var_6289_end_mask_0, x = k_73_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_6290_cast_fp16 = mul(x = var_6289_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_6290_cast_fp16")];
            tensor<int32, [4]> var_6295_begin_0 = const()[name = string("op_6295_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6295_end_0 = const()[name = string("op_6295_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_6295_end_mask_0 = const()[name = string("op_6295_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_6295_cast_fp16 = slice_by_index(begin = var_6295_begin_0, end = var_6295_end_0, end_mask = var_6295_end_mask_0, x = k_73_cast_fp16)[name = string("op_6295_cast_fp16")];
            int32 var_6297 = const()[name = string("op_6297"), val = int32(-1)];
            bool var_6298_interleave_0 = const()[name = string("op_6298_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_6298_cast_fp16 = concat(axis = var_6297, interleave = var_6298_interleave_0, values = (var_6290_cast_fp16, var_6295_cast_fp16))[name = string("op_6298_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6299_cast_fp16 = mul(x = var_6298_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6299_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_75_cast_fp16 = add(x = var_6284_cast_fp16, y = var_6299_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_24_reps_0 = const()[name = string("tile_24_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = k_75_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [4, 1, 48, 256]> tile_24_cast_fp16 = tile(reps = tile_24_reps_0, x = transpose_48_cast_fp16)[name = string("tile_24_cast_fp16")];
            tensor<int32, [5]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_48_cast_fp16 = reshape(shape = concat_48, x = tile_24_cast_fp16)[name = string("reshape_48_cast_fp16")];
            tensor<int32, [5]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = reshape_48_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [4, 1, 48, 256]> reshape_49_cast_fp16 = reshape(shape = concat_49, x = transpose_49_cast_fp16)[name = string("reshape_49_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_25_reps_0 = const()[name = string("tile_25_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = v_49_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [4, 1, 48, 256]> tile_25_cast_fp16 = tile(reps = tile_25_reps_0, x = transpose_50_cast_fp16)[name = string("tile_25_cast_fp16")];
            tensor<int32, [5]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_50_cast_fp16 = reshape(shape = concat_50, x = tile_25_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [5]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [4, 1, 48, 256]> reshape_51_cast_fp16 = reshape(shape = concat_51, x = transpose_51_cast_fp16)[name = string("reshape_51_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6313_transpose_x_0 = const()[name = string("op_6313_transpose_x_0"), val = bool(false)];
            bool var_6313_transpose_y_0 = const()[name = string("op_6313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = reshape_49_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 4, 48, 48]> var_6313_cast_fp16 = matmul(transpose_x = var_6313_transpose_x_0, transpose_y = var_6313_transpose_y_0, x = q_51_cast_fp16, y = transpose_108_cast_fp16)[name = string("op_6313_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_6315_cast_fp16 = add(x = var_6313_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6315_cast_fp16")];
            int32 var_6316 = const()[name = string("op_6316"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_201_cast_fp16 = softmax(axis = var_6316, x = var_6315_cast_fp16)[name = string("w_201_cast_fp16")];
            bool var_6319_transpose_x_0 = const()[name = string("op_6319_transpose_x_0"), val = bool(false)];
            bool var_6319_transpose_y_0 = const()[name = string("op_6319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = reshape_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 4, 48, 256]> var_6319_cast_fp16 = matmul(transpose_x = var_6319_transpose_x_0, transpose_y = var_6319_transpose_y_0, x = w_201_cast_fp16, y = v_51_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<int32, [4]> var_6322_perm_0 = const()[name = string("op_6322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6326 = const()[name = string("op_6326"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_6322_cast_fp16 = transpose(perm = var_6322_perm_0, x = var_6319_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 48, 1024]> input_147_cast_fp16 = reshape(shape = var_6326, x = var_6322_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72710208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73234560))))[name = string("tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6331_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6331_cast_fp16")];
            bool var_6334 = const()[name = string("op_6334"), val = bool(true)];
            tensor<int32, [1]> var_6335_axes_0 = const()[name = string("op_6335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6335_cast_fp16 = reduce_max(axes = var_6335_axes_0, keep_dims = var_6334, x = var_6331_cast_fp16)[name = string("op_6335_cast_fp16")];
            fp16 var_6336_to_fp16 = const()[name = string("op_6336_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_203_cast_fp16 = maximum(x = var_6335_cast_fp16, y = var_6336_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_203_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_203_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6339_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6339_cast_fp16")];
            tensor<int32, [1]> var_6344_axes_0 = const()[name = string("op_6344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6344_keep_dims_0 = const()[name = string("op_6344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6344_cast_fp16 = reduce_mean(axes = var_6344_axes_0, keep_dims = var_6344_keep_dims_0, x = var_6339_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6345_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6345_cast_fp16")];
            fp32 var_6346_epsilon_0 = const()[name = string("op_6346_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6346_cast_fp16 = inverse(epsilon = var_6346_epsilon_0, x = var_6345_cast_fp16)[name = string("op_6346_cast_fp16")];
            fp16 var_6347_to_fp16 = const()[name = string("op_6347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6348_cast_fp16 = mul(x = var_6346_cast_fp16, y = var_6347_to_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6350_cast_fp16 = add(x = var_6344_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            fp32 var_6351_epsilon_0 = const()[name = string("op_6351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6351_cast_fp16 = rsqrt(epsilon = var_6351_epsilon_0, x = var_6350_cast_fp16)[name = string("op_6351_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_177_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6351_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73235648)))];
            tensor<fp16, [1, 48, 512]> var_6353_cast_fp16 = mul(x = out_177_cast_fp16, y = tm_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = var_6353_cast_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6356_cast_fp16 = abs(x = x_227_cast_fp16)[name = string("op_6356_cast_fp16")];
            bool var_6359 = const()[name = string("op_6359"), val = bool(true)];
            tensor<int32, [1]> var_6360_axes_0 = const()[name = string("op_6360_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6360_cast_fp16 = reduce_max(axes = var_6360_axes_0, keep_dims = var_6359, x = var_6356_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_205_cast_fp16 = maximum(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_205_cast_fp16 = real_div(x = x_227_cast_fp16, y = s_205_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6364_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<int32, [1]> var_6369_axes_0 = const()[name = string("op_6369_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6369_keep_dims_0 = const()[name = string("op_6369_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6369_cast_fp16 = reduce_mean(axes = var_6369_axes_0, keep_dims = var_6369_keep_dims_0, x = var_6364_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6370_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6370_cast_fp16")];
            fp32 var_6371_epsilon_0 = const()[name = string("op_6371_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6371_cast_fp16 = inverse(epsilon = var_6371_epsilon_0, x = var_6370_cast_fp16)[name = string("op_6371_cast_fp16")];
            fp16 var_6372_to_fp16 = const()[name = string("op_6372_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6373_cast_fp16 = mul(x = var_6371_cast_fp16, y = var_6372_to_fp16)[name = string("op_6373_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6375_cast_fp16 = add(x = var_6369_cast_fp16, y = var_6373_cast_fp16)[name = string("op_6375_cast_fp16")];
            fp32 var_6376_epsilon_0 = const()[name = string("op_6376_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6376_cast_fp16 = rsqrt(epsilon = var_6376_epsilon_0, x = var_6375_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_179_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6376_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73236736)))];
            tensor<fp16, [1, 48, 512]> input_149_cast_fp16 = mul(x = out_179_cast_fp16, y = tm_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73237824))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74286464))))[name = string("tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_113_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_113_cast_fp16")];
            string var_6383_mode_0 = const()[name = string("op_6383_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_6383_cast_fp16 = gelu(mode = var_6383_mode_0, x = linear_113_cast_fp16)[name = string("op_6383_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74290624))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339264))))[name = string("tm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_114_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_151_cast_fp16 = mul(x = var_6383_cast_fp16, y = linear_114_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75343424))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76392064))))[name = string("tm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6391_cast_fp16 = abs(x = linear_115_cast_fp16)[name = string("op_6391_cast_fp16")];
            bool var_6394 = const()[name = string("op_6394"), val = bool(true)];
            tensor<int32, [1]> var_6395_axes_0 = const()[name = string("op_6395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6395_cast_fp16 = reduce_max(axes = var_6395_axes_0, keep_dims = var_6394, x = var_6391_cast_fp16)[name = string("op_6395_cast_fp16")];
            fp16 var_6396_to_fp16 = const()[name = string("op_6396_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_207_cast_fp16 = maximum(x = var_6395_cast_fp16, y = var_6396_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_207_cast_fp16 = real_div(x = linear_115_cast_fp16, y = s_207_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6399_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6399_cast_fp16")];
            tensor<int32, [1]> var_6404_axes_0 = const()[name = string("op_6404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6404_keep_dims_0 = const()[name = string("op_6404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6404_cast_fp16 = reduce_mean(axes = var_6404_axes_0, keep_dims = var_6404_keep_dims_0, x = var_6399_cast_fp16)[name = string("op_6404_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6405_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6405_cast_fp16")];
            fp32 var_6406_epsilon_0 = const()[name = string("op_6406_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6406_cast_fp16 = inverse(epsilon = var_6406_epsilon_0, x = var_6405_cast_fp16)[name = string("op_6406_cast_fp16")];
            fp16 var_6407_to_fp16 = const()[name = string("op_6407_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6408_cast_fp16 = mul(x = var_6406_cast_fp16, y = var_6407_to_fp16)[name = string("op_6408_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6410_cast_fp16 = add(x = var_6404_cast_fp16, y = var_6408_cast_fp16)[name = string("op_6410_cast_fp16")];
            fp32 var_6411_epsilon_0 = const()[name = string("op_6411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6411_cast_fp16 = rsqrt(epsilon = var_6411_epsilon_0, x = var_6410_cast_fp16)[name = string("op_6411_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_181_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6411_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76393152)))];
            tensor<fp16, [1, 48, 512]> var_6413_cast_fp16 = mul(x = out_181_cast_fp16, y = tm_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6413_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_153_cast_fp16 = add(x = x_227_cast_fp16, y = var_6413_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76394240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76656448))))[name = string("tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_116_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_116_cast_fp16")];
            string x_231_mode_0 = const()[name = string("x_231_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_231_cast_fp16 = gelu(mode = x_231_mode_0, x = linear_116_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_6433_begin_0 = const()[name = string("op_6433_begin_0"), val = tensor<int32, [4]>([0, 0, 12, 0])];
            tensor<int32, [4]> var_6433_end_0 = const()[name = string("op_6433_end_0"), val = tensor<int32, [4]>([1, 48, 13, 512])];
            tensor<bool, [4]> var_6433_end_mask_0 = const()[name = string("op_6433_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6433_squeeze_mask_0 = const()[name = string("op_6433_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_6433_cast_fp16 = slice_by_index(begin = var_6433_begin_0, end = var_6433_end_0, end_mask = var_6433_end_mask_0, squeeze_mask = var_6433_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6433_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_155_cast_fp16 = mul(x = x_231_cast_fp16, y = var_6433_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76657536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76919744))))[name = string("tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6443_cast_fp16 = abs(x = linear_117_cast_fp16)[name = string("op_6443_cast_fp16")];
            bool var_6446 = const()[name = string("op_6446"), val = bool(true)];
            tensor<int32, [1]> var_6447_axes_0 = const()[name = string("op_6447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6447_cast_fp16 = reduce_max(axes = var_6447_axes_0, keep_dims = var_6446, x = var_6443_cast_fp16)[name = string("op_6447_cast_fp16")];
            fp16 var_6448_to_fp16 = const()[name = string("op_6448_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_209_cast_fp16 = maximum(x = var_6447_cast_fp16, y = var_6448_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_209_cast_fp16 = real_div(x = linear_117_cast_fp16, y = s_209_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6451_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6451_cast_fp16")];
            tensor<int32, [1]> var_6456_axes_0 = const()[name = string("op_6456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6456_keep_dims_0 = const()[name = string("op_6456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6456_cast_fp16 = reduce_mean(axes = var_6456_axes_0, keep_dims = var_6456_keep_dims_0, x = var_6451_cast_fp16)[name = string("op_6456_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6457_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6457_cast_fp16")];
            fp32 var_6458_epsilon_0 = const()[name = string("op_6458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6458_cast_fp16 = inverse(epsilon = var_6458_epsilon_0, x = var_6457_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6459_to_fp16 = const()[name = string("op_6459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6460_cast_fp16 = mul(x = var_6458_cast_fp16, y = var_6459_to_fp16)[name = string("op_6460_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6462_cast_fp16 = add(x = var_6456_cast_fp16, y = var_6460_cast_fp16)[name = string("op_6462_cast_fp16")];
            fp32 var_6463_epsilon_0 = const()[name = string("op_6463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6463_cast_fp16 = rsqrt(epsilon = var_6463_epsilon_0, x = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_183_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6463_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76920832)))];
            tensor<fp16, [1, 48, 512]> var_6465_cast_fp16 = mul(x = out_183_cast_fp16, y = tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6465_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_25_cast_fp16 = add(x = input_153_cast_fp16, y = var_6465_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<fp16, [1]> tm_layers_12_layer_scalar_to_fp16 = const()[name = string("tm_layers_12_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.8cp-1])];
            tensor<fp16, [1, 48, 512]> x_235_cast_fp16 = mul(x = h_25_cast_fp16, y = tm_layers_12_layer_scalar_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6469_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_6469_cast_fp16")];
            bool var_6472 = const()[name = string("op_6472"), val = bool(true)];
            tensor<int32, [1]> var_6473_axes_0 = const()[name = string("op_6473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6473_cast_fp16 = reduce_max(axes = var_6473_axes_0, keep_dims = var_6472, x = var_6469_cast_fp16)[name = string("op_6473_cast_fp16")];
            fp16 var_6474_to_fp16 = const()[name = string("op_6474_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_211_cast_fp16 = maximum(x = var_6473_cast_fp16, y = var_6474_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_211_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_211_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6477_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6477_cast_fp16")];
            tensor<int32, [1]> var_6482_axes_0 = const()[name = string("op_6482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6482_keep_dims_0 = const()[name = string("op_6482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6482_cast_fp16 = reduce_mean(axes = var_6482_axes_0, keep_dims = var_6482_keep_dims_0, x = var_6477_cast_fp16)[name = string("op_6482_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6483_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6483_cast_fp16")];
            fp32 var_6484_epsilon_0 = const()[name = string("op_6484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6484_cast_fp16 = inverse(epsilon = var_6484_epsilon_0, x = var_6483_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp16 var_6485_to_fp16 = const()[name = string("op_6485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6486_cast_fp16 = mul(x = var_6484_cast_fp16, y = var_6485_to_fp16)[name = string("op_6486_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6488_cast_fp16 = add(x = var_6482_cast_fp16, y = var_6486_cast_fp16)[name = string("op_6488_cast_fp16")];
            fp32 var_6489_epsilon_0 = const()[name = string("op_6489_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6489_cast_fp16 = rsqrt(epsilon = var_6489_epsilon_0, x = var_6488_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_185_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6489_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76921920)))];
            tensor<fp16, [1, 48, 512]> input_157_cast_fp16 = mul(x = out_185_cast_fp16, y = tm_layers_13_input_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76923008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77447360))))[name = string("tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<int32, [4]> var_6499 = const()[name = string("op_6499"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_6500_cast_fp16 = reshape(shape = var_6499, x = linear_118_cast_fp16)[name = string("op_6500_cast_fp16")];
            tensor<int32, [4]> x_237_perm_0 = const()[name = string("x_237_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77449472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77711680))))[name = string("tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6511 = const()[name = string("op_6511"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_6512_cast_fp16 = reshape(shape = var_6511, x = linear_119_cast_fp16)[name = string("op_6512_cast_fp16")];
            tensor<int32, [4]> x_239_perm_0 = const()[name = string("x_239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77712768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77974976))))[name = string("tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6523 = const()[name = string("op_6523"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_6524_cast_fp16 = reshape(shape = var_6523, x = linear_120_cast_fp16)[name = string("op_6524_cast_fp16")];
            tensor<int32, [4]> x_241_perm_0 = const()[name = string("x_241_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_237_cast_fp16 = transpose(perm = x_237_perm_0, x = var_6500_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 4, 48, 256]> var_6528_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_6528_cast_fp16")];
            bool var_6531 = const()[name = string("op_6531"), val = bool(true)];
            tensor<int32, [1]> var_6532_axes_0 = const()[name = string("op_6532_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_6532_cast_fp16 = reduce_max(axes = var_6532_axes_0, keep_dims = var_6531, x = var_6528_cast_fp16)[name = string("op_6532_cast_fp16")];
            fp16 var_6533_to_fp16 = const()[name = string("op_6533_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_213_cast_fp16 = maximum(x = var_6532_cast_fp16, y = var_6533_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_213_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_213_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6536_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6536_cast_fp16")];
            tensor<int32, [1]> var_6541_axes_0 = const()[name = string("op_6541_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6541_keep_dims_0 = const()[name = string("op_6541_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_6541_cast_fp16 = reduce_mean(axes = var_6541_axes_0, keep_dims = var_6541_keep_dims_0, x = var_6536_cast_fp16)[name = string("op_6541_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_6542_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp32 var_6543_epsilon_0 = const()[name = string("op_6543_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_6543_cast_fp16 = inverse(epsilon = var_6543_epsilon_0, x = var_6542_cast_fp16)[name = string("op_6543_cast_fp16")];
            fp16 var_6544_to_fp16 = const()[name = string("op_6544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_6545_cast_fp16 = mul(x = var_6543_cast_fp16, y = var_6544_to_fp16)[name = string("op_6545_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_6547_cast_fp16 = add(x = var_6541_cast_fp16, y = var_6545_cast_fp16)[name = string("op_6547_cast_fp16")];
            fp32 var_6548_epsilon_0 = const()[name = string("op_6548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_6548_cast_fp16 = rsqrt(epsilon = var_6548_epsilon_0, x = var_6547_cast_fp16)[name = string("op_6548_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_187_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6548_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976064)))];
            tensor<fp16, [1, 4, 48, 256]> q_53_cast_fp16 = mul(x = out_187_cast_fp16, y = tm_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("q_53_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_239_cast_fp16 = transpose(perm = x_239_perm_0, x = var_6512_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2, 48, 256]> var_6551_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_6551_cast_fp16")];
            bool var_6554 = const()[name = string("op_6554"), val = bool(true)];
            tensor<int32, [1]> var_6555_axes_0 = const()[name = string("op_6555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_6555_cast_fp16 = reduce_max(axes = var_6555_axes_0, keep_dims = var_6554, x = var_6551_cast_fp16)[name = string("op_6555_cast_fp16")];
            fp16 var_6556_to_fp16 = const()[name = string("op_6556_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_215_cast_fp16 = maximum(x = var_6555_cast_fp16, y = var_6556_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_215_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_215_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6559_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6559_cast_fp16")];
            tensor<int32, [1]> var_6564_axes_0 = const()[name = string("op_6564_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6564_keep_dims_0 = const()[name = string("op_6564_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_6564_cast_fp16 = reduce_mean(axes = var_6564_axes_0, keep_dims = var_6564_keep_dims_0, x = var_6559_cast_fp16)[name = string("op_6564_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6565_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6565_cast_fp16")];
            fp32 var_6566_epsilon_0 = const()[name = string("op_6566_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_6566_cast_fp16 = inverse(epsilon = var_6566_epsilon_0, x = var_6565_cast_fp16)[name = string("op_6566_cast_fp16")];
            fp16 var_6567_to_fp16 = const()[name = string("op_6567_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_6568_cast_fp16 = mul(x = var_6566_cast_fp16, y = var_6567_to_fp16)[name = string("op_6568_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6570_cast_fp16 = add(x = var_6564_cast_fp16, y = var_6568_cast_fp16)[name = string("op_6570_cast_fp16")];
            fp32 var_6571_epsilon_0 = const()[name = string("op_6571_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_6571_cast_fp16 = rsqrt(epsilon = var_6571_epsilon_0, x = var_6570_cast_fp16)[name = string("op_6571_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_189_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6571_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976640)))];
            tensor<fp16, [1, 2, 48, 256]> k_79_cast_fp16 = mul(x = out_189_cast_fp16, y = tm_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("k_79_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_241_cast_fp16 = transpose(perm = x_241_perm_0, x = var_6524_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 2, 48, 256]> var_6574_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_6574_cast_fp16")];
            bool var_6577 = const()[name = string("op_6577"), val = bool(true)];
            tensor<int32, [1]> var_6578_axes_0 = const()[name = string("op_6578_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_6578_cast_fp16 = reduce_max(axes = var_6578_axes_0, keep_dims = var_6577, x = var_6574_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_217_cast_fp16 = maximum(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_217_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_217_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6582_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6587_axes_0 = const()[name = string("op_6587_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6587_keep_dims_0 = const()[name = string("op_6587_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_6587_cast_fp16 = reduce_mean(axes = var_6587_axes_0, keep_dims = var_6587_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6588_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6588_cast_fp16")];
            fp32 var_6589_epsilon_0 = const()[name = string("op_6589_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_6589_cast_fp16 = inverse(epsilon = var_6589_epsilon_0, x = var_6588_cast_fp16)[name = string("op_6589_cast_fp16")];
            fp16 var_6590_to_fp16 = const()[name = string("op_6590_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_6591_cast_fp16 = mul(x = var_6589_cast_fp16, y = var_6590_to_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6593_cast_fp16 = add(x = var_6587_cast_fp16, y = var_6591_cast_fp16)[name = string("op_6593_cast_fp16")];
            fp32 var_6594_epsilon_0 = const()[name = string("op_6594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_6594_cast_fp16 = rsqrt(epsilon = var_6594_epsilon_0, x = var_6593_cast_fp16)[name = string("op_6594_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_53_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6594_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6596_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6596_cast_fp16")];
            tensor<int32, [4]> var_6601_begin_0 = const()[name = string("op_6601_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6601_end_0 = const()[name = string("op_6601_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_6601_end_mask_0 = const()[name = string("op_6601_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_6601_cast_fp16 = slice_by_index(begin = var_6601_begin_0, end = var_6601_end_0, end_mask = var_6601_end_mask_0, x = q_53_cast_fp16)[name = string("op_6601_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_6602_cast_fp16 = mul(x = var_6601_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_6602_cast_fp16")];
            tensor<int32, [4]> var_6607_begin_0 = const()[name = string("op_6607_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6607_end_0 = const()[name = string("op_6607_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_6607_end_mask_0 = const()[name = string("op_6607_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_6607_cast_fp16 = slice_by_index(begin = var_6607_begin_0, end = var_6607_end_0, end_mask = var_6607_end_mask_0, x = q_53_cast_fp16)[name = string("op_6607_cast_fp16")];
            int32 var_6609 = const()[name = string("op_6609"), val = int32(-1)];
            bool var_6610_interleave_0 = const()[name = string("op_6610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_6610_cast_fp16 = concat(axis = var_6609, interleave = var_6610_interleave_0, values = (var_6602_cast_fp16, var_6607_cast_fp16))[name = string("op_6610_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6611_cast_fp16 = mul(x = var_6610_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_55_cast_fp16 = add(x = var_6596_cast_fp16, y = var_6611_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6614_cast_fp16 = mul(x = k_79_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = k_79_cast_fp16)[name = string("op_6619_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_6620_cast_fp16 = mul(x = var_6619_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_6620_cast_fp16")];
            tensor<int32, [4]> var_6625_begin_0 = const()[name = string("op_6625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6625_end_0 = const()[name = string("op_6625_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_6625_end_mask_0 = const()[name = string("op_6625_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_6625_cast_fp16 = slice_by_index(begin = var_6625_begin_0, end = var_6625_end_0, end_mask = var_6625_end_mask_0, x = k_79_cast_fp16)[name = string("op_6625_cast_fp16")];
            int32 var_6627 = const()[name = string("op_6627"), val = int32(-1)];
            bool var_6628_interleave_0 = const()[name = string("op_6628_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_6628_cast_fp16 = concat(axis = var_6627, interleave = var_6628_interleave_0, values = (var_6620_cast_fp16, var_6625_cast_fp16))[name = string("op_6628_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6629_cast_fp16 = mul(x = var_6628_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6629_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_81_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6629_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_26_reps_0 = const()[name = string("tile_26_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = k_81_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [4, 1, 48, 256]> tile_26_cast_fp16 = tile(reps = tile_26_reps_0, x = transpose_52_cast_fp16)[name = string("tile_26_cast_fp16")];
            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_52_cast_fp16 = reshape(shape = concat_52, x = tile_26_cast_fp16)[name = string("reshape_52_cast_fp16")];
            tensor<int32, [5]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = reshape_52_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [4, 1, 48, 256]> reshape_53_cast_fp16 = reshape(shape = concat_53, x = transpose_53_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_27_reps_0 = const()[name = string("tile_27_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = v_53_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [4, 1, 48, 256]> tile_27_cast_fp16 = tile(reps = tile_27_reps_0, x = transpose_54_cast_fp16)[name = string("tile_27_cast_fp16")];
            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_54_cast_fp16 = reshape(shape = concat_54, x = tile_27_cast_fp16)[name = string("reshape_54_cast_fp16")];
            tensor<int32, [5]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = reshape_54_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [4, 1, 48, 256]> reshape_55_cast_fp16 = reshape(shape = concat_55, x = transpose_55_cast_fp16)[name = string("reshape_55_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6643_transpose_x_0 = const()[name = string("op_6643_transpose_x_0"), val = bool(false)];
            bool var_6643_transpose_y_0 = const()[name = string("op_6643_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 4, 48, 48]> var_6643_cast_fp16 = matmul(transpose_x = var_6643_transpose_x_0, transpose_y = var_6643_transpose_y_0, x = q_55_cast_fp16, y = transpose_109_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_6645_cast_fp16 = add(x = var_6643_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6646 = const()[name = string("op_6646"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_217_cast_fp16 = softmax(axis = var_6646, x = var_6645_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6649_transpose_x_0 = const()[name = string("op_6649_transpose_x_0"), val = bool(false)];
            bool var_6649_transpose_y_0 = const()[name = string("op_6649_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = reshape_55_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 4, 48, 256]> var_6649_cast_fp16 = matmul(transpose_x = var_6649_transpose_x_0, transpose_y = var_6649_transpose_y_0, x = w_217_cast_fp16, y = v_55_cast_fp16)[name = string("op_6649_cast_fp16")];
            tensor<int32, [4]> var_6652_perm_0 = const()[name = string("op_6652_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6656 = const()[name = string("op_6656"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_6652_cast_fp16 = transpose(perm = var_6652_perm_0, x = var_6649_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 48, 1024]> input_159_cast_fp16 = reshape(shape = var_6656, x = var_6652_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77977216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78501568))))[name = string("tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6661_cast_fp16 = abs(x = linear_121_cast_fp16)[name = string("op_6661_cast_fp16")];
            bool var_6664 = const()[name = string("op_6664"), val = bool(true)];
            tensor<int32, [1]> var_6665_axes_0 = const()[name = string("op_6665_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6665_cast_fp16 = reduce_max(axes = var_6665_axes_0, keep_dims = var_6664, x = var_6661_cast_fp16)[name = string("op_6665_cast_fp16")];
            fp16 var_6666_to_fp16 = const()[name = string("op_6666_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_219_cast_fp16 = maximum(x = var_6665_cast_fp16, y = var_6666_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_219_cast_fp16 = real_div(x = linear_121_cast_fp16, y = s_219_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6669_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6669_cast_fp16")];
            tensor<int32, [1]> var_6674_axes_0 = const()[name = string("op_6674_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6674_keep_dims_0 = const()[name = string("op_6674_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6674_cast_fp16 = reduce_mean(axes = var_6674_axes_0, keep_dims = var_6674_keep_dims_0, x = var_6669_cast_fp16)[name = string("op_6674_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6675_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6675_cast_fp16")];
            fp32 var_6676_epsilon_0 = const()[name = string("op_6676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6676_cast_fp16 = inverse(epsilon = var_6676_epsilon_0, x = var_6675_cast_fp16)[name = string("op_6676_cast_fp16")];
            fp16 var_6677_to_fp16 = const()[name = string("op_6677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6678_cast_fp16 = mul(x = var_6676_cast_fp16, y = var_6677_to_fp16)[name = string("op_6678_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6680_cast_fp16 = add(x = var_6674_cast_fp16, y = var_6678_cast_fp16)[name = string("op_6680_cast_fp16")];
            fp32 var_6681_epsilon_0 = const()[name = string("op_6681_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6681_cast_fp16 = rsqrt(epsilon = var_6681_epsilon_0, x = var_6680_cast_fp16)[name = string("op_6681_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_191_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6681_cast_fp16)[name = string("out_191_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78502656)))];
            tensor<fp16, [1, 48, 512]> var_6683_cast_fp16 = mul(x = out_191_cast_fp16, y = tm_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6683_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_245_cast_fp16 = add(x = x_235_cast_fp16, y = var_6683_cast_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6686_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_6686_cast_fp16")];
            bool var_6689 = const()[name = string("op_6689"), val = bool(true)];
            tensor<int32, [1]> var_6690_axes_0 = const()[name = string("op_6690_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6690_cast_fp16 = reduce_max(axes = var_6690_axes_0, keep_dims = var_6689, x = var_6686_cast_fp16)[name = string("op_6690_cast_fp16")];
            fp16 var_6691_to_fp16 = const()[name = string("op_6691_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_221_cast_fp16 = maximum(x = var_6690_cast_fp16, y = var_6691_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_221_cast_fp16 = real_div(x = x_245_cast_fp16, y = s_221_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6694_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6694_cast_fp16")];
            tensor<int32, [1]> var_6699_axes_0 = const()[name = string("op_6699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6699_keep_dims_0 = const()[name = string("op_6699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6699_cast_fp16 = reduce_mean(axes = var_6699_axes_0, keep_dims = var_6699_keep_dims_0, x = var_6694_cast_fp16)[name = string("op_6699_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6700_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6700_cast_fp16")];
            fp32 var_6701_epsilon_0 = const()[name = string("op_6701_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6701_cast_fp16 = inverse(epsilon = var_6701_epsilon_0, x = var_6700_cast_fp16)[name = string("op_6701_cast_fp16")];
            fp16 var_6702_to_fp16 = const()[name = string("op_6702_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6703_cast_fp16 = mul(x = var_6701_cast_fp16, y = var_6702_to_fp16)[name = string("op_6703_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6705_cast_fp16 = add(x = var_6699_cast_fp16, y = var_6703_cast_fp16)[name = string("op_6705_cast_fp16")];
            fp32 var_6706_epsilon_0 = const()[name = string("op_6706_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6706_cast_fp16 = rsqrt(epsilon = var_6706_epsilon_0, x = var_6705_cast_fp16)[name = string("op_6706_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_193_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6706_cast_fp16)[name = string("out_193_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78503744)))];
            tensor<fp16, [1, 48, 512]> input_161_cast_fp16 = mul(x = out_193_cast_fp16, y = tm_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78504832))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79553472))))[name = string("tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_122_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_122_cast_fp16")];
            string var_6713_mode_0 = const()[name = string("op_6713_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_6713_cast_fp16 = gelu(mode = var_6713_mode_0, x = linear_122_cast_fp16)[name = string("op_6713_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79557632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80606272))))[name = string("tm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_123_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_163_cast_fp16 = mul(x = var_6713_cast_fp16, y = linear_123_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80610432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81659072))))[name = string("tm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6721_cast_fp16 = abs(x = linear_124_cast_fp16)[name = string("op_6721_cast_fp16")];
            bool var_6724 = const()[name = string("op_6724"), val = bool(true)];
            tensor<int32, [1]> var_6725_axes_0 = const()[name = string("op_6725_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6725_cast_fp16 = reduce_max(axes = var_6725_axes_0, keep_dims = var_6724, x = var_6721_cast_fp16)[name = string("op_6725_cast_fp16")];
            fp16 var_6726_to_fp16 = const()[name = string("op_6726_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_223_cast_fp16 = maximum(x = var_6725_cast_fp16, y = var_6726_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_223_cast_fp16 = real_div(x = linear_124_cast_fp16, y = s_223_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6729_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [1]> var_6734_axes_0 = const()[name = string("op_6734_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6734_keep_dims_0 = const()[name = string("op_6734_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6734_cast_fp16 = reduce_mean(axes = var_6734_axes_0, keep_dims = var_6734_keep_dims_0, x = var_6729_cast_fp16)[name = string("op_6734_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6735_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6735_cast_fp16")];
            fp32 var_6736_epsilon_0 = const()[name = string("op_6736_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6736_cast_fp16 = inverse(epsilon = var_6736_epsilon_0, x = var_6735_cast_fp16)[name = string("op_6736_cast_fp16")];
            fp16 var_6737_to_fp16 = const()[name = string("op_6737_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6738_cast_fp16 = mul(x = var_6736_cast_fp16, y = var_6737_to_fp16)[name = string("op_6738_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6740_cast_fp16 = add(x = var_6734_cast_fp16, y = var_6738_cast_fp16)[name = string("op_6740_cast_fp16")];
            fp32 var_6741_epsilon_0 = const()[name = string("op_6741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6741_cast_fp16 = rsqrt(epsilon = var_6741_epsilon_0, x = var_6740_cast_fp16)[name = string("op_6741_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_195_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6741_cast_fp16)[name = string("out_195_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81660160)))];
            tensor<fp16, [1, 48, 512]> var_6743_cast_fp16 = mul(x = out_195_cast_fp16, y = tm_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6743_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_165_cast_fp16 = add(x = x_245_cast_fp16, y = var_6743_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81661248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81923456))))[name = string("tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_125_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_125_cast_fp16")];
            string x_249_mode_0 = const()[name = string("x_249_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_249_cast_fp16 = gelu(mode = x_249_mode_0, x = linear_125_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<int32, [4]> var_6763_begin_0 = const()[name = string("op_6763_begin_0"), val = tensor<int32, [4]>([0, 0, 13, 0])];
            tensor<int32, [4]> var_6763_end_0 = const()[name = string("op_6763_end_0"), val = tensor<int32, [4]>([1, 48, 14, 512])];
            tensor<bool, [4]> var_6763_end_mask_0 = const()[name = string("op_6763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6763_squeeze_mask_0 = const()[name = string("op_6763_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_6763_cast_fp16 = slice_by_index(begin = var_6763_begin_0, end = var_6763_end_0, end_mask = var_6763_end_mask_0, squeeze_mask = var_6763_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_167_cast_fp16 = mul(x = x_249_cast_fp16, y = var_6763_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81924544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82186752))))[name = string("tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6773_cast_fp16 = abs(x = linear_126_cast_fp16)[name = string("op_6773_cast_fp16")];
            bool var_6776 = const()[name = string("op_6776"), val = bool(true)];
            tensor<int32, [1]> var_6777_axes_0 = const()[name = string("op_6777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6777_cast_fp16 = reduce_max(axes = var_6777_axes_0, keep_dims = var_6776, x = var_6773_cast_fp16)[name = string("op_6777_cast_fp16")];
            fp16 var_6778_to_fp16 = const()[name = string("op_6778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_225_cast_fp16 = maximum(x = var_6777_cast_fp16, y = var_6778_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_225_cast_fp16 = real_div(x = linear_126_cast_fp16, y = s_225_cast_fp16)[name = string("y_225_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6781_cast_fp16 = mul(x = y_225_cast_fp16, y = y_225_cast_fp16)[name = string("op_6781_cast_fp16")];
            tensor<int32, [1]> var_6786_axes_0 = const()[name = string("op_6786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6786_keep_dims_0 = const()[name = string("op_6786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6786_cast_fp16 = reduce_mean(axes = var_6786_axes_0, keep_dims = var_6786_keep_dims_0, x = var_6781_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6787_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6787_cast_fp16")];
            fp32 var_6788_epsilon_0 = const()[name = string("op_6788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6788_cast_fp16 = inverse(epsilon = var_6788_epsilon_0, x = var_6787_cast_fp16)[name = string("op_6788_cast_fp16")];
            fp16 var_6789_to_fp16 = const()[name = string("op_6789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6790_cast_fp16 = mul(x = var_6788_cast_fp16, y = var_6789_to_fp16)[name = string("op_6790_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6792_cast_fp16 = add(x = var_6786_cast_fp16, y = var_6790_cast_fp16)[name = string("op_6792_cast_fp16")];
            fp32 var_6793_epsilon_0 = const()[name = string("op_6793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6793_cast_fp16 = rsqrt(epsilon = var_6793_epsilon_0, x = var_6792_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_197_cast_fp16 = mul(x = y_225_cast_fp16, y = var_6793_cast_fp16)[name = string("out_197_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82187840)))];
            tensor<fp16, [1, 48, 512]> var_6795_cast_fp16 = mul(x = out_197_cast_fp16, y = tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6795_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_27_cast_fp16 = add(x = input_165_cast_fp16, y = var_6795_cast_fp16)[name = string("h_27_cast_fp16")];
            tensor<fp16, [1]> tm_layers_13_layer_scalar_to_fp16 = const()[name = string("tm_layers_13_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9cp-1])];
            tensor<fp16, [1, 48, 512]> x_253_cast_fp16 = mul(x = h_27_cast_fp16, y = tm_layers_13_layer_scalar_to_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6799_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6799_cast_fp16")];
            bool var_6802 = const()[name = string("op_6802"), val = bool(true)];
            tensor<int32, [1]> var_6803_axes_0 = const()[name = string("op_6803_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6803_cast_fp16 = reduce_max(axes = var_6803_axes_0, keep_dims = var_6802, x = var_6799_cast_fp16)[name = string("op_6803_cast_fp16")];
            fp16 var_6804_to_fp16 = const()[name = string("op_6804_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_227_cast_fp16 = maximum(x = var_6803_cast_fp16, y = var_6804_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_227_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_227_cast_fp16)[name = string("y_227_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6807_cast_fp16 = mul(x = y_227_cast_fp16, y = y_227_cast_fp16)[name = string("op_6807_cast_fp16")];
            tensor<int32, [1]> var_6812_axes_0 = const()[name = string("op_6812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6812_keep_dims_0 = const()[name = string("op_6812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_6812_cast_fp16 = reduce_mean(axes = var_6812_axes_0, keep_dims = var_6812_keep_dims_0, x = var_6807_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6813_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6813_cast_fp16")];
            fp32 var_6814_epsilon_0 = const()[name = string("op_6814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_6814_cast_fp16 = inverse(epsilon = var_6814_epsilon_0, x = var_6813_cast_fp16)[name = string("op_6814_cast_fp16")];
            fp16 var_6815_to_fp16 = const()[name = string("op_6815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_6816_cast_fp16 = mul(x = var_6814_cast_fp16, y = var_6815_to_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_6818_cast_fp16 = add(x = var_6812_cast_fp16, y = var_6816_cast_fp16)[name = string("op_6818_cast_fp16")];
            fp32 var_6819_epsilon_0 = const()[name = string("op_6819_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_6819_cast_fp16 = rsqrt(epsilon = var_6819_epsilon_0, x = var_6818_cast_fp16)[name = string("op_6819_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_199_cast_fp16 = mul(x = y_227_cast_fp16, y = var_6819_cast_fp16)[name = string("out_199_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82188928)))];
            tensor<fp16, [1, 48, 512]> input_169_cast_fp16 = mul(x = out_199_cast_fp16, y = tm_layers_14_input_layernorm_weight_to_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82190016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82714368))))[name = string("tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6829 = const()[name = string("op_6829"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_6830_cast_fp16 = reshape(shape = var_6829, x = linear_127_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82716480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82978688))))[name = string("tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6841 = const()[name = string("op_6841"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_6842_cast_fp16 = reshape(shape = var_6841, x = linear_128_cast_fp16)[name = string("op_6842_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82979776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83241984))))[name = string("tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_6853 = const()[name = string("op_6853"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_6854_cast_fp16 = reshape(shape = var_6853, x = linear_129_cast_fp16)[name = string("op_6854_cast_fp16")];
            tensor<int32, [4]> x_259_perm_0 = const()[name = string("x_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_6830_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 4, 48, 256]> var_6858_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6858_cast_fp16")];
            bool var_6861 = const()[name = string("op_6861"), val = bool(true)];
            tensor<int32, [1]> var_6862_axes_0 = const()[name = string("op_6862_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_6862_cast_fp16 = reduce_max(axes = var_6862_axes_0, keep_dims = var_6861, x = var_6858_cast_fp16)[name = string("op_6862_cast_fp16")];
            fp16 var_6863_to_fp16 = const()[name = string("op_6863_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_229_cast_fp16 = maximum(x = var_6862_cast_fp16, y = var_6863_to_fp16)[name = string("s_229_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_229_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_229_cast_fp16)[name = string("y_229_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6866_cast_fp16 = mul(x = y_229_cast_fp16, y = y_229_cast_fp16)[name = string("op_6866_cast_fp16")];
            tensor<int32, [1]> var_6871_axes_0 = const()[name = string("op_6871_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6871_keep_dims_0 = const()[name = string("op_6871_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_6871_cast_fp16 = reduce_mean(axes = var_6871_axes_0, keep_dims = var_6871_keep_dims_0, x = var_6866_cast_fp16)[name = string("op_6871_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_6872_cast_fp16 = mul(x = s_229_cast_fp16, y = s_229_cast_fp16)[name = string("op_6872_cast_fp16")];
            fp32 var_6873_epsilon_0 = const()[name = string("op_6873_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_6873_cast_fp16 = inverse(epsilon = var_6873_epsilon_0, x = var_6872_cast_fp16)[name = string("op_6873_cast_fp16")];
            fp16 var_6874_to_fp16 = const()[name = string("op_6874_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_6875_cast_fp16 = mul(x = var_6873_cast_fp16, y = var_6874_to_fp16)[name = string("op_6875_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_6877_cast_fp16 = add(x = var_6871_cast_fp16, y = var_6875_cast_fp16)[name = string("op_6877_cast_fp16")];
            fp32 var_6878_epsilon_0 = const()[name = string("op_6878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_6878_cast_fp16 = rsqrt(epsilon = var_6878_epsilon_0, x = var_6877_cast_fp16)[name = string("op_6878_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_201_cast_fp16 = mul(x = y_229_cast_fp16, y = var_6878_cast_fp16)[name = string("out_201_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243072)))];
            tensor<fp16, [1, 4, 48, 256]> q_57_cast_fp16 = mul(x = out_201_cast_fp16, y = tm_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("q_57_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_6842_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, 48, 256]> var_6881_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6884 = const()[name = string("op_6884"), val = bool(true)];
            tensor<int32, [1]> var_6885_axes_0 = const()[name = string("op_6885_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_6885_cast_fp16 = reduce_max(axes = var_6885_axes_0, keep_dims = var_6884, x = var_6881_cast_fp16)[name = string("op_6885_cast_fp16")];
            fp16 var_6886_to_fp16 = const()[name = string("op_6886_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_231_cast_fp16 = maximum(x = var_6885_cast_fp16, y = var_6886_to_fp16)[name = string("s_231_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_231_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_231_cast_fp16)[name = string("y_231_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6889_cast_fp16 = mul(x = y_231_cast_fp16, y = y_231_cast_fp16)[name = string("op_6889_cast_fp16")];
            tensor<int32, [1]> var_6894_axes_0 = const()[name = string("op_6894_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6894_keep_dims_0 = const()[name = string("op_6894_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_6894_cast_fp16 = reduce_mean(axes = var_6894_axes_0, keep_dims = var_6894_keep_dims_0, x = var_6889_cast_fp16)[name = string("op_6894_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6895_cast_fp16 = mul(x = s_231_cast_fp16, y = s_231_cast_fp16)[name = string("op_6895_cast_fp16")];
            fp32 var_6896_epsilon_0 = const()[name = string("op_6896_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_6896_cast_fp16 = inverse(epsilon = var_6896_epsilon_0, x = var_6895_cast_fp16)[name = string("op_6896_cast_fp16")];
            fp16 var_6897_to_fp16 = const()[name = string("op_6897_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_6898_cast_fp16 = mul(x = var_6896_cast_fp16, y = var_6897_to_fp16)[name = string("op_6898_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6900_cast_fp16 = add(x = var_6894_cast_fp16, y = var_6898_cast_fp16)[name = string("op_6900_cast_fp16")];
            fp32 var_6901_epsilon_0 = const()[name = string("op_6901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_6901_cast_fp16 = rsqrt(epsilon = var_6901_epsilon_0, x = var_6900_cast_fp16)[name = string("op_6901_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_203_cast_fp16 = mul(x = y_231_cast_fp16, y = var_6901_cast_fp16)[name = string("out_203_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243648)))];
            tensor<fp16, [1, 2, 48, 256]> k_85_cast_fp16 = mul(x = out_203_cast_fp16, y = tm_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("k_85_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_259_cast_fp16 = transpose(perm = x_259_perm_0, x = var_6854_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 2, 48, 256]> var_6904_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6904_cast_fp16")];
            bool var_6907 = const()[name = string("op_6907"), val = bool(true)];
            tensor<int32, [1]> var_6908_axes_0 = const()[name = string("op_6908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_6908_cast_fp16 = reduce_max(axes = var_6908_axes_0, keep_dims = var_6907, x = var_6904_cast_fp16)[name = string("op_6908_cast_fp16")];
            fp16 var_6909_to_fp16 = const()[name = string("op_6909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_233_cast_fp16 = maximum(x = var_6908_cast_fp16, y = var_6909_to_fp16)[name = string("s_233_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_233_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_233_cast_fp16)[name = string("y_233_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6912_cast_fp16 = mul(x = y_233_cast_fp16, y = y_233_cast_fp16)[name = string("op_6912_cast_fp16")];
            tensor<int32, [1]> var_6917_axes_0 = const()[name = string("op_6917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6917_keep_dims_0 = const()[name = string("op_6917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_6917_cast_fp16 = reduce_mean(axes = var_6917_axes_0, keep_dims = var_6917_keep_dims_0, x = var_6912_cast_fp16)[name = string("op_6917_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6918_cast_fp16 = mul(x = s_233_cast_fp16, y = s_233_cast_fp16)[name = string("op_6918_cast_fp16")];
            fp32 var_6919_epsilon_0 = const()[name = string("op_6919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_6919_cast_fp16 = inverse(epsilon = var_6919_epsilon_0, x = var_6918_cast_fp16)[name = string("op_6919_cast_fp16")];
            fp16 var_6920_to_fp16 = const()[name = string("op_6920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_6921_cast_fp16 = mul(x = var_6919_cast_fp16, y = var_6920_to_fp16)[name = string("op_6921_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_6923_cast_fp16 = add(x = var_6917_cast_fp16, y = var_6921_cast_fp16)[name = string("op_6923_cast_fp16")];
            fp32 var_6924_epsilon_0 = const()[name = string("op_6924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_6924_cast_fp16 = rsqrt(epsilon = var_6924_epsilon_0, x = var_6923_cast_fp16)[name = string("op_6924_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_57_cast_fp16 = mul(x = y_233_cast_fp16, y = var_6924_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6926_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6926_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = q_57_cast_fp16)[name = string("op_6931_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_6932_cast_fp16 = mul(x = var_6931_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_6932_cast_fp16")];
            tensor<int32, [4]> var_6937_begin_0 = const()[name = string("op_6937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6937_end_0 = const()[name = string("op_6937_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_6937_end_mask_0 = const()[name = string("op_6937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_6937_cast_fp16 = slice_by_index(begin = var_6937_begin_0, end = var_6937_end_0, end_mask = var_6937_end_mask_0, x = q_57_cast_fp16)[name = string("op_6937_cast_fp16")];
            int32 var_6939 = const()[name = string("op_6939"), val = int32(-1)];
            bool var_6940_interleave_0 = const()[name = string("op_6940_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_6940_cast_fp16 = concat(axis = var_6939, interleave = var_6940_interleave_0, values = (var_6932_cast_fp16, var_6937_cast_fp16))[name = string("op_6940_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_6941_cast_fp16 = mul(x = var_6940_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6941_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_59_cast_fp16 = add(x = var_6926_cast_fp16, y = var_6941_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6944_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6944_cast_fp16")];
            tensor<int32, [4]> var_6949_begin_0 = const()[name = string("op_6949_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6949_end_0 = const()[name = string("op_6949_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_6949_end_mask_0 = const()[name = string("op_6949_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_6949_cast_fp16 = slice_by_index(begin = var_6949_begin_0, end = var_6949_end_0, end_mask = var_6949_end_mask_0, x = k_85_cast_fp16)[name = string("op_6949_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_6950_cast_fp16 = mul(x = var_6949_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_6950_cast_fp16")];
            tensor<int32, [4]> var_6955_begin_0 = const()[name = string("op_6955_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6955_end_0 = const()[name = string("op_6955_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_6955_end_mask_0 = const()[name = string("op_6955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_6955_cast_fp16 = slice_by_index(begin = var_6955_begin_0, end = var_6955_end_0, end_mask = var_6955_end_mask_0, x = k_85_cast_fp16)[name = string("op_6955_cast_fp16")];
            int32 var_6957 = const()[name = string("op_6957"), val = int32(-1)];
            bool var_6958_interleave_0 = const()[name = string("op_6958_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_6958_cast_fp16 = concat(axis = var_6957, interleave = var_6958_interleave_0, values = (var_6950_cast_fp16, var_6955_cast_fp16))[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_6959_cast_fp16 = mul(x = var_6958_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_87_cast_fp16 = add(x = var_6944_cast_fp16, y = var_6959_cast_fp16)[name = string("k_87_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_28_reps_0 = const()[name = string("tile_28_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = k_87_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 1, 48, 256]> tile_28_cast_fp16 = tile(reps = tile_28_reps_0, x = transpose_56_cast_fp16)[name = string("tile_28_cast_fp16")];
            tensor<int32, [5]> concat_56 = const()[name = string("concat_56"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_56_cast_fp16 = reshape(shape = concat_56, x = tile_28_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [5]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_57 = const()[name = string("concat_57"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 1, 48, 256]> reshape_57_cast_fp16 = reshape(shape = concat_57, x = transpose_57_cast_fp16)[name = string("reshape_57_cast_fp16")];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_29_reps_0 = const()[name = string("tile_29_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = v_57_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 1, 48, 256]> tile_29_cast_fp16 = tile(reps = tile_29_reps_0, x = transpose_58_cast_fp16)[name = string("tile_29_cast_fp16")];
            tensor<int32, [5]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_58_cast_fp16 = reshape(shape = concat_58, x = tile_29_cast_fp16)[name = string("reshape_58_cast_fp16")];
            tensor<int32, [5]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = reshape_58_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 1, 48, 256]> reshape_59_cast_fp16 = reshape(shape = concat_59, x = transpose_59_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6973_transpose_x_0 = const()[name = string("op_6973_transpose_x_0"), val = bool(false)];
            bool var_6973_transpose_y_0 = const()[name = string("op_6973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_110_cast_fp16 = transpose(perm = transpose_110_perm_0, x = reshape_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 4, 48, 48]> var_6973_cast_fp16 = matmul(transpose_x = var_6973_transpose_x_0, transpose_y = var_6973_transpose_y_0, x = q_59_cast_fp16, y = transpose_110_cast_fp16)[name = string("op_6973_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_6975_cast_fp16 = add(x = var_6973_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6976 = const()[name = string("op_6976"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_233_cast_fp16 = softmax(axis = var_6976, x = var_6975_cast_fp16)[name = string("w_233_cast_fp16")];
            bool var_6979_transpose_x_0 = const()[name = string("op_6979_transpose_x_0"), val = bool(false)];
            bool var_6979_transpose_y_0 = const()[name = string("op_6979_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 4, 48, 256]> var_6979_cast_fp16 = matmul(transpose_x = var_6979_transpose_x_0, transpose_y = var_6979_transpose_y_0, x = w_233_cast_fp16, y = v_59_cast_fp16)[name = string("op_6979_cast_fp16")];
            tensor<int32, [4]> var_6982_perm_0 = const()[name = string("op_6982_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6986 = const()[name = string("op_6986"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_6982_cast_fp16 = transpose(perm = var_6982_perm_0, x = var_6979_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 48, 1024]> input_171_cast_fp16 = reshape(shape = var_6986, x = var_6982_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83244224))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83768576))))[name = string("tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6991_cast_fp16 = abs(x = linear_130_cast_fp16)[name = string("op_6991_cast_fp16")];
            bool var_6994 = const()[name = string("op_6994"), val = bool(true)];
            tensor<int32, [1]> var_6995_axes_0 = const()[name = string("op_6995_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_6995_cast_fp16 = reduce_max(axes = var_6995_axes_0, keep_dims = var_6994, x = var_6991_cast_fp16)[name = string("op_6995_cast_fp16")];
            fp16 var_6996_to_fp16 = const()[name = string("op_6996_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_235_cast_fp16 = maximum(x = var_6995_cast_fp16, y = var_6996_to_fp16)[name = string("s_235_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_235_cast_fp16 = real_div(x = linear_130_cast_fp16, y = s_235_cast_fp16)[name = string("y_235_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_6999_cast_fp16 = mul(x = y_235_cast_fp16, y = y_235_cast_fp16)[name = string("op_6999_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_6999_cast_fp16)[name = string("op_7004_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7005_cast_fp16 = mul(x = s_235_cast_fp16, y = s_235_cast_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7010_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7010_cast_fp16")];
            fp32 var_7011_epsilon_0 = const()[name = string("op_7011_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7011_cast_fp16 = rsqrt(epsilon = var_7011_epsilon_0, x = var_7010_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_205_cast_fp16 = mul(x = y_235_cast_fp16, y = var_7011_cast_fp16)[name = string("out_205_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83769664)))];
            tensor<fp16, [1, 48, 512]> var_7013_cast_fp16 = mul(x = out_205_cast_fp16, y = tm_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_7013_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = var_7013_cast_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7016_cast_fp16 = abs(x = x_263_cast_fp16)[name = string("op_7016_cast_fp16")];
            bool var_7019 = const()[name = string("op_7019"), val = bool(true)];
            tensor<int32, [1]> var_7020_axes_0 = const()[name = string("op_7020_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7020_cast_fp16 = reduce_max(axes = var_7020_axes_0, keep_dims = var_7019, x = var_7016_cast_fp16)[name = string("op_7020_cast_fp16")];
            fp16 var_7021_to_fp16 = const()[name = string("op_7021_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_237_cast_fp16 = maximum(x = var_7020_cast_fp16, y = var_7021_to_fp16)[name = string("s_237_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_237_cast_fp16 = real_div(x = x_263_cast_fp16, y = s_237_cast_fp16)[name = string("y_237_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7024_cast_fp16 = mul(x = y_237_cast_fp16, y = y_237_cast_fp16)[name = string("op_7024_cast_fp16")];
            tensor<int32, [1]> var_7029_axes_0 = const()[name = string("op_7029_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7029_keep_dims_0 = const()[name = string("op_7029_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7029_cast_fp16 = reduce_mean(axes = var_7029_axes_0, keep_dims = var_7029_keep_dims_0, x = var_7024_cast_fp16)[name = string("op_7029_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7030_cast_fp16 = mul(x = s_237_cast_fp16, y = s_237_cast_fp16)[name = string("op_7030_cast_fp16")];
            fp32 var_7031_epsilon_0 = const()[name = string("op_7031_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7031_cast_fp16 = inverse(epsilon = var_7031_epsilon_0, x = var_7030_cast_fp16)[name = string("op_7031_cast_fp16")];
            fp16 var_7032_to_fp16 = const()[name = string("op_7032_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7033_cast_fp16 = mul(x = var_7031_cast_fp16, y = var_7032_to_fp16)[name = string("op_7033_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7035_cast_fp16 = add(x = var_7029_cast_fp16, y = var_7033_cast_fp16)[name = string("op_7035_cast_fp16")];
            fp32 var_7036_epsilon_0 = const()[name = string("op_7036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7036_cast_fp16 = rsqrt(epsilon = var_7036_epsilon_0, x = var_7035_cast_fp16)[name = string("op_7036_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_207_cast_fp16 = mul(x = y_237_cast_fp16, y = var_7036_cast_fp16)[name = string("out_207_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83770752)))];
            tensor<fp16, [1, 48, 512]> input_173_cast_fp16 = mul(x = out_207_cast_fp16, y = tm_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83771840))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84820480))))[name = string("tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_131_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_131_cast_fp16")];
            string var_7043_mode_0 = const()[name = string("op_7043_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_7043_cast_fp16 = gelu(mode = var_7043_mode_0, x = linear_131_cast_fp16)[name = string("op_7043_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84824640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85873280))))[name = string("tm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_132_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_175_cast_fp16 = mul(x = var_7043_cast_fp16, y = linear_132_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85877440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86926080))))[name = string("tm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7051_cast_fp16 = abs(x = linear_133_cast_fp16)[name = string("op_7051_cast_fp16")];
            bool var_7054 = const()[name = string("op_7054"), val = bool(true)];
            tensor<int32, [1]> var_7055_axes_0 = const()[name = string("op_7055_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7055_cast_fp16 = reduce_max(axes = var_7055_axes_0, keep_dims = var_7054, x = var_7051_cast_fp16)[name = string("op_7055_cast_fp16")];
            fp16 var_7056_to_fp16 = const()[name = string("op_7056_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_239_cast_fp16 = maximum(x = var_7055_cast_fp16, y = var_7056_to_fp16)[name = string("s_239_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_239_cast_fp16 = real_div(x = linear_133_cast_fp16, y = s_239_cast_fp16)[name = string("y_239_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7059_cast_fp16 = mul(x = y_239_cast_fp16, y = y_239_cast_fp16)[name = string("op_7059_cast_fp16")];
            tensor<int32, [1]> var_7064_axes_0 = const()[name = string("op_7064_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7064_keep_dims_0 = const()[name = string("op_7064_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7064_cast_fp16 = reduce_mean(axes = var_7064_axes_0, keep_dims = var_7064_keep_dims_0, x = var_7059_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7065_cast_fp16 = mul(x = s_239_cast_fp16, y = s_239_cast_fp16)[name = string("op_7065_cast_fp16")];
            fp32 var_7066_epsilon_0 = const()[name = string("op_7066_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7066_cast_fp16 = inverse(epsilon = var_7066_epsilon_0, x = var_7065_cast_fp16)[name = string("op_7066_cast_fp16")];
            fp16 var_7067_to_fp16 = const()[name = string("op_7067_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7068_cast_fp16 = mul(x = var_7066_cast_fp16, y = var_7067_to_fp16)[name = string("op_7068_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7070_cast_fp16 = add(x = var_7064_cast_fp16, y = var_7068_cast_fp16)[name = string("op_7070_cast_fp16")];
            fp32 var_7071_epsilon_0 = const()[name = string("op_7071_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7071_cast_fp16 = rsqrt(epsilon = var_7071_epsilon_0, x = var_7070_cast_fp16)[name = string("op_7071_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_209_cast_fp16 = mul(x = y_239_cast_fp16, y = var_7071_cast_fp16)[name = string("out_209_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86927168)))];
            tensor<fp16, [1, 48, 512]> var_7073_cast_fp16 = mul(x = out_209_cast_fp16, y = tm_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7073_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_177_cast_fp16 = add(x = x_263_cast_fp16, y = var_7073_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86928256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87190464))))[name = string("tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_134_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_134_cast_fp16")];
            string x_267_mode_0 = const()[name = string("x_267_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_267_cast_fp16 = gelu(mode = x_267_mode_0, x = linear_134_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<int32, [4]> var_7093_begin_0 = const()[name = string("op_7093_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
            tensor<int32, [4]> var_7093_end_0 = const()[name = string("op_7093_end_0"), val = tensor<int32, [4]>([1, 48, 15, 512])];
            tensor<bool, [4]> var_7093_end_mask_0 = const()[name = string("op_7093_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7093_squeeze_mask_0 = const()[name = string("op_7093_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_7093_cast_fp16 = slice_by_index(begin = var_7093_begin_0, end = var_7093_end_0, end_mask = var_7093_end_mask_0, squeeze_mask = var_7093_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7093_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_179_cast_fp16 = mul(x = x_267_cast_fp16, y = var_7093_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87191552))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87453760))))[name = string("tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7103_cast_fp16 = abs(x = linear_135_cast_fp16)[name = string("op_7103_cast_fp16")];
            bool var_7106 = const()[name = string("op_7106"), val = bool(true)];
            tensor<int32, [1]> var_7107_axes_0 = const()[name = string("op_7107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7107_cast_fp16 = reduce_max(axes = var_7107_axes_0, keep_dims = var_7106, x = var_7103_cast_fp16)[name = string("op_7107_cast_fp16")];
            fp16 var_7108_to_fp16 = const()[name = string("op_7108_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_241_cast_fp16 = maximum(x = var_7107_cast_fp16, y = var_7108_to_fp16)[name = string("s_241_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_241_cast_fp16 = real_div(x = linear_135_cast_fp16, y = s_241_cast_fp16)[name = string("y_241_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7111_cast_fp16 = mul(x = y_241_cast_fp16, y = y_241_cast_fp16)[name = string("op_7111_cast_fp16")];
            tensor<int32, [1]> var_7116_axes_0 = const()[name = string("op_7116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7116_keep_dims_0 = const()[name = string("op_7116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7116_cast_fp16 = reduce_mean(axes = var_7116_axes_0, keep_dims = var_7116_keep_dims_0, x = var_7111_cast_fp16)[name = string("op_7116_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7117_cast_fp16 = mul(x = s_241_cast_fp16, y = s_241_cast_fp16)[name = string("op_7117_cast_fp16")];
            fp32 var_7118_epsilon_0 = const()[name = string("op_7118_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7118_cast_fp16 = inverse(epsilon = var_7118_epsilon_0, x = var_7117_cast_fp16)[name = string("op_7118_cast_fp16")];
            fp16 var_7119_to_fp16 = const()[name = string("op_7119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7120_cast_fp16 = mul(x = var_7118_cast_fp16, y = var_7119_to_fp16)[name = string("op_7120_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7122_cast_fp16 = add(x = var_7116_cast_fp16, y = var_7120_cast_fp16)[name = string("op_7122_cast_fp16")];
            fp32 var_7123_epsilon_0 = const()[name = string("op_7123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7123_cast_fp16 = rsqrt(epsilon = var_7123_epsilon_0, x = var_7122_cast_fp16)[name = string("op_7123_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_211_cast_fp16 = mul(x = y_241_cast_fp16, y = var_7123_cast_fp16)[name = string("out_211_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87454848)))];
            tensor<fp16, [1, 48, 512]> var_7125_cast_fp16 = mul(x = out_211_cast_fp16, y = tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_29_cast_fp16 = add(x = input_177_cast_fp16, y = var_7125_cast_fp16)[name = string("h_29_cast_fp16")];
            tensor<fp16, [1]> tm_layers_14_layer_scalar_to_fp16 = const()[name = string("tm_layers_14_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1p-1])];
            tensor<fp16, [1, 48, 512]> x_271_cast_fp16 = mul(x = h_29_cast_fp16, y = tm_layers_14_layer_scalar_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7129_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_7129_cast_fp16")];
            bool var_7132 = const()[name = string("op_7132"), val = bool(true)];
            tensor<int32, [1]> var_7133_axes_0 = const()[name = string("op_7133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7133_cast_fp16 = reduce_max(axes = var_7133_axes_0, keep_dims = var_7132, x = var_7129_cast_fp16)[name = string("op_7133_cast_fp16")];
            fp16 var_7134_to_fp16 = const()[name = string("op_7134_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_243_cast_fp16 = maximum(x = var_7133_cast_fp16, y = var_7134_to_fp16)[name = string("s_243_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_243_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_243_cast_fp16)[name = string("y_243_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7137_cast_fp16 = mul(x = y_243_cast_fp16, y = y_243_cast_fp16)[name = string("op_7137_cast_fp16")];
            tensor<int32, [1]> var_7142_axes_0 = const()[name = string("op_7142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7142_keep_dims_0 = const()[name = string("op_7142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7142_cast_fp16 = reduce_mean(axes = var_7142_axes_0, keep_dims = var_7142_keep_dims_0, x = var_7137_cast_fp16)[name = string("op_7142_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7143_cast_fp16 = mul(x = s_243_cast_fp16, y = s_243_cast_fp16)[name = string("op_7143_cast_fp16")];
            fp32 var_7144_epsilon_0 = const()[name = string("op_7144_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7144_cast_fp16 = inverse(epsilon = var_7144_epsilon_0, x = var_7143_cast_fp16)[name = string("op_7144_cast_fp16")];
            fp16 var_7145_to_fp16 = const()[name = string("op_7145_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7146_cast_fp16 = mul(x = var_7144_cast_fp16, y = var_7145_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7148_cast_fp16 = add(x = var_7142_cast_fp16, y = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp32 var_7149_epsilon_0 = const()[name = string("op_7149_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7149_cast_fp16 = rsqrt(epsilon = var_7149_epsilon_0, x = var_7148_cast_fp16)[name = string("op_7149_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_213_cast_fp16 = mul(x = y_243_cast_fp16, y = var_7149_cast_fp16)[name = string("out_213_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87455936)))];
            tensor<fp16, [1, 48, 512]> input_181_cast_fp16 = mul(x = out_213_cast_fp16, y = tm_layers_15_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87457024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87981376))))[name = string("tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<int32, [4]> var_7159 = const()[name = string("op_7159"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_7160_cast_fp16 = reshape(shape = var_7159, x = linear_136_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<int32, [4]> x_273_perm_0 = const()[name = string("x_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87983488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88245696))))[name = string("tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<int32, [4]> var_7171 = const()[name = string("op_7171"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_7172_cast_fp16 = reshape(shape = var_7171, x = linear_137_cast_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [4]> x_275_perm_0 = const()[name = string("x_275_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88246784))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88508992))))[name = string("tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_7183 = const()[name = string("op_7183"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_7184_cast_fp16 = reshape(shape = var_7183, x = linear_138_cast_fp16)[name = string("op_7184_cast_fp16")];
            tensor<int32, [4]> x_277_perm_0 = const()[name = string("x_277_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_273_cast_fp16 = transpose(perm = x_273_perm_0, x = var_7160_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 4, 48, 256]> var_7188_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_7188_cast_fp16")];
            bool var_7191 = const()[name = string("op_7191"), val = bool(true)];
            tensor<int32, [1]> var_7192_axes_0 = const()[name = string("op_7192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_7192_cast_fp16 = reduce_max(axes = var_7192_axes_0, keep_dims = var_7191, x = var_7188_cast_fp16)[name = string("op_7192_cast_fp16")];
            fp16 var_7193_to_fp16 = const()[name = string("op_7193_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_245_cast_fp16 = maximum(x = var_7192_cast_fp16, y = var_7193_to_fp16)[name = string("s_245_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_245_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_245_cast_fp16)[name = string("y_245_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_7196_cast_fp16 = mul(x = y_245_cast_fp16, y = y_245_cast_fp16)[name = string("op_7196_cast_fp16")];
            tensor<int32, [1]> var_7201_axes_0 = const()[name = string("op_7201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7201_keep_dims_0 = const()[name = string("op_7201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_7201_cast_fp16 = reduce_mean(axes = var_7201_axes_0, keep_dims = var_7201_keep_dims_0, x = var_7196_cast_fp16)[name = string("op_7201_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_7202_cast_fp16 = mul(x = s_245_cast_fp16, y = s_245_cast_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_7203_cast_fp16 = inverse(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            fp16 var_7204_to_fp16 = const()[name = string("op_7204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_7205_cast_fp16 = mul(x = var_7203_cast_fp16, y = var_7204_to_fp16)[name = string("op_7205_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_7207_cast_fp16 = add(x = var_7201_cast_fp16, y = var_7205_cast_fp16)[name = string("op_7207_cast_fp16")];
            fp32 var_7208_epsilon_0 = const()[name = string("op_7208_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_7208_cast_fp16 = rsqrt(epsilon = var_7208_epsilon_0, x = var_7207_cast_fp16)[name = string("op_7208_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_215_cast_fp16 = mul(x = y_245_cast_fp16, y = var_7208_cast_fp16)[name = string("out_215_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_61_cast_fp16 = mul(x = out_215_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_61_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_275_cast_fp16 = transpose(perm = x_275_perm_0, x = var_7172_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 2, 48, 256]> var_7211_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_7211_cast_fp16")];
            bool var_7214 = const()[name = string("op_7214"), val = bool(true)];
            tensor<int32, [1]> var_7215_axes_0 = const()[name = string("op_7215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_7215_cast_fp16 = reduce_max(axes = var_7215_axes_0, keep_dims = var_7214, x = var_7211_cast_fp16)[name = string("op_7215_cast_fp16")];
            fp16 var_7216_to_fp16 = const()[name = string("op_7216_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_247_cast_fp16 = maximum(x = var_7215_cast_fp16, y = var_7216_to_fp16)[name = string("s_247_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_247_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_247_cast_fp16)[name = string("y_247_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_7219_cast_fp16 = mul(x = y_247_cast_fp16, y = y_247_cast_fp16)[name = string("op_7219_cast_fp16")];
            tensor<int32, [1]> var_7224_axes_0 = const()[name = string("op_7224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7224_keep_dims_0 = const()[name = string("op_7224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_7224_cast_fp16 = reduce_mean(axes = var_7224_axes_0, keep_dims = var_7224_keep_dims_0, x = var_7219_cast_fp16)[name = string("op_7224_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_7225_cast_fp16 = mul(x = s_247_cast_fp16, y = s_247_cast_fp16)[name = string("op_7225_cast_fp16")];
            fp32 var_7226_epsilon_0 = const()[name = string("op_7226_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_7226_cast_fp16 = inverse(epsilon = var_7226_epsilon_0, x = var_7225_cast_fp16)[name = string("op_7226_cast_fp16")];
            fp16 var_7227_to_fp16 = const()[name = string("op_7227_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_7228_cast_fp16 = mul(x = var_7226_cast_fp16, y = var_7227_to_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_7230_cast_fp16 = add(x = var_7224_cast_fp16, y = var_7228_cast_fp16)[name = string("op_7230_cast_fp16")];
            fp32 var_7231_epsilon_0 = const()[name = string("op_7231_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_7231_cast_fp16 = rsqrt(epsilon = var_7231_epsilon_0, x = var_7230_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_217_cast_fp16 = mul(x = y_247_cast_fp16, y = var_7231_cast_fp16)[name = string("out_217_cast_fp16")];
            tensor<fp16, [256]> tm_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510080)))];
            tensor<fp16, [1, 2, 48, 256]> k_91_cast_fp16 = mul(x = out_217_cast_fp16, y = tm_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_277_cast_fp16 = transpose(perm = x_277_perm_0, x = var_7184_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2, 48, 256]> var_7234_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_7234_cast_fp16")];
            bool var_7237 = const()[name = string("op_7237"), val = bool(true)];
            tensor<int32, [1]> var_7238_axes_0 = const()[name = string("op_7238_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_7238_cast_fp16 = reduce_max(axes = var_7238_axes_0, keep_dims = var_7237, x = var_7234_cast_fp16)[name = string("op_7238_cast_fp16")];
            fp16 var_7239_to_fp16 = const()[name = string("op_7239_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_249_cast_fp16 = maximum(x = var_7238_cast_fp16, y = var_7239_to_fp16)[name = string("s_249_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_249_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_249_cast_fp16)[name = string("y_249_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_7242_cast_fp16 = mul(x = y_249_cast_fp16, y = y_249_cast_fp16)[name = string("op_7242_cast_fp16")];
            tensor<int32, [1]> var_7247_axes_0 = const()[name = string("op_7247_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7247_keep_dims_0 = const()[name = string("op_7247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_7247_cast_fp16 = reduce_mean(axes = var_7247_axes_0, keep_dims = var_7247_keep_dims_0, x = var_7242_cast_fp16)[name = string("op_7247_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_7248_cast_fp16 = mul(x = s_249_cast_fp16, y = s_249_cast_fp16)[name = string("op_7248_cast_fp16")];
            fp32 var_7249_epsilon_0 = const()[name = string("op_7249_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_7249_cast_fp16 = inverse(epsilon = var_7249_epsilon_0, x = var_7248_cast_fp16)[name = string("op_7249_cast_fp16")];
            fp16 var_7250_to_fp16 = const()[name = string("op_7250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_7251_cast_fp16 = mul(x = var_7249_cast_fp16, y = var_7250_to_fp16)[name = string("op_7251_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_7253_cast_fp16 = add(x = var_7247_cast_fp16, y = var_7251_cast_fp16)[name = string("op_7253_cast_fp16")];
            fp32 var_7254_epsilon_0 = const()[name = string("op_7254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_7254_cast_fp16 = rsqrt(epsilon = var_7254_epsilon_0, x = var_7253_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_61_cast_fp16 = mul(x = y_249_cast_fp16, y = var_7254_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_7256_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7256_cast_fp16")];
            tensor<int32, [4]> var_7261_begin_0 = const()[name = string("op_7261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7261_end_0 = const()[name = string("op_7261_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_7261_end_mask_0 = const()[name = string("op_7261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_7261_cast_fp16 = slice_by_index(begin = var_7261_begin_0, end = var_7261_end_0, end_mask = var_7261_end_mask_0, x = q_61_cast_fp16)[name = string("op_7261_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_7262_cast_fp16 = mul(x = var_7261_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_7262_cast_fp16")];
            tensor<int32, [4]> var_7267_begin_0 = const()[name = string("op_7267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7267_end_0 = const()[name = string("op_7267_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_7267_end_mask_0 = const()[name = string("op_7267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_7267_cast_fp16 = slice_by_index(begin = var_7267_begin_0, end = var_7267_end_0, end_mask = var_7267_end_mask_0, x = q_61_cast_fp16)[name = string("op_7267_cast_fp16")];
            int32 var_7269 = const()[name = string("op_7269"), val = int32(-1)];
            bool var_7270_interleave_0 = const()[name = string("op_7270_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_7270_cast_fp16 = concat(axis = var_7269, interleave = var_7270_interleave_0, values = (var_7262_cast_fp16, var_7267_cast_fp16))[name = string("op_7270_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_7271_cast_fp16 = mul(x = var_7270_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7271_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_63_cast_fp16 = add(x = var_7256_cast_fp16, y = var_7271_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_7274_cast_fp16 = mul(x = k_91_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7274_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, x = k_91_cast_fp16)[name = string("op_7279_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_7280_cast_fp16 = mul(x = var_7279_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_7280_cast_fp16")];
            tensor<int32, [4]> var_7285_begin_0 = const()[name = string("op_7285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7285_end_0 = const()[name = string("op_7285_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_7285_end_mask_0 = const()[name = string("op_7285_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, x = k_91_cast_fp16)[name = string("op_7285_cast_fp16")];
            int32 var_7287 = const()[name = string("op_7287"), val = int32(-1)];
            bool var_7288_interleave_0 = const()[name = string("op_7288_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_7288_cast_fp16 = concat(axis = var_7287, interleave = var_7288_interleave_0, values = (var_7280_cast_fp16, var_7285_cast_fp16))[name = string("op_7288_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_7289_cast_fp16 = mul(x = var_7288_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7289_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_93_cast_fp16 = add(x = var_7274_cast_fp16, y = var_7289_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_30_reps_0 = const()[name = string("tile_30_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = k_93_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 1, 48, 256]> tile_30_cast_fp16 = tile(reps = tile_30_reps_0, x = transpose_60_cast_fp16)[name = string("tile_30_cast_fp16")];
            tensor<int32, [5]> concat_60 = const()[name = string("concat_60"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_60_cast_fp16 = reshape(shape = concat_60, x = tile_30_cast_fp16)[name = string("reshape_60_cast_fp16")];
            tensor<int32, [5]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = reshape_60_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 1, 48, 256]> reshape_61_cast_fp16 = reshape(shape = concat_61, x = transpose_61_cast_fp16)[name = string("reshape_61_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_31_reps_0 = const()[name = string("tile_31_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = v_61_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 1, 48, 256]> tile_31_cast_fp16 = tile(reps = tile_31_reps_0, x = transpose_62_cast_fp16)[name = string("tile_31_cast_fp16")];
            tensor<int32, [5]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_62_cast_fp16 = reshape(shape = concat_62, x = tile_31_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [5]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 1, 48, 256]> reshape_63_cast_fp16 = reshape(shape = concat_63, x = transpose_63_cast_fp16)[name = string("reshape_63_cast_fp16")];
            tensor<int32, [4]> v_63_perm_0 = const()[name = string("v_63_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7303_transpose_x_0 = const()[name = string("op_7303_transpose_x_0"), val = bool(false)];
            bool var_7303_transpose_y_0 = const()[name = string("op_7303_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = reshape_61_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 4, 48, 48]> var_7303_cast_fp16 = matmul(transpose_x = var_7303_transpose_x_0, transpose_y = var_7303_transpose_y_0, x = q_63_cast_fp16, y = transpose_111_cast_fp16)[name = string("op_7303_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_7305_cast_fp16 = add(x = var_7303_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7305_cast_fp16")];
            int32 var_7306 = const()[name = string("op_7306"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_249_cast_fp16 = softmax(axis = var_7306, x = var_7305_cast_fp16)[name = string("w_249_cast_fp16")];
            bool var_7309_transpose_x_0 = const()[name = string("op_7309_transpose_x_0"), val = bool(false)];
            bool var_7309_transpose_y_0 = const()[name = string("op_7309_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = reshape_63_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 4, 48, 256]> var_7309_cast_fp16 = matmul(transpose_x = var_7309_transpose_x_0, transpose_y = var_7309_transpose_y_0, x = w_249_cast_fp16, y = v_63_cast_fp16)[name = string("op_7309_cast_fp16")];
            tensor<int32, [4]> var_7312_perm_0 = const()[name = string("op_7312_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7316 = const()[name = string("op_7316"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_7312_cast_fp16 = transpose(perm = var_7312_perm_0, x = var_7309_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 48, 1024]> input_183_cast_fp16 = reshape(shape = var_7316, x = var_7312_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510656))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89035008))))[name = string("tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7321_cast_fp16 = abs(x = linear_139_cast_fp16)[name = string("op_7321_cast_fp16")];
            bool var_7324 = const()[name = string("op_7324"), val = bool(true)];
            tensor<int32, [1]> var_7325_axes_0 = const()[name = string("op_7325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7325_cast_fp16 = reduce_max(axes = var_7325_axes_0, keep_dims = var_7324, x = var_7321_cast_fp16)[name = string("op_7325_cast_fp16")];
            fp16 var_7326_to_fp16 = const()[name = string("op_7326_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_251_cast_fp16 = maximum(x = var_7325_cast_fp16, y = var_7326_to_fp16)[name = string("s_251_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_251_cast_fp16 = real_div(x = linear_139_cast_fp16, y = s_251_cast_fp16)[name = string("y_251_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7329_cast_fp16 = mul(x = y_251_cast_fp16, y = y_251_cast_fp16)[name = string("op_7329_cast_fp16")];
            tensor<int32, [1]> var_7334_axes_0 = const()[name = string("op_7334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7334_keep_dims_0 = const()[name = string("op_7334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7334_cast_fp16 = reduce_mean(axes = var_7334_axes_0, keep_dims = var_7334_keep_dims_0, x = var_7329_cast_fp16)[name = string("op_7334_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7335_cast_fp16 = mul(x = s_251_cast_fp16, y = s_251_cast_fp16)[name = string("op_7335_cast_fp16")];
            fp32 var_7336_epsilon_0 = const()[name = string("op_7336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7336_cast_fp16 = inverse(epsilon = var_7336_epsilon_0, x = var_7335_cast_fp16)[name = string("op_7336_cast_fp16")];
            fp16 var_7337_to_fp16 = const()[name = string("op_7337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7338_cast_fp16 = mul(x = var_7336_cast_fp16, y = var_7337_to_fp16)[name = string("op_7338_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7340_cast_fp16 = add(x = var_7334_cast_fp16, y = var_7338_cast_fp16)[name = string("op_7340_cast_fp16")];
            fp32 var_7341_epsilon_0 = const()[name = string("op_7341_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7341_cast_fp16 = rsqrt(epsilon = var_7341_epsilon_0, x = var_7340_cast_fp16)[name = string("op_7341_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_219_cast_fp16 = mul(x = y_251_cast_fp16, y = var_7341_cast_fp16)[name = string("out_219_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89036096)))];
            tensor<fp16, [1, 48, 512]> var_7343_cast_fp16 = mul(x = out_219_cast_fp16, y = tm_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_7343_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_281_cast_fp16 = add(x = x_271_cast_fp16, y = var_7343_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7346_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_7346_cast_fp16")];
            bool var_7349 = const()[name = string("op_7349"), val = bool(true)];
            tensor<int32, [1]> var_7350_axes_0 = const()[name = string("op_7350_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7350_cast_fp16 = reduce_max(axes = var_7350_axes_0, keep_dims = var_7349, x = var_7346_cast_fp16)[name = string("op_7350_cast_fp16")];
            fp16 var_7351_to_fp16 = const()[name = string("op_7351_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_253_cast_fp16 = maximum(x = var_7350_cast_fp16, y = var_7351_to_fp16)[name = string("s_253_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_253_cast_fp16 = real_div(x = x_281_cast_fp16, y = s_253_cast_fp16)[name = string("y_253_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7354_cast_fp16 = mul(x = y_253_cast_fp16, y = y_253_cast_fp16)[name = string("op_7354_cast_fp16")];
            tensor<int32, [1]> var_7359_axes_0 = const()[name = string("op_7359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7359_keep_dims_0 = const()[name = string("op_7359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7359_cast_fp16 = reduce_mean(axes = var_7359_axes_0, keep_dims = var_7359_keep_dims_0, x = var_7354_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7360_cast_fp16 = mul(x = s_253_cast_fp16, y = s_253_cast_fp16)[name = string("op_7360_cast_fp16")];
            fp32 var_7361_epsilon_0 = const()[name = string("op_7361_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7361_cast_fp16 = inverse(epsilon = var_7361_epsilon_0, x = var_7360_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7362_to_fp16 = const()[name = string("op_7362_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7363_cast_fp16 = mul(x = var_7361_cast_fp16, y = var_7362_to_fp16)[name = string("op_7363_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7365_cast_fp16 = add(x = var_7359_cast_fp16, y = var_7363_cast_fp16)[name = string("op_7365_cast_fp16")];
            fp32 var_7366_epsilon_0 = const()[name = string("op_7366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7366_cast_fp16 = rsqrt(epsilon = var_7366_epsilon_0, x = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_221_cast_fp16 = mul(x = y_253_cast_fp16, y = var_7366_cast_fp16)[name = string("out_221_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89037184)))];
            tensor<fp16, [1, 48, 512]> input_185_cast_fp16 = mul(x = out_221_cast_fp16, y = tm_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89038272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90086912))))[name = string("tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_140_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_140_cast_fp16")];
            string var_7373_mode_0 = const()[name = string("op_7373_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_7373_cast_fp16 = gelu(mode = var_7373_mode_0, x = linear_140_cast_fp16)[name = string("op_7373_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90091072))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91139712))))[name = string("tm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_141_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_187_cast_fp16 = mul(x = var_7373_cast_fp16, y = linear_141_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91143872))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92192512))))[name = string("tm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7381_cast_fp16 = abs(x = linear_142_cast_fp16)[name = string("op_7381_cast_fp16")];
            bool var_7384 = const()[name = string("op_7384"), val = bool(true)];
            tensor<int32, [1]> var_7385_axes_0 = const()[name = string("op_7385_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7385_cast_fp16 = reduce_max(axes = var_7385_axes_0, keep_dims = var_7384, x = var_7381_cast_fp16)[name = string("op_7385_cast_fp16")];
            fp16 var_7386_to_fp16 = const()[name = string("op_7386_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_255_cast_fp16 = maximum(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = string("s_255_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_255_cast_fp16 = real_div(x = linear_142_cast_fp16, y = s_255_cast_fp16)[name = string("y_255_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7389_cast_fp16 = mul(x = y_255_cast_fp16, y = y_255_cast_fp16)[name = string("op_7389_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7389_cast_fp16)[name = string("op_7394_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7395_cast_fp16 = mul(x = s_255_cast_fp16, y = s_255_cast_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7400_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7400_cast_fp16")];
            fp32 var_7401_epsilon_0 = const()[name = string("op_7401_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7401_cast_fp16 = rsqrt(epsilon = var_7401_epsilon_0, x = var_7400_cast_fp16)[name = string("op_7401_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_223_cast_fp16 = mul(x = y_255_cast_fp16, y = var_7401_cast_fp16)[name = string("out_223_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92193600)))];
            tensor<fp16, [1, 48, 512]> var_7403_cast_fp16 = mul(x = out_223_cast_fp16, y = tm_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7403_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_189_cast_fp16 = add(x = x_281_cast_fp16, y = var_7403_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92194688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92456896))))[name = string("tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_143_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_143_cast_fp16")];
            string x_285_mode_0 = const()[name = string("x_285_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_285_cast_fp16 = gelu(mode = x_285_mode_0, x = linear_143_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [4]> var_7423_begin_0 = const()[name = string("op_7423_begin_0"), val = tensor<int32, [4]>([0, 0, 15, 0])];
            tensor<int32, [4]> var_7423_end_0 = const()[name = string("op_7423_end_0"), val = tensor<int32, [4]>([1, 48, 16, 512])];
            tensor<bool, [4]> var_7423_end_mask_0 = const()[name = string("op_7423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7423_squeeze_mask_0 = const()[name = string("op_7423_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_7423_cast_fp16 = slice_by_index(begin = var_7423_begin_0, end = var_7423_end_0, end_mask = var_7423_end_mask_0, squeeze_mask = var_7423_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7423_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_191_cast_fp16 = mul(x = x_285_cast_fp16, y = var_7423_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92457984))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92720192))))[name = string("tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7433_cast_fp16 = abs(x = linear_144_cast_fp16)[name = string("op_7433_cast_fp16")];
            bool var_7436 = const()[name = string("op_7436"), val = bool(true)];
            tensor<int32, [1]> var_7437_axes_0 = const()[name = string("op_7437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7437_cast_fp16 = reduce_max(axes = var_7437_axes_0, keep_dims = var_7436, x = var_7433_cast_fp16)[name = string("op_7437_cast_fp16")];
            fp16 var_7438_to_fp16 = const()[name = string("op_7438_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_257_cast_fp16 = maximum(x = var_7437_cast_fp16, y = var_7438_to_fp16)[name = string("s_257_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_257_cast_fp16 = real_div(x = linear_144_cast_fp16, y = s_257_cast_fp16)[name = string("y_257_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7441_cast_fp16 = mul(x = y_257_cast_fp16, y = y_257_cast_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7446_axes_0 = const()[name = string("op_7446_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7446_keep_dims_0 = const()[name = string("op_7446_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7446_cast_fp16 = reduce_mean(axes = var_7446_axes_0, keep_dims = var_7446_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7447_cast_fp16 = mul(x = s_257_cast_fp16, y = s_257_cast_fp16)[name = string("op_7447_cast_fp16")];
            fp32 var_7448_epsilon_0 = const()[name = string("op_7448_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7448_cast_fp16 = inverse(epsilon = var_7448_epsilon_0, x = var_7447_cast_fp16)[name = string("op_7448_cast_fp16")];
            fp16 var_7449_to_fp16 = const()[name = string("op_7449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7450_cast_fp16 = mul(x = var_7448_cast_fp16, y = var_7449_to_fp16)[name = string("op_7450_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7452_cast_fp16 = add(x = var_7446_cast_fp16, y = var_7450_cast_fp16)[name = string("op_7452_cast_fp16")];
            fp32 var_7453_epsilon_0 = const()[name = string("op_7453_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7453_cast_fp16 = rsqrt(epsilon = var_7453_epsilon_0, x = var_7452_cast_fp16)[name = string("op_7453_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_225_cast_fp16 = mul(x = y_257_cast_fp16, y = var_7453_cast_fp16)[name = string("out_225_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92721280)))];
            tensor<fp16, [1, 48, 512]> var_7455_cast_fp16 = mul(x = out_225_cast_fp16, y = tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7455_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_31_cast_fp16 = add(x = input_189_cast_fp16, y = var_7455_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<fp16, [1]> tm_layers_15_layer_scalar_to_fp16 = const()[name = string("tm_layers_15_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7ap-1])];
            tensor<fp16, [1, 48, 512]> x_289_cast_fp16 = mul(x = h_31_cast_fp16, y = tm_layers_15_layer_scalar_to_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7459_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_7459_cast_fp16")];
            bool var_7462 = const()[name = string("op_7462"), val = bool(true)];
            tensor<int32, [1]> var_7463_axes_0 = const()[name = string("op_7463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7463_cast_fp16 = reduce_max(axes = var_7463_axes_0, keep_dims = var_7462, x = var_7459_cast_fp16)[name = string("op_7463_cast_fp16")];
            fp16 var_7464_to_fp16 = const()[name = string("op_7464_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_259_cast_fp16 = maximum(x = var_7463_cast_fp16, y = var_7464_to_fp16)[name = string("s_259_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_259_cast_fp16 = real_div(x = x_289_cast_fp16, y = s_259_cast_fp16)[name = string("y_259_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7467_cast_fp16 = mul(x = y_259_cast_fp16, y = y_259_cast_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7472_axes_0 = const()[name = string("op_7472_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7472_keep_dims_0 = const()[name = string("op_7472_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7472_cast_fp16 = reduce_mean(axes = var_7472_axes_0, keep_dims = var_7472_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7473_cast_fp16 = mul(x = s_259_cast_fp16, y = s_259_cast_fp16)[name = string("op_7473_cast_fp16")];
            fp32 var_7474_epsilon_0 = const()[name = string("op_7474_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7474_cast_fp16 = inverse(epsilon = var_7474_epsilon_0, x = var_7473_cast_fp16)[name = string("op_7474_cast_fp16")];
            fp16 var_7475_to_fp16 = const()[name = string("op_7475_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7476_cast_fp16 = mul(x = var_7474_cast_fp16, y = var_7475_to_fp16)[name = string("op_7476_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7478_cast_fp16 = add(x = var_7472_cast_fp16, y = var_7476_cast_fp16)[name = string("op_7478_cast_fp16")];
            fp32 var_7479_epsilon_0 = const()[name = string("op_7479_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7479_cast_fp16 = rsqrt(epsilon = var_7479_epsilon_0, x = var_7478_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_227_cast_fp16 = mul(x = y_259_cast_fp16, y = var_7479_cast_fp16)[name = string("out_227_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92722368)))];
            tensor<fp16, [1, 48, 512]> input_193_cast_fp16 = mul(x = out_227_cast_fp16, y = tm_layers_16_input_layernorm_weight_to_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92723456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93247808))))[name = string("tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<int32, [4]> var_7489 = const()[name = string("op_7489"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_7490_cast_fp16 = reshape(shape = var_7489, x = linear_145_cast_fp16)[name = string("op_7490_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93249920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93512128))))[name = string("tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<int32, [4]> var_7501 = const()[name = string("op_7501"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_7502_cast_fp16 = reshape(shape = var_7501, x = linear_146_cast_fp16)[name = string("op_7502_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93513216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93775424))))[name = string("tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7513 = const()[name = string("op_7513"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_7514_cast_fp16 = reshape(shape = var_7513, x = linear_147_cast_fp16)[name = string("op_7514_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_7490_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 4, 48, 256]> var_7518_cast_fp16 = abs(x = x_291_cast_fp16)[name = string("op_7518_cast_fp16")];
            bool var_7521 = const()[name = string("op_7521"), val = bool(true)];
            tensor<int32, [1]> var_7522_axes_0 = const()[name = string("op_7522_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_7522_cast_fp16 = reduce_max(axes = var_7522_axes_0, keep_dims = var_7521, x = var_7518_cast_fp16)[name = string("op_7522_cast_fp16")];
            fp16 var_7523_to_fp16 = const()[name = string("op_7523_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_261_cast_fp16 = maximum(x = var_7522_cast_fp16, y = var_7523_to_fp16)[name = string("s_261_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_261_cast_fp16 = real_div(x = x_291_cast_fp16, y = s_261_cast_fp16)[name = string("y_261_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_7526_cast_fp16 = mul(x = y_261_cast_fp16, y = y_261_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<int32, [1]> var_7531_axes_0 = const()[name = string("op_7531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7531_keep_dims_0 = const()[name = string("op_7531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_7531_cast_fp16 = reduce_mean(axes = var_7531_axes_0, keep_dims = var_7531_keep_dims_0, x = var_7526_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_7532_cast_fp16 = mul(x = s_261_cast_fp16, y = s_261_cast_fp16)[name = string("op_7532_cast_fp16")];
            fp32 var_7533_epsilon_0 = const()[name = string("op_7533_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_7533_cast_fp16 = inverse(epsilon = var_7533_epsilon_0, x = var_7532_cast_fp16)[name = string("op_7533_cast_fp16")];
            fp16 var_7534_to_fp16 = const()[name = string("op_7534_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_7535_cast_fp16 = mul(x = var_7533_cast_fp16, y = var_7534_to_fp16)[name = string("op_7535_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_7537_cast_fp16 = add(x = var_7531_cast_fp16, y = var_7535_cast_fp16)[name = string("op_7537_cast_fp16")];
            fp32 var_7538_epsilon_0 = const()[name = string("op_7538_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_7538_cast_fp16 = rsqrt(epsilon = var_7538_epsilon_0, x = var_7537_cast_fp16)[name = string("op_7538_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_229_cast_fp16 = mul(x = y_261_cast_fp16, y = var_7538_cast_fp16)[name = string("out_229_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_65_cast_fp16 = mul(x = out_229_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_65_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_7502_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 48, 256]> var_7541_cast_fp16 = abs(x = x_293_cast_fp16)[name = string("op_7541_cast_fp16")];
            bool var_7544 = const()[name = string("op_7544"), val = bool(true)];
            tensor<int32, [1]> var_7545_axes_0 = const()[name = string("op_7545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_7545_cast_fp16 = reduce_max(axes = var_7545_axes_0, keep_dims = var_7544, x = var_7541_cast_fp16)[name = string("op_7545_cast_fp16")];
            fp16 var_7546_to_fp16 = const()[name = string("op_7546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_263_cast_fp16 = maximum(x = var_7545_cast_fp16, y = var_7546_to_fp16)[name = string("s_263_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_263_cast_fp16 = real_div(x = x_293_cast_fp16, y = s_263_cast_fp16)[name = string("y_263_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_7549_cast_fp16 = mul(x = y_263_cast_fp16, y = y_263_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<int32, [1]> var_7554_axes_0 = const()[name = string("op_7554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7554_keep_dims_0 = const()[name = string("op_7554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_7554_cast_fp16 = reduce_mean(axes = var_7554_axes_0, keep_dims = var_7554_keep_dims_0, x = var_7549_cast_fp16)[name = string("op_7554_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_7555_cast_fp16 = mul(x = s_263_cast_fp16, y = s_263_cast_fp16)[name = string("op_7555_cast_fp16")];
            fp32 var_7556_epsilon_0 = const()[name = string("op_7556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_7556_cast_fp16 = inverse(epsilon = var_7556_epsilon_0, x = var_7555_cast_fp16)[name = string("op_7556_cast_fp16")];
            fp16 var_7557_to_fp16 = const()[name = string("op_7557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_7558_cast_fp16 = mul(x = var_7556_cast_fp16, y = var_7557_to_fp16)[name = string("op_7558_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_7560_cast_fp16 = add(x = var_7554_cast_fp16, y = var_7558_cast_fp16)[name = string("op_7560_cast_fp16")];
            fp32 var_7561_epsilon_0 = const()[name = string("op_7561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_7561_cast_fp16 = rsqrt(epsilon = var_7561_epsilon_0, x = var_7560_cast_fp16)[name = string("op_7561_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_231_cast_fp16 = mul(x = y_263_cast_fp16, y = var_7561_cast_fp16)[name = string("out_231_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_97_cast_fp16 = mul(x = out_231_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_97_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_7514_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 48, 256]> var_7564_cast_fp16 = abs(x = x_295_cast_fp16)[name = string("op_7564_cast_fp16")];
            bool var_7567 = const()[name = string("op_7567"), val = bool(true)];
            tensor<int32, [1]> var_7568_axes_0 = const()[name = string("op_7568_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_7568_cast_fp16 = reduce_max(axes = var_7568_axes_0, keep_dims = var_7567, x = var_7564_cast_fp16)[name = string("op_7568_cast_fp16")];
            fp16 var_7569_to_fp16 = const()[name = string("op_7569_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_265_cast_fp16 = maximum(x = var_7568_cast_fp16, y = var_7569_to_fp16)[name = string("s_265_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_265_cast_fp16 = real_div(x = x_295_cast_fp16, y = s_265_cast_fp16)[name = string("y_265_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_7572_cast_fp16 = mul(x = y_265_cast_fp16, y = y_265_cast_fp16)[name = string("op_7572_cast_fp16")];
            tensor<int32, [1]> var_7577_axes_0 = const()[name = string("op_7577_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7577_keep_dims_0 = const()[name = string("op_7577_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_7577_cast_fp16 = reduce_mean(axes = var_7577_axes_0, keep_dims = var_7577_keep_dims_0, x = var_7572_cast_fp16)[name = string("op_7577_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_7578_cast_fp16 = mul(x = s_265_cast_fp16, y = s_265_cast_fp16)[name = string("op_7578_cast_fp16")];
            fp32 var_7579_epsilon_0 = const()[name = string("op_7579_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_7579_cast_fp16 = inverse(epsilon = var_7579_epsilon_0, x = var_7578_cast_fp16)[name = string("op_7579_cast_fp16")];
            fp16 var_7580_to_fp16 = const()[name = string("op_7580_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_7581_cast_fp16 = mul(x = var_7579_cast_fp16, y = var_7580_to_fp16)[name = string("op_7581_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_7583_cast_fp16 = add(x = var_7577_cast_fp16, y = var_7581_cast_fp16)[name = string("op_7583_cast_fp16")];
            fp32 var_7584_epsilon_0 = const()[name = string("op_7584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_7584_cast_fp16 = rsqrt(epsilon = var_7584_epsilon_0, x = var_7583_cast_fp16)[name = string("op_7584_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_65_cast_fp16 = mul(x = y_265_cast_fp16, y = var_7584_cast_fp16)[name = string("v_65_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_7586_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7586_cast_fp16")];
            tensor<int32, [4]> var_7591_begin_0 = const()[name = string("op_7591_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7591_end_0 = const()[name = string("op_7591_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_7591_end_mask_0 = const()[name = string("op_7591_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_7591_cast_fp16 = slice_by_index(begin = var_7591_begin_0, end = var_7591_end_0, end_mask = var_7591_end_mask_0, x = q_65_cast_fp16)[name = string("op_7591_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_7592_cast_fp16 = mul(x = var_7591_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_7592_cast_fp16")];
            tensor<int32, [4]> var_7597_begin_0 = const()[name = string("op_7597_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7597_end_0 = const()[name = string("op_7597_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_7597_end_mask_0 = const()[name = string("op_7597_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, x = q_65_cast_fp16)[name = string("op_7597_cast_fp16")];
            int32 var_7599 = const()[name = string("op_7599"), val = int32(-1)];
            bool var_7600_interleave_0 = const()[name = string("op_7600_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_7600_cast_fp16 = concat(axis = var_7599, interleave = var_7600_interleave_0, values = (var_7592_cast_fp16, var_7597_cast_fp16))[name = string("op_7600_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_7601_cast_fp16 = mul(x = var_7600_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7601_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_67_cast_fp16 = add(x = var_7586_cast_fp16, y = var_7601_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_7604_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7604_cast_fp16")];
            tensor<int32, [4]> var_7609_begin_0 = const()[name = string("op_7609_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7609_end_0 = const()[name = string("op_7609_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_7609_end_mask_0 = const()[name = string("op_7609_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_7609_cast_fp16 = slice_by_index(begin = var_7609_begin_0, end = var_7609_end_0, end_mask = var_7609_end_mask_0, x = k_97_cast_fp16)[name = string("op_7609_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_7610_cast_fp16 = mul(x = var_7609_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_7610_cast_fp16")];
            tensor<int32, [4]> var_7615_begin_0 = const()[name = string("op_7615_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7615_end_0 = const()[name = string("op_7615_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_7615_end_mask_0 = const()[name = string("op_7615_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_7615_cast_fp16 = slice_by_index(begin = var_7615_begin_0, end = var_7615_end_0, end_mask = var_7615_end_mask_0, x = k_97_cast_fp16)[name = string("op_7615_cast_fp16")];
            int32 var_7617 = const()[name = string("op_7617"), val = int32(-1)];
            bool var_7618_interleave_0 = const()[name = string("op_7618_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_7618_cast_fp16 = concat(axis = var_7617, interleave = var_7618_interleave_0, values = (var_7610_cast_fp16, var_7615_cast_fp16))[name = string("op_7618_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_7619_cast_fp16 = mul(x = var_7618_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7619_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_99_cast_fp16 = add(x = var_7604_cast_fp16, y = var_7619_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_32_reps_0 = const()[name = string("tile_32_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = k_99_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 1, 48, 256]> tile_32_cast_fp16 = tile(reps = tile_32_reps_0, x = transpose_64_cast_fp16)[name = string("tile_32_cast_fp16")];
            tensor<int32, [5]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_64_cast_fp16 = reshape(shape = concat_64, x = tile_32_cast_fp16)[name = string("reshape_64_cast_fp16")];
            tensor<int32, [5]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = reshape_64_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 1, 48, 256]> reshape_65_cast_fp16 = reshape(shape = concat_65, x = transpose_65_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_33_reps_0 = const()[name = string("tile_33_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = v_65_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 1, 48, 256]> tile_33_cast_fp16 = tile(reps = tile_33_reps_0, x = transpose_66_cast_fp16)[name = string("tile_33_cast_fp16")];
            tensor<int32, [5]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_66_cast_fp16 = reshape(shape = concat_66, x = tile_33_cast_fp16)[name = string("reshape_66_cast_fp16")];
            tensor<int32, [5]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = reshape_66_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 1, 48, 256]> reshape_67_cast_fp16 = reshape(shape = concat_67, x = transpose_67_cast_fp16)[name = string("reshape_67_cast_fp16")];
            tensor<int32, [4]> v_67_perm_0 = const()[name = string("v_67_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7633_transpose_x_0 = const()[name = string("op_7633_transpose_x_0"), val = bool(false)];
            bool var_7633_transpose_y_0 = const()[name = string("op_7633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_112_cast_fp16 = transpose(perm = transpose_112_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 4, 48, 48]> var_7633_cast_fp16 = matmul(transpose_x = var_7633_transpose_x_0, transpose_y = var_7633_transpose_y_0, x = q_67_cast_fp16, y = transpose_112_cast_fp16)[name = string("op_7633_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_7635_cast_fp16 = add(x = var_7633_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7635_cast_fp16")];
            int32 var_7636 = const()[name = string("op_7636"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_265_cast_fp16 = softmax(axis = var_7636, x = var_7635_cast_fp16)[name = string("w_265_cast_fp16")];
            bool var_7639_transpose_x_0 = const()[name = string("op_7639_transpose_x_0"), val = bool(false)];
            bool var_7639_transpose_y_0 = const()[name = string("op_7639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = reshape_67_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 4, 48, 256]> var_7639_cast_fp16 = matmul(transpose_x = var_7639_transpose_x_0, transpose_y = var_7639_transpose_y_0, x = w_265_cast_fp16, y = v_67_cast_fp16)[name = string("op_7639_cast_fp16")];
            tensor<int32, [4]> var_7642_perm_0 = const()[name = string("op_7642_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7646 = const()[name = string("op_7646"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_7642_cast_fp16 = transpose(perm = var_7642_perm_0, x = var_7639_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 48, 1024]> input_195_cast_fp16 = reshape(shape = var_7646, x = var_7642_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93776512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94300864))))[name = string("tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7651_cast_fp16 = abs(x = linear_148_cast_fp16)[name = string("op_7651_cast_fp16")];
            bool var_7654 = const()[name = string("op_7654"), val = bool(true)];
            tensor<int32, [1]> var_7655_axes_0 = const()[name = string("op_7655_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7655_cast_fp16 = reduce_max(axes = var_7655_axes_0, keep_dims = var_7654, x = var_7651_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp16 var_7656_to_fp16 = const()[name = string("op_7656_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_267_cast_fp16 = maximum(x = var_7655_cast_fp16, y = var_7656_to_fp16)[name = string("s_267_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_267_cast_fp16 = real_div(x = linear_148_cast_fp16, y = s_267_cast_fp16)[name = string("y_267_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7659_cast_fp16 = mul(x = y_267_cast_fp16, y = y_267_cast_fp16)[name = string("op_7659_cast_fp16")];
            tensor<int32, [1]> var_7664_axes_0 = const()[name = string("op_7664_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7664_keep_dims_0 = const()[name = string("op_7664_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7664_cast_fp16 = reduce_mean(axes = var_7664_axes_0, keep_dims = var_7664_keep_dims_0, x = var_7659_cast_fp16)[name = string("op_7664_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7665_cast_fp16 = mul(x = s_267_cast_fp16, y = s_267_cast_fp16)[name = string("op_7665_cast_fp16")];
            fp32 var_7666_epsilon_0 = const()[name = string("op_7666_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7666_cast_fp16 = inverse(epsilon = var_7666_epsilon_0, x = var_7665_cast_fp16)[name = string("op_7666_cast_fp16")];
            fp16 var_7667_to_fp16 = const()[name = string("op_7667_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7668_cast_fp16 = mul(x = var_7666_cast_fp16, y = var_7667_to_fp16)[name = string("op_7668_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7670_cast_fp16 = add(x = var_7664_cast_fp16, y = var_7668_cast_fp16)[name = string("op_7670_cast_fp16")];
            fp32 var_7671_epsilon_0 = const()[name = string("op_7671_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7671_cast_fp16 = rsqrt(epsilon = var_7671_epsilon_0, x = var_7670_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_233_cast_fp16 = mul(x = y_267_cast_fp16, y = var_7671_cast_fp16)[name = string("out_233_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94301952)))];
            tensor<fp16, [1, 48, 512]> var_7673_cast_fp16 = mul(x = out_233_cast_fp16, y = tm_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("op_7673_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = var_7673_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7676_cast_fp16 = abs(x = x_299_cast_fp16)[name = string("op_7676_cast_fp16")];
            bool var_7679 = const()[name = string("op_7679"), val = bool(true)];
            tensor<int32, [1]> var_7680_axes_0 = const()[name = string("op_7680_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7680_cast_fp16 = reduce_max(axes = var_7680_axes_0, keep_dims = var_7679, x = var_7676_cast_fp16)[name = string("op_7680_cast_fp16")];
            fp16 var_7681_to_fp16 = const()[name = string("op_7681_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_269_cast_fp16 = maximum(x = var_7680_cast_fp16, y = var_7681_to_fp16)[name = string("s_269_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_269_cast_fp16 = real_div(x = x_299_cast_fp16, y = s_269_cast_fp16)[name = string("y_269_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7684_cast_fp16 = mul(x = y_269_cast_fp16, y = y_269_cast_fp16)[name = string("op_7684_cast_fp16")];
            tensor<int32, [1]> var_7689_axes_0 = const()[name = string("op_7689_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7689_keep_dims_0 = const()[name = string("op_7689_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7689_cast_fp16 = reduce_mean(axes = var_7689_axes_0, keep_dims = var_7689_keep_dims_0, x = var_7684_cast_fp16)[name = string("op_7689_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7690_cast_fp16 = mul(x = s_269_cast_fp16, y = s_269_cast_fp16)[name = string("op_7690_cast_fp16")];
            fp32 var_7691_epsilon_0 = const()[name = string("op_7691_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7691_cast_fp16 = inverse(epsilon = var_7691_epsilon_0, x = var_7690_cast_fp16)[name = string("op_7691_cast_fp16")];
            fp16 var_7692_to_fp16 = const()[name = string("op_7692_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7693_cast_fp16 = mul(x = var_7691_cast_fp16, y = var_7692_to_fp16)[name = string("op_7693_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7695_cast_fp16 = add(x = var_7689_cast_fp16, y = var_7693_cast_fp16)[name = string("op_7695_cast_fp16")];
            fp32 var_7696_epsilon_0 = const()[name = string("op_7696_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7696_cast_fp16 = rsqrt(epsilon = var_7696_epsilon_0, x = var_7695_cast_fp16)[name = string("op_7696_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_235_cast_fp16 = mul(x = y_269_cast_fp16, y = var_7696_cast_fp16)[name = string("out_235_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94303040)))];
            tensor<fp16, [1, 48, 512]> input_197_cast_fp16 = mul(x = out_235_cast_fp16, y = tm_layers_16_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94304128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95352768))))[name = string("tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_149_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_149_cast_fp16")];
            string var_7703_mode_0 = const()[name = string("op_7703_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_7703_cast_fp16 = gelu(mode = var_7703_mode_0, x = linear_149_cast_fp16)[name = string("op_7703_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95356928))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96405568))))[name = string("tm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_150_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_199_cast_fp16 = mul(x = var_7703_cast_fp16, y = linear_150_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96409728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97458368))))[name = string("tm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7711_cast_fp16 = abs(x = linear_151_cast_fp16)[name = string("op_7711_cast_fp16")];
            bool var_7714 = const()[name = string("op_7714"), val = bool(true)];
            tensor<int32, [1]> var_7715_axes_0 = const()[name = string("op_7715_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7715_cast_fp16 = reduce_max(axes = var_7715_axes_0, keep_dims = var_7714, x = var_7711_cast_fp16)[name = string("op_7715_cast_fp16")];
            fp16 var_7716_to_fp16 = const()[name = string("op_7716_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_271_cast_fp16 = maximum(x = var_7715_cast_fp16, y = var_7716_to_fp16)[name = string("s_271_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_271_cast_fp16 = real_div(x = linear_151_cast_fp16, y = s_271_cast_fp16)[name = string("y_271_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7719_cast_fp16 = mul(x = y_271_cast_fp16, y = y_271_cast_fp16)[name = string("op_7719_cast_fp16")];
            tensor<int32, [1]> var_7724_axes_0 = const()[name = string("op_7724_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7724_keep_dims_0 = const()[name = string("op_7724_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7724_cast_fp16 = reduce_mean(axes = var_7724_axes_0, keep_dims = var_7724_keep_dims_0, x = var_7719_cast_fp16)[name = string("op_7724_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7725_cast_fp16 = mul(x = s_271_cast_fp16, y = s_271_cast_fp16)[name = string("op_7725_cast_fp16")];
            fp32 var_7726_epsilon_0 = const()[name = string("op_7726_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7726_cast_fp16 = inverse(epsilon = var_7726_epsilon_0, x = var_7725_cast_fp16)[name = string("op_7726_cast_fp16")];
            fp16 var_7727_to_fp16 = const()[name = string("op_7727_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7728_cast_fp16 = mul(x = var_7726_cast_fp16, y = var_7727_to_fp16)[name = string("op_7728_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7730_cast_fp16 = add(x = var_7724_cast_fp16, y = var_7728_cast_fp16)[name = string("op_7730_cast_fp16")];
            fp32 var_7731_epsilon_0 = const()[name = string("op_7731_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7731_cast_fp16 = rsqrt(epsilon = var_7731_epsilon_0, x = var_7730_cast_fp16)[name = string("op_7731_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_237_cast_fp16 = mul(x = y_271_cast_fp16, y = var_7731_cast_fp16)[name = string("out_237_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97459456)))];
            tensor<fp16, [1, 48, 512]> var_7733_cast_fp16 = mul(x = out_237_cast_fp16, y = tm_layers_16_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7733_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_201_cast_fp16 = add(x = x_299_cast_fp16, y = var_7733_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97460544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97722752))))[name = string("tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_152_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_152_cast_fp16")];
            string x_303_mode_0 = const()[name = string("x_303_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_303_cast_fp16 = gelu(mode = x_303_mode_0, x = linear_152_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [4]> var_7753_begin_0 = const()[name = string("op_7753_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7753_end_0 = const()[name = string("op_7753_end_0"), val = tensor<int32, [4]>([1, 48, 17, 512])];
            tensor<bool, [4]> var_7753_end_mask_0 = const()[name = string("op_7753_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7753_squeeze_mask_0 = const()[name = string("op_7753_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_7753_cast_fp16 = slice_by_index(begin = var_7753_begin_0, end = var_7753_end_0, end_mask = var_7753_end_mask_0, squeeze_mask = var_7753_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7753_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_203_cast_fp16 = mul(x = x_303_cast_fp16, y = var_7753_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97723840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97986048))))[name = string("tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7763_cast_fp16 = abs(x = linear_153_cast_fp16)[name = string("op_7763_cast_fp16")];
            bool var_7766 = const()[name = string("op_7766"), val = bool(true)];
            tensor<int32, [1]> var_7767_axes_0 = const()[name = string("op_7767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7767_cast_fp16 = reduce_max(axes = var_7767_axes_0, keep_dims = var_7766, x = var_7763_cast_fp16)[name = string("op_7767_cast_fp16")];
            fp16 var_7768_to_fp16 = const()[name = string("op_7768_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_273_cast_fp16 = maximum(x = var_7767_cast_fp16, y = var_7768_to_fp16)[name = string("s_273_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_273_cast_fp16 = real_div(x = linear_153_cast_fp16, y = s_273_cast_fp16)[name = string("y_273_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7771_cast_fp16 = mul(x = y_273_cast_fp16, y = y_273_cast_fp16)[name = string("op_7771_cast_fp16")];
            tensor<int32, [1]> var_7776_axes_0 = const()[name = string("op_7776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7776_keep_dims_0 = const()[name = string("op_7776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7776_cast_fp16 = reduce_mean(axes = var_7776_axes_0, keep_dims = var_7776_keep_dims_0, x = var_7771_cast_fp16)[name = string("op_7776_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7777_cast_fp16 = mul(x = s_273_cast_fp16, y = s_273_cast_fp16)[name = string("op_7777_cast_fp16")];
            fp32 var_7778_epsilon_0 = const()[name = string("op_7778_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7778_cast_fp16 = inverse(epsilon = var_7778_epsilon_0, x = var_7777_cast_fp16)[name = string("op_7778_cast_fp16")];
            fp16 var_7779_to_fp16 = const()[name = string("op_7779_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7780_cast_fp16 = mul(x = var_7778_cast_fp16, y = var_7779_to_fp16)[name = string("op_7780_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7782_cast_fp16 = add(x = var_7776_cast_fp16, y = var_7780_cast_fp16)[name = string("op_7782_cast_fp16")];
            fp32 var_7783_epsilon_0 = const()[name = string("op_7783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7783_cast_fp16 = rsqrt(epsilon = var_7783_epsilon_0, x = var_7782_cast_fp16)[name = string("op_7783_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_239_cast_fp16 = mul(x = y_273_cast_fp16, y = var_7783_cast_fp16)[name = string("out_239_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97987136)))];
            tensor<fp16, [1, 48, 512]> var_7785_cast_fp16 = mul(x = out_239_cast_fp16, y = tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_33_cast_fp16 = add(x = input_201_cast_fp16, y = var_7785_cast_fp16)[name = string("h_33_cast_fp16")];
            tensor<fp16, [1]> tm_layers_16_layer_scalar_to_fp16 = const()[name = string("tm_layers_16_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7p-2])];
            tensor<fp16, [1, 48, 512]> x_307_cast_fp16 = mul(x = h_33_cast_fp16, y = tm_layers_16_layer_scalar_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7789_cast_fp16 = abs(x = x_307_cast_fp16)[name = string("op_7789_cast_fp16")];
            bool var_7792 = const()[name = string("op_7792"), val = bool(true)];
            tensor<int32, [1]> var_7793_axes_0 = const()[name = string("op_7793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7793_cast_fp16 = reduce_max(axes = var_7793_axes_0, keep_dims = var_7792, x = var_7789_cast_fp16)[name = string("op_7793_cast_fp16")];
            fp16 var_7794_to_fp16 = const()[name = string("op_7794_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_275_cast_fp16 = maximum(x = var_7793_cast_fp16, y = var_7794_to_fp16)[name = string("s_275_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_275_cast_fp16 = real_div(x = x_307_cast_fp16, y = s_275_cast_fp16)[name = string("y_275_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7797_cast_fp16 = mul(x = y_275_cast_fp16, y = y_275_cast_fp16)[name = string("op_7797_cast_fp16")];
            tensor<int32, [1]> var_7802_axes_0 = const()[name = string("op_7802_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7802_keep_dims_0 = const()[name = string("op_7802_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7802_cast_fp16 = reduce_mean(axes = var_7802_axes_0, keep_dims = var_7802_keep_dims_0, x = var_7797_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7803_cast_fp16 = mul(x = s_275_cast_fp16, y = s_275_cast_fp16)[name = string("op_7803_cast_fp16")];
            fp32 var_7804_epsilon_0 = const()[name = string("op_7804_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7804_cast_fp16 = inverse(epsilon = var_7804_epsilon_0, x = var_7803_cast_fp16)[name = string("op_7804_cast_fp16")];
            fp16 var_7805_to_fp16 = const()[name = string("op_7805_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7806_cast_fp16 = mul(x = var_7804_cast_fp16, y = var_7805_to_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7808_cast_fp16 = add(x = var_7802_cast_fp16, y = var_7806_cast_fp16)[name = string("op_7808_cast_fp16")];
            fp32 var_7809_epsilon_0 = const()[name = string("op_7809_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_7809_cast_fp16 = rsqrt(epsilon = var_7809_epsilon_0, x = var_7808_cast_fp16)[name = string("op_7809_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_241_cast_fp16 = mul(x = y_275_cast_fp16, y = var_7809_cast_fp16)[name = string("out_241_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97988224)))];
            tensor<fp16, [1, 48, 512]> input_205_cast_fp16 = mul(x = out_241_cast_fp16, y = tm_layers_17_input_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97989312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99037952))))[name = string("tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_154_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_7819 = const()[name = string("op_7819"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 48, 4, 512]> var_7820_cast_fp16 = reshape(shape = var_7819, x = linear_154_cast_fp16)[name = string("op_7820_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99042112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99304320))))[name = string("tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_7831 = const()[name = string("op_7831"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 48, 1, 512]> var_7832_cast_fp16 = reshape(shape = var_7831, x = linear_155_cast_fp16)[name = string("op_7832_cast_fp16")];
            tensor<int32, [4]> x_311_perm_0 = const()[name = string("x_311_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99305408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99567616))))[name = string("tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_7843 = const()[name = string("op_7843"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 48, 1, 512]> var_7844_cast_fp16 = reshape(shape = var_7843, x = linear_156_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<int32, [4]> x_313_perm_0 = const()[name = string("x_313_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 512]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_7820_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 4, 48, 512]> var_7848_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_7848_cast_fp16")];
            bool var_7851 = const()[name = string("op_7851"), val = bool(true)];
            tensor<int32, [1]> var_7852_axes_0 = const()[name = string("op_7852_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_7852_cast_fp16 = reduce_max(axes = var_7852_axes_0, keep_dims = var_7851, x = var_7848_cast_fp16)[name = string("op_7852_cast_fp16")];
            fp16 var_7853_to_fp16 = const()[name = string("op_7853_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_277_cast_fp16 = maximum(x = var_7852_cast_fp16, y = var_7853_to_fp16)[name = string("s_277_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> y_277_cast_fp16 = real_div(x = x_309_cast_fp16, y = s_277_cast_fp16)[name = string("y_277_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_7856_cast_fp16 = mul(x = y_277_cast_fp16, y = y_277_cast_fp16)[name = string("op_7856_cast_fp16")];
            tensor<int32, [1]> var_7861_axes_0 = const()[name = string("op_7861_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7861_keep_dims_0 = const()[name = string("op_7861_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_7861_cast_fp16 = reduce_mean(axes = var_7861_axes_0, keep_dims = var_7861_keep_dims_0, x = var_7856_cast_fp16)[name = string("op_7861_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_7862_cast_fp16 = mul(x = s_277_cast_fp16, y = s_277_cast_fp16)[name = string("op_7862_cast_fp16")];
            fp32 var_7863_epsilon_0 = const()[name = string("op_7863_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_7863_cast_fp16 = inverse(epsilon = var_7863_epsilon_0, x = var_7862_cast_fp16)[name = string("op_7863_cast_fp16")];
            fp16 var_7864_to_fp16 = const()[name = string("op_7864_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_7865_cast_fp16 = mul(x = var_7863_cast_fp16, y = var_7864_to_fp16)[name = string("op_7865_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_7867_cast_fp16 = add(x = var_7861_cast_fp16, y = var_7865_cast_fp16)[name = string("op_7867_cast_fp16")];
            fp32 var_7868_epsilon_0 = const()[name = string("op_7868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_7868_cast_fp16 = rsqrt(epsilon = var_7868_epsilon_0, x = var_7867_cast_fp16)[name = string("op_7868_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> out_243_cast_fp16 = mul(x = y_277_cast_fp16, y = var_7868_cast_fp16)[name = string("out_243_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99568704)))];
            tensor<fp16, [1, 4, 48, 512]> q_69_cast_fp16 = mul(x = out_243_cast_fp16, y = tm_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("q_69_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> x_311_cast_fp16 = transpose(perm = x_311_perm_0, x = var_7832_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 1, 48, 512]> var_7871_cast_fp16 = abs(x = x_311_cast_fp16)[name = string("op_7871_cast_fp16")];
            bool var_7874 = const()[name = string("op_7874"), val = bool(true)];
            tensor<int32, [1]> var_7875_axes_0 = const()[name = string("op_7875_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 48, 1]> var_7875_cast_fp16 = reduce_max(axes = var_7875_axes_0, keep_dims = var_7874, x = var_7871_cast_fp16)[name = string("op_7875_cast_fp16")];
            fp16 var_7876_to_fp16 = const()[name = string("op_7876_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 48, 1]> s_279_cast_fp16 = maximum(x = var_7875_cast_fp16, y = var_7876_to_fp16)[name = string("s_279_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> y_279_cast_fp16 = real_div(x = x_311_cast_fp16, y = s_279_cast_fp16)[name = string("y_279_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_7879_cast_fp16 = mul(x = y_279_cast_fp16, y = y_279_cast_fp16)[name = string("op_7879_cast_fp16")];
            tensor<int32, [1]> var_7884_axes_0 = const()[name = string("op_7884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7884_keep_dims_0 = const()[name = string("op_7884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 48, 1]> var_7884_cast_fp16 = reduce_mean(axes = var_7884_axes_0, keep_dims = var_7884_keep_dims_0, x = var_7879_cast_fp16)[name = string("op_7884_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_7885_cast_fp16 = mul(x = s_279_cast_fp16, y = s_279_cast_fp16)[name = string("op_7885_cast_fp16")];
            fp32 var_7886_epsilon_0 = const()[name = string("op_7886_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 48, 1]> var_7886_cast_fp16 = inverse(epsilon = var_7886_epsilon_0, x = var_7885_cast_fp16)[name = string("op_7886_cast_fp16")];
            fp16 var_7887_to_fp16 = const()[name = string("op_7887_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 48, 1]> var_7888_cast_fp16 = mul(x = var_7886_cast_fp16, y = var_7887_to_fp16)[name = string("op_7888_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_7890_cast_fp16 = add(x = var_7884_cast_fp16, y = var_7888_cast_fp16)[name = string("op_7890_cast_fp16")];
            fp32 var_7891_epsilon_0 = const()[name = string("op_7891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 48, 1]> var_7891_cast_fp16 = rsqrt(epsilon = var_7891_epsilon_0, x = var_7890_cast_fp16)[name = string("op_7891_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> out_245_cast_fp16 = mul(x = y_279_cast_fp16, y = var_7891_cast_fp16)[name = string("out_245_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99569792)))];
            tensor<fp16, [1, 1, 48, 512]> k_103_cast_fp16 = mul(x = out_245_cast_fp16, y = tm_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("k_103_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> x_313_cast_fp16 = transpose(perm = x_313_perm_0, x = var_7844_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1, 48, 512]> var_7894_cast_fp16 = abs(x = x_313_cast_fp16)[name = string("op_7894_cast_fp16")];
            bool var_7897 = const()[name = string("op_7897"), val = bool(true)];
            tensor<int32, [1]> var_7898_axes_0 = const()[name = string("op_7898_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 48, 1]> var_7898_cast_fp16 = reduce_max(axes = var_7898_axes_0, keep_dims = var_7897, x = var_7894_cast_fp16)[name = string("op_7898_cast_fp16")];
            fp16 var_7899_to_fp16 = const()[name = string("op_7899_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 48, 1]> s_281_cast_fp16 = maximum(x = var_7898_cast_fp16, y = var_7899_to_fp16)[name = string("s_281_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> y_281_cast_fp16 = real_div(x = x_313_cast_fp16, y = s_281_cast_fp16)[name = string("y_281_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_7902_cast_fp16 = mul(x = y_281_cast_fp16, y = y_281_cast_fp16)[name = string("op_7902_cast_fp16")];
            tensor<int32, [1]> var_7907_axes_0 = const()[name = string("op_7907_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7907_keep_dims_0 = const()[name = string("op_7907_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 48, 1]> var_7907_cast_fp16 = reduce_mean(axes = var_7907_axes_0, keep_dims = var_7907_keep_dims_0, x = var_7902_cast_fp16)[name = string("op_7907_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_7908_cast_fp16 = mul(x = s_281_cast_fp16, y = s_281_cast_fp16)[name = string("op_7908_cast_fp16")];
            fp32 var_7909_epsilon_0 = const()[name = string("op_7909_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 48, 1]> var_7909_cast_fp16 = inverse(epsilon = var_7909_epsilon_0, x = var_7908_cast_fp16)[name = string("op_7909_cast_fp16")];
            fp16 var_7910_to_fp16 = const()[name = string("op_7910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 48, 1]> var_7911_cast_fp16 = mul(x = var_7909_cast_fp16, y = var_7910_to_fp16)[name = string("op_7911_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_7913_cast_fp16 = add(x = var_7907_cast_fp16, y = var_7911_cast_fp16)[name = string("op_7913_cast_fp16")];
            fp32 var_7914_epsilon_0 = const()[name = string("op_7914_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 48, 1]> var_7914_cast_fp16 = rsqrt(epsilon = var_7914_epsilon_0, x = var_7913_cast_fp16)[name = string("op_7914_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> v_69_cast_fp16 = mul(x = y_281_cast_fp16, y = var_7914_cast_fp16)[name = string("v_69_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_7916_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7916_cast_fp16")];
            tensor<int32, [4]> var_7921_begin_0 = const()[name = string("op_7921_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7921_end_0 = const()[name = string("op_7921_end_0"), val = tensor<int32, [4]>([1, 4, 48, 512])];
            tensor<bool, [4]> var_7921_end_mask_0 = const()[name = string("op_7921_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 256]> var_7921_cast_fp16 = slice_by_index(begin = var_7921_begin_0, end = var_7921_end_0, end_mask = var_7921_end_mask_0, x = q_69_cast_fp16)[name = string("op_7921_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 256]> var_7922_cast_fp16 = mul(x = var_7921_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 256]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = q_69_cast_fp16)[name = string("op_7927_cast_fp16")];
            int32 var_7929 = const()[name = string("op_7929"), val = int32(-1)];
            bool var_7930_interleave_0 = const()[name = string("op_7930_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 512]> var_7930_cast_fp16 = concat(axis = var_7929, interleave = var_7930_interleave_0, values = (var_7922_cast_fp16, var_7927_cast_fp16))[name = string("op_7930_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_7931_cast_fp16 = mul(x = var_7930_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7931_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> q_71_cast_fp16 = add(x = var_7916_cast_fp16, y = var_7931_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_7934_cast_fp16 = mul(x = k_103_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7934_cast_fp16")];
            tensor<int32, [4]> var_7939_begin_0 = const()[name = string("op_7939_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7939_end_0 = const()[name = string("op_7939_end_0"), val = tensor<int32, [4]>([1, 1, 48, 512])];
            tensor<bool, [4]> var_7939_end_mask_0 = const()[name = string("op_7939_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 48, 256]> var_7939_cast_fp16 = slice_by_index(begin = var_7939_begin_0, end = var_7939_end_0, end_mask = var_7939_end_mask_0, x = k_103_cast_fp16)[name = string("op_7939_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 48, 256]> var_7940_cast_fp16 = mul(x = var_7939_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_7940_cast_fp16")];
            tensor<int32, [4]> var_7945_begin_0 = const()[name = string("op_7945_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7945_end_0 = const()[name = string("op_7945_end_0"), val = tensor<int32, [4]>([1, 1, 48, 256])];
            tensor<bool, [4]> var_7945_end_mask_0 = const()[name = string("op_7945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 48, 256]> var_7945_cast_fp16 = slice_by_index(begin = var_7945_begin_0, end = var_7945_end_0, end_mask = var_7945_end_mask_0, x = k_103_cast_fp16)[name = string("op_7945_cast_fp16")];
            int32 var_7947 = const()[name = string("op_7947"), val = int32(-1)];
            bool var_7948_interleave_0 = const()[name = string("op_7948_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 48, 512]> var_7948_cast_fp16 = concat(axis = var_7947, interleave = var_7948_interleave_0, values = (var_7940_cast_fp16, var_7945_cast_fp16))[name = string("op_7948_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_7949_cast_fp16 = mul(x = var_7948_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> k_105_cast_fp16 = add(x = var_7934_cast_fp16, y = var_7949_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_34_reps_0 = const()[name = string("tile_34_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 48, 512]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = k_105_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 1, 48, 512]> tile_34_cast_fp16 = tile(reps = tile_34_reps_0, x = transpose_68_cast_fp16)[name = string("tile_34_cast_fp16")];
            tensor<int32, [5]> concat_68 = const()[name = string("concat_68"), val = tensor<int32, [5]>([4, 1, 1, 48, 512])];
            tensor<fp16, [4, 1, 1, 48, 512]> reshape_68_cast_fp16 = reshape(shape = concat_68, x = tile_34_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [5]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([-1, 1, 48, 512])];
            tensor<fp16, [1, 4, 1, 48, 512]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 1, 48, 512]> reshape_69_cast_fp16 = reshape(shape = concat_69, x = transpose_69_cast_fp16)[name = string("reshape_69_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_35_reps_0 = const()[name = string("tile_35_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 48, 512]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = v_69_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 1, 48, 512]> tile_35_cast_fp16 = tile(reps = tile_35_reps_0, x = transpose_70_cast_fp16)[name = string("tile_35_cast_fp16")];
            tensor<int32, [5]> concat_70 = const()[name = string("concat_70"), val = tensor<int32, [5]>([4, 1, 1, 48, 512])];
            tensor<fp16, [4, 1, 1, 48, 512]> reshape_70_cast_fp16 = reshape(shape = concat_70, x = tile_35_cast_fp16)[name = string("reshape_70_cast_fp16")];
            tensor<int32, [5]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([-1, 1, 48, 512])];
            tensor<fp16, [1, 4, 1, 48, 512]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = reshape_70_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 1, 48, 512]> reshape_71_cast_fp16 = reshape(shape = concat_71, x = transpose_71_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7963_transpose_x_0 = const()[name = string("op_7963_transpose_x_0"), val = bool(false)];
            bool var_7963_transpose_y_0 = const()[name = string("op_7963_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 48]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = reshape_69_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 4, 48, 48]> var_7963_cast_fp16 = matmul(transpose_x = var_7963_transpose_x_0, transpose_y = var_7963_transpose_y_0, x = q_71_cast_fp16, y = transpose_113_cast_fp16)[name = string("op_7963_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_7965_cast_fp16 = add(x = var_7963_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7965_cast_fp16")];
            int32 var_7966 = const()[name = string("op_7966"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_281_cast_fp16 = softmax(axis = var_7966, x = var_7965_cast_fp16)[name = string("w_281_cast_fp16")];
            bool var_7969_transpose_x_0 = const()[name = string("op_7969_transpose_x_0"), val = bool(false)];
            bool var_7969_transpose_y_0 = const()[name = string("op_7969_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 512]> v_71_cast_fp16 = transpose(perm = v_71_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 4, 48, 512]> var_7969_cast_fp16 = matmul(transpose_x = var_7969_transpose_x_0, transpose_y = var_7969_transpose_y_0, x = w_281_cast_fp16, y = v_71_cast_fp16)[name = string("op_7969_cast_fp16")];
            tensor<int32, [4]> var_7972_perm_0 = const()[name = string("op_7972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7976 = const()[name = string("op_7976"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 48, 4, 512]> var_7972_cast_fp16 = transpose(perm = var_7972_perm_0, x = var_7969_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 48, 2048]> input_207_cast_fp16 = reshape(shape = var_7976, x = var_7972_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99570880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619520))))[name = string("tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7981_cast_fp16 = abs(x = linear_157_cast_fp16)[name = string("op_7981_cast_fp16")];
            bool var_7984 = const()[name = string("op_7984"), val = bool(true)];
            tensor<int32, [1]> var_7985_axes_0 = const()[name = string("op_7985_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_7985_cast_fp16 = reduce_max(axes = var_7985_axes_0, keep_dims = var_7984, x = var_7981_cast_fp16)[name = string("op_7985_cast_fp16")];
            fp16 var_7986_to_fp16 = const()[name = string("op_7986_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_283_cast_fp16 = maximum(x = var_7985_cast_fp16, y = var_7986_to_fp16)[name = string("s_283_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_283_cast_fp16 = real_div(x = linear_157_cast_fp16, y = s_283_cast_fp16)[name = string("y_283_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_7989_cast_fp16 = mul(x = y_283_cast_fp16, y = y_283_cast_fp16)[name = string("op_7989_cast_fp16")];
            tensor<int32, [1]> var_7994_axes_0 = const()[name = string("op_7994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7994_keep_dims_0 = const()[name = string("op_7994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_7994_cast_fp16 = reduce_mean(axes = var_7994_axes_0, keep_dims = var_7994_keep_dims_0, x = var_7989_cast_fp16)[name = string("op_7994_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_7995_cast_fp16 = mul(x = s_283_cast_fp16, y = s_283_cast_fp16)[name = string("op_7995_cast_fp16")];
            fp32 var_7996_epsilon_0 = const()[name = string("op_7996_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_7996_cast_fp16 = inverse(epsilon = var_7996_epsilon_0, x = var_7995_cast_fp16)[name = string("op_7996_cast_fp16")];
            fp16 var_7997_to_fp16 = const()[name = string("op_7997_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_7998_cast_fp16 = mul(x = var_7996_cast_fp16, y = var_7997_to_fp16)[name = string("op_7998_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8000_cast_fp16 = add(x = var_7994_cast_fp16, y = var_7998_cast_fp16)[name = string("op_8000_cast_fp16")];
            fp32 var_8001_epsilon_0 = const()[name = string("op_8001_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8001_cast_fp16 = rsqrt(epsilon = var_8001_epsilon_0, x = var_8000_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_247_cast_fp16 = mul(x = y_283_cast_fp16, y = var_8001_cast_fp16)[name = string("out_247_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100620608)))];
            tensor<fp16, [1, 48, 512]> var_8003_cast_fp16 = mul(x = out_247_cast_fp16, y = tm_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("op_8003_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_317_cast_fp16 = add(x = x_307_cast_fp16, y = var_8003_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8006_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_8006_cast_fp16")];
            bool var_8009 = const()[name = string("op_8009"), val = bool(true)];
            tensor<int32, [1]> var_8010_axes_0 = const()[name = string("op_8010_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8010_cast_fp16 = reduce_max(axes = var_8010_axes_0, keep_dims = var_8009, x = var_8006_cast_fp16)[name = string("op_8010_cast_fp16")];
            fp16 var_8011_to_fp16 = const()[name = string("op_8011_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_285_cast_fp16 = maximum(x = var_8010_cast_fp16, y = var_8011_to_fp16)[name = string("s_285_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_285_cast_fp16 = real_div(x = x_317_cast_fp16, y = s_285_cast_fp16)[name = string("y_285_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8014_cast_fp16 = mul(x = y_285_cast_fp16, y = y_285_cast_fp16)[name = string("op_8014_cast_fp16")];
            tensor<int32, [1]> var_8019_axes_0 = const()[name = string("op_8019_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8019_keep_dims_0 = const()[name = string("op_8019_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8019_cast_fp16 = reduce_mean(axes = var_8019_axes_0, keep_dims = var_8019_keep_dims_0, x = var_8014_cast_fp16)[name = string("op_8019_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8020_cast_fp16 = mul(x = s_285_cast_fp16, y = s_285_cast_fp16)[name = string("op_8020_cast_fp16")];
            fp32 var_8021_epsilon_0 = const()[name = string("op_8021_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8021_cast_fp16 = inverse(epsilon = var_8021_epsilon_0, x = var_8020_cast_fp16)[name = string("op_8021_cast_fp16")];
            fp16 var_8022_to_fp16 = const()[name = string("op_8022_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8023_cast_fp16 = mul(x = var_8021_cast_fp16, y = var_8022_to_fp16)[name = string("op_8023_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8025_cast_fp16 = add(x = var_8019_cast_fp16, y = var_8023_cast_fp16)[name = string("op_8025_cast_fp16")];
            fp32 var_8026_epsilon_0 = const()[name = string("op_8026_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8026_cast_fp16 = rsqrt(epsilon = var_8026_epsilon_0, x = var_8025_cast_fp16)[name = string("op_8026_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_249_cast_fp16 = mul(x = y_285_cast_fp16, y = var_8026_cast_fp16)[name = string("out_249_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621696)))];
            tensor<fp16, [1, 48, 512]> input_209_cast_fp16 = mul(x = out_249_cast_fp16, y = tm_layers_17_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100622784))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101671424))))[name = string("tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_158_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_158_cast_fp16")];
            string var_8033_mode_0 = const()[name = string("op_8033_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_8033_cast_fp16 = gelu(mode = var_8033_mode_0, x = linear_158_cast_fp16)[name = string("op_8033_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101675584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102724224))))[name = string("tm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_159_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_211_cast_fp16 = mul(x = var_8033_cast_fp16, y = linear_159_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102728384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103777024))))[name = string("tm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8041_cast_fp16 = abs(x = linear_160_cast_fp16)[name = string("op_8041_cast_fp16")];
            bool var_8044 = const()[name = string("op_8044"), val = bool(true)];
            tensor<int32, [1]> var_8045_axes_0 = const()[name = string("op_8045_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8045_cast_fp16 = reduce_max(axes = var_8045_axes_0, keep_dims = var_8044, x = var_8041_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp16 var_8046_to_fp16 = const()[name = string("op_8046_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_287_cast_fp16 = maximum(x = var_8045_cast_fp16, y = var_8046_to_fp16)[name = string("s_287_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_287_cast_fp16 = real_div(x = linear_160_cast_fp16, y = s_287_cast_fp16)[name = string("y_287_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8049_cast_fp16 = mul(x = y_287_cast_fp16, y = y_287_cast_fp16)[name = string("op_8049_cast_fp16")];
            tensor<int32, [1]> var_8054_axes_0 = const()[name = string("op_8054_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8054_keep_dims_0 = const()[name = string("op_8054_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8054_cast_fp16 = reduce_mean(axes = var_8054_axes_0, keep_dims = var_8054_keep_dims_0, x = var_8049_cast_fp16)[name = string("op_8054_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8055_cast_fp16 = mul(x = s_287_cast_fp16, y = s_287_cast_fp16)[name = string("op_8055_cast_fp16")];
            fp32 var_8056_epsilon_0 = const()[name = string("op_8056_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8056_cast_fp16 = inverse(epsilon = var_8056_epsilon_0, x = var_8055_cast_fp16)[name = string("op_8056_cast_fp16")];
            fp16 var_8057_to_fp16 = const()[name = string("op_8057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8058_cast_fp16 = mul(x = var_8056_cast_fp16, y = var_8057_to_fp16)[name = string("op_8058_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8060_cast_fp16 = add(x = var_8054_cast_fp16, y = var_8058_cast_fp16)[name = string("op_8060_cast_fp16")];
            fp32 var_8061_epsilon_0 = const()[name = string("op_8061_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8061_cast_fp16 = rsqrt(epsilon = var_8061_epsilon_0, x = var_8060_cast_fp16)[name = string("op_8061_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_251_cast_fp16 = mul(x = y_287_cast_fp16, y = var_8061_cast_fp16)[name = string("out_251_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103778112)))];
            tensor<fp16, [1, 48, 512]> var_8063_cast_fp16 = mul(x = out_251_cast_fp16, y = tm_layers_17_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_213_cast_fp16 = add(x = x_317_cast_fp16, y = var_8063_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103779200))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041408))))[name = string("tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_161_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_161_cast_fp16")];
            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_321_cast_fp16 = gelu(mode = x_321_mode_0, x = linear_161_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [4]> var_8083_begin_0 = const()[name = string("op_8083_begin_0"), val = tensor<int32, [4]>([0, 0, 17, 0])];
            tensor<int32, [4]> var_8083_end_0 = const()[name = string("op_8083_end_0"), val = tensor<int32, [4]>([1, 48, 18, 512])];
            tensor<bool, [4]> var_8083_end_mask_0 = const()[name = string("op_8083_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8083_squeeze_mask_0 = const()[name = string("op_8083_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_8083_cast_fp16 = slice_by_index(begin = var_8083_begin_0, end = var_8083_end_0, end_mask = var_8083_end_mask_0, squeeze_mask = var_8083_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8083_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_215_cast_fp16 = mul(x = x_321_cast_fp16, y = var_8083_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104042496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104304704))))[name = string("tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8093_cast_fp16 = abs(x = linear_162_cast_fp16)[name = string("op_8093_cast_fp16")];
            bool var_8096 = const()[name = string("op_8096"), val = bool(true)];
            tensor<int32, [1]> var_8097_axes_0 = const()[name = string("op_8097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8097_cast_fp16 = reduce_max(axes = var_8097_axes_0, keep_dims = var_8096, x = var_8093_cast_fp16)[name = string("op_8097_cast_fp16")];
            fp16 var_8098_to_fp16 = const()[name = string("op_8098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_289_cast_fp16 = maximum(x = var_8097_cast_fp16, y = var_8098_to_fp16)[name = string("s_289_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_289_cast_fp16 = real_div(x = linear_162_cast_fp16, y = s_289_cast_fp16)[name = string("y_289_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8101_cast_fp16 = mul(x = y_289_cast_fp16, y = y_289_cast_fp16)[name = string("op_8101_cast_fp16")];
            tensor<int32, [1]> var_8106_axes_0 = const()[name = string("op_8106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8106_keep_dims_0 = const()[name = string("op_8106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8106_cast_fp16 = reduce_mean(axes = var_8106_axes_0, keep_dims = var_8106_keep_dims_0, x = var_8101_cast_fp16)[name = string("op_8106_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8107_cast_fp16 = mul(x = s_289_cast_fp16, y = s_289_cast_fp16)[name = string("op_8107_cast_fp16")];
            fp32 var_8108_epsilon_0 = const()[name = string("op_8108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8108_cast_fp16 = inverse(epsilon = var_8108_epsilon_0, x = var_8107_cast_fp16)[name = string("op_8108_cast_fp16")];
            fp16 var_8109_to_fp16 = const()[name = string("op_8109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8110_cast_fp16 = mul(x = var_8108_cast_fp16, y = var_8109_to_fp16)[name = string("op_8110_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8112_cast_fp16 = add(x = var_8106_cast_fp16, y = var_8110_cast_fp16)[name = string("op_8112_cast_fp16")];
            fp32 var_8113_epsilon_0 = const()[name = string("op_8113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8113_cast_fp16 = rsqrt(epsilon = var_8113_epsilon_0, x = var_8112_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_253_cast_fp16 = mul(x = y_289_cast_fp16, y = var_8113_cast_fp16)[name = string("out_253_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104305792)))];
            tensor<fp16, [1, 48, 512]> var_8115_cast_fp16 = mul(x = out_253_cast_fp16, y = tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8115_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_35_cast_fp16 = add(x = input_213_cast_fp16, y = var_8115_cast_fp16)[name = string("h_35_cast_fp16")];
            tensor<fp16, [1]> tm_layers_17_layer_scalar_to_fp16 = const()[name = string("tm_layers_17_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e6p-1])];
            tensor<fp16, [1, 48, 512]> x_325_cast_fp16 = mul(x = h_35_cast_fp16, y = tm_layers_17_layer_scalar_to_fp16)[name = string("x_325_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8119_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_8119_cast_fp16")];
            bool var_8122 = const()[name = string("op_8122"), val = bool(true)];
            tensor<int32, [1]> var_8123_axes_0 = const()[name = string("op_8123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8123_cast_fp16 = reduce_max(axes = var_8123_axes_0, keep_dims = var_8122, x = var_8119_cast_fp16)[name = string("op_8123_cast_fp16")];
            fp16 var_8124_to_fp16 = const()[name = string("op_8124_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_291_cast_fp16 = maximum(x = var_8123_cast_fp16, y = var_8124_to_fp16)[name = string("s_291_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_291_cast_fp16 = real_div(x = x_325_cast_fp16, y = s_291_cast_fp16)[name = string("y_291_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8127_cast_fp16 = mul(x = y_291_cast_fp16, y = y_291_cast_fp16)[name = string("op_8127_cast_fp16")];
            tensor<int32, [1]> var_8132_axes_0 = const()[name = string("op_8132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8132_keep_dims_0 = const()[name = string("op_8132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8132_cast_fp16 = reduce_mean(axes = var_8132_axes_0, keep_dims = var_8132_keep_dims_0, x = var_8127_cast_fp16)[name = string("op_8132_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8133_cast_fp16 = mul(x = s_291_cast_fp16, y = s_291_cast_fp16)[name = string("op_8133_cast_fp16")];
            fp32 var_8134_epsilon_0 = const()[name = string("op_8134_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8134_cast_fp16 = inverse(epsilon = var_8134_epsilon_0, x = var_8133_cast_fp16)[name = string("op_8134_cast_fp16")];
            fp16 var_8135_to_fp16 = const()[name = string("op_8135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = string("op_8136_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8138_cast_fp16 = add(x = var_8132_cast_fp16, y = var_8136_cast_fp16)[name = string("op_8138_cast_fp16")];
            fp32 var_8139_epsilon_0 = const()[name = string("op_8139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8139_cast_fp16 = rsqrt(epsilon = var_8139_epsilon_0, x = var_8138_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_255_cast_fp16 = mul(x = y_291_cast_fp16, y = var_8139_cast_fp16)[name = string("out_255_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104306880)))];
            tensor<fp16, [1, 48, 512]> input_217_cast_fp16 = mul(x = out_255_cast_fp16, y = tm_layers_18_input_layernorm_weight_to_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104307968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104832320))))[name = string("tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8149 = const()[name = string("op_8149"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_8150_cast_fp16 = reshape(shape = var_8149, x = linear_163_cast_fp16)[name = string("op_8150_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104834432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105096640))))[name = string("tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [4]> var_8161 = const()[name = string("op_8161"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_8162_cast_fp16 = reshape(shape = var_8161, x = linear_164_cast_fp16)[name = string("op_8162_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105097728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105359936))))[name = string("tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<int32, [4]> var_8173 = const()[name = string("op_8173"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_8174_cast_fp16 = reshape(shape = var_8173, x = linear_165_cast_fp16)[name = string("op_8174_cast_fp16")];
            tensor<int32, [4]> x_331_perm_0 = const()[name = string("x_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_8150_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 4, 48, 256]> var_8178_cast_fp16 = abs(x = x_327_cast_fp16)[name = string("op_8178_cast_fp16")];
            bool var_8181 = const()[name = string("op_8181"), val = bool(true)];
            tensor<int32, [1]> var_8182_axes_0 = const()[name = string("op_8182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_8182_cast_fp16 = reduce_max(axes = var_8182_axes_0, keep_dims = var_8181, x = var_8178_cast_fp16)[name = string("op_8182_cast_fp16")];
            fp16 var_8183_to_fp16 = const()[name = string("op_8183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_293_cast_fp16 = maximum(x = var_8182_cast_fp16, y = var_8183_to_fp16)[name = string("s_293_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_293_cast_fp16 = real_div(x = x_327_cast_fp16, y = s_293_cast_fp16)[name = string("y_293_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8186_cast_fp16 = mul(x = y_293_cast_fp16, y = y_293_cast_fp16)[name = string("op_8186_cast_fp16")];
            tensor<int32, [1]> var_8191_axes_0 = const()[name = string("op_8191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8191_keep_dims_0 = const()[name = string("op_8191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_8191_cast_fp16 = reduce_mean(axes = var_8191_axes_0, keep_dims = var_8191_keep_dims_0, x = var_8186_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_8192_cast_fp16 = mul(x = s_293_cast_fp16, y = s_293_cast_fp16)[name = string("op_8192_cast_fp16")];
            fp32 var_8193_epsilon_0 = const()[name = string("op_8193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_8193_cast_fp16 = inverse(epsilon = var_8193_epsilon_0, x = var_8192_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_8197_cast_fp16 = add(x = var_8191_cast_fp16, y = var_8195_cast_fp16)[name = string("op_8197_cast_fp16")];
            fp32 var_8198_epsilon_0 = const()[name = string("op_8198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_8198_cast_fp16 = rsqrt(epsilon = var_8198_epsilon_0, x = var_8197_cast_fp16)[name = string("op_8198_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_257_cast_fp16 = mul(x = y_293_cast_fp16, y = var_8198_cast_fp16)[name = string("out_257_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_73_cast_fp16 = mul(x = out_257_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_73_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_8162_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 48, 256]> var_8201_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_8201_cast_fp16")];
            bool var_8204 = const()[name = string("op_8204"), val = bool(true)];
            tensor<int32, [1]> var_8205_axes_0 = const()[name = string("op_8205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_8205_cast_fp16 = reduce_max(axes = var_8205_axes_0, keep_dims = var_8204, x = var_8201_cast_fp16)[name = string("op_8205_cast_fp16")];
            fp16 var_8206_to_fp16 = const()[name = string("op_8206_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_295_cast_fp16 = maximum(x = var_8205_cast_fp16, y = var_8206_to_fp16)[name = string("s_295_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_295_cast_fp16 = real_div(x = x_329_cast_fp16, y = s_295_cast_fp16)[name = string("y_295_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8209_cast_fp16 = mul(x = y_295_cast_fp16, y = y_295_cast_fp16)[name = string("op_8209_cast_fp16")];
            tensor<int32, [1]> var_8214_axes_0 = const()[name = string("op_8214_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8214_keep_dims_0 = const()[name = string("op_8214_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_8214_cast_fp16 = reduce_mean(axes = var_8214_axes_0, keep_dims = var_8214_keep_dims_0, x = var_8209_cast_fp16)[name = string("op_8214_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8215_cast_fp16 = mul(x = s_295_cast_fp16, y = s_295_cast_fp16)[name = string("op_8215_cast_fp16")];
            fp32 var_8216_epsilon_0 = const()[name = string("op_8216_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_8216_cast_fp16 = inverse(epsilon = var_8216_epsilon_0, x = var_8215_cast_fp16)[name = string("op_8216_cast_fp16")];
            fp16 var_8217_to_fp16 = const()[name = string("op_8217_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_8218_cast_fp16 = mul(x = var_8216_cast_fp16, y = var_8217_to_fp16)[name = string("op_8218_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8220_cast_fp16 = add(x = var_8214_cast_fp16, y = var_8218_cast_fp16)[name = string("op_8220_cast_fp16")];
            fp32 var_8221_epsilon_0 = const()[name = string("op_8221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_8221_cast_fp16 = rsqrt(epsilon = var_8221_epsilon_0, x = var_8220_cast_fp16)[name = string("op_8221_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_259_cast_fp16 = mul(x = y_295_cast_fp16, y = var_8221_cast_fp16)[name = string("out_259_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_109_cast_fp16 = mul(x = out_259_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_109_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_331_cast_fp16 = transpose(perm = x_331_perm_0, x = var_8174_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 48, 256]> var_8224_cast_fp16 = abs(x = x_331_cast_fp16)[name = string("op_8224_cast_fp16")];
            bool var_8227 = const()[name = string("op_8227"), val = bool(true)];
            tensor<int32, [1]> var_8228_axes_0 = const()[name = string("op_8228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_8228_cast_fp16 = reduce_max(axes = var_8228_axes_0, keep_dims = var_8227, x = var_8224_cast_fp16)[name = string("op_8228_cast_fp16")];
            fp16 var_8229_to_fp16 = const()[name = string("op_8229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_297_cast_fp16 = maximum(x = var_8228_cast_fp16, y = var_8229_to_fp16)[name = string("s_297_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_297_cast_fp16 = real_div(x = x_331_cast_fp16, y = s_297_cast_fp16)[name = string("y_297_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8232_cast_fp16 = mul(x = y_297_cast_fp16, y = y_297_cast_fp16)[name = string("op_8232_cast_fp16")];
            tensor<int32, [1]> var_8237_axes_0 = const()[name = string("op_8237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8237_keep_dims_0 = const()[name = string("op_8237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_8237_cast_fp16 = reduce_mean(axes = var_8237_axes_0, keep_dims = var_8237_keep_dims_0, x = var_8232_cast_fp16)[name = string("op_8237_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8238_cast_fp16 = mul(x = s_297_cast_fp16, y = s_297_cast_fp16)[name = string("op_8238_cast_fp16")];
            fp32 var_8239_epsilon_0 = const()[name = string("op_8239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_8239_cast_fp16 = inverse(epsilon = var_8239_epsilon_0, x = var_8238_cast_fp16)[name = string("op_8239_cast_fp16")];
            fp16 var_8240_to_fp16 = const()[name = string("op_8240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_8241_cast_fp16 = mul(x = var_8239_cast_fp16, y = var_8240_to_fp16)[name = string("op_8241_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8243_cast_fp16 = add(x = var_8237_cast_fp16, y = var_8241_cast_fp16)[name = string("op_8243_cast_fp16")];
            fp32 var_8244_epsilon_0 = const()[name = string("op_8244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_8244_cast_fp16 = rsqrt(epsilon = var_8244_epsilon_0, x = var_8243_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_73_cast_fp16 = mul(x = y_297_cast_fp16, y = var_8244_cast_fp16)[name = string("v_73_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8246_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8246_cast_fp16")];
            tensor<int32, [4]> var_8251_begin_0 = const()[name = string("op_8251_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8251_end_0 = const()[name = string("op_8251_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_8251_end_mask_0 = const()[name = string("op_8251_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_8251_cast_fp16 = slice_by_index(begin = var_8251_begin_0, end = var_8251_end_0, end_mask = var_8251_end_mask_0, x = q_73_cast_fp16)[name = string("op_8251_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_8252_cast_fp16 = mul(x = var_8251_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8257_begin_0 = const()[name = string("op_8257_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8257_end_0 = const()[name = string("op_8257_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_8257_end_mask_0 = const()[name = string("op_8257_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_8257_cast_fp16 = slice_by_index(begin = var_8257_begin_0, end = var_8257_end_0, end_mask = var_8257_end_mask_0, x = q_73_cast_fp16)[name = string("op_8257_cast_fp16")];
            int32 var_8259 = const()[name = string("op_8259"), val = int32(-1)];
            bool var_8260_interleave_0 = const()[name = string("op_8260_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_8260_cast_fp16 = concat(axis = var_8259, interleave = var_8260_interleave_0, values = (var_8252_cast_fp16, var_8257_cast_fp16))[name = string("op_8260_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8261_cast_fp16 = mul(x = var_8260_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8261_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_75_cast_fp16 = add(x = var_8246_cast_fp16, y = var_8261_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8264_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8264_cast_fp16")];
            tensor<int32, [4]> var_8269_begin_0 = const()[name = string("op_8269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8269_end_0 = const()[name = string("op_8269_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_8269_end_mask_0 = const()[name = string("op_8269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_8269_cast_fp16 = slice_by_index(begin = var_8269_begin_0, end = var_8269_end_0, end_mask = var_8269_end_mask_0, x = k_109_cast_fp16)[name = string("op_8269_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_8270_cast_fp16 = mul(x = var_8269_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_8270_cast_fp16")];
            tensor<int32, [4]> var_8275_begin_0 = const()[name = string("op_8275_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8275_end_0 = const()[name = string("op_8275_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_8275_end_mask_0 = const()[name = string("op_8275_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_8275_cast_fp16 = slice_by_index(begin = var_8275_begin_0, end = var_8275_end_0, end_mask = var_8275_end_mask_0, x = k_109_cast_fp16)[name = string("op_8275_cast_fp16")];
            int32 var_8277 = const()[name = string("op_8277"), val = int32(-1)];
            bool var_8278_interleave_0 = const()[name = string("op_8278_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_8278_cast_fp16 = concat(axis = var_8277, interleave = var_8278_interleave_0, values = (var_8270_cast_fp16, var_8275_cast_fp16))[name = string("op_8278_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8279_cast_fp16 = mul(x = var_8278_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8279_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_111_cast_fp16 = add(x = var_8264_cast_fp16, y = var_8279_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_36_reps_0 = const()[name = string("tile_36_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_72_cast_fp16 = transpose(perm = transpose_72_perm_0, x = k_111_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 1, 48, 256]> tile_36_cast_fp16 = tile(reps = tile_36_reps_0, x = transpose_72_cast_fp16)[name = string("tile_36_cast_fp16")];
            tensor<int32, [5]> concat_72 = const()[name = string("concat_72"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_72_cast_fp16 = reshape(shape = concat_72, x = tile_36_cast_fp16)[name = string("reshape_72_cast_fp16")];
            tensor<int32, [5]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_73 = const()[name = string("concat_73"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = reshape_72_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 1, 48, 256]> reshape_73_cast_fp16 = reshape(shape = concat_73, x = transpose_73_cast_fp16)[name = string("reshape_73_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_37_reps_0 = const()[name = string("tile_37_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = v_73_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 1, 48, 256]> tile_37_cast_fp16 = tile(reps = tile_37_reps_0, x = transpose_74_cast_fp16)[name = string("tile_37_cast_fp16")];
            tensor<int32, [5]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_74_cast_fp16 = reshape(shape = concat_74, x = tile_37_cast_fp16)[name = string("reshape_74_cast_fp16")];
            tensor<int32, [5]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = reshape_74_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 1, 48, 256]> reshape_75_cast_fp16 = reshape(shape = concat_75, x = transpose_75_cast_fp16)[name = string("reshape_75_cast_fp16")];
            tensor<int32, [4]> v_75_perm_0 = const()[name = string("v_75_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8293_transpose_x_0 = const()[name = string("op_8293_transpose_x_0"), val = bool(false)];
            bool var_8293_transpose_y_0 = const()[name = string("op_8293_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = reshape_73_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 4, 48, 48]> var_8293_cast_fp16 = matmul(transpose_x = var_8293_transpose_x_0, transpose_y = var_8293_transpose_y_0, x = q_75_cast_fp16, y = transpose_114_cast_fp16)[name = string("op_8293_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_8295_cast_fp16 = add(x = var_8293_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8295_cast_fp16")];
            int32 var_8296 = const()[name = string("op_8296"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_297_cast_fp16 = softmax(axis = var_8296, x = var_8295_cast_fp16)[name = string("w_297_cast_fp16")];
            bool var_8299_transpose_x_0 = const()[name = string("op_8299_transpose_x_0"), val = bool(false)];
            bool var_8299_transpose_y_0 = const()[name = string("op_8299_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_75_cast_fp16 = transpose(perm = v_75_perm_0, x = reshape_75_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 4, 48, 256]> var_8299_cast_fp16 = matmul(transpose_x = var_8299_transpose_x_0, transpose_y = var_8299_transpose_y_0, x = w_297_cast_fp16, y = v_75_cast_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [4]> var_8302_perm_0 = const()[name = string("op_8302_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8306 = const()[name = string("op_8306"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_8302_cast_fp16 = transpose(perm = var_8302_perm_0, x = var_8299_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 48, 1024]> input_219_cast_fp16 = reshape(shape = var_8306, x = var_8302_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105361024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105885376))))[name = string("tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8311_cast_fp16 = abs(x = linear_166_cast_fp16)[name = string("op_8311_cast_fp16")];
            bool var_8314 = const()[name = string("op_8314"), val = bool(true)];
            tensor<int32, [1]> var_8315_axes_0 = const()[name = string("op_8315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8315_cast_fp16 = reduce_max(axes = var_8315_axes_0, keep_dims = var_8314, x = var_8311_cast_fp16)[name = string("op_8315_cast_fp16")];
            fp16 var_8316_to_fp16 = const()[name = string("op_8316_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_299_cast_fp16 = maximum(x = var_8315_cast_fp16, y = var_8316_to_fp16)[name = string("s_299_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_299_cast_fp16 = real_div(x = linear_166_cast_fp16, y = s_299_cast_fp16)[name = string("y_299_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8319_cast_fp16 = mul(x = y_299_cast_fp16, y = y_299_cast_fp16)[name = string("op_8319_cast_fp16")];
            tensor<int32, [1]> var_8324_axes_0 = const()[name = string("op_8324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8324_keep_dims_0 = const()[name = string("op_8324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8324_cast_fp16 = reduce_mean(axes = var_8324_axes_0, keep_dims = var_8324_keep_dims_0, x = var_8319_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8325_cast_fp16 = mul(x = s_299_cast_fp16, y = s_299_cast_fp16)[name = string("op_8325_cast_fp16")];
            fp32 var_8326_epsilon_0 = const()[name = string("op_8326_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8326_cast_fp16 = inverse(epsilon = var_8326_epsilon_0, x = var_8325_cast_fp16)[name = string("op_8326_cast_fp16")];
            fp16 var_8327_to_fp16 = const()[name = string("op_8327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8328_cast_fp16 = mul(x = var_8326_cast_fp16, y = var_8327_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8330_cast_fp16 = add(x = var_8324_cast_fp16, y = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_261_cast_fp16 = mul(x = y_299_cast_fp16, y = var_8331_cast_fp16)[name = string("out_261_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105886464)))];
            tensor<fp16, [1, 48, 512]> var_8333_cast_fp16 = mul(x = out_261_cast_fp16, y = tm_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("op_8333_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = var_8333_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8336_cast_fp16 = abs(x = x_335_cast_fp16)[name = string("op_8336_cast_fp16")];
            bool var_8339 = const()[name = string("op_8339"), val = bool(true)];
            tensor<int32, [1]> var_8340_axes_0 = const()[name = string("op_8340_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8340_cast_fp16 = reduce_max(axes = var_8340_axes_0, keep_dims = var_8339, x = var_8336_cast_fp16)[name = string("op_8340_cast_fp16")];
            fp16 var_8341_to_fp16 = const()[name = string("op_8341_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_301_cast_fp16 = maximum(x = var_8340_cast_fp16, y = var_8341_to_fp16)[name = string("s_301_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_301_cast_fp16 = real_div(x = x_335_cast_fp16, y = s_301_cast_fp16)[name = string("y_301_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8344_cast_fp16 = mul(x = y_301_cast_fp16, y = y_301_cast_fp16)[name = string("op_8344_cast_fp16")];
            tensor<int32, [1]> var_8349_axes_0 = const()[name = string("op_8349_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8349_keep_dims_0 = const()[name = string("op_8349_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8349_cast_fp16 = reduce_mean(axes = var_8349_axes_0, keep_dims = var_8349_keep_dims_0, x = var_8344_cast_fp16)[name = string("op_8349_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8350_cast_fp16 = mul(x = s_301_cast_fp16, y = s_301_cast_fp16)[name = string("op_8350_cast_fp16")];
            fp32 var_8351_epsilon_0 = const()[name = string("op_8351_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8351_cast_fp16 = inverse(epsilon = var_8351_epsilon_0, x = var_8350_cast_fp16)[name = string("op_8351_cast_fp16")];
            fp16 var_8352_to_fp16 = const()[name = string("op_8352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8353_cast_fp16 = mul(x = var_8351_cast_fp16, y = var_8352_to_fp16)[name = string("op_8353_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8355_cast_fp16 = add(x = var_8349_cast_fp16, y = var_8353_cast_fp16)[name = string("op_8355_cast_fp16")];
            fp32 var_8356_epsilon_0 = const()[name = string("op_8356_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8356_cast_fp16 = rsqrt(epsilon = var_8356_epsilon_0, x = var_8355_cast_fp16)[name = string("op_8356_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_263_cast_fp16 = mul(x = y_301_cast_fp16, y = var_8356_cast_fp16)[name = string("out_263_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105887552)))];
            tensor<fp16, [1, 48, 512]> input_221_cast_fp16 = mul(x = out_263_cast_fp16, y = tm_layers_18_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105888640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106937280))))[name = string("tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_167_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_167_cast_fp16")];
            string var_8363_mode_0 = const()[name = string("op_8363_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_8363_cast_fp16 = gelu(mode = var_8363_mode_0, x = linear_167_cast_fp16)[name = string("op_8363_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106941440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107990080))))[name = string("tm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_168_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_223_cast_fp16 = mul(x = var_8363_cast_fp16, y = linear_168_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107994240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109042880))))[name = string("tm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8371_cast_fp16 = abs(x = linear_169_cast_fp16)[name = string("op_8371_cast_fp16")];
            bool var_8374 = const()[name = string("op_8374"), val = bool(true)];
            tensor<int32, [1]> var_8375_axes_0 = const()[name = string("op_8375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8375_cast_fp16 = reduce_max(axes = var_8375_axes_0, keep_dims = var_8374, x = var_8371_cast_fp16)[name = string("op_8375_cast_fp16")];
            fp16 var_8376_to_fp16 = const()[name = string("op_8376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_303_cast_fp16 = maximum(x = var_8375_cast_fp16, y = var_8376_to_fp16)[name = string("s_303_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_303_cast_fp16 = real_div(x = linear_169_cast_fp16, y = s_303_cast_fp16)[name = string("y_303_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8379_cast_fp16 = mul(x = y_303_cast_fp16, y = y_303_cast_fp16)[name = string("op_8379_cast_fp16")];
            tensor<int32, [1]> var_8384_axes_0 = const()[name = string("op_8384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8384_keep_dims_0 = const()[name = string("op_8384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8384_cast_fp16 = reduce_mean(axes = var_8384_axes_0, keep_dims = var_8384_keep_dims_0, x = var_8379_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8385_cast_fp16 = mul(x = s_303_cast_fp16, y = s_303_cast_fp16)[name = string("op_8385_cast_fp16")];
            fp32 var_8386_epsilon_0 = const()[name = string("op_8386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8386_cast_fp16 = inverse(epsilon = var_8386_epsilon_0, x = var_8385_cast_fp16)[name = string("op_8386_cast_fp16")];
            fp16 var_8387_to_fp16 = const()[name = string("op_8387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8388_cast_fp16 = mul(x = var_8386_cast_fp16, y = var_8387_to_fp16)[name = string("op_8388_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8390_cast_fp16 = add(x = var_8384_cast_fp16, y = var_8388_cast_fp16)[name = string("op_8390_cast_fp16")];
            fp32 var_8391_epsilon_0 = const()[name = string("op_8391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8391_cast_fp16 = rsqrt(epsilon = var_8391_epsilon_0, x = var_8390_cast_fp16)[name = string("op_8391_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_265_cast_fp16 = mul(x = y_303_cast_fp16, y = var_8391_cast_fp16)[name = string("out_265_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109043968)))];
            tensor<fp16, [1, 48, 512]> var_8393_cast_fp16 = mul(x = out_265_cast_fp16, y = tm_layers_18_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8393_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_225_cast_fp16 = add(x = x_335_cast_fp16, y = var_8393_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109045056))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109307264))))[name = string("tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_170_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_170_cast_fp16")];
            string x_339_mode_0 = const()[name = string("x_339_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_339_cast_fp16 = gelu(mode = x_339_mode_0, x = linear_170_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [4]> var_8413_begin_0 = const()[name = string("op_8413_begin_0"), val = tensor<int32, [4]>([0, 0, 18, 0])];
            tensor<int32, [4]> var_8413_end_0 = const()[name = string("op_8413_end_0"), val = tensor<int32, [4]>([1, 48, 19, 512])];
            tensor<bool, [4]> var_8413_end_mask_0 = const()[name = string("op_8413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8413_squeeze_mask_0 = const()[name = string("op_8413_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_8413_cast_fp16 = slice_by_index(begin = var_8413_begin_0, end = var_8413_end_0, end_mask = var_8413_end_mask_0, squeeze_mask = var_8413_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_227_cast_fp16 = mul(x = x_339_cast_fp16, y = var_8413_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109308352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109570560))))[name = string("tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8423_cast_fp16 = abs(x = linear_171_cast_fp16)[name = string("op_8423_cast_fp16")];
            bool var_8426 = const()[name = string("op_8426"), val = bool(true)];
            tensor<int32, [1]> var_8427_axes_0 = const()[name = string("op_8427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8427_cast_fp16 = reduce_max(axes = var_8427_axes_0, keep_dims = var_8426, x = var_8423_cast_fp16)[name = string("op_8427_cast_fp16")];
            fp16 var_8428_to_fp16 = const()[name = string("op_8428_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_305_cast_fp16 = maximum(x = var_8427_cast_fp16, y = var_8428_to_fp16)[name = string("s_305_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_305_cast_fp16 = real_div(x = linear_171_cast_fp16, y = s_305_cast_fp16)[name = string("y_305_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8431_cast_fp16 = mul(x = y_305_cast_fp16, y = y_305_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<int32, [1]> var_8436_axes_0 = const()[name = string("op_8436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8436_keep_dims_0 = const()[name = string("op_8436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8436_cast_fp16 = reduce_mean(axes = var_8436_axes_0, keep_dims = var_8436_keep_dims_0, x = var_8431_cast_fp16)[name = string("op_8436_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8437_cast_fp16 = mul(x = s_305_cast_fp16, y = s_305_cast_fp16)[name = string("op_8437_cast_fp16")];
            fp32 var_8438_epsilon_0 = const()[name = string("op_8438_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8438_cast_fp16 = inverse(epsilon = var_8438_epsilon_0, x = var_8437_cast_fp16)[name = string("op_8438_cast_fp16")];
            fp16 var_8439_to_fp16 = const()[name = string("op_8439_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8440_cast_fp16 = mul(x = var_8438_cast_fp16, y = var_8439_to_fp16)[name = string("op_8440_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8442_cast_fp16 = add(x = var_8436_cast_fp16, y = var_8440_cast_fp16)[name = string("op_8442_cast_fp16")];
            fp32 var_8443_epsilon_0 = const()[name = string("op_8443_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8443_cast_fp16 = rsqrt(epsilon = var_8443_epsilon_0, x = var_8442_cast_fp16)[name = string("op_8443_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_267_cast_fp16 = mul(x = y_305_cast_fp16, y = var_8443_cast_fp16)[name = string("out_267_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109571648)))];
            tensor<fp16, [1, 48, 512]> var_8445_cast_fp16 = mul(x = out_267_cast_fp16, y = tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8445_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_37_cast_fp16 = add(x = input_225_cast_fp16, y = var_8445_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<fp16, [1]> tm_layers_18_layer_scalar_to_fp16 = const()[name = string("tm_layers_18_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.68p-1])];
            tensor<fp16, [1, 48, 512]> x_343_cast_fp16 = mul(x = h_37_cast_fp16, y = tm_layers_18_layer_scalar_to_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8449_cast_fp16 = abs(x = x_343_cast_fp16)[name = string("op_8449_cast_fp16")];
            bool var_8452 = const()[name = string("op_8452"), val = bool(true)];
            tensor<int32, [1]> var_8453_axes_0 = const()[name = string("op_8453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8453_cast_fp16 = reduce_max(axes = var_8453_axes_0, keep_dims = var_8452, x = var_8449_cast_fp16)[name = string("op_8453_cast_fp16")];
            fp16 var_8454_to_fp16 = const()[name = string("op_8454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_307_cast_fp16 = maximum(x = var_8453_cast_fp16, y = var_8454_to_fp16)[name = string("s_307_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_307_cast_fp16 = real_div(x = x_343_cast_fp16, y = s_307_cast_fp16)[name = string("y_307_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8457_cast_fp16 = mul(x = y_307_cast_fp16, y = y_307_cast_fp16)[name = string("op_8457_cast_fp16")];
            tensor<int32, [1]> var_8462_axes_0 = const()[name = string("op_8462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8462_keep_dims_0 = const()[name = string("op_8462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8462_cast_fp16 = reduce_mean(axes = var_8462_axes_0, keep_dims = var_8462_keep_dims_0, x = var_8457_cast_fp16)[name = string("op_8462_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8463_cast_fp16 = mul(x = s_307_cast_fp16, y = s_307_cast_fp16)[name = string("op_8463_cast_fp16")];
            fp32 var_8464_epsilon_0 = const()[name = string("op_8464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8464_cast_fp16 = inverse(epsilon = var_8464_epsilon_0, x = var_8463_cast_fp16)[name = string("op_8464_cast_fp16")];
            fp16 var_8465_to_fp16 = const()[name = string("op_8465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8466_cast_fp16 = mul(x = var_8464_cast_fp16, y = var_8465_to_fp16)[name = string("op_8466_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8468_cast_fp16 = add(x = var_8462_cast_fp16, y = var_8466_cast_fp16)[name = string("op_8468_cast_fp16")];
            fp32 var_8469_epsilon_0 = const()[name = string("op_8469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8469_cast_fp16 = rsqrt(epsilon = var_8469_epsilon_0, x = var_8468_cast_fp16)[name = string("op_8469_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_269_cast_fp16 = mul(x = y_307_cast_fp16, y = var_8469_cast_fp16)[name = string("out_269_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109572736)))];
            tensor<fp16, [1, 48, 512]> input_229_cast_fp16 = mul(x = out_269_cast_fp16, y = tm_layers_19_input_layernorm_weight_to_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109573824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110098176))))[name = string("tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [4]> var_8479 = const()[name = string("op_8479"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_8480_cast_fp16 = reshape(shape = var_8479, x = linear_172_cast_fp16)[name = string("op_8480_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110100288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110362496))))[name = string("tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<int32, [4]> var_8491 = const()[name = string("op_8491"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_8492_cast_fp16 = reshape(shape = var_8491, x = linear_173_cast_fp16)[name = string("op_8492_cast_fp16")];
            tensor<int32, [4]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110363584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110625792))))[name = string("tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<int32, [4]> var_8503 = const()[name = string("op_8503"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_8504_cast_fp16 = reshape(shape = var_8503, x = linear_174_cast_fp16)[name = string("op_8504_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_8480_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 4, 48, 256]> var_8508_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_8508_cast_fp16")];
            bool var_8511 = const()[name = string("op_8511"), val = bool(true)];
            tensor<int32, [1]> var_8512_axes_0 = const()[name = string("op_8512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_8512_cast_fp16 = reduce_max(axes = var_8512_axes_0, keep_dims = var_8511, x = var_8508_cast_fp16)[name = string("op_8512_cast_fp16")];
            fp16 var_8513_to_fp16 = const()[name = string("op_8513_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_309_cast_fp16 = maximum(x = var_8512_cast_fp16, y = var_8513_to_fp16)[name = string("s_309_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_309_cast_fp16 = real_div(x = x_345_cast_fp16, y = s_309_cast_fp16)[name = string("y_309_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8516_cast_fp16 = mul(x = y_309_cast_fp16, y = y_309_cast_fp16)[name = string("op_8516_cast_fp16")];
            tensor<int32, [1]> var_8521_axes_0 = const()[name = string("op_8521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8521_keep_dims_0 = const()[name = string("op_8521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_8521_cast_fp16 = reduce_mean(axes = var_8521_axes_0, keep_dims = var_8521_keep_dims_0, x = var_8516_cast_fp16)[name = string("op_8521_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_8522_cast_fp16 = mul(x = s_309_cast_fp16, y = s_309_cast_fp16)[name = string("op_8522_cast_fp16")];
            fp32 var_8523_epsilon_0 = const()[name = string("op_8523_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_8523_cast_fp16 = inverse(epsilon = var_8523_epsilon_0, x = var_8522_cast_fp16)[name = string("op_8523_cast_fp16")];
            fp16 var_8524_to_fp16 = const()[name = string("op_8524_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_8525_cast_fp16 = mul(x = var_8523_cast_fp16, y = var_8524_to_fp16)[name = string("op_8525_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_8527_cast_fp16 = add(x = var_8521_cast_fp16, y = var_8525_cast_fp16)[name = string("op_8527_cast_fp16")];
            fp32 var_8528_epsilon_0 = const()[name = string("op_8528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_8528_cast_fp16 = rsqrt(epsilon = var_8528_epsilon_0, x = var_8527_cast_fp16)[name = string("op_8528_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_271_cast_fp16 = mul(x = y_309_cast_fp16, y = var_8528_cast_fp16)[name = string("out_271_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_77_cast_fp16 = mul(x = out_271_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_77_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = var_8492_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 48, 256]> var_8531_cast_fp16 = abs(x = x_347_cast_fp16)[name = string("op_8531_cast_fp16")];
            bool var_8534 = const()[name = string("op_8534"), val = bool(true)];
            tensor<int32, [1]> var_8535_axes_0 = const()[name = string("op_8535_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_8535_cast_fp16 = reduce_max(axes = var_8535_axes_0, keep_dims = var_8534, x = var_8531_cast_fp16)[name = string("op_8535_cast_fp16")];
            fp16 var_8536_to_fp16 = const()[name = string("op_8536_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_311_cast_fp16 = maximum(x = var_8535_cast_fp16, y = var_8536_to_fp16)[name = string("s_311_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_311_cast_fp16 = real_div(x = x_347_cast_fp16, y = s_311_cast_fp16)[name = string("y_311_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8539_cast_fp16 = mul(x = y_311_cast_fp16, y = y_311_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<int32, [1]> var_8544_axes_0 = const()[name = string("op_8544_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8544_keep_dims_0 = const()[name = string("op_8544_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_8544_cast_fp16 = reduce_mean(axes = var_8544_axes_0, keep_dims = var_8544_keep_dims_0, x = var_8539_cast_fp16)[name = string("op_8544_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8545_cast_fp16 = mul(x = s_311_cast_fp16, y = s_311_cast_fp16)[name = string("op_8545_cast_fp16")];
            fp32 var_8546_epsilon_0 = const()[name = string("op_8546_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_8546_cast_fp16 = inverse(epsilon = var_8546_epsilon_0, x = var_8545_cast_fp16)[name = string("op_8546_cast_fp16")];
            fp16 var_8547_to_fp16 = const()[name = string("op_8547_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_8548_cast_fp16 = mul(x = var_8546_cast_fp16, y = var_8547_to_fp16)[name = string("op_8548_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8550_cast_fp16 = add(x = var_8544_cast_fp16, y = var_8548_cast_fp16)[name = string("op_8550_cast_fp16")];
            fp32 var_8551_epsilon_0 = const()[name = string("op_8551_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_8551_cast_fp16 = rsqrt(epsilon = var_8551_epsilon_0, x = var_8550_cast_fp16)[name = string("op_8551_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_273_cast_fp16 = mul(x = y_311_cast_fp16, y = var_8551_cast_fp16)[name = string("out_273_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_115_cast_fp16 = mul(x = out_273_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_115_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_8504_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 2, 48, 256]> var_8554_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_8554_cast_fp16")];
            bool var_8557 = const()[name = string("op_8557"), val = bool(true)];
            tensor<int32, [1]> var_8558_axes_0 = const()[name = string("op_8558_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_8558_cast_fp16 = reduce_max(axes = var_8558_axes_0, keep_dims = var_8557, x = var_8554_cast_fp16)[name = string("op_8558_cast_fp16")];
            fp16 var_8559_to_fp16 = const()[name = string("op_8559_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_313_cast_fp16 = maximum(x = var_8558_cast_fp16, y = var_8559_to_fp16)[name = string("s_313_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_313_cast_fp16 = real_div(x = x_349_cast_fp16, y = s_313_cast_fp16)[name = string("y_313_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8562_cast_fp16 = mul(x = y_313_cast_fp16, y = y_313_cast_fp16)[name = string("op_8562_cast_fp16")];
            tensor<int32, [1]> var_8567_axes_0 = const()[name = string("op_8567_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8567_keep_dims_0 = const()[name = string("op_8567_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_8567_cast_fp16 = reduce_mean(axes = var_8567_axes_0, keep_dims = var_8567_keep_dims_0, x = var_8562_cast_fp16)[name = string("op_8567_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8568_cast_fp16 = mul(x = s_313_cast_fp16, y = s_313_cast_fp16)[name = string("op_8568_cast_fp16")];
            fp32 var_8569_epsilon_0 = const()[name = string("op_8569_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_8569_cast_fp16 = inverse(epsilon = var_8569_epsilon_0, x = var_8568_cast_fp16)[name = string("op_8569_cast_fp16")];
            fp16 var_8570_to_fp16 = const()[name = string("op_8570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_8571_cast_fp16 = mul(x = var_8569_cast_fp16, y = var_8570_to_fp16)[name = string("op_8571_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8573_cast_fp16 = add(x = var_8567_cast_fp16, y = var_8571_cast_fp16)[name = string("op_8573_cast_fp16")];
            fp32 var_8574_epsilon_0 = const()[name = string("op_8574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_8574_cast_fp16 = rsqrt(epsilon = var_8574_epsilon_0, x = var_8573_cast_fp16)[name = string("op_8574_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_77_cast_fp16 = mul(x = y_313_cast_fp16, y = var_8574_cast_fp16)[name = string("v_77_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8576_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8576_cast_fp16")];
            tensor<int32, [4]> var_8581_begin_0 = const()[name = string("op_8581_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8581_end_0 = const()[name = string("op_8581_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_8581_end_mask_0 = const()[name = string("op_8581_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_8581_cast_fp16 = slice_by_index(begin = var_8581_begin_0, end = var_8581_end_0, end_mask = var_8581_end_mask_0, x = q_77_cast_fp16)[name = string("op_8581_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_8582_cast_fp16 = mul(x = var_8581_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_8582_cast_fp16")];
            tensor<int32, [4]> var_8587_begin_0 = const()[name = string("op_8587_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8587_end_0 = const()[name = string("op_8587_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_8587_end_mask_0 = const()[name = string("op_8587_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_8587_cast_fp16 = slice_by_index(begin = var_8587_begin_0, end = var_8587_end_0, end_mask = var_8587_end_mask_0, x = q_77_cast_fp16)[name = string("op_8587_cast_fp16")];
            int32 var_8589 = const()[name = string("op_8589"), val = int32(-1)];
            bool var_8590_interleave_0 = const()[name = string("op_8590_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_8590_cast_fp16 = concat(axis = var_8589, interleave = var_8590_interleave_0, values = (var_8582_cast_fp16, var_8587_cast_fp16))[name = string("op_8590_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8591_cast_fp16 = mul(x = var_8590_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8591_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_79_cast_fp16 = add(x = var_8576_cast_fp16, y = var_8591_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8594_cast_fp16 = mul(x = k_115_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8594_cast_fp16")];
            tensor<int32, [4]> var_8599_begin_0 = const()[name = string("op_8599_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8599_end_0 = const()[name = string("op_8599_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_8599_end_mask_0 = const()[name = string("op_8599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_8599_cast_fp16 = slice_by_index(begin = var_8599_begin_0, end = var_8599_end_0, end_mask = var_8599_end_mask_0, x = k_115_cast_fp16)[name = string("op_8599_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_8600_cast_fp16 = mul(x = var_8599_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_8600_cast_fp16")];
            tensor<int32, [4]> var_8605_begin_0 = const()[name = string("op_8605_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8605_end_0 = const()[name = string("op_8605_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_8605_end_mask_0 = const()[name = string("op_8605_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_8605_cast_fp16 = slice_by_index(begin = var_8605_begin_0, end = var_8605_end_0, end_mask = var_8605_end_mask_0, x = k_115_cast_fp16)[name = string("op_8605_cast_fp16")];
            int32 var_8607 = const()[name = string("op_8607"), val = int32(-1)];
            bool var_8608_interleave_0 = const()[name = string("op_8608_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_8608_cast_fp16 = concat(axis = var_8607, interleave = var_8608_interleave_0, values = (var_8600_cast_fp16, var_8605_cast_fp16))[name = string("op_8608_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8609_cast_fp16 = mul(x = var_8608_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8609_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_117_cast_fp16 = add(x = var_8594_cast_fp16, y = var_8609_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_38_reps_0 = const()[name = string("tile_38_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = k_117_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 1, 48, 256]> tile_38_cast_fp16 = tile(reps = tile_38_reps_0, x = transpose_76_cast_fp16)[name = string("tile_38_cast_fp16")];
            tensor<int32, [5]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_76_cast_fp16 = reshape(shape = concat_76, x = tile_38_cast_fp16)[name = string("reshape_76_cast_fp16")];
            tensor<int32, [5]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_77 = const()[name = string("concat_77"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_77_cast_fp16 = transpose(perm = transpose_77_perm_0, x = reshape_76_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 1, 48, 256]> reshape_77_cast_fp16 = reshape(shape = concat_77, x = transpose_77_cast_fp16)[name = string("reshape_77_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_39_reps_0 = const()[name = string("tile_39_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_78_cast_fp16 = transpose(perm = transpose_78_perm_0, x = v_77_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 1, 48, 256]> tile_39_cast_fp16 = tile(reps = tile_39_reps_0, x = transpose_78_cast_fp16)[name = string("tile_39_cast_fp16")];
            tensor<int32, [5]> concat_78 = const()[name = string("concat_78"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_78_cast_fp16 = reshape(shape = concat_78, x = tile_39_cast_fp16)[name = string("reshape_78_cast_fp16")];
            tensor<int32, [5]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = reshape_78_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 1, 48, 256]> reshape_79_cast_fp16 = reshape(shape = concat_79, x = transpose_79_cast_fp16)[name = string("reshape_79_cast_fp16")];
            tensor<int32, [4]> v_79_perm_0 = const()[name = string("v_79_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8623_transpose_x_0 = const()[name = string("op_8623_transpose_x_0"), val = bool(false)];
            bool var_8623_transpose_y_0 = const()[name = string("op_8623_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = reshape_77_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4, 48, 48]> var_8623_cast_fp16 = matmul(transpose_x = var_8623_transpose_x_0, transpose_y = var_8623_transpose_y_0, x = q_79_cast_fp16, y = transpose_115_cast_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8625_cast_fp16")];
            int32 var_8626 = const()[name = string("op_8626"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_313_cast_fp16 = softmax(axis = var_8626, x = var_8625_cast_fp16)[name = string("w_313_cast_fp16")];
            bool var_8629_transpose_x_0 = const()[name = string("op_8629_transpose_x_0"), val = bool(false)];
            bool var_8629_transpose_y_0 = const()[name = string("op_8629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = reshape_79_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 4, 48, 256]> var_8629_cast_fp16 = matmul(transpose_x = var_8629_transpose_x_0, transpose_y = var_8629_transpose_y_0, x = w_313_cast_fp16, y = v_79_cast_fp16)[name = string("op_8629_cast_fp16")];
            tensor<int32, [4]> var_8632_perm_0 = const()[name = string("op_8632_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_8632_cast_fp16 = transpose(perm = var_8632_perm_0, x = var_8629_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 48, 1024]> input_231_cast_fp16 = reshape(shape = var_8636, x = var_8632_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110626880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111151232))))[name = string("tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8641_cast_fp16 = abs(x = linear_175_cast_fp16)[name = string("op_8641_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<int32, [1]> var_8645_axes_0 = const()[name = string("op_8645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8645_cast_fp16 = reduce_max(axes = var_8645_axes_0, keep_dims = var_8644, x = var_8641_cast_fp16)[name = string("op_8645_cast_fp16")];
            fp16 var_8646_to_fp16 = const()[name = string("op_8646_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_315_cast_fp16 = maximum(x = var_8645_cast_fp16, y = var_8646_to_fp16)[name = string("s_315_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_315_cast_fp16 = real_div(x = linear_175_cast_fp16, y = s_315_cast_fp16)[name = string("y_315_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8649_cast_fp16 = mul(x = y_315_cast_fp16, y = y_315_cast_fp16)[name = string("op_8649_cast_fp16")];
            tensor<int32, [1]> var_8654_axes_0 = const()[name = string("op_8654_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8654_keep_dims_0 = const()[name = string("op_8654_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8654_cast_fp16 = reduce_mean(axes = var_8654_axes_0, keep_dims = var_8654_keep_dims_0, x = var_8649_cast_fp16)[name = string("op_8654_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8655_cast_fp16 = mul(x = s_315_cast_fp16, y = s_315_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp32 var_8656_epsilon_0 = const()[name = string("op_8656_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8656_cast_fp16 = inverse(epsilon = var_8656_epsilon_0, x = var_8655_cast_fp16)[name = string("op_8656_cast_fp16")];
            fp16 var_8657_to_fp16 = const()[name = string("op_8657_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8658_cast_fp16 = mul(x = var_8656_cast_fp16, y = var_8657_to_fp16)[name = string("op_8658_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8660_cast_fp16 = add(x = var_8654_cast_fp16, y = var_8658_cast_fp16)[name = string("op_8660_cast_fp16")];
            fp32 var_8661_epsilon_0 = const()[name = string("op_8661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8661_cast_fp16 = rsqrt(epsilon = var_8661_epsilon_0, x = var_8660_cast_fp16)[name = string("op_8661_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_275_cast_fp16 = mul(x = y_315_cast_fp16, y = var_8661_cast_fp16)[name = string("out_275_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111152320)))];
            tensor<fp16, [1, 48, 512]> var_8663_cast_fp16 = mul(x = out_275_cast_fp16, y = tm_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("op_8663_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_353_cast_fp16 = add(x = x_343_cast_fp16, y = var_8663_cast_fp16)[name = string("x_353_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8666_cast_fp16 = abs(x = x_353_cast_fp16)[name = string("op_8666_cast_fp16")];
            bool var_8669 = const()[name = string("op_8669"), val = bool(true)];
            tensor<int32, [1]> var_8670_axes_0 = const()[name = string("op_8670_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8670_cast_fp16 = reduce_max(axes = var_8670_axes_0, keep_dims = var_8669, x = var_8666_cast_fp16)[name = string("op_8670_cast_fp16")];
            fp16 var_8671_to_fp16 = const()[name = string("op_8671_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_317_cast_fp16 = maximum(x = var_8670_cast_fp16, y = var_8671_to_fp16)[name = string("s_317_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_317_cast_fp16 = real_div(x = x_353_cast_fp16, y = s_317_cast_fp16)[name = string("y_317_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8674_cast_fp16 = mul(x = y_317_cast_fp16, y = y_317_cast_fp16)[name = string("op_8674_cast_fp16")];
            tensor<int32, [1]> var_8679_axes_0 = const()[name = string("op_8679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8679_keep_dims_0 = const()[name = string("op_8679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8679_cast_fp16 = reduce_mean(axes = var_8679_axes_0, keep_dims = var_8679_keep_dims_0, x = var_8674_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8680_cast_fp16 = mul(x = s_317_cast_fp16, y = s_317_cast_fp16)[name = string("op_8680_cast_fp16")];
            fp32 var_8681_epsilon_0 = const()[name = string("op_8681_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8681_cast_fp16 = inverse(epsilon = var_8681_epsilon_0, x = var_8680_cast_fp16)[name = string("op_8681_cast_fp16")];
            fp16 var_8682_to_fp16 = const()[name = string("op_8682_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8683_cast_fp16 = mul(x = var_8681_cast_fp16, y = var_8682_to_fp16)[name = string("op_8683_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8685_cast_fp16 = add(x = var_8679_cast_fp16, y = var_8683_cast_fp16)[name = string("op_8685_cast_fp16")];
            fp32 var_8686_epsilon_0 = const()[name = string("op_8686_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8686_cast_fp16 = rsqrt(epsilon = var_8686_epsilon_0, x = var_8685_cast_fp16)[name = string("op_8686_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_277_cast_fp16 = mul(x = y_317_cast_fp16, y = var_8686_cast_fp16)[name = string("out_277_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111153408)))];
            tensor<fp16, [1, 48, 512]> input_233_cast_fp16 = mul(x = out_277_cast_fp16, y = tm_layers_19_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111154496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112203136))))[name = string("tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_176_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_176_cast_fp16")];
            string var_8693_mode_0 = const()[name = string("op_8693_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_8693_cast_fp16 = gelu(mode = var_8693_mode_0, x = linear_176_cast_fp16)[name = string("op_8693_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112207296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113255936))))[name = string("tm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_177_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_235_cast_fp16 = mul(x = var_8693_cast_fp16, y = linear_177_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113260096))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114308736))))[name = string("tm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8701_cast_fp16 = abs(x = linear_178_cast_fp16)[name = string("op_8701_cast_fp16")];
            bool var_8704 = const()[name = string("op_8704"), val = bool(true)];
            tensor<int32, [1]> var_8705_axes_0 = const()[name = string("op_8705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8705_cast_fp16 = reduce_max(axes = var_8705_axes_0, keep_dims = var_8704, x = var_8701_cast_fp16)[name = string("op_8705_cast_fp16")];
            fp16 var_8706_to_fp16 = const()[name = string("op_8706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_319_cast_fp16 = maximum(x = var_8705_cast_fp16, y = var_8706_to_fp16)[name = string("s_319_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_319_cast_fp16 = real_div(x = linear_178_cast_fp16, y = s_319_cast_fp16)[name = string("y_319_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8709_cast_fp16 = mul(x = y_319_cast_fp16, y = y_319_cast_fp16)[name = string("op_8709_cast_fp16")];
            tensor<int32, [1]> var_8714_axes_0 = const()[name = string("op_8714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8714_keep_dims_0 = const()[name = string("op_8714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8714_cast_fp16 = reduce_mean(axes = var_8714_axes_0, keep_dims = var_8714_keep_dims_0, x = var_8709_cast_fp16)[name = string("op_8714_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8715_cast_fp16 = mul(x = s_319_cast_fp16, y = s_319_cast_fp16)[name = string("op_8715_cast_fp16")];
            fp32 var_8716_epsilon_0 = const()[name = string("op_8716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8716_cast_fp16 = inverse(epsilon = var_8716_epsilon_0, x = var_8715_cast_fp16)[name = string("op_8716_cast_fp16")];
            fp16 var_8717_to_fp16 = const()[name = string("op_8717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8718_cast_fp16 = mul(x = var_8716_cast_fp16, y = var_8717_to_fp16)[name = string("op_8718_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8720_cast_fp16 = add(x = var_8714_cast_fp16, y = var_8718_cast_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8721_cast_fp16 = rsqrt(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_279_cast_fp16 = mul(x = y_319_cast_fp16, y = var_8721_cast_fp16)[name = string("out_279_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114309824)))];
            tensor<fp16, [1, 48, 512]> var_8723_cast_fp16 = mul(x = out_279_cast_fp16, y = tm_layers_19_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_237_cast_fp16 = add(x = x_353_cast_fp16, y = var_8723_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114310912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114573120))))[name = string("tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_179_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_237_cast_fp16)[name = string("linear_179_cast_fp16")];
            string x_357_mode_0 = const()[name = string("x_357_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_357_cast_fp16 = gelu(mode = x_357_mode_0, x = linear_179_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [4]> var_8743_begin_0 = const()[name = string("op_8743_begin_0"), val = tensor<int32, [4]>([0, 0, 19, 0])];
            tensor<int32, [4]> var_8743_end_0 = const()[name = string("op_8743_end_0"), val = tensor<int32, [4]>([1, 48, 20, 512])];
            tensor<bool, [4]> var_8743_end_mask_0 = const()[name = string("op_8743_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8743_squeeze_mask_0 = const()[name = string("op_8743_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_8743_cast_fp16 = slice_by_index(begin = var_8743_begin_0, end = var_8743_end_0, end_mask = var_8743_end_mask_0, squeeze_mask = var_8743_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8743_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_239_cast_fp16 = mul(x = x_357_cast_fp16, y = var_8743_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114574208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114836416))))[name = string("tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8753_cast_fp16 = abs(x = linear_180_cast_fp16)[name = string("op_8753_cast_fp16")];
            bool var_8756 = const()[name = string("op_8756"), val = bool(true)];
            tensor<int32, [1]> var_8757_axes_0 = const()[name = string("op_8757_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8757_cast_fp16 = reduce_max(axes = var_8757_axes_0, keep_dims = var_8756, x = var_8753_cast_fp16)[name = string("op_8757_cast_fp16")];
            fp16 var_8758_to_fp16 = const()[name = string("op_8758_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_321_cast_fp16 = maximum(x = var_8757_cast_fp16, y = var_8758_to_fp16)[name = string("s_321_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_321_cast_fp16 = real_div(x = linear_180_cast_fp16, y = s_321_cast_fp16)[name = string("y_321_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8761_cast_fp16 = mul(x = y_321_cast_fp16, y = y_321_cast_fp16)[name = string("op_8761_cast_fp16")];
            tensor<int32, [1]> var_8766_axes_0 = const()[name = string("op_8766_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8766_keep_dims_0 = const()[name = string("op_8766_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8766_cast_fp16 = reduce_mean(axes = var_8766_axes_0, keep_dims = var_8766_keep_dims_0, x = var_8761_cast_fp16)[name = string("op_8766_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8767_cast_fp16 = mul(x = s_321_cast_fp16, y = s_321_cast_fp16)[name = string("op_8767_cast_fp16")];
            fp32 var_8768_epsilon_0 = const()[name = string("op_8768_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8768_cast_fp16 = inverse(epsilon = var_8768_epsilon_0, x = var_8767_cast_fp16)[name = string("op_8768_cast_fp16")];
            fp16 var_8769_to_fp16 = const()[name = string("op_8769_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8770_cast_fp16 = mul(x = var_8768_cast_fp16, y = var_8769_to_fp16)[name = string("op_8770_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8772_cast_fp16 = add(x = var_8766_cast_fp16, y = var_8770_cast_fp16)[name = string("op_8772_cast_fp16")];
            fp32 var_8773_epsilon_0 = const()[name = string("op_8773_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8773_cast_fp16 = rsqrt(epsilon = var_8773_epsilon_0, x = var_8772_cast_fp16)[name = string("op_8773_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_281_cast_fp16 = mul(x = y_321_cast_fp16, y = var_8773_cast_fp16)[name = string("out_281_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114837504)))];
            tensor<fp16, [1, 48, 512]> var_8775_cast_fp16 = mul(x = out_281_cast_fp16, y = tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_39_cast_fp16 = add(x = input_237_cast_fp16, y = var_8775_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<fp16, [1]> tm_layers_19_layer_scalar_to_fp16 = const()[name = string("tm_layers_19_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fp-1])];
            tensor<fp16, [1, 48, 512]> x_361_cast_fp16 = mul(x = h_39_cast_fp16, y = tm_layers_19_layer_scalar_to_fp16)[name = string("x_361_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8779_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_8779_cast_fp16")];
            bool var_8782 = const()[name = string("op_8782"), val = bool(true)];
            tensor<int32, [1]> var_8783_axes_0 = const()[name = string("op_8783_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8783_cast_fp16 = reduce_max(axes = var_8783_axes_0, keep_dims = var_8782, x = var_8779_cast_fp16)[name = string("op_8783_cast_fp16")];
            fp16 var_8784_to_fp16 = const()[name = string("op_8784_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_323_cast_fp16 = maximum(x = var_8783_cast_fp16, y = var_8784_to_fp16)[name = string("s_323_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_323_cast_fp16 = real_div(x = x_361_cast_fp16, y = s_323_cast_fp16)[name = string("y_323_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8787_cast_fp16 = mul(x = y_323_cast_fp16, y = y_323_cast_fp16)[name = string("op_8787_cast_fp16")];
            tensor<int32, [1]> var_8792_axes_0 = const()[name = string("op_8792_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8792_keep_dims_0 = const()[name = string("op_8792_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8792_cast_fp16 = reduce_mean(axes = var_8792_axes_0, keep_dims = var_8792_keep_dims_0, x = var_8787_cast_fp16)[name = string("op_8792_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8793_cast_fp16 = mul(x = s_323_cast_fp16, y = s_323_cast_fp16)[name = string("op_8793_cast_fp16")];
            fp32 var_8794_epsilon_0 = const()[name = string("op_8794_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8794_cast_fp16 = inverse(epsilon = var_8794_epsilon_0, x = var_8793_cast_fp16)[name = string("op_8794_cast_fp16")];
            fp16 var_8795_to_fp16 = const()[name = string("op_8795_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8796_cast_fp16 = mul(x = var_8794_cast_fp16, y = var_8795_to_fp16)[name = string("op_8796_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8798_cast_fp16 = add(x = var_8792_cast_fp16, y = var_8796_cast_fp16)[name = string("op_8798_cast_fp16")];
            fp32 var_8799_epsilon_0 = const()[name = string("op_8799_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8799_cast_fp16 = rsqrt(epsilon = var_8799_epsilon_0, x = var_8798_cast_fp16)[name = string("op_8799_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_283_cast_fp16 = mul(x = y_323_cast_fp16, y = var_8799_cast_fp16)[name = string("out_283_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114838592)))];
            tensor<fp16, [1, 48, 512]> input_241_cast_fp16 = mul(x = out_283_cast_fp16, y = tm_layers_20_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114839680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115364032))))[name = string("tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<int32, [4]> var_8809 = const()[name = string("op_8809"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_8810_cast_fp16 = reshape(shape = var_8809, x = linear_181_cast_fp16)[name = string("op_8810_cast_fp16")];
            tensor<int32, [4]> x_363_perm_0 = const()[name = string("x_363_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115366144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115628352))))[name = string("tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_8821 = const()[name = string("op_8821"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_8822_cast_fp16 = reshape(shape = var_8821, x = linear_182_cast_fp16)[name = string("op_8822_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115629440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115891648))))[name = string("tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_8833 = const()[name = string("op_8833"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_8834_cast_fp16 = reshape(shape = var_8833, x = linear_183_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<int32, [4]> x_367_perm_0 = const()[name = string("x_367_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_363_cast_fp16 = transpose(perm = x_363_perm_0, x = var_8810_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 4, 48, 256]> var_8838_cast_fp16 = abs(x = x_363_cast_fp16)[name = string("op_8838_cast_fp16")];
            bool var_8841 = const()[name = string("op_8841"), val = bool(true)];
            tensor<int32, [1]> var_8842_axes_0 = const()[name = string("op_8842_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_8842_cast_fp16 = reduce_max(axes = var_8842_axes_0, keep_dims = var_8841, x = var_8838_cast_fp16)[name = string("op_8842_cast_fp16")];
            fp16 var_8843_to_fp16 = const()[name = string("op_8843_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_325_cast_fp16 = maximum(x = var_8842_cast_fp16, y = var_8843_to_fp16)[name = string("s_325_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_325_cast_fp16 = real_div(x = x_363_cast_fp16, y = s_325_cast_fp16)[name = string("y_325_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8846_cast_fp16 = mul(x = y_325_cast_fp16, y = y_325_cast_fp16)[name = string("op_8846_cast_fp16")];
            tensor<int32, [1]> var_8851_axes_0 = const()[name = string("op_8851_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8851_keep_dims_0 = const()[name = string("op_8851_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_8851_cast_fp16 = reduce_mean(axes = var_8851_axes_0, keep_dims = var_8851_keep_dims_0, x = var_8846_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_8852_cast_fp16 = mul(x = s_325_cast_fp16, y = s_325_cast_fp16)[name = string("op_8852_cast_fp16")];
            fp32 var_8853_epsilon_0 = const()[name = string("op_8853_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_8853_cast_fp16 = inverse(epsilon = var_8853_epsilon_0, x = var_8852_cast_fp16)[name = string("op_8853_cast_fp16")];
            fp16 var_8854_to_fp16 = const()[name = string("op_8854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_8855_cast_fp16 = mul(x = var_8853_cast_fp16, y = var_8854_to_fp16)[name = string("op_8855_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_8857_cast_fp16 = add(x = var_8851_cast_fp16, y = var_8855_cast_fp16)[name = string("op_8857_cast_fp16")];
            fp32 var_8858_epsilon_0 = const()[name = string("op_8858_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_8858_cast_fp16 = rsqrt(epsilon = var_8858_epsilon_0, x = var_8857_cast_fp16)[name = string("op_8858_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_285_cast_fp16 = mul(x = y_325_cast_fp16, y = var_8858_cast_fp16)[name = string("out_285_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_81_cast_fp16 = mul(x = out_285_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_81_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_8822_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 48, 256]> var_8861_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_8861_cast_fp16")];
            bool var_8864 = const()[name = string("op_8864"), val = bool(true)];
            tensor<int32, [1]> var_8865_axes_0 = const()[name = string("op_8865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_8865_cast_fp16 = reduce_max(axes = var_8865_axes_0, keep_dims = var_8864, x = var_8861_cast_fp16)[name = string("op_8865_cast_fp16")];
            fp16 var_8866_to_fp16 = const()[name = string("op_8866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_327_cast_fp16 = maximum(x = var_8865_cast_fp16, y = var_8866_to_fp16)[name = string("s_327_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_327_cast_fp16 = real_div(x = x_365_cast_fp16, y = s_327_cast_fp16)[name = string("y_327_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8869_cast_fp16 = mul(x = y_327_cast_fp16, y = y_327_cast_fp16)[name = string("op_8869_cast_fp16")];
            tensor<int32, [1]> var_8874_axes_0 = const()[name = string("op_8874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8874_keep_dims_0 = const()[name = string("op_8874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_8874_cast_fp16 = reduce_mean(axes = var_8874_axes_0, keep_dims = var_8874_keep_dims_0, x = var_8869_cast_fp16)[name = string("op_8874_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8875_cast_fp16 = mul(x = s_327_cast_fp16, y = s_327_cast_fp16)[name = string("op_8875_cast_fp16")];
            fp32 var_8876_epsilon_0 = const()[name = string("op_8876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_8876_cast_fp16 = inverse(epsilon = var_8876_epsilon_0, x = var_8875_cast_fp16)[name = string("op_8876_cast_fp16")];
            fp16 var_8877_to_fp16 = const()[name = string("op_8877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_8878_cast_fp16 = mul(x = var_8876_cast_fp16, y = var_8877_to_fp16)[name = string("op_8878_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8880_cast_fp16 = add(x = var_8874_cast_fp16, y = var_8878_cast_fp16)[name = string("op_8880_cast_fp16")];
            fp32 var_8881_epsilon_0 = const()[name = string("op_8881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_8881_cast_fp16 = rsqrt(epsilon = var_8881_epsilon_0, x = var_8880_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_287_cast_fp16 = mul(x = y_327_cast_fp16, y = var_8881_cast_fp16)[name = string("out_287_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_121_cast_fp16 = mul(x = out_287_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_121_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_367_cast_fp16 = transpose(perm = x_367_perm_0, x = var_8834_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 48, 256]> var_8884_cast_fp16 = abs(x = x_367_cast_fp16)[name = string("op_8884_cast_fp16")];
            bool var_8887 = const()[name = string("op_8887"), val = bool(true)];
            tensor<int32, [1]> var_8888_axes_0 = const()[name = string("op_8888_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_8888_cast_fp16 = reduce_max(axes = var_8888_axes_0, keep_dims = var_8887, x = var_8884_cast_fp16)[name = string("op_8888_cast_fp16")];
            fp16 var_8889_to_fp16 = const()[name = string("op_8889_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_329_cast_fp16 = maximum(x = var_8888_cast_fp16, y = var_8889_to_fp16)[name = string("s_329_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_329_cast_fp16 = real_div(x = x_367_cast_fp16, y = s_329_cast_fp16)[name = string("y_329_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8892_cast_fp16 = mul(x = y_329_cast_fp16, y = y_329_cast_fp16)[name = string("op_8892_cast_fp16")];
            tensor<int32, [1]> var_8897_axes_0 = const()[name = string("op_8897_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8897_keep_dims_0 = const()[name = string("op_8897_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_8897_cast_fp16 = reduce_mean(axes = var_8897_axes_0, keep_dims = var_8897_keep_dims_0, x = var_8892_cast_fp16)[name = string("op_8897_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8898_cast_fp16 = mul(x = s_329_cast_fp16, y = s_329_cast_fp16)[name = string("op_8898_cast_fp16")];
            fp32 var_8899_epsilon_0 = const()[name = string("op_8899_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_8899_cast_fp16 = inverse(epsilon = var_8899_epsilon_0, x = var_8898_cast_fp16)[name = string("op_8899_cast_fp16")];
            fp16 var_8900_to_fp16 = const()[name = string("op_8900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_8901_cast_fp16 = mul(x = var_8899_cast_fp16, y = var_8900_to_fp16)[name = string("op_8901_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_8903_cast_fp16 = add(x = var_8897_cast_fp16, y = var_8901_cast_fp16)[name = string("op_8903_cast_fp16")];
            fp32 var_8904_epsilon_0 = const()[name = string("op_8904_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_8904_cast_fp16 = rsqrt(epsilon = var_8904_epsilon_0, x = var_8903_cast_fp16)[name = string("op_8904_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_81_cast_fp16 = mul(x = y_329_cast_fp16, y = var_8904_cast_fp16)[name = string("v_81_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8906_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8906_cast_fp16")];
            tensor<int32, [4]> var_8911_begin_0 = const()[name = string("op_8911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8911_end_0 = const()[name = string("op_8911_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_8911_end_mask_0 = const()[name = string("op_8911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_8911_cast_fp16 = slice_by_index(begin = var_8911_begin_0, end = var_8911_end_0, end_mask = var_8911_end_mask_0, x = q_81_cast_fp16)[name = string("op_8911_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_8912_cast_fp16 = mul(x = var_8911_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_8912_cast_fp16")];
            tensor<int32, [4]> var_8917_begin_0 = const()[name = string("op_8917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8917_end_0 = const()[name = string("op_8917_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_8917_end_mask_0 = const()[name = string("op_8917_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_8917_cast_fp16 = slice_by_index(begin = var_8917_begin_0, end = var_8917_end_0, end_mask = var_8917_end_mask_0, x = q_81_cast_fp16)[name = string("op_8917_cast_fp16")];
            int32 var_8919 = const()[name = string("op_8919"), val = int32(-1)];
            bool var_8920_interleave_0 = const()[name = string("op_8920_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_8920_cast_fp16 = concat(axis = var_8919, interleave = var_8920_interleave_0, values = (var_8912_cast_fp16, var_8917_cast_fp16))[name = string("op_8920_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_8921_cast_fp16 = mul(x = var_8920_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_83_cast_fp16 = add(x = var_8906_cast_fp16, y = var_8921_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8924_cast_fp16 = mul(x = k_121_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8924_cast_fp16")];
            tensor<int32, [4]> var_8929_begin_0 = const()[name = string("op_8929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8929_end_0 = const()[name = string("op_8929_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_8929_end_mask_0 = const()[name = string("op_8929_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, x = k_121_cast_fp16)[name = string("op_8929_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_8930_cast_fp16 = mul(x = var_8929_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_8930_cast_fp16")];
            tensor<int32, [4]> var_8935_begin_0 = const()[name = string("op_8935_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8935_end_0 = const()[name = string("op_8935_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_8935_end_mask_0 = const()[name = string("op_8935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, x = k_121_cast_fp16)[name = string("op_8935_cast_fp16")];
            int32 var_8937 = const()[name = string("op_8937"), val = int32(-1)];
            bool var_8938_interleave_0 = const()[name = string("op_8938_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_8938_cast_fp16 = concat(axis = var_8937, interleave = var_8938_interleave_0, values = (var_8930_cast_fp16, var_8935_cast_fp16))[name = string("op_8938_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_8939_cast_fp16 = mul(x = var_8938_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8939_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_123_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8939_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_40_reps_0 = const()[name = string("tile_40_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = k_123_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 1, 48, 256]> tile_40_cast_fp16 = tile(reps = tile_40_reps_0, x = transpose_80_cast_fp16)[name = string("tile_40_cast_fp16")];
            tensor<int32, [5]> concat_80 = const()[name = string("concat_80"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_80_cast_fp16 = reshape(shape = concat_80, x = tile_40_cast_fp16)[name = string("reshape_80_cast_fp16")];
            tensor<int32, [5]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_81 = const()[name = string("concat_81"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = reshape_80_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 1, 48, 256]> reshape_81_cast_fp16 = reshape(shape = concat_81, x = transpose_81_cast_fp16)[name = string("reshape_81_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_41_reps_0 = const()[name = string("tile_41_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_82_cast_fp16 = transpose(perm = transpose_82_perm_0, x = v_81_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 1, 48, 256]> tile_41_cast_fp16 = tile(reps = tile_41_reps_0, x = transpose_82_cast_fp16)[name = string("tile_41_cast_fp16")];
            tensor<int32, [5]> concat_82 = const()[name = string("concat_82"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_82_cast_fp16 = reshape(shape = concat_82, x = tile_41_cast_fp16)[name = string("reshape_82_cast_fp16")];
            tensor<int32, [5]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_83 = const()[name = string("concat_83"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = reshape_82_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 1, 48, 256]> reshape_83_cast_fp16 = reshape(shape = concat_83, x = transpose_83_cast_fp16)[name = string("reshape_83_cast_fp16")];
            tensor<int32, [4]> v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8953_transpose_x_0 = const()[name = string("op_8953_transpose_x_0"), val = bool(false)];
            bool var_8953_transpose_y_0 = const()[name = string("op_8953_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = reshape_81_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 4, 48, 48]> var_8953_cast_fp16 = matmul(transpose_x = var_8953_transpose_x_0, transpose_y = var_8953_transpose_y_0, x = q_83_cast_fp16, y = transpose_116_cast_fp16)[name = string("op_8953_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_8955_cast_fp16 = add(x = var_8953_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8955_cast_fp16")];
            int32 var_8956 = const()[name = string("op_8956"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_329_cast_fp16 = softmax(axis = var_8956, x = var_8955_cast_fp16)[name = string("w_329_cast_fp16")];
            bool var_8959_transpose_x_0 = const()[name = string("op_8959_transpose_x_0"), val = bool(false)];
            bool var_8959_transpose_y_0 = const()[name = string("op_8959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_83_cast_fp16 = transpose(perm = v_83_perm_0, x = reshape_83_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 4, 48, 256]> var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = w_329_cast_fp16, y = v_83_cast_fp16)[name = string("op_8959_cast_fp16")];
            tensor<int32, [4]> var_8962_perm_0 = const()[name = string("op_8962_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8966 = const()[name = string("op_8966"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_8962_cast_fp16 = transpose(perm = var_8962_perm_0, x = var_8959_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 48, 1024]> input_243_cast_fp16 = reshape(shape = var_8966, x = var_8962_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115892736))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116417088))))[name = string("tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8971_cast_fp16 = abs(x = linear_184_cast_fp16)[name = string("op_8971_cast_fp16")];
            bool var_8974 = const()[name = string("op_8974"), val = bool(true)];
            tensor<int32, [1]> var_8975_axes_0 = const()[name = string("op_8975_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_8975_cast_fp16 = reduce_max(axes = var_8975_axes_0, keep_dims = var_8974, x = var_8971_cast_fp16)[name = string("op_8975_cast_fp16")];
            fp16 var_8976_to_fp16 = const()[name = string("op_8976_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_331_cast_fp16 = maximum(x = var_8975_cast_fp16, y = var_8976_to_fp16)[name = string("s_331_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_331_cast_fp16 = real_div(x = linear_184_cast_fp16, y = s_331_cast_fp16)[name = string("y_331_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8979_cast_fp16 = mul(x = y_331_cast_fp16, y = y_331_cast_fp16)[name = string("op_8979_cast_fp16")];
            tensor<int32, [1]> var_8984_axes_0 = const()[name = string("op_8984_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8984_keep_dims_0 = const()[name = string("op_8984_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_8984_cast_fp16 = reduce_mean(axes = var_8984_axes_0, keep_dims = var_8984_keep_dims_0, x = var_8979_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8985_cast_fp16 = mul(x = s_331_cast_fp16, y = s_331_cast_fp16)[name = string("op_8985_cast_fp16")];
            fp32 var_8986_epsilon_0 = const()[name = string("op_8986_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_8986_cast_fp16 = inverse(epsilon = var_8986_epsilon_0, x = var_8985_cast_fp16)[name = string("op_8986_cast_fp16")];
            fp16 var_8987_to_fp16 = const()[name = string("op_8987_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_8988_cast_fp16 = mul(x = var_8986_cast_fp16, y = var_8987_to_fp16)[name = string("op_8988_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_8990_cast_fp16 = add(x = var_8984_cast_fp16, y = var_8988_cast_fp16)[name = string("op_8990_cast_fp16")];
            fp32 var_8991_epsilon_0 = const()[name = string("op_8991_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_8991_cast_fp16 = rsqrt(epsilon = var_8991_epsilon_0, x = var_8990_cast_fp16)[name = string("op_8991_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_289_cast_fp16 = mul(x = y_331_cast_fp16, y = var_8991_cast_fp16)[name = string("out_289_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116418176)))];
            tensor<fp16, [1, 48, 512]> var_8993_cast_fp16 = mul(x = out_289_cast_fp16, y = tm_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("op_8993_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_371_cast_fp16 = add(x = x_361_cast_fp16, y = var_8993_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_8996_cast_fp16 = abs(x = x_371_cast_fp16)[name = string("op_8996_cast_fp16")];
            bool var_8999 = const()[name = string("op_8999"), val = bool(true)];
            tensor<int32, [1]> var_9000_axes_0 = const()[name = string("op_9000_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9000_cast_fp16 = reduce_max(axes = var_9000_axes_0, keep_dims = var_8999, x = var_8996_cast_fp16)[name = string("op_9000_cast_fp16")];
            fp16 var_9001_to_fp16 = const()[name = string("op_9001_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_333_cast_fp16 = maximum(x = var_9000_cast_fp16, y = var_9001_to_fp16)[name = string("s_333_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_333_cast_fp16 = real_div(x = x_371_cast_fp16, y = s_333_cast_fp16)[name = string("y_333_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9004_cast_fp16 = mul(x = y_333_cast_fp16, y = y_333_cast_fp16)[name = string("op_9004_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9004_cast_fp16)[name = string("op_9009_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9010_cast_fp16 = mul(x = s_333_cast_fp16, y = s_333_cast_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9015_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9015_cast_fp16")];
            fp32 var_9016_epsilon_0 = const()[name = string("op_9016_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9016_cast_fp16 = rsqrt(epsilon = var_9016_epsilon_0, x = var_9015_cast_fp16)[name = string("op_9016_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_291_cast_fp16 = mul(x = y_333_cast_fp16, y = var_9016_cast_fp16)[name = string("out_291_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116419264)))];
            tensor<fp16, [1, 48, 512]> input_245_cast_fp16 = mul(x = out_291_cast_fp16, y = tm_layers_20_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116420352))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117468992))))[name = string("tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_185_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_185_cast_fp16")];
            string var_9023_mode_0 = const()[name = string("op_9023_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_9023_cast_fp16 = gelu(mode = var_9023_mode_0, x = linear_185_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117473152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118521792))))[name = string("tm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_186_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_247_cast_fp16 = mul(x = var_9023_cast_fp16, y = linear_186_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118525952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119574592))))[name = string("tm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9031_cast_fp16 = abs(x = linear_187_cast_fp16)[name = string("op_9031_cast_fp16")];
            bool var_9034 = const()[name = string("op_9034"), val = bool(true)];
            tensor<int32, [1]> var_9035_axes_0 = const()[name = string("op_9035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9035_cast_fp16 = reduce_max(axes = var_9035_axes_0, keep_dims = var_9034, x = var_9031_cast_fp16)[name = string("op_9035_cast_fp16")];
            fp16 var_9036_to_fp16 = const()[name = string("op_9036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_335_cast_fp16 = maximum(x = var_9035_cast_fp16, y = var_9036_to_fp16)[name = string("s_335_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_335_cast_fp16 = real_div(x = linear_187_cast_fp16, y = s_335_cast_fp16)[name = string("y_335_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9039_cast_fp16 = mul(x = y_335_cast_fp16, y = y_335_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [1]> var_9044_axes_0 = const()[name = string("op_9044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9044_keep_dims_0 = const()[name = string("op_9044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9044_cast_fp16 = reduce_mean(axes = var_9044_axes_0, keep_dims = var_9044_keep_dims_0, x = var_9039_cast_fp16)[name = string("op_9044_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9045_cast_fp16 = mul(x = s_335_cast_fp16, y = s_335_cast_fp16)[name = string("op_9045_cast_fp16")];
            fp32 var_9046_epsilon_0 = const()[name = string("op_9046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9046_cast_fp16 = inverse(epsilon = var_9046_epsilon_0, x = var_9045_cast_fp16)[name = string("op_9046_cast_fp16")];
            fp16 var_9047_to_fp16 = const()[name = string("op_9047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9048_cast_fp16 = mul(x = var_9046_cast_fp16, y = var_9047_to_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9050_cast_fp16 = add(x = var_9044_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9050_cast_fp16")];
            fp32 var_9051_epsilon_0 = const()[name = string("op_9051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9051_cast_fp16 = rsqrt(epsilon = var_9051_epsilon_0, x = var_9050_cast_fp16)[name = string("op_9051_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_293_cast_fp16 = mul(x = y_335_cast_fp16, y = var_9051_cast_fp16)[name = string("out_293_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119575680)))];
            tensor<fp16, [1, 48, 512]> var_9053_cast_fp16 = mul(x = out_293_cast_fp16, y = tm_layers_20_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9053_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_249_cast_fp16 = add(x = x_371_cast_fp16, y = var_9053_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119576768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119838976))))[name = string("tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_188_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_188_cast_fp16")];
            string x_375_mode_0 = const()[name = string("x_375_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_375_cast_fp16 = gelu(mode = x_375_mode_0, x = linear_188_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [4]> var_9073_begin_0 = const()[name = string("op_9073_begin_0"), val = tensor<int32, [4]>([0, 0, 20, 0])];
            tensor<int32, [4]> var_9073_end_0 = const()[name = string("op_9073_end_0"), val = tensor<int32, [4]>([1, 48, 21, 512])];
            tensor<bool, [4]> var_9073_end_mask_0 = const()[name = string("op_9073_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9073_squeeze_mask_0 = const()[name = string("op_9073_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_9073_cast_fp16 = slice_by_index(begin = var_9073_begin_0, end = var_9073_end_0, end_mask = var_9073_end_mask_0, squeeze_mask = var_9073_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_251_cast_fp16 = mul(x = x_375_cast_fp16, y = var_9073_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119840064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120102272))))[name = string("tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9083_cast_fp16 = abs(x = linear_189_cast_fp16)[name = string("op_9083_cast_fp16")];
            bool var_9086 = const()[name = string("op_9086"), val = bool(true)];
            tensor<int32, [1]> var_9087_axes_0 = const()[name = string("op_9087_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9087_cast_fp16 = reduce_max(axes = var_9087_axes_0, keep_dims = var_9086, x = var_9083_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_337_cast_fp16 = maximum(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("s_337_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_337_cast_fp16 = real_div(x = linear_189_cast_fp16, y = s_337_cast_fp16)[name = string("y_337_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9091_cast_fp16 = mul(x = y_337_cast_fp16, y = y_337_cast_fp16)[name = string("op_9091_cast_fp16")];
            tensor<int32, [1]> var_9096_axes_0 = const()[name = string("op_9096_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9096_keep_dims_0 = const()[name = string("op_9096_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9096_cast_fp16 = reduce_mean(axes = var_9096_axes_0, keep_dims = var_9096_keep_dims_0, x = var_9091_cast_fp16)[name = string("op_9096_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9097_cast_fp16 = mul(x = s_337_cast_fp16, y = s_337_cast_fp16)[name = string("op_9097_cast_fp16")];
            fp32 var_9098_epsilon_0 = const()[name = string("op_9098_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9098_cast_fp16 = inverse(epsilon = var_9098_epsilon_0, x = var_9097_cast_fp16)[name = string("op_9098_cast_fp16")];
            fp16 var_9099_to_fp16 = const()[name = string("op_9099_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9100_cast_fp16 = mul(x = var_9098_cast_fp16, y = var_9099_to_fp16)[name = string("op_9100_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9102_cast_fp16 = add(x = var_9096_cast_fp16, y = var_9100_cast_fp16)[name = string("op_9102_cast_fp16")];
            fp32 var_9103_epsilon_0 = const()[name = string("op_9103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9103_cast_fp16 = rsqrt(epsilon = var_9103_epsilon_0, x = var_9102_cast_fp16)[name = string("op_9103_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_295_cast_fp16 = mul(x = y_337_cast_fp16, y = var_9103_cast_fp16)[name = string("out_295_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120103360)))];
            tensor<fp16, [1, 48, 512]> var_9105_cast_fp16 = mul(x = out_295_cast_fp16, y = tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9105_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_41_cast_fp16 = add(x = input_249_cast_fp16, y = var_9105_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<fp16, [1]> tm_layers_20_layer_scalar_to_fp16 = const()[name = string("tm_layers_20_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ecp-1])];
            tensor<fp16, [1, 48, 512]> x_379_cast_fp16 = mul(x = h_41_cast_fp16, y = tm_layers_20_layer_scalar_to_fp16)[name = string("x_379_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9109_cast_fp16 = abs(x = x_379_cast_fp16)[name = string("op_9109_cast_fp16")];
            bool var_9112 = const()[name = string("op_9112"), val = bool(true)];
            tensor<int32, [1]> var_9113_axes_0 = const()[name = string("op_9113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9113_cast_fp16 = reduce_max(axes = var_9113_axes_0, keep_dims = var_9112, x = var_9109_cast_fp16)[name = string("op_9113_cast_fp16")];
            fp16 var_9114_to_fp16 = const()[name = string("op_9114_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_339_cast_fp16 = maximum(x = var_9113_cast_fp16, y = var_9114_to_fp16)[name = string("s_339_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_339_cast_fp16 = real_div(x = x_379_cast_fp16, y = s_339_cast_fp16)[name = string("y_339_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9117_cast_fp16 = mul(x = y_339_cast_fp16, y = y_339_cast_fp16)[name = string("op_9117_cast_fp16")];
            tensor<int32, [1]> var_9122_axes_0 = const()[name = string("op_9122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9122_keep_dims_0 = const()[name = string("op_9122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9122_cast_fp16 = reduce_mean(axes = var_9122_axes_0, keep_dims = var_9122_keep_dims_0, x = var_9117_cast_fp16)[name = string("op_9122_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9123_cast_fp16 = mul(x = s_339_cast_fp16, y = s_339_cast_fp16)[name = string("op_9123_cast_fp16")];
            fp32 var_9124_epsilon_0 = const()[name = string("op_9124_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9124_cast_fp16 = inverse(epsilon = var_9124_epsilon_0, x = var_9123_cast_fp16)[name = string("op_9124_cast_fp16")];
            fp16 var_9125_to_fp16 = const()[name = string("op_9125_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9126_cast_fp16 = mul(x = var_9124_cast_fp16, y = var_9125_to_fp16)[name = string("op_9126_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9128_cast_fp16 = add(x = var_9122_cast_fp16, y = var_9126_cast_fp16)[name = string("op_9128_cast_fp16")];
            fp32 var_9129_epsilon_0 = const()[name = string("op_9129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9129_cast_fp16 = rsqrt(epsilon = var_9129_epsilon_0, x = var_9128_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_297_cast_fp16 = mul(x = y_339_cast_fp16, y = var_9129_cast_fp16)[name = string("out_297_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120104448)))];
            tensor<fp16, [1, 48, 512]> input_253_cast_fp16 = mul(x = out_297_cast_fp16, y = tm_layers_21_input_layernorm_weight_to_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120105536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120629888))))[name = string("tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9139 = const()[name = string("op_9139"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_9140_cast_fp16 = reshape(shape = var_9139, x = linear_190_cast_fp16)[name = string("op_9140_cast_fp16")];
            tensor<int32, [4]> x_381_perm_0 = const()[name = string("x_381_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120632000))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120894208))))[name = string("tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9151 = const()[name = string("op_9151"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_9152_cast_fp16 = reshape(shape = var_9151, x = linear_191_cast_fp16)[name = string("op_9152_cast_fp16")];
            tensor<int32, [4]> x_383_perm_0 = const()[name = string("x_383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120895296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121157504))))[name = string("tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_9163 = const()[name = string("op_9163"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_9164_cast_fp16 = reshape(shape = var_9163, x = linear_192_cast_fp16)[name = string("op_9164_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_381_cast_fp16 = transpose(perm = x_381_perm_0, x = var_9140_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 4, 48, 256]> var_9168_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_9168_cast_fp16")];
            bool var_9171 = const()[name = string("op_9171"), val = bool(true)];
            tensor<int32, [1]> var_9172_axes_0 = const()[name = string("op_9172_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_9172_cast_fp16 = reduce_max(axes = var_9172_axes_0, keep_dims = var_9171, x = var_9168_cast_fp16)[name = string("op_9172_cast_fp16")];
            fp16 var_9173_to_fp16 = const()[name = string("op_9173_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_341_cast_fp16 = maximum(x = var_9172_cast_fp16, y = var_9173_to_fp16)[name = string("s_341_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_341_cast_fp16 = real_div(x = x_381_cast_fp16, y = s_341_cast_fp16)[name = string("y_341_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_9176_cast_fp16 = mul(x = y_341_cast_fp16, y = y_341_cast_fp16)[name = string("op_9176_cast_fp16")];
            tensor<int32, [1]> var_9181_axes_0 = const()[name = string("op_9181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9181_keep_dims_0 = const()[name = string("op_9181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_9181_cast_fp16 = reduce_mean(axes = var_9181_axes_0, keep_dims = var_9181_keep_dims_0, x = var_9176_cast_fp16)[name = string("op_9181_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_9182_cast_fp16 = mul(x = s_341_cast_fp16, y = s_341_cast_fp16)[name = string("op_9182_cast_fp16")];
            fp32 var_9183_epsilon_0 = const()[name = string("op_9183_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_9183_cast_fp16 = inverse(epsilon = var_9183_epsilon_0, x = var_9182_cast_fp16)[name = string("op_9183_cast_fp16")];
            fp16 var_9184_to_fp16 = const()[name = string("op_9184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_9185_cast_fp16 = mul(x = var_9183_cast_fp16, y = var_9184_to_fp16)[name = string("op_9185_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_9187_cast_fp16 = add(x = var_9181_cast_fp16, y = var_9185_cast_fp16)[name = string("op_9187_cast_fp16")];
            fp32 var_9188_epsilon_0 = const()[name = string("op_9188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_9188_cast_fp16 = rsqrt(epsilon = var_9188_epsilon_0, x = var_9187_cast_fp16)[name = string("op_9188_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_299_cast_fp16 = mul(x = y_341_cast_fp16, y = var_9188_cast_fp16)[name = string("out_299_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_85_cast_fp16 = mul(x = out_299_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_85_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_383_cast_fp16 = transpose(perm = x_383_perm_0, x = var_9152_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2, 48, 256]> var_9191_cast_fp16 = abs(x = x_383_cast_fp16)[name = string("op_9191_cast_fp16")];
            bool var_9194 = const()[name = string("op_9194"), val = bool(true)];
            tensor<int32, [1]> var_9195_axes_0 = const()[name = string("op_9195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_9195_cast_fp16 = reduce_max(axes = var_9195_axes_0, keep_dims = var_9194, x = var_9191_cast_fp16)[name = string("op_9195_cast_fp16")];
            fp16 var_9196_to_fp16 = const()[name = string("op_9196_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_343_cast_fp16 = maximum(x = var_9195_cast_fp16, y = var_9196_to_fp16)[name = string("s_343_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_343_cast_fp16 = real_div(x = x_383_cast_fp16, y = s_343_cast_fp16)[name = string("y_343_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_9199_cast_fp16 = mul(x = y_343_cast_fp16, y = y_343_cast_fp16)[name = string("op_9199_cast_fp16")];
            tensor<int32, [1]> var_9204_axes_0 = const()[name = string("op_9204_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9204_keep_dims_0 = const()[name = string("op_9204_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_9204_cast_fp16 = reduce_mean(axes = var_9204_axes_0, keep_dims = var_9204_keep_dims_0, x = var_9199_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_9205_cast_fp16 = mul(x = s_343_cast_fp16, y = s_343_cast_fp16)[name = string("op_9205_cast_fp16")];
            fp32 var_9206_epsilon_0 = const()[name = string("op_9206_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_9206_cast_fp16 = inverse(epsilon = var_9206_epsilon_0, x = var_9205_cast_fp16)[name = string("op_9206_cast_fp16")];
            fp16 var_9207_to_fp16 = const()[name = string("op_9207_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_9208_cast_fp16 = mul(x = var_9206_cast_fp16, y = var_9207_to_fp16)[name = string("op_9208_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_9210_cast_fp16 = add(x = var_9204_cast_fp16, y = var_9208_cast_fp16)[name = string("op_9210_cast_fp16")];
            fp32 var_9211_epsilon_0 = const()[name = string("op_9211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_9211_cast_fp16 = rsqrt(epsilon = var_9211_epsilon_0, x = var_9210_cast_fp16)[name = string("op_9211_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_301_cast_fp16 = mul(x = y_343_cast_fp16, y = var_9211_cast_fp16)[name = string("out_301_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_127_cast_fp16 = mul(x = out_301_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_127_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_9164_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2, 48, 256]> var_9214_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_9214_cast_fp16")];
            bool var_9217 = const()[name = string("op_9217"), val = bool(true)];
            tensor<int32, [1]> var_9218_axes_0 = const()[name = string("op_9218_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_9218_cast_fp16 = reduce_max(axes = var_9218_axes_0, keep_dims = var_9217, x = var_9214_cast_fp16)[name = string("op_9218_cast_fp16")];
            fp16 var_9219_to_fp16 = const()[name = string("op_9219_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_345_cast_fp16 = maximum(x = var_9218_cast_fp16, y = var_9219_to_fp16)[name = string("s_345_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_345_cast_fp16 = real_div(x = x_385_cast_fp16, y = s_345_cast_fp16)[name = string("y_345_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_9222_cast_fp16 = mul(x = y_345_cast_fp16, y = y_345_cast_fp16)[name = string("op_9222_cast_fp16")];
            tensor<int32, [1]> var_9227_axes_0 = const()[name = string("op_9227_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9227_keep_dims_0 = const()[name = string("op_9227_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_9227_cast_fp16 = reduce_mean(axes = var_9227_axes_0, keep_dims = var_9227_keep_dims_0, x = var_9222_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_9228_cast_fp16 = mul(x = s_345_cast_fp16, y = s_345_cast_fp16)[name = string("op_9228_cast_fp16")];
            fp32 var_9229_epsilon_0 = const()[name = string("op_9229_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_9229_cast_fp16 = inverse(epsilon = var_9229_epsilon_0, x = var_9228_cast_fp16)[name = string("op_9229_cast_fp16")];
            fp16 var_9230_to_fp16 = const()[name = string("op_9230_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_9231_cast_fp16 = mul(x = var_9229_cast_fp16, y = var_9230_to_fp16)[name = string("op_9231_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_9233_cast_fp16 = add(x = var_9227_cast_fp16, y = var_9231_cast_fp16)[name = string("op_9233_cast_fp16")];
            fp32 var_9234_epsilon_0 = const()[name = string("op_9234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_9234_cast_fp16 = rsqrt(epsilon = var_9234_epsilon_0, x = var_9233_cast_fp16)[name = string("op_9234_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_85_cast_fp16 = mul(x = y_345_cast_fp16, y = var_9234_cast_fp16)[name = string("v_85_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_9236_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9236_cast_fp16")];
            tensor<int32, [4]> var_9241_begin_0 = const()[name = string("op_9241_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9241_end_0 = const()[name = string("op_9241_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_9241_end_mask_0 = const()[name = string("op_9241_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_9241_cast_fp16 = slice_by_index(begin = var_9241_begin_0, end = var_9241_end_0, end_mask = var_9241_end_mask_0, x = q_85_cast_fp16)[name = string("op_9241_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_9242_cast_fp16 = mul(x = var_9241_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_9242_cast_fp16")];
            tensor<int32, [4]> var_9247_begin_0 = const()[name = string("op_9247_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9247_end_0 = const()[name = string("op_9247_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_9247_end_mask_0 = const()[name = string("op_9247_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_9247_cast_fp16 = slice_by_index(begin = var_9247_begin_0, end = var_9247_end_0, end_mask = var_9247_end_mask_0, x = q_85_cast_fp16)[name = string("op_9247_cast_fp16")];
            int32 var_9249 = const()[name = string("op_9249"), val = int32(-1)];
            bool var_9250_interleave_0 = const()[name = string("op_9250_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_9250_cast_fp16 = concat(axis = var_9249, interleave = var_9250_interleave_0, values = (var_9242_cast_fp16, var_9247_cast_fp16))[name = string("op_9250_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_9251_cast_fp16 = mul(x = var_9250_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9251_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_87_cast_fp16 = add(x = var_9236_cast_fp16, y = var_9251_cast_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_9254_cast_fp16 = mul(x = k_127_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9254_cast_fp16")];
            tensor<int32, [4]> var_9259_begin_0 = const()[name = string("op_9259_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9259_end_0 = const()[name = string("op_9259_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_9259_end_mask_0 = const()[name = string("op_9259_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_9259_cast_fp16 = slice_by_index(begin = var_9259_begin_0, end = var_9259_end_0, end_mask = var_9259_end_mask_0, x = k_127_cast_fp16)[name = string("op_9259_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_9260_cast_fp16 = mul(x = var_9259_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_9260_cast_fp16")];
            tensor<int32, [4]> var_9265_begin_0 = const()[name = string("op_9265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9265_end_0 = const()[name = string("op_9265_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_9265_end_mask_0 = const()[name = string("op_9265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_9265_cast_fp16 = slice_by_index(begin = var_9265_begin_0, end = var_9265_end_0, end_mask = var_9265_end_mask_0, x = k_127_cast_fp16)[name = string("op_9265_cast_fp16")];
            int32 var_9267 = const()[name = string("op_9267"), val = int32(-1)];
            bool var_9268_interleave_0 = const()[name = string("op_9268_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_9268_cast_fp16 = concat(axis = var_9267, interleave = var_9268_interleave_0, values = (var_9260_cast_fp16, var_9265_cast_fp16))[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_9269_cast_fp16 = mul(x = var_9268_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_129_cast_fp16 = add(x = var_9254_cast_fp16, y = var_9269_cast_fp16)[name = string("k_129_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_42_reps_0 = const()[name = string("tile_42_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = k_129_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 1, 48, 256]> tile_42_cast_fp16 = tile(reps = tile_42_reps_0, x = transpose_84_cast_fp16)[name = string("tile_42_cast_fp16")];
            tensor<int32, [5]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_84_cast_fp16 = reshape(shape = concat_84, x = tile_42_cast_fp16)[name = string("reshape_84_cast_fp16")];
            tensor<int32, [5]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = reshape_84_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 1, 48, 256]> reshape_85_cast_fp16 = reshape(shape = concat_85, x = transpose_85_cast_fp16)[name = string("reshape_85_cast_fp16")];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_43_reps_0 = const()[name = string("tile_43_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_86_cast_fp16 = transpose(perm = transpose_86_perm_0, x = v_85_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 1, 48, 256]> tile_43_cast_fp16 = tile(reps = tile_43_reps_0, x = transpose_86_cast_fp16)[name = string("tile_43_cast_fp16")];
            tensor<int32, [5]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_86_cast_fp16 = reshape(shape = concat_86, x = tile_43_cast_fp16)[name = string("reshape_86_cast_fp16")];
            tensor<int32, [5]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_87_cast_fp16 = transpose(perm = transpose_87_perm_0, x = reshape_86_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 1, 48, 256]> reshape_87_cast_fp16 = reshape(shape = concat_87, x = transpose_87_cast_fp16)[name = string("reshape_87_cast_fp16")];
            tensor<int32, [4]> v_87_perm_0 = const()[name = string("v_87_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9283_transpose_x_0 = const()[name = string("op_9283_transpose_x_0"), val = bool(false)];
            bool var_9283_transpose_y_0 = const()[name = string("op_9283_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_117_cast_fp16 = transpose(perm = transpose_117_perm_0, x = reshape_85_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 4, 48, 48]> var_9283_cast_fp16 = matmul(transpose_x = var_9283_transpose_x_0, transpose_y = var_9283_transpose_y_0, x = q_87_cast_fp16, y = transpose_117_cast_fp16)[name = string("op_9283_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_9285_cast_fp16 = add(x = var_9283_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9285_cast_fp16")];
            int32 var_9286 = const()[name = string("op_9286"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_345_cast_fp16 = softmax(axis = var_9286, x = var_9285_cast_fp16)[name = string("w_345_cast_fp16")];
            bool var_9289_transpose_x_0 = const()[name = string("op_9289_transpose_x_0"), val = bool(false)];
            bool var_9289_transpose_y_0 = const()[name = string("op_9289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_87_cast_fp16 = transpose(perm = v_87_perm_0, x = reshape_87_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 4, 48, 256]> var_9289_cast_fp16 = matmul(transpose_x = var_9289_transpose_x_0, transpose_y = var_9289_transpose_y_0, x = w_345_cast_fp16, y = v_87_cast_fp16)[name = string("op_9289_cast_fp16")];
            tensor<int32, [4]> var_9292_perm_0 = const()[name = string("op_9292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9296 = const()[name = string("op_9296"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_9292_cast_fp16 = transpose(perm = var_9292_perm_0, x = var_9289_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 48, 1024]> input_255_cast_fp16 = reshape(shape = var_9296, x = var_9292_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121158592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121682944))))[name = string("tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9301_cast_fp16 = abs(x = linear_193_cast_fp16)[name = string("op_9301_cast_fp16")];
            bool var_9304 = const()[name = string("op_9304"), val = bool(true)];
            tensor<int32, [1]> var_9305_axes_0 = const()[name = string("op_9305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9305_cast_fp16 = reduce_max(axes = var_9305_axes_0, keep_dims = var_9304, x = var_9301_cast_fp16)[name = string("op_9305_cast_fp16")];
            fp16 var_9306_to_fp16 = const()[name = string("op_9306_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_347_cast_fp16 = maximum(x = var_9305_cast_fp16, y = var_9306_to_fp16)[name = string("s_347_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_347_cast_fp16 = real_div(x = linear_193_cast_fp16, y = s_347_cast_fp16)[name = string("y_347_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9309_cast_fp16 = mul(x = y_347_cast_fp16, y = y_347_cast_fp16)[name = string("op_9309_cast_fp16")];
            tensor<int32, [1]> var_9314_axes_0 = const()[name = string("op_9314_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9314_keep_dims_0 = const()[name = string("op_9314_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9314_cast_fp16 = reduce_mean(axes = var_9314_axes_0, keep_dims = var_9314_keep_dims_0, x = var_9309_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9315_cast_fp16 = mul(x = s_347_cast_fp16, y = s_347_cast_fp16)[name = string("op_9315_cast_fp16")];
            fp32 var_9316_epsilon_0 = const()[name = string("op_9316_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9316_cast_fp16 = inverse(epsilon = var_9316_epsilon_0, x = var_9315_cast_fp16)[name = string("op_9316_cast_fp16")];
            fp16 var_9317_to_fp16 = const()[name = string("op_9317_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9318_cast_fp16 = mul(x = var_9316_cast_fp16, y = var_9317_to_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9320_cast_fp16 = add(x = var_9314_cast_fp16, y = var_9318_cast_fp16)[name = string("op_9320_cast_fp16")];
            fp32 var_9321_epsilon_0 = const()[name = string("op_9321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9321_cast_fp16 = rsqrt(epsilon = var_9321_epsilon_0, x = var_9320_cast_fp16)[name = string("op_9321_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_303_cast_fp16 = mul(x = y_347_cast_fp16, y = var_9321_cast_fp16)[name = string("out_303_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121684032)))];
            tensor<fp16, [1, 48, 512]> var_9323_cast_fp16 = mul(x = out_303_cast_fp16, y = tm_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("op_9323_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_389_cast_fp16 = add(x = x_379_cast_fp16, y = var_9323_cast_fp16)[name = string("x_389_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9326_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_9326_cast_fp16")];
            bool var_9329 = const()[name = string("op_9329"), val = bool(true)];
            tensor<int32, [1]> var_9330_axes_0 = const()[name = string("op_9330_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9330_cast_fp16 = reduce_max(axes = var_9330_axes_0, keep_dims = var_9329, x = var_9326_cast_fp16)[name = string("op_9330_cast_fp16")];
            fp16 var_9331_to_fp16 = const()[name = string("op_9331_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_349_cast_fp16 = maximum(x = var_9330_cast_fp16, y = var_9331_to_fp16)[name = string("s_349_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_349_cast_fp16 = real_div(x = x_389_cast_fp16, y = s_349_cast_fp16)[name = string("y_349_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9334_cast_fp16 = mul(x = y_349_cast_fp16, y = y_349_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [1]> var_9339_axes_0 = const()[name = string("op_9339_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9339_keep_dims_0 = const()[name = string("op_9339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9339_cast_fp16 = reduce_mean(axes = var_9339_axes_0, keep_dims = var_9339_keep_dims_0, x = var_9334_cast_fp16)[name = string("op_9339_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9340_cast_fp16 = mul(x = s_349_cast_fp16, y = s_349_cast_fp16)[name = string("op_9340_cast_fp16")];
            fp32 var_9341_epsilon_0 = const()[name = string("op_9341_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9341_cast_fp16 = inverse(epsilon = var_9341_epsilon_0, x = var_9340_cast_fp16)[name = string("op_9341_cast_fp16")];
            fp16 var_9342_to_fp16 = const()[name = string("op_9342_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9343_cast_fp16 = mul(x = var_9341_cast_fp16, y = var_9342_to_fp16)[name = string("op_9343_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9345_cast_fp16 = add(x = var_9339_cast_fp16, y = var_9343_cast_fp16)[name = string("op_9345_cast_fp16")];
            fp32 var_9346_epsilon_0 = const()[name = string("op_9346_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9346_cast_fp16 = rsqrt(epsilon = var_9346_epsilon_0, x = var_9345_cast_fp16)[name = string("op_9346_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_305_cast_fp16 = mul(x = y_349_cast_fp16, y = var_9346_cast_fp16)[name = string("out_305_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121685120)))];
            tensor<fp16, [1, 48, 512]> input_257_cast_fp16 = mul(x = out_305_cast_fp16, y = tm_layers_21_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121686208))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122734848))))[name = string("tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_194_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_194_cast_fp16")];
            string var_9353_mode_0 = const()[name = string("op_9353_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_9353_cast_fp16 = gelu(mode = var_9353_mode_0, x = linear_194_cast_fp16)[name = string("op_9353_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122739008))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123787648))))[name = string("tm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_195_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_259_cast_fp16 = mul(x = var_9353_cast_fp16, y = linear_195_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123791808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124840448))))[name = string("tm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9361_cast_fp16 = abs(x = linear_196_cast_fp16)[name = string("op_9361_cast_fp16")];
            bool var_9364 = const()[name = string("op_9364"), val = bool(true)];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9365_cast_fp16 = reduce_max(axes = var_9365_axes_0, keep_dims = var_9364, x = var_9361_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9366_to_fp16 = const()[name = string("op_9366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_351_cast_fp16 = maximum(x = var_9365_cast_fp16, y = var_9366_to_fp16)[name = string("s_351_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_351_cast_fp16 = real_div(x = linear_196_cast_fp16, y = s_351_cast_fp16)[name = string("y_351_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9369_cast_fp16 = mul(x = y_351_cast_fp16, y = y_351_cast_fp16)[name = string("op_9369_cast_fp16")];
            tensor<int32, [1]> var_9374_axes_0 = const()[name = string("op_9374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9374_keep_dims_0 = const()[name = string("op_9374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9374_cast_fp16 = reduce_mean(axes = var_9374_axes_0, keep_dims = var_9374_keep_dims_0, x = var_9369_cast_fp16)[name = string("op_9374_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9375_cast_fp16 = mul(x = s_351_cast_fp16, y = s_351_cast_fp16)[name = string("op_9375_cast_fp16")];
            fp32 var_9376_epsilon_0 = const()[name = string("op_9376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9376_cast_fp16 = inverse(epsilon = var_9376_epsilon_0, x = var_9375_cast_fp16)[name = string("op_9376_cast_fp16")];
            fp16 var_9377_to_fp16 = const()[name = string("op_9377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9378_cast_fp16 = mul(x = var_9376_cast_fp16, y = var_9377_to_fp16)[name = string("op_9378_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9380_cast_fp16 = add(x = var_9374_cast_fp16, y = var_9378_cast_fp16)[name = string("op_9380_cast_fp16")];
            fp32 var_9381_epsilon_0 = const()[name = string("op_9381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9381_cast_fp16 = rsqrt(epsilon = var_9381_epsilon_0, x = var_9380_cast_fp16)[name = string("op_9381_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_307_cast_fp16 = mul(x = y_351_cast_fp16, y = var_9381_cast_fp16)[name = string("out_307_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124841536)))];
            tensor<fp16, [1, 48, 512]> var_9383_cast_fp16 = mul(x = out_307_cast_fp16, y = tm_layers_21_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9383_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_261_cast_fp16 = add(x = x_389_cast_fp16, y = var_9383_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124842624))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125104832))))[name = string("tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_197_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_197_cast_fp16")];
            string x_393_mode_0 = const()[name = string("x_393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_393_cast_fp16 = gelu(mode = x_393_mode_0, x = linear_197_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<int32, [4]> var_9403_begin_0 = const()[name = string("op_9403_begin_0"), val = tensor<int32, [4]>([0, 0, 21, 0])];
            tensor<int32, [4]> var_9403_end_0 = const()[name = string("op_9403_end_0"), val = tensor<int32, [4]>([1, 48, 22, 512])];
            tensor<bool, [4]> var_9403_end_mask_0 = const()[name = string("op_9403_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9403_squeeze_mask_0 = const()[name = string("op_9403_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_9403_cast_fp16 = slice_by_index(begin = var_9403_begin_0, end = var_9403_end_0, end_mask = var_9403_end_mask_0, squeeze_mask = var_9403_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9403_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_263_cast_fp16 = mul(x = x_393_cast_fp16, y = var_9403_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125105920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125368128))))[name = string("tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9413_cast_fp16 = abs(x = linear_198_cast_fp16)[name = string("op_9413_cast_fp16")];
            bool var_9416 = const()[name = string("op_9416"), val = bool(true)];
            tensor<int32, [1]> var_9417_axes_0 = const()[name = string("op_9417_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9417_cast_fp16 = reduce_max(axes = var_9417_axes_0, keep_dims = var_9416, x = var_9413_cast_fp16)[name = string("op_9417_cast_fp16")];
            fp16 var_9418_to_fp16 = const()[name = string("op_9418_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_353_cast_fp16 = maximum(x = var_9417_cast_fp16, y = var_9418_to_fp16)[name = string("s_353_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_353_cast_fp16 = real_div(x = linear_198_cast_fp16, y = s_353_cast_fp16)[name = string("y_353_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9421_cast_fp16 = mul(x = y_353_cast_fp16, y = y_353_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<int32, [1]> var_9426_axes_0 = const()[name = string("op_9426_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9426_keep_dims_0 = const()[name = string("op_9426_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9426_cast_fp16 = reduce_mean(axes = var_9426_axes_0, keep_dims = var_9426_keep_dims_0, x = var_9421_cast_fp16)[name = string("op_9426_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9427_cast_fp16 = mul(x = s_353_cast_fp16, y = s_353_cast_fp16)[name = string("op_9427_cast_fp16")];
            fp32 var_9428_epsilon_0 = const()[name = string("op_9428_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9428_cast_fp16 = inverse(epsilon = var_9428_epsilon_0, x = var_9427_cast_fp16)[name = string("op_9428_cast_fp16")];
            fp16 var_9429_to_fp16 = const()[name = string("op_9429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9430_cast_fp16 = mul(x = var_9428_cast_fp16, y = var_9429_to_fp16)[name = string("op_9430_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9432_cast_fp16 = add(x = var_9426_cast_fp16, y = var_9430_cast_fp16)[name = string("op_9432_cast_fp16")];
            fp32 var_9433_epsilon_0 = const()[name = string("op_9433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9433_cast_fp16 = rsqrt(epsilon = var_9433_epsilon_0, x = var_9432_cast_fp16)[name = string("op_9433_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_309_cast_fp16 = mul(x = y_353_cast_fp16, y = var_9433_cast_fp16)[name = string("out_309_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125369216)))];
            tensor<fp16, [1, 48, 512]> var_9435_cast_fp16 = mul(x = out_309_cast_fp16, y = tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9435_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_43_cast_fp16 = add(x = input_261_cast_fp16, y = var_9435_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<fp16, [1]> tm_layers_21_layer_scalar_to_fp16 = const()[name = string("tm_layers_21_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e8p-1])];
            tensor<fp16, [1, 48, 512]> x_397_cast_fp16 = mul(x = h_43_cast_fp16, y = tm_layers_21_layer_scalar_to_fp16)[name = string("x_397_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9439_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_9439_cast_fp16")];
            bool var_9442 = const()[name = string("op_9442"), val = bool(true)];
            tensor<int32, [1]> var_9443_axes_0 = const()[name = string("op_9443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9443_cast_fp16 = reduce_max(axes = var_9443_axes_0, keep_dims = var_9442, x = var_9439_cast_fp16)[name = string("op_9443_cast_fp16")];
            fp16 var_9444_to_fp16 = const()[name = string("op_9444_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_355_cast_fp16 = maximum(x = var_9443_cast_fp16, y = var_9444_to_fp16)[name = string("s_355_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_355_cast_fp16 = real_div(x = x_397_cast_fp16, y = s_355_cast_fp16)[name = string("y_355_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9447_cast_fp16 = mul(x = y_355_cast_fp16, y = y_355_cast_fp16)[name = string("op_9447_cast_fp16")];
            tensor<int32, [1]> var_9452_axes_0 = const()[name = string("op_9452_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9452_keep_dims_0 = const()[name = string("op_9452_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9452_cast_fp16 = reduce_mean(axes = var_9452_axes_0, keep_dims = var_9452_keep_dims_0, x = var_9447_cast_fp16)[name = string("op_9452_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9453_cast_fp16 = mul(x = s_355_cast_fp16, y = s_355_cast_fp16)[name = string("op_9453_cast_fp16")];
            fp32 var_9454_epsilon_0 = const()[name = string("op_9454_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9454_cast_fp16 = inverse(epsilon = var_9454_epsilon_0, x = var_9453_cast_fp16)[name = string("op_9454_cast_fp16")];
            fp16 var_9455_to_fp16 = const()[name = string("op_9455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9456_cast_fp16 = mul(x = var_9454_cast_fp16, y = var_9455_to_fp16)[name = string("op_9456_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9458_cast_fp16 = add(x = var_9452_cast_fp16, y = var_9456_cast_fp16)[name = string("op_9458_cast_fp16")];
            fp32 var_9459_epsilon_0 = const()[name = string("op_9459_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9459_cast_fp16 = rsqrt(epsilon = var_9459_epsilon_0, x = var_9458_cast_fp16)[name = string("op_9459_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_311_cast_fp16 = mul(x = y_355_cast_fp16, y = var_9459_cast_fp16)[name = string("out_311_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125370304)))];
            tensor<fp16, [1, 48, 512]> input_265_cast_fp16 = mul(x = out_311_cast_fp16, y = tm_layers_22_input_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125371392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125895744))))[name = string("tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 1024]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<int32, [4]> var_9469 = const()[name = string("op_9469"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 48, 4, 256]> var_9470_cast_fp16 = reshape(shape = var_9469, x = linear_199_cast_fp16)[name = string("op_9470_cast_fp16")];
            tensor<int32, [4]> x_399_perm_0 = const()[name = string("x_399_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125897856))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126160064))))[name = string("tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<int32, [4]> var_9481 = const()[name = string("op_9481"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_9482_cast_fp16 = reshape(shape = var_9481, x = linear_200_cast_fp16)[name = string("op_9482_cast_fp16")];
            tensor<int32, [4]> x_401_perm_0 = const()[name = string("x_401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126161152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126423360))))[name = string("tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<int32, [4]> var_9493 = const()[name = string("op_9493"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 48, 2, 256]> var_9494_cast_fp16 = reshape(shape = var_9493, x = linear_201_cast_fp16)[name = string("op_9494_cast_fp16")];
            tensor<int32, [4]> x_403_perm_0 = const()[name = string("x_403_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 256]> x_399_cast_fp16 = transpose(perm = x_399_perm_0, x = var_9470_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 4, 48, 256]> var_9498_cast_fp16 = abs(x = x_399_cast_fp16)[name = string("op_9498_cast_fp16")];
            bool var_9501 = const()[name = string("op_9501"), val = bool(true)];
            tensor<int32, [1]> var_9502_axes_0 = const()[name = string("op_9502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_9502_cast_fp16 = reduce_max(axes = var_9502_axes_0, keep_dims = var_9501, x = var_9498_cast_fp16)[name = string("op_9502_cast_fp16")];
            fp16 var_9503_to_fp16 = const()[name = string("op_9503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_357_cast_fp16 = maximum(x = var_9502_cast_fp16, y = var_9503_to_fp16)[name = string("s_357_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> y_357_cast_fp16 = real_div(x = x_399_cast_fp16, y = s_357_cast_fp16)[name = string("y_357_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_9506_cast_fp16 = mul(x = y_357_cast_fp16, y = y_357_cast_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9511_axes_0 = const()[name = string("op_9511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9511_keep_dims_0 = const()[name = string("op_9511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_9511_cast_fp16 = reduce_mean(axes = var_9511_axes_0, keep_dims = var_9511_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_9512_cast_fp16 = mul(x = s_357_cast_fp16, y = s_357_cast_fp16)[name = string("op_9512_cast_fp16")];
            fp32 var_9513_epsilon_0 = const()[name = string("op_9513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_9513_cast_fp16 = inverse(epsilon = var_9513_epsilon_0, x = var_9512_cast_fp16)[name = string("op_9513_cast_fp16")];
            fp16 var_9514_to_fp16 = const()[name = string("op_9514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_9515_cast_fp16 = mul(x = var_9513_cast_fp16, y = var_9514_to_fp16)[name = string("op_9515_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_9517_cast_fp16 = add(x = var_9511_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9517_cast_fp16")];
            fp32 var_9518_epsilon_0 = const()[name = string("op_9518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_9518_cast_fp16 = rsqrt(epsilon = var_9518_epsilon_0, x = var_9517_cast_fp16)[name = string("op_9518_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> out_313_cast_fp16 = mul(x = y_357_cast_fp16, y = var_9518_cast_fp16)[name = string("out_313_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_89_cast_fp16 = mul(x = out_313_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_89_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_401_cast_fp16 = transpose(perm = x_401_perm_0, x = var_9482_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 48, 256]> var_9521_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_9521_cast_fp16")];
            bool var_9524 = const()[name = string("op_9524"), val = bool(true)];
            tensor<int32, [1]> var_9525_axes_0 = const()[name = string("op_9525_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_9525_cast_fp16 = reduce_max(axes = var_9525_axes_0, keep_dims = var_9524, x = var_9521_cast_fp16)[name = string("op_9525_cast_fp16")];
            fp16 var_9526_to_fp16 = const()[name = string("op_9526_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_359_cast_fp16 = maximum(x = var_9525_cast_fp16, y = var_9526_to_fp16)[name = string("s_359_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_359_cast_fp16 = real_div(x = x_401_cast_fp16, y = s_359_cast_fp16)[name = string("y_359_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_9529_cast_fp16 = mul(x = y_359_cast_fp16, y = y_359_cast_fp16)[name = string("op_9529_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9529_cast_fp16)[name = string("op_9534_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_9535_cast_fp16 = mul(x = s_359_cast_fp16, y = s_359_cast_fp16)[name = string("op_9535_cast_fp16")];
            fp32 var_9536_epsilon_0 = const()[name = string("op_9536_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_9536_cast_fp16 = inverse(epsilon = var_9536_epsilon_0, x = var_9535_cast_fp16)[name = string("op_9536_cast_fp16")];
            fp16 var_9537_to_fp16 = const()[name = string("op_9537_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_9538_cast_fp16 = mul(x = var_9536_cast_fp16, y = var_9537_to_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_9540_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9538_cast_fp16)[name = string("op_9540_cast_fp16")];
            fp32 var_9541_epsilon_0 = const()[name = string("op_9541_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_9541_cast_fp16 = rsqrt(epsilon = var_9541_epsilon_0, x = var_9540_cast_fp16)[name = string("op_9541_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> out_315_cast_fp16 = mul(x = y_359_cast_fp16, y = var_9541_cast_fp16)[name = string("out_315_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_133_cast_fp16 = mul(x = out_315_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_133_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> x_403_cast_fp16 = transpose(perm = x_403_perm_0, x = var_9494_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 48, 256]> var_9544_cast_fp16 = abs(x = x_403_cast_fp16)[name = string("op_9544_cast_fp16")];
            bool var_9547 = const()[name = string("op_9547"), val = bool(true)];
            tensor<int32, [1]> var_9548_axes_0 = const()[name = string("op_9548_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 48, 1]> var_9548_cast_fp16 = reduce_max(axes = var_9548_axes_0, keep_dims = var_9547, x = var_9544_cast_fp16)[name = string("op_9548_cast_fp16")];
            fp16 var_9549_to_fp16 = const()[name = string("op_9549_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 48, 1]> s_361_cast_fp16 = maximum(x = var_9548_cast_fp16, y = var_9549_to_fp16)[name = string("s_361_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> y_361_cast_fp16 = real_div(x = x_403_cast_fp16, y = s_361_cast_fp16)[name = string("y_361_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_9552_cast_fp16 = mul(x = y_361_cast_fp16, y = y_361_cast_fp16)[name = string("op_9552_cast_fp16")];
            tensor<int32, [1]> var_9557_axes_0 = const()[name = string("op_9557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9557_keep_dims_0 = const()[name = string("op_9557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 48, 1]> var_9557_cast_fp16 = reduce_mean(axes = var_9557_axes_0, keep_dims = var_9557_keep_dims_0, x = var_9552_cast_fp16)[name = string("op_9557_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_9558_cast_fp16 = mul(x = s_361_cast_fp16, y = s_361_cast_fp16)[name = string("op_9558_cast_fp16")];
            fp32 var_9559_epsilon_0 = const()[name = string("op_9559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 48, 1]> var_9559_cast_fp16 = inverse(epsilon = var_9559_epsilon_0, x = var_9558_cast_fp16)[name = string("op_9559_cast_fp16")];
            fp16 var_9560_to_fp16 = const()[name = string("op_9560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 48, 1]> var_9561_cast_fp16 = mul(x = var_9559_cast_fp16, y = var_9560_to_fp16)[name = string("op_9561_cast_fp16")];
            tensor<fp16, [1, 2, 48, 1]> var_9563_cast_fp16 = add(x = var_9557_cast_fp16, y = var_9561_cast_fp16)[name = string("op_9563_cast_fp16")];
            fp32 var_9564_epsilon_0 = const()[name = string("op_9564_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 48, 1]> var_9564_cast_fp16 = rsqrt(epsilon = var_9564_epsilon_0, x = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> v_89_cast_fp16 = mul(x = y_361_cast_fp16, y = var_9564_cast_fp16)[name = string("v_89_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_9566_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9566_cast_fp16")];
            tensor<int32, [4]> var_9571_begin_0 = const()[name = string("op_9571_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9571_end_0 = const()[name = string("op_9571_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_9571_end_mask_0 = const()[name = string("op_9571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 128]> var_9571_cast_fp16 = slice_by_index(begin = var_9571_begin_0, end = var_9571_end_0, end_mask = var_9571_end_mask_0, x = q_89_cast_fp16)[name = string("op_9571_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 128]> var_9572_cast_fp16 = mul(x = var_9571_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_9572_cast_fp16")];
            tensor<int32, [4]> var_9577_begin_0 = const()[name = string("op_9577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9577_end_0 = const()[name = string("op_9577_end_0"), val = tensor<int32, [4]>([1, 4, 48, 128])];
            tensor<bool, [4]> var_9577_end_mask_0 = const()[name = string("op_9577_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 128]> var_9577_cast_fp16 = slice_by_index(begin = var_9577_begin_0, end = var_9577_end_0, end_mask = var_9577_end_mask_0, x = q_89_cast_fp16)[name = string("op_9577_cast_fp16")];
            int32 var_9579 = const()[name = string("op_9579"), val = int32(-1)];
            bool var_9580_interleave_0 = const()[name = string("op_9580_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> var_9580_cast_fp16 = concat(axis = var_9579, interleave = var_9580_interleave_0, values = (var_9572_cast_fp16, var_9577_cast_fp16))[name = string("op_9580_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> var_9581_cast_fp16 = mul(x = var_9580_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9581_cast_fp16")];
            tensor<fp16, [1, 4, 48, 256]> q_91_cast_fp16 = add(x = var_9566_cast_fp16, y = var_9581_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_9584_cast_fp16 = mul(x = k_133_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9584_cast_fp16")];
            tensor<int32, [4]> var_9589_begin_0 = const()[name = string("op_9589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9589_end_0 = const()[name = string("op_9589_end_0"), val = tensor<int32, [4]>([1, 2, 48, 256])];
            tensor<bool, [4]> var_9589_end_mask_0 = const()[name = string("op_9589_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 48, 128]> var_9589_cast_fp16 = slice_by_index(begin = var_9589_begin_0, end = var_9589_end_0, end_mask = var_9589_end_mask_0, x = k_133_cast_fp16)[name = string("op_9589_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 48, 128]> var_9590_cast_fp16 = mul(x = var_9589_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_9590_cast_fp16")];
            tensor<int32, [4]> var_9595_begin_0 = const()[name = string("op_9595_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9595_end_0 = const()[name = string("op_9595_end_0"), val = tensor<int32, [4]>([1, 2, 48, 128])];
            tensor<bool, [4]> var_9595_end_mask_0 = const()[name = string("op_9595_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 48, 128]> var_9595_cast_fp16 = slice_by_index(begin = var_9595_begin_0, end = var_9595_end_0, end_mask = var_9595_end_mask_0, x = k_133_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool var_9598_interleave_0 = const()[name = string("op_9598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 48, 256]> var_9598_cast_fp16 = concat(axis = var_9597, interleave = var_9598_interleave_0, values = (var_9590_cast_fp16, var_9595_cast_fp16))[name = string("op_9598_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> var_9599_cast_fp16 = mul(x = var_9598_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9599_cast_fp16")];
            tensor<fp16, [1, 2, 48, 256]> k_135_cast_fp16 = add(x = var_9584_cast_fp16, y = var_9599_cast_fp16)[name = string("k_135_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_44_reps_0 = const()[name = string("tile_44_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = k_135_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 1, 48, 256]> tile_44_cast_fp16 = tile(reps = tile_44_reps_0, x = transpose_88_cast_fp16)[name = string("tile_44_cast_fp16")];
            tensor<int32, [5]> concat_88 = const()[name = string("concat_88"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_88_cast_fp16 = reshape(shape = concat_88, x = tile_44_cast_fp16)[name = string("reshape_88_cast_fp16")];
            tensor<int32, [5]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = reshape_88_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 1, 48, 256]> reshape_89_cast_fp16 = reshape(shape = concat_89, x = transpose_89_cast_fp16)[name = string("reshape_89_cast_fp16")];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_45_reps_0 = const()[name = string("tile_45_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 48, 256]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = v_89_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 1, 48, 256]> tile_45_cast_fp16 = tile(reps = tile_45_reps_0, x = transpose_90_cast_fp16)[name = string("tile_45_cast_fp16")];
            tensor<int32, [5]> concat_90 = const()[name = string("concat_90"), val = tensor<int32, [5]>([2, 2, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> reshape_90_cast_fp16 = reshape(shape = concat_90, x = tile_45_cast_fp16)[name = string("reshape_90_cast_fp16")];
            tensor<int32, [5]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_91 = const()[name = string("concat_91"), val = tensor<int32, [4]>([-1, 1, 48, 256])];
            tensor<fp16, [2, 2, 1, 48, 256]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = reshape_90_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 1, 48, 256]> reshape_91_cast_fp16 = reshape(shape = concat_91, x = transpose_91_cast_fp16)[name = string("reshape_91_cast_fp16")];
            tensor<int32, [4]> v_91_perm_0 = const()[name = string("v_91_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9613_transpose_x_0 = const()[name = string("op_9613_transpose_x_0"), val = bool(false)];
            bool var_9613_transpose_y_0 = const()[name = string("op_9613_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 48]> transpose_118_cast_fp16 = transpose(perm = transpose_118_perm_0, x = reshape_89_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 4, 48, 48]> var_9613_cast_fp16 = matmul(transpose_x = var_9613_transpose_x_0, transpose_y = var_9613_transpose_y_0, x = q_91_cast_fp16, y = transpose_118_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_9615_cast_fp16 = add(x = var_9613_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9615_cast_fp16")];
            int32 var_9616 = const()[name = string("op_9616"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_361_cast_fp16 = softmax(axis = var_9616, x = var_9615_cast_fp16)[name = string("w_361_cast_fp16")];
            bool var_9619_transpose_x_0 = const()[name = string("op_9619_transpose_x_0"), val = bool(false)];
            bool var_9619_transpose_y_0 = const()[name = string("op_9619_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 256]> v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = reshape_91_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 4, 48, 256]> var_9619_cast_fp16 = matmul(transpose_x = var_9619_transpose_x_0, transpose_y = var_9619_transpose_y_0, x = w_361_cast_fp16, y = v_91_cast_fp16)[name = string("op_9619_cast_fp16")];
            tensor<int32, [4]> var_9622_perm_0 = const()[name = string("op_9622_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9626 = const()[name = string("op_9626"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 48, 4, 256]> var_9622_cast_fp16 = transpose(perm = var_9622_perm_0, x = var_9619_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 48, 1024]> input_267_cast_fp16 = reshape(shape = var_9626, x = var_9622_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126424448))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126948800))))[name = string("tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9631_cast_fp16 = abs(x = linear_202_cast_fp16)[name = string("op_9631_cast_fp16")];
            bool var_9634 = const()[name = string("op_9634"), val = bool(true)];
            tensor<int32, [1]> var_9635_axes_0 = const()[name = string("op_9635_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9635_cast_fp16 = reduce_max(axes = var_9635_axes_0, keep_dims = var_9634, x = var_9631_cast_fp16)[name = string("op_9635_cast_fp16")];
            fp16 var_9636_to_fp16 = const()[name = string("op_9636_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_363_cast_fp16 = maximum(x = var_9635_cast_fp16, y = var_9636_to_fp16)[name = string("s_363_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_363_cast_fp16 = real_div(x = linear_202_cast_fp16, y = s_363_cast_fp16)[name = string("y_363_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9639_cast_fp16 = mul(x = y_363_cast_fp16, y = y_363_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [1]> var_9644_axes_0 = const()[name = string("op_9644_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9644_keep_dims_0 = const()[name = string("op_9644_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9644_cast_fp16 = reduce_mean(axes = var_9644_axes_0, keep_dims = var_9644_keep_dims_0, x = var_9639_cast_fp16)[name = string("op_9644_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9645_cast_fp16 = mul(x = s_363_cast_fp16, y = s_363_cast_fp16)[name = string("op_9645_cast_fp16")];
            fp32 var_9646_epsilon_0 = const()[name = string("op_9646_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9646_cast_fp16 = inverse(epsilon = var_9646_epsilon_0, x = var_9645_cast_fp16)[name = string("op_9646_cast_fp16")];
            fp16 var_9647_to_fp16 = const()[name = string("op_9647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9648_cast_fp16 = mul(x = var_9646_cast_fp16, y = var_9647_to_fp16)[name = string("op_9648_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9650_cast_fp16 = add(x = var_9644_cast_fp16, y = var_9648_cast_fp16)[name = string("op_9650_cast_fp16")];
            fp32 var_9651_epsilon_0 = const()[name = string("op_9651_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9651_cast_fp16 = rsqrt(epsilon = var_9651_epsilon_0, x = var_9650_cast_fp16)[name = string("op_9651_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_317_cast_fp16 = mul(x = y_363_cast_fp16, y = var_9651_cast_fp16)[name = string("out_317_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126949888)))];
            tensor<fp16, [1, 48, 512]> var_9653_cast_fp16 = mul(x = out_317_cast_fp16, y = tm_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_407_cast_fp16 = add(x = x_397_cast_fp16, y = var_9653_cast_fp16)[name = string("x_407_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9656_cast_fp16 = abs(x = x_407_cast_fp16)[name = string("op_9656_cast_fp16")];
            bool var_9659 = const()[name = string("op_9659"), val = bool(true)];
            tensor<int32, [1]> var_9660_axes_0 = const()[name = string("op_9660_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9660_cast_fp16 = reduce_max(axes = var_9660_axes_0, keep_dims = var_9659, x = var_9656_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp16 var_9661_to_fp16 = const()[name = string("op_9661_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_365_cast_fp16 = maximum(x = var_9660_cast_fp16, y = var_9661_to_fp16)[name = string("s_365_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_365_cast_fp16 = real_div(x = x_407_cast_fp16, y = s_365_cast_fp16)[name = string("y_365_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9664_cast_fp16 = mul(x = y_365_cast_fp16, y = y_365_cast_fp16)[name = string("op_9664_cast_fp16")];
            tensor<int32, [1]> var_9669_axes_0 = const()[name = string("op_9669_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9669_keep_dims_0 = const()[name = string("op_9669_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9669_cast_fp16 = reduce_mean(axes = var_9669_axes_0, keep_dims = var_9669_keep_dims_0, x = var_9664_cast_fp16)[name = string("op_9669_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9670_cast_fp16 = mul(x = s_365_cast_fp16, y = s_365_cast_fp16)[name = string("op_9670_cast_fp16")];
            fp32 var_9671_epsilon_0 = const()[name = string("op_9671_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9671_cast_fp16 = inverse(epsilon = var_9671_epsilon_0, x = var_9670_cast_fp16)[name = string("op_9671_cast_fp16")];
            fp16 var_9672_to_fp16 = const()[name = string("op_9672_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9673_cast_fp16 = mul(x = var_9671_cast_fp16, y = var_9672_to_fp16)[name = string("op_9673_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9675_cast_fp16 = add(x = var_9669_cast_fp16, y = var_9673_cast_fp16)[name = string("op_9675_cast_fp16")];
            fp32 var_9676_epsilon_0 = const()[name = string("op_9676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9676_cast_fp16 = rsqrt(epsilon = var_9676_epsilon_0, x = var_9675_cast_fp16)[name = string("op_9676_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_319_cast_fp16 = mul(x = y_365_cast_fp16, y = var_9676_cast_fp16)[name = string("out_319_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126950976)))];
            tensor<fp16, [1, 48, 512]> input_269_cast_fp16 = mul(x = out_319_cast_fp16, y = tm_layers_22_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126952064))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128000704))))[name = string("tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_203_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_203_cast_fp16")];
            string var_9683_mode_0 = const()[name = string("op_9683_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_9683_cast_fp16 = gelu(mode = var_9683_mode_0, x = linear_203_cast_fp16)[name = string("op_9683_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128004864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129053504))))[name = string("tm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_204_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_271_cast_fp16 = mul(x = var_9683_cast_fp16, y = linear_204_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129057664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130106304))))[name = string("tm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9691_cast_fp16 = abs(x = linear_205_cast_fp16)[name = string("op_9691_cast_fp16")];
            bool var_9694 = const()[name = string("op_9694"), val = bool(true)];
            tensor<int32, [1]> var_9695_axes_0 = const()[name = string("op_9695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9695_cast_fp16 = reduce_max(axes = var_9695_axes_0, keep_dims = var_9694, x = var_9691_cast_fp16)[name = string("op_9695_cast_fp16")];
            fp16 var_9696_to_fp16 = const()[name = string("op_9696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_367_cast_fp16 = maximum(x = var_9695_cast_fp16, y = var_9696_to_fp16)[name = string("s_367_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_367_cast_fp16 = real_div(x = linear_205_cast_fp16, y = s_367_cast_fp16)[name = string("y_367_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9699_cast_fp16 = mul(x = y_367_cast_fp16, y = y_367_cast_fp16)[name = string("op_9699_cast_fp16")];
            tensor<int32, [1]> var_9704_axes_0 = const()[name = string("op_9704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9704_keep_dims_0 = const()[name = string("op_9704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9704_cast_fp16 = reduce_mean(axes = var_9704_axes_0, keep_dims = var_9704_keep_dims_0, x = var_9699_cast_fp16)[name = string("op_9704_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9705_cast_fp16 = mul(x = s_367_cast_fp16, y = s_367_cast_fp16)[name = string("op_9705_cast_fp16")];
            fp32 var_9706_epsilon_0 = const()[name = string("op_9706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9706_cast_fp16 = inverse(epsilon = var_9706_epsilon_0, x = var_9705_cast_fp16)[name = string("op_9706_cast_fp16")];
            fp16 var_9707_to_fp16 = const()[name = string("op_9707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9708_cast_fp16 = mul(x = var_9706_cast_fp16, y = var_9707_to_fp16)[name = string("op_9708_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9710_cast_fp16 = add(x = var_9704_cast_fp16, y = var_9708_cast_fp16)[name = string("op_9710_cast_fp16")];
            fp32 var_9711_epsilon_0 = const()[name = string("op_9711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9711_cast_fp16 = rsqrt(epsilon = var_9711_epsilon_0, x = var_9710_cast_fp16)[name = string("op_9711_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_321_cast_fp16 = mul(x = y_367_cast_fp16, y = var_9711_cast_fp16)[name = string("out_321_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130107392)))];
            tensor<fp16, [1, 48, 512]> var_9713_cast_fp16 = mul(x = out_321_cast_fp16, y = tm_layers_22_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9713_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_273_cast_fp16 = add(x = x_407_cast_fp16, y = var_9713_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130108480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130370688))))[name = string("tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_206_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_206_cast_fp16")];
            string x_411_mode_0 = const()[name = string("x_411_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_411_cast_fp16 = gelu(mode = x_411_mode_0, x = linear_206_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [4]> var_9733_begin_0 = const()[name = string("op_9733_begin_0"), val = tensor<int32, [4]>([0, 0, 22, 0])];
            tensor<int32, [4]> var_9733_end_0 = const()[name = string("op_9733_end_0"), val = tensor<int32, [4]>([1, 48, 23, 512])];
            tensor<bool, [4]> var_9733_end_mask_0 = const()[name = string("op_9733_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9733_squeeze_mask_0 = const()[name = string("op_9733_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_9733_cast_fp16 = slice_by_index(begin = var_9733_begin_0, end = var_9733_end_0, end_mask = var_9733_end_mask_0, squeeze_mask = var_9733_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9733_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_275_cast_fp16 = mul(x = x_411_cast_fp16, y = var_9733_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130371776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130633984))))[name = string("tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9743_cast_fp16 = abs(x = linear_207_cast_fp16)[name = string("op_9743_cast_fp16")];
            bool var_9746 = const()[name = string("op_9746"), val = bool(true)];
            tensor<int32, [1]> var_9747_axes_0 = const()[name = string("op_9747_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9747_cast_fp16 = reduce_max(axes = var_9747_axes_0, keep_dims = var_9746, x = var_9743_cast_fp16)[name = string("op_9747_cast_fp16")];
            fp16 var_9748_to_fp16 = const()[name = string("op_9748_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_369_cast_fp16 = maximum(x = var_9747_cast_fp16, y = var_9748_to_fp16)[name = string("s_369_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_369_cast_fp16 = real_div(x = linear_207_cast_fp16, y = s_369_cast_fp16)[name = string("y_369_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9751_cast_fp16 = mul(x = y_369_cast_fp16, y = y_369_cast_fp16)[name = string("op_9751_cast_fp16")];
            tensor<int32, [1]> var_9756_axes_0 = const()[name = string("op_9756_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9756_keep_dims_0 = const()[name = string("op_9756_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9756_cast_fp16 = reduce_mean(axes = var_9756_axes_0, keep_dims = var_9756_keep_dims_0, x = var_9751_cast_fp16)[name = string("op_9756_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9757_cast_fp16 = mul(x = s_369_cast_fp16, y = s_369_cast_fp16)[name = string("op_9757_cast_fp16")];
            fp32 var_9758_epsilon_0 = const()[name = string("op_9758_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9758_cast_fp16 = inverse(epsilon = var_9758_epsilon_0, x = var_9757_cast_fp16)[name = string("op_9758_cast_fp16")];
            fp16 var_9759_to_fp16 = const()[name = string("op_9759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9760_cast_fp16 = mul(x = var_9758_cast_fp16, y = var_9759_to_fp16)[name = string("op_9760_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9762_cast_fp16 = add(x = var_9756_cast_fp16, y = var_9760_cast_fp16)[name = string("op_9762_cast_fp16")];
            fp32 var_9763_epsilon_0 = const()[name = string("op_9763_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9763_cast_fp16 = rsqrt(epsilon = var_9763_epsilon_0, x = var_9762_cast_fp16)[name = string("op_9763_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_323_cast_fp16 = mul(x = y_369_cast_fp16, y = var_9763_cast_fp16)[name = string("out_323_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130635072)))];
            tensor<fp16, [1, 48, 512]> var_9765_cast_fp16 = mul(x = out_323_cast_fp16, y = tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9765_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_45_cast_fp16 = add(x = input_273_cast_fp16, y = var_9765_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<fp16, [1]> tm_layers_22_layer_scalar_to_fp16 = const()[name = string("tm_layers_22_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 48, 512]> x_415_cast_fp16 = mul(x = h_45_cast_fp16, y = tm_layers_22_layer_scalar_to_fp16)[name = string("x_415_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9769_cast_fp16 = abs(x = x_415_cast_fp16)[name = string("op_9769_cast_fp16")];
            bool var_9772 = const()[name = string("op_9772"), val = bool(true)];
            tensor<int32, [1]> var_9773_axes_0 = const()[name = string("op_9773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9773_cast_fp16 = reduce_max(axes = var_9773_axes_0, keep_dims = var_9772, x = var_9769_cast_fp16)[name = string("op_9773_cast_fp16")];
            fp16 var_9774_to_fp16 = const()[name = string("op_9774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_371_cast_fp16 = maximum(x = var_9773_cast_fp16, y = var_9774_to_fp16)[name = string("s_371_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_371_cast_fp16 = real_div(x = x_415_cast_fp16, y = s_371_cast_fp16)[name = string("y_371_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9777_cast_fp16 = mul(x = y_371_cast_fp16, y = y_371_cast_fp16)[name = string("op_9777_cast_fp16")];
            tensor<int32, [1]> var_9782_axes_0 = const()[name = string("op_9782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9782_keep_dims_0 = const()[name = string("op_9782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9782_cast_fp16 = reduce_mean(axes = var_9782_axes_0, keep_dims = var_9782_keep_dims_0, x = var_9777_cast_fp16)[name = string("op_9782_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9783_cast_fp16 = mul(x = s_371_cast_fp16, y = s_371_cast_fp16)[name = string("op_9783_cast_fp16")];
            fp32 var_9784_epsilon_0 = const()[name = string("op_9784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9784_cast_fp16 = inverse(epsilon = var_9784_epsilon_0, x = var_9783_cast_fp16)[name = string("op_9784_cast_fp16")];
            fp16 var_9785_to_fp16 = const()[name = string("op_9785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9786_cast_fp16 = mul(x = var_9784_cast_fp16, y = var_9785_to_fp16)[name = string("op_9786_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9788_cast_fp16 = add(x = var_9782_cast_fp16, y = var_9786_cast_fp16)[name = string("op_9788_cast_fp16")];
            fp32 var_9789_epsilon_0 = const()[name = string("op_9789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9789_cast_fp16 = rsqrt(epsilon = var_9789_epsilon_0, x = var_9788_cast_fp16)[name = string("op_9789_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_325_cast_fp16 = mul(x = y_371_cast_fp16, y = var_9789_cast_fp16)[name = string("out_325_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130636160)))];
            tensor<fp16, [1, 48, 512]> input_277_cast_fp16 = mul(x = out_325_cast_fp16, y = tm_layers_23_input_layernorm_weight_to_fp16)[name = string("input_277_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130637248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131685888))))[name = string("tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_208_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<int32, [4]> var_9799 = const()[name = string("op_9799"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 48, 4, 512]> var_9800_cast_fp16 = reshape(shape = var_9799, x = linear_208_cast_fp16)[name = string("op_9800_cast_fp16")];
            tensor<int32, [4]> x_417_perm_0 = const()[name = string("x_417_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131690048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131952256))))[name = string("tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<int32, [4]> var_9811 = const()[name = string("op_9811"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 48, 1, 512]> var_9812_cast_fp16 = reshape(shape = var_9811, x = linear_209_cast_fp16)[name = string("op_9812_cast_fp16")];
            tensor<int32, [4]> x_419_perm_0 = const()[name = string("x_419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131953344))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132215552))))[name = string("tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<int32, [4]> var_9823 = const()[name = string("op_9823"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 48, 1, 512]> var_9824_cast_fp16 = reshape(shape = var_9823, x = linear_210_cast_fp16)[name = string("op_9824_cast_fp16")];
            tensor<int32, [4]> x_421_perm_0 = const()[name = string("x_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 48, 512]> x_417_cast_fp16 = transpose(perm = x_417_perm_0, x = var_9800_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 4, 48, 512]> var_9828_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_9828_cast_fp16")];
            bool var_9831 = const()[name = string("op_9831"), val = bool(true)];
            tensor<int32, [1]> var_9832_axes_0 = const()[name = string("op_9832_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 48, 1]> var_9832_cast_fp16 = reduce_max(axes = var_9832_axes_0, keep_dims = var_9831, x = var_9828_cast_fp16)[name = string("op_9832_cast_fp16")];
            fp16 var_9833_to_fp16 = const()[name = string("op_9833_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 48, 1]> s_373_cast_fp16 = maximum(x = var_9832_cast_fp16, y = var_9833_to_fp16)[name = string("s_373_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> y_373_cast_fp16 = real_div(x = x_417_cast_fp16, y = s_373_cast_fp16)[name = string("y_373_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_9836_cast_fp16 = mul(x = y_373_cast_fp16, y = y_373_cast_fp16)[name = string("op_9836_cast_fp16")];
            tensor<int32, [1]> var_9841_axes_0 = const()[name = string("op_9841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9841_keep_dims_0 = const()[name = string("op_9841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 48, 1]> var_9841_cast_fp16 = reduce_mean(axes = var_9841_axes_0, keep_dims = var_9841_keep_dims_0, x = var_9836_cast_fp16)[name = string("op_9841_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_9842_cast_fp16 = mul(x = s_373_cast_fp16, y = s_373_cast_fp16)[name = string("op_9842_cast_fp16")];
            fp32 var_9843_epsilon_0 = const()[name = string("op_9843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 48, 1]> var_9843_cast_fp16 = inverse(epsilon = var_9843_epsilon_0, x = var_9842_cast_fp16)[name = string("op_9843_cast_fp16")];
            fp16 var_9844_to_fp16 = const()[name = string("op_9844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 48, 1]> var_9845_cast_fp16 = mul(x = var_9843_cast_fp16, y = var_9844_to_fp16)[name = string("op_9845_cast_fp16")];
            tensor<fp16, [1, 4, 48, 1]> var_9847_cast_fp16 = add(x = var_9841_cast_fp16, y = var_9845_cast_fp16)[name = string("op_9847_cast_fp16")];
            fp32 var_9848_epsilon_0 = const()[name = string("op_9848_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 48, 1]> var_9848_cast_fp16 = rsqrt(epsilon = var_9848_epsilon_0, x = var_9847_cast_fp16)[name = string("op_9848_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> out_327_cast_fp16 = mul(x = y_373_cast_fp16, y = var_9848_cast_fp16)[name = string("out_327_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132216640)))];
            tensor<fp16, [1, 4, 48, 512]> q_93_cast_fp16 = mul(x = out_327_cast_fp16, y = tm_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("q_93_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> x_419_cast_fp16 = transpose(perm = x_419_perm_0, x = var_9812_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1, 48, 512]> var_9851_cast_fp16 = abs(x = x_419_cast_fp16)[name = string("op_9851_cast_fp16")];
            bool var_9854 = const()[name = string("op_9854"), val = bool(true)];
            tensor<int32, [1]> var_9855_axes_0 = const()[name = string("op_9855_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 48, 1]> var_9855_cast_fp16 = reduce_max(axes = var_9855_axes_0, keep_dims = var_9854, x = var_9851_cast_fp16)[name = string("op_9855_cast_fp16")];
            fp16 var_9856_to_fp16 = const()[name = string("op_9856_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 48, 1]> s_375_cast_fp16 = maximum(x = var_9855_cast_fp16, y = var_9856_to_fp16)[name = string("s_375_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> y_375_cast_fp16 = real_div(x = x_419_cast_fp16, y = s_375_cast_fp16)[name = string("y_375_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_9859_cast_fp16 = mul(x = y_375_cast_fp16, y = y_375_cast_fp16)[name = string("op_9859_cast_fp16")];
            tensor<int32, [1]> var_9864_axes_0 = const()[name = string("op_9864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9864_keep_dims_0 = const()[name = string("op_9864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 48, 1]> var_9864_cast_fp16 = reduce_mean(axes = var_9864_axes_0, keep_dims = var_9864_keep_dims_0, x = var_9859_cast_fp16)[name = string("op_9864_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_9865_cast_fp16 = mul(x = s_375_cast_fp16, y = s_375_cast_fp16)[name = string("op_9865_cast_fp16")];
            fp32 var_9866_epsilon_0 = const()[name = string("op_9866_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 48, 1]> var_9866_cast_fp16 = inverse(epsilon = var_9866_epsilon_0, x = var_9865_cast_fp16)[name = string("op_9866_cast_fp16")];
            fp16 var_9867_to_fp16 = const()[name = string("op_9867_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 48, 1]> var_9868_cast_fp16 = mul(x = var_9866_cast_fp16, y = var_9867_to_fp16)[name = string("op_9868_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_9870_cast_fp16 = add(x = var_9864_cast_fp16, y = var_9868_cast_fp16)[name = string("op_9870_cast_fp16")];
            fp32 var_9871_epsilon_0 = const()[name = string("op_9871_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 48, 1]> var_9871_cast_fp16 = rsqrt(epsilon = var_9871_epsilon_0, x = var_9870_cast_fp16)[name = string("op_9871_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> out_329_cast_fp16 = mul(x = y_375_cast_fp16, y = var_9871_cast_fp16)[name = string("out_329_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132217728)))];
            tensor<fp16, [1, 1, 48, 512]> k_139_cast_fp16 = mul(x = out_329_cast_fp16, y = tm_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("k_139_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> x_421_cast_fp16 = transpose(perm = x_421_perm_0, x = var_9824_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 1, 48, 512]> var_9874_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_9874_cast_fp16")];
            bool var_9877 = const()[name = string("op_9877"), val = bool(true)];
            tensor<int32, [1]> var_9878_axes_0 = const()[name = string("op_9878_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 48, 1]> var_9878_cast_fp16 = reduce_max(axes = var_9878_axes_0, keep_dims = var_9877, x = var_9874_cast_fp16)[name = string("op_9878_cast_fp16")];
            fp16 var_9879_to_fp16 = const()[name = string("op_9879_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 48, 1]> s_377_cast_fp16 = maximum(x = var_9878_cast_fp16, y = var_9879_to_fp16)[name = string("s_377_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> y_377_cast_fp16 = real_div(x = x_421_cast_fp16, y = s_377_cast_fp16)[name = string("y_377_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_9882_cast_fp16 = mul(x = y_377_cast_fp16, y = y_377_cast_fp16)[name = string("op_9882_cast_fp16")];
            tensor<int32, [1]> var_9887_axes_0 = const()[name = string("op_9887_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9887_keep_dims_0 = const()[name = string("op_9887_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 48, 1]> var_9887_cast_fp16 = reduce_mean(axes = var_9887_axes_0, keep_dims = var_9887_keep_dims_0, x = var_9882_cast_fp16)[name = string("op_9887_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_9888_cast_fp16 = mul(x = s_377_cast_fp16, y = s_377_cast_fp16)[name = string("op_9888_cast_fp16")];
            fp32 var_9889_epsilon_0 = const()[name = string("op_9889_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 48, 1]> var_9889_cast_fp16 = inverse(epsilon = var_9889_epsilon_0, x = var_9888_cast_fp16)[name = string("op_9889_cast_fp16")];
            fp16 var_9890_to_fp16 = const()[name = string("op_9890_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 48, 1]> var_9891_cast_fp16 = mul(x = var_9889_cast_fp16, y = var_9890_to_fp16)[name = string("op_9891_cast_fp16")];
            tensor<fp16, [1, 1, 48, 1]> var_9893_cast_fp16 = add(x = var_9887_cast_fp16, y = var_9891_cast_fp16)[name = string("op_9893_cast_fp16")];
            fp32 var_9894_epsilon_0 = const()[name = string("op_9894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 48, 1]> var_9894_cast_fp16 = rsqrt(epsilon = var_9894_epsilon_0, x = var_9893_cast_fp16)[name = string("op_9894_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> v_93_cast_fp16 = mul(x = y_377_cast_fp16, y = var_9894_cast_fp16)[name = string("v_93_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_9896_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9896_cast_fp16")];
            tensor<int32, [4]> var_9901_begin_0 = const()[name = string("op_9901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9901_end_0 = const()[name = string("op_9901_end_0"), val = tensor<int32, [4]>([1, 4, 48, 512])];
            tensor<bool, [4]> var_9901_end_mask_0 = const()[name = string("op_9901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 48, 256]> var_9901_cast_fp16 = slice_by_index(begin = var_9901_begin_0, end = var_9901_end_0, end_mask = var_9901_end_mask_0, x = q_93_cast_fp16)[name = string("op_9901_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 48, 256]> var_9902_cast_fp16 = mul(x = var_9901_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_9902_cast_fp16")];
            tensor<int32, [4]> var_9907_begin_0 = const()[name = string("op_9907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9907_end_0 = const()[name = string("op_9907_end_0"), val = tensor<int32, [4]>([1, 4, 48, 256])];
            tensor<bool, [4]> var_9907_end_mask_0 = const()[name = string("op_9907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 48, 256]> var_9907_cast_fp16 = slice_by_index(begin = var_9907_begin_0, end = var_9907_end_0, end_mask = var_9907_end_mask_0, x = q_93_cast_fp16)[name = string("op_9907_cast_fp16")];
            int32 var_9909 = const()[name = string("op_9909"), val = int32(-1)];
            bool var_9910_interleave_0 = const()[name = string("op_9910_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 512]> var_9910_cast_fp16 = concat(axis = var_9909, interleave = var_9910_interleave_0, values = (var_9902_cast_fp16, var_9907_cast_fp16))[name = string("op_9910_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> var_9911_cast_fp16 = mul(x = var_9910_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9911_cast_fp16")];
            tensor<fp16, [1, 4, 48, 512]> q_cast_fp16 = add(x = var_9896_cast_fp16, y = var_9911_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_9914_cast_fp16 = mul(x = k_139_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9914_cast_fp16")];
            tensor<int32, [4]> var_9919_begin_0 = const()[name = string("op_9919_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9919_end_0 = const()[name = string("op_9919_end_0"), val = tensor<int32, [4]>([1, 1, 48, 512])];
            tensor<bool, [4]> var_9919_end_mask_0 = const()[name = string("op_9919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 48, 256]> var_9919_cast_fp16 = slice_by_index(begin = var_9919_begin_0, end = var_9919_end_0, end_mask = var_9919_end_mask_0, x = k_139_cast_fp16)[name = string("op_9919_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 48, 256]> var_9920_cast_fp16 = mul(x = var_9919_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_9920_cast_fp16")];
            tensor<int32, [4]> var_9925_begin_0 = const()[name = string("op_9925_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9925_end_0 = const()[name = string("op_9925_end_0"), val = tensor<int32, [4]>([1, 1, 48, 256])];
            tensor<bool, [4]> var_9925_end_mask_0 = const()[name = string("op_9925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 48, 256]> var_9925_cast_fp16 = slice_by_index(begin = var_9925_begin_0, end = var_9925_end_0, end_mask = var_9925_end_mask_0, x = k_139_cast_fp16)[name = string("op_9925_cast_fp16")];
            int32 var_9927 = const()[name = string("op_9927"), val = int32(-1)];
            bool var_9928_interleave_0 = const()[name = string("op_9928_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 48, 512]> var_9928_cast_fp16 = concat(axis = var_9927, interleave = var_9928_interleave_0, values = (var_9920_cast_fp16, var_9925_cast_fp16))[name = string("op_9928_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> var_9929_cast_fp16 = mul(x = var_9928_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9929_cast_fp16")];
            tensor<fp16, [1, 1, 48, 512]> k_141_cast_fp16 = add(x = var_9914_cast_fp16, y = var_9929_cast_fp16)[name = string("k_141_cast_fp16")];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_46_reps_0 = const()[name = string("tile_46_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 48, 512]> transpose_92_cast_fp16 = transpose(perm = transpose_92_perm_0, x = k_141_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 1, 48, 512]> tile_46_cast_fp16 = tile(reps = tile_46_reps_0, x = transpose_92_cast_fp16)[name = string("tile_46_cast_fp16")];
            tensor<int32, [5]> concat_92 = const()[name = string("concat_92"), val = tensor<int32, [5]>([4, 1, 1, 48, 512])];
            tensor<fp16, [4, 1, 1, 48, 512]> reshape_92_cast_fp16 = reshape(shape = concat_92, x = tile_46_cast_fp16)[name = string("reshape_92_cast_fp16")];
            tensor<int32, [5]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_93 = const()[name = string("concat_93"), val = tensor<int32, [4]>([-1, 1, 48, 512])];
            tensor<fp16, [1, 4, 1, 48, 512]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = reshape_92_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 1, 48, 512]> reshape_93_cast_fp16 = reshape(shape = concat_93, x = transpose_93_cast_fp16)[name = string("reshape_93_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_47_reps_0 = const()[name = string("tile_47_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 48, 512]> transpose_94_cast_fp16 = transpose(perm = transpose_94_perm_0, x = v_93_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 1, 48, 512]> tile_47_cast_fp16 = tile(reps = tile_47_reps_0, x = transpose_94_cast_fp16)[name = string("tile_47_cast_fp16")];
            tensor<int32, [5]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [5]>([4, 1, 1, 48, 512])];
            tensor<fp16, [4, 1, 1, 48, 512]> reshape_94_cast_fp16 = reshape(shape = concat_94, x = tile_47_cast_fp16)[name = string("reshape_94_cast_fp16")];
            tensor<int32, [5]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [4]>([-1, 1, 48, 512])];
            tensor<fp16, [1, 4, 1, 48, 512]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = reshape_94_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 1, 48, 512]> reshape_95_cast_fp16 = reshape(shape = concat_95, x = transpose_95_cast_fp16)[name = string("reshape_95_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9943_transpose_x_0 = const()[name = string("op_9943_transpose_x_0"), val = bool(false)];
            bool var_9943_transpose_y_0 = const()[name = string("op_9943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 48]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = reshape_93_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 4, 48, 48]> var_9943_cast_fp16 = matmul(transpose_x = var_9943_transpose_x_0, transpose_y = var_9943_transpose_y_0, x = q_cast_fp16, y = transpose_119_cast_fp16)[name = string("op_9943_cast_fp16")];
            tensor<fp16, [1, 4, 48, 48]> var_9945_cast_fp16 = add(x = var_9943_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9945_cast_fp16")];
            int32 var_9946 = const()[name = string("op_9946"), val = int32(-1)];
            tensor<fp16, [1, 4, 48, 48]> w_377_cast_fp16 = softmax(axis = var_9946, x = var_9945_cast_fp16)[name = string("w_377_cast_fp16")];
            bool var_9949_transpose_x_0 = const()[name = string("op_9949_transpose_x_0"), val = bool(false)];
            bool var_9949_transpose_y_0 = const()[name = string("op_9949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 48, 512]> v_cast_fp16 = transpose(perm = v_perm_0, x = reshape_95_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 4, 48, 512]> var_9949_cast_fp16 = matmul(transpose_x = var_9949_transpose_x_0, transpose_y = var_9949_transpose_y_0, x = w_377_cast_fp16, y = v_cast_fp16)[name = string("op_9949_cast_fp16")];
            tensor<int32, [4]> var_9952_perm_0 = const()[name = string("op_9952_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9956 = const()[name = string("op_9956"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 48, 4, 512]> var_9952_cast_fp16 = transpose(perm = var_9952_perm_0, x = var_9949_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 48, 2048]> input_279_cast_fp16 = reshape(shape = var_9956, x = var_9952_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132218816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133267456))))[name = string("tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9961_cast_fp16 = abs(x = linear_211_cast_fp16)[name = string("op_9961_cast_fp16")];
            bool var_9964 = const()[name = string("op_9964"), val = bool(true)];
            tensor<int32, [1]> var_9965_axes_0 = const()[name = string("op_9965_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9965_cast_fp16 = reduce_max(axes = var_9965_axes_0, keep_dims = var_9964, x = var_9961_cast_fp16)[name = string("op_9965_cast_fp16")];
            fp16 var_9966_to_fp16 = const()[name = string("op_9966_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_379_cast_fp16 = maximum(x = var_9965_cast_fp16, y = var_9966_to_fp16)[name = string("s_379_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_379_cast_fp16 = real_div(x = linear_211_cast_fp16, y = s_379_cast_fp16)[name = string("y_379_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9969_cast_fp16 = mul(x = y_379_cast_fp16, y = y_379_cast_fp16)[name = string("op_9969_cast_fp16")];
            tensor<int32, [1]> var_9974_axes_0 = const()[name = string("op_9974_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9974_keep_dims_0 = const()[name = string("op_9974_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9974_cast_fp16 = reduce_mean(axes = var_9974_axes_0, keep_dims = var_9974_keep_dims_0, x = var_9969_cast_fp16)[name = string("op_9974_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9975_cast_fp16 = mul(x = s_379_cast_fp16, y = s_379_cast_fp16)[name = string("op_9975_cast_fp16")];
            fp32 var_9976_epsilon_0 = const()[name = string("op_9976_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_9976_cast_fp16 = inverse(epsilon = var_9976_epsilon_0, x = var_9975_cast_fp16)[name = string("op_9976_cast_fp16")];
            fp16 var_9977_to_fp16 = const()[name = string("op_9977_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_9978_cast_fp16 = mul(x = var_9976_cast_fp16, y = var_9977_to_fp16)[name = string("op_9978_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_9980_cast_fp16 = add(x = var_9974_cast_fp16, y = var_9978_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp32 var_9981_epsilon_0 = const()[name = string("op_9981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_9981_cast_fp16 = rsqrt(epsilon = var_9981_epsilon_0, x = var_9980_cast_fp16)[name = string("op_9981_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_331_cast_fp16 = mul(x = y_379_cast_fp16, y = var_9981_cast_fp16)[name = string("out_331_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133268544)))];
            tensor<fp16, [1, 48, 512]> var_9983_cast_fp16 = mul(x = out_331_cast_fp16, y = tm_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("op_9983_cast_fp16")];
            tensor<fp16, [1, 48, 512]> x_425_cast_fp16 = add(x = x_415_cast_fp16, y = var_9983_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9986_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_9986_cast_fp16")];
            bool var_9989 = const()[name = string("op_9989"), val = bool(true)];
            tensor<int32, [1]> var_9990_axes_0 = const()[name = string("op_9990_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_9990_cast_fp16 = reduce_max(axes = var_9990_axes_0, keep_dims = var_9989, x = var_9986_cast_fp16)[name = string("op_9990_cast_fp16")];
            fp16 var_9991_to_fp16 = const()[name = string("op_9991_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_381_cast_fp16 = maximum(x = var_9990_cast_fp16, y = var_9991_to_fp16)[name = string("s_381_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_381_cast_fp16 = real_div(x = x_425_cast_fp16, y = s_381_cast_fp16)[name = string("y_381_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_9994_cast_fp16 = mul(x = y_381_cast_fp16, y = y_381_cast_fp16)[name = string("op_9994_cast_fp16")];
            tensor<int32, [1]> var_9999_axes_0 = const()[name = string("op_9999_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9999_keep_dims_0 = const()[name = string("op_9999_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_9999_cast_fp16 = reduce_mean(axes = var_9999_axes_0, keep_dims = var_9999_keep_dims_0, x = var_9994_cast_fp16)[name = string("op_9999_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_10000_cast_fp16 = mul(x = s_381_cast_fp16, y = s_381_cast_fp16)[name = string("op_10000_cast_fp16")];
            fp32 var_10001_epsilon_0 = const()[name = string("op_10001_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_10001_cast_fp16 = inverse(epsilon = var_10001_epsilon_0, x = var_10000_cast_fp16)[name = string("op_10001_cast_fp16")];
            fp16 var_10002_to_fp16 = const()[name = string("op_10002_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_10003_cast_fp16 = mul(x = var_10001_cast_fp16, y = var_10002_to_fp16)[name = string("op_10003_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_10005_cast_fp16 = add(x = var_9999_cast_fp16, y = var_10003_cast_fp16)[name = string("op_10005_cast_fp16")];
            fp32 var_10006_epsilon_0 = const()[name = string("op_10006_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_10006_cast_fp16 = rsqrt(epsilon = var_10006_epsilon_0, x = var_10005_cast_fp16)[name = string("op_10006_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_333_cast_fp16 = mul(x = y_381_cast_fp16, y = var_10006_cast_fp16)[name = string("out_333_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133269632)))];
            tensor<fp16, [1, 48, 512]> input_281_cast_fp16 = mul(x = out_333_cast_fp16, y = tm_layers_23_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133270720))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134319360))))[name = string("tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_212_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_212_cast_fp16")];
            string var_10013_mode_0 = const()[name = string("op_10013_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 2048]> var_10013_cast_fp16 = gelu(mode = var_10013_mode_0, x = linear_212_cast_fp16)[name = string("op_10013_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134323520))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135372160))))[name = string("tm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 2048]> linear_213_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_213_cast_fp16")];
            tensor<fp16, [1, 48, 2048]> input_283_cast_fp16 = mul(x = var_10013_cast_fp16, y = linear_213_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135376320))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136424960))))[name = string("tm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_283_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_10021_cast_fp16 = abs(x = linear_214_cast_fp16)[name = string("op_10021_cast_fp16")];
            bool var_10024 = const()[name = string("op_10024"), val = bool(true)];
            tensor<int32, [1]> var_10025_axes_0 = const()[name = string("op_10025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_10025_cast_fp16 = reduce_max(axes = var_10025_axes_0, keep_dims = var_10024, x = var_10021_cast_fp16)[name = string("op_10025_cast_fp16")];
            fp16 var_10026_to_fp16 = const()[name = string("op_10026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_383_cast_fp16 = maximum(x = var_10025_cast_fp16, y = var_10026_to_fp16)[name = string("s_383_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_383_cast_fp16 = real_div(x = linear_214_cast_fp16, y = s_383_cast_fp16)[name = string("y_383_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_10029_cast_fp16 = mul(x = y_383_cast_fp16, y = y_383_cast_fp16)[name = string("op_10029_cast_fp16")];
            tensor<int32, [1]> var_10034_axes_0 = const()[name = string("op_10034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10034_keep_dims_0 = const()[name = string("op_10034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_10034_cast_fp16 = reduce_mean(axes = var_10034_axes_0, keep_dims = var_10034_keep_dims_0, x = var_10029_cast_fp16)[name = string("op_10034_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_10035_cast_fp16 = mul(x = s_383_cast_fp16, y = s_383_cast_fp16)[name = string("op_10035_cast_fp16")];
            fp32 var_10036_epsilon_0 = const()[name = string("op_10036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_10036_cast_fp16 = inverse(epsilon = var_10036_epsilon_0, x = var_10035_cast_fp16)[name = string("op_10036_cast_fp16")];
            fp16 var_10037_to_fp16 = const()[name = string("op_10037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_10038_cast_fp16 = mul(x = var_10036_cast_fp16, y = var_10037_to_fp16)[name = string("op_10038_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_10040_cast_fp16 = add(x = var_10034_cast_fp16, y = var_10038_cast_fp16)[name = string("op_10040_cast_fp16")];
            fp32 var_10041_epsilon_0 = const()[name = string("op_10041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_10041_cast_fp16 = rsqrt(epsilon = var_10041_epsilon_0, x = var_10040_cast_fp16)[name = string("op_10041_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_335_cast_fp16 = mul(x = y_383_cast_fp16, y = var_10041_cast_fp16)[name = string("out_335_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136426048)))];
            tensor<fp16, [1, 48, 512]> var_10043_cast_fp16 = mul(x = out_335_cast_fp16, y = tm_layers_23_post_feedforward_layernorm_weight_to_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_285_cast_fp16 = add(x = x_425_cast_fp16, y = var_10043_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136427136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136689344))))[name = string("tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_215_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = string("linear_215_cast_fp16")];
            string x_429_mode_0 = const()[name = string("x_429_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 48, 512]> x_429_cast_fp16 = gelu(mode = x_429_mode_0, x = linear_215_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [4]> var_10063_begin_0 = const()[name = string("op_10063_begin_0"), val = tensor<int32, [4]>([0, 0, 23, 0])];
            tensor<int32, [4]> var_10063_end_0 = const()[name = string("op_10063_end_0"), val = tensor<int32, [4]>([1, 48, 24, 512])];
            tensor<bool, [4]> var_10063_end_mask_0 = const()[name = string("op_10063_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_10063_squeeze_mask_0 = const()[name = string("op_10063_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 48, 512]> var_10063_cast_fp16 = slice_by_index(begin = var_10063_begin_0, end = var_10063_end_0, end_mask = var_10063_end_mask_0, squeeze_mask = var_10063_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 48, 512]> input_287_cast_fp16 = mul(x = x_429_cast_fp16, y = var_10063_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136690432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136952640))))[name = string("tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 48, 512]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_10073_cast_fp16 = abs(x = linear_216_cast_fp16)[name = string("op_10073_cast_fp16")];
            bool var_10076 = const()[name = string("op_10076"), val = bool(true)];
            tensor<int32, [1]> var_10077_axes_0 = const()[name = string("op_10077_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_10077_cast_fp16 = reduce_max(axes = var_10077_axes_0, keep_dims = var_10076, x = var_10073_cast_fp16)[name = string("op_10077_cast_fp16")];
            fp16 var_10078_to_fp16 = const()[name = string("op_10078_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_385_cast_fp16 = maximum(x = var_10077_cast_fp16, y = var_10078_to_fp16)[name = string("s_385_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_385_cast_fp16 = real_div(x = linear_216_cast_fp16, y = s_385_cast_fp16)[name = string("y_385_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_10081_cast_fp16 = mul(x = y_385_cast_fp16, y = y_385_cast_fp16)[name = string("op_10081_cast_fp16")];
            tensor<int32, [1]> var_10086_axes_0 = const()[name = string("op_10086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10086_keep_dims_0 = const()[name = string("op_10086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_10086_cast_fp16 = reduce_mean(axes = var_10086_axes_0, keep_dims = var_10086_keep_dims_0, x = var_10081_cast_fp16)[name = string("op_10086_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_10087_cast_fp16 = mul(x = s_385_cast_fp16, y = s_385_cast_fp16)[name = string("op_10087_cast_fp16")];
            fp32 var_10088_epsilon_0 = const()[name = string("op_10088_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_10088_cast_fp16 = inverse(epsilon = var_10088_epsilon_0, x = var_10087_cast_fp16)[name = string("op_10088_cast_fp16")];
            fp16 var_10089_to_fp16 = const()[name = string("op_10089_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_10090_cast_fp16 = mul(x = var_10088_cast_fp16, y = var_10089_to_fp16)[name = string("op_10090_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_10092_cast_fp16 = add(x = var_10086_cast_fp16, y = var_10090_cast_fp16)[name = string("op_10092_cast_fp16")];
            fp32 var_10093_epsilon_0 = const()[name = string("op_10093_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_10093_cast_fp16 = rsqrt(epsilon = var_10093_epsilon_0, x = var_10092_cast_fp16)[name = string("op_10093_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_337_cast_fp16 = mul(x = y_385_cast_fp16, y = var_10093_cast_fp16)[name = string("out_337_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136953728)))];
            tensor<fp16, [1, 48, 512]> var_10095_cast_fp16 = mul(x = out_337_cast_fp16, y = tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_10095_cast_fp16")];
            tensor<fp16, [1, 48, 512]> h_47_cast_fp16 = add(x = input_285_cast_fp16, y = var_10095_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<fp16, [1]> tm_layers_23_layer_scalar_to_fp16 = const()[name = string("tm_layers_23_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d6p-4])];
            tensor<fp16, [1, 48, 512]> x_cast_fp16 = mul(x = h_47_cast_fp16, y = tm_layers_23_layer_scalar_to_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_10099_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_10099_cast_fp16")];
            bool var_10102 = const()[name = string("op_10102"), val = bool(true)];
            tensor<int32, [1]> var_10103_axes_0 = const()[name = string("op_10103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 48, 1]> var_10103_cast_fp16 = reduce_max(axes = var_10103_axes_0, keep_dims = var_10102, x = var_10099_cast_fp16)[name = string("op_10103_cast_fp16")];
            fp16 var_10104_to_fp16 = const()[name = string("op_10104_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 48, 1]> s_387_cast_fp16 = maximum(x = var_10103_cast_fp16, y = var_10104_to_fp16)[name = string("s_387_cast_fp16")];
            tensor<fp16, [1, 48, 512]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_387_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 48, 512]> var_10107_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_10107_cast_fp16")];
            tensor<int32, [1]> var_10112_axes_0 = const()[name = string("op_10112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10112_keep_dims_0 = const()[name = string("op_10112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 48, 1]> var_10112_cast_fp16 = reduce_mean(axes = var_10112_axes_0, keep_dims = var_10112_keep_dims_0, x = var_10107_cast_fp16)[name = string("op_10112_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_10113_cast_fp16 = mul(x = s_387_cast_fp16, y = s_387_cast_fp16)[name = string("op_10113_cast_fp16")];
            fp32 var_10114_epsilon_0 = const()[name = string("op_10114_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 48, 1]> var_10114_cast_fp16 = inverse(epsilon = var_10114_epsilon_0, x = var_10113_cast_fp16)[name = string("op_10114_cast_fp16")];
            fp16 var_10115_to_fp16 = const()[name = string("op_10115_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 48, 1]> var_10116_cast_fp16 = mul(x = var_10114_cast_fp16, y = var_10115_to_fp16)[name = string("op_10116_cast_fp16")];
            tensor<fp16, [1, 48, 1]> var_10118_cast_fp16 = add(x = var_10112_cast_fp16, y = var_10116_cast_fp16)[name = string("op_10118_cast_fp16")];
            fp32 var_10119_epsilon_0 = const()[name = string("op_10119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 48, 1]> var_10119_cast_fp16 = rsqrt(epsilon = var_10119_epsilon_0, x = var_10118_cast_fp16)[name = string("op_10119_cast_fp16")];
            tensor<fp16, [1, 48, 512]> out_cast_fp16 = mul(x = y_cast_fp16, y = var_10119_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [512]> tm_norm_weight_to_fp16 = const()[name = string("tm_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136954816)))];
            tensor<fp16, [1, 48, 512]> h_cast_fp16 = mul(x = out_cast_fp16, y = tm_norm_weight_to_fp16)[name = string("h_cast_fp16")];
            tensor<int32, [1]> var_10126_axes_0 = const()[name = string("op_10126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10126_keep_dims_0 = const()[name = string("op_10126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10126_cast_fp16 = reduce_sum(axes = var_10126_axes_0, keep_dims = var_10126_keep_dims_0, x = attention_mask)[name = string("op_10126_cast_fp16")];
            tensor<fp16, [1, 48]> var_10127_cast_fp16 = real_div(x = attention_mask, y = var_10126_cast_fp16)[name = string("op_10127_cast_fp16")];
            tensor<int32, [1]> var_10134_axes_0 = const()[name = string("op_10134_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 48]> var_10134_cast_fp16 = expand_dims(axes = var_10134_axes_0, x = var_10127_cast_fp16)[name = string("op_10134_cast_fp16")];
            bool input_transpose_x_0 = const()[name = string("input_transpose_x_0"), val = bool(false)];
            bool input_transpose_y_0 = const()[name = string("input_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512]> input_cast_fp16 = matmul(transpose_x = input_transpose_x_0, transpose_y = input_transpose_y_0, x = var_10134_cast_fp16, y = h_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [768, 512]> tm_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136955904))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137349184))))[name = string("tm_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_217_bias_0_to_fp16 = const()[name = string("linear_217_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137350784)))];
            tensor<fp16, [1, 1, 768]> linear_217_cast_fp16 = linear(bias = linear_217_bias_0_to_fp16, weight = tm_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<int32, [3]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_10151_squeeze_mask_0 = const()[name = string("op_10151_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, squeeze_mask = var_10151_squeeze_mask_0, x = linear_217_cast_fp16)[name = string("op_10151_cast_fp16")];
            tensor<fp16, [1, 768]> var_10157_cast_fp16 = abs(x = var_10151_cast_fp16)[name = string("op_10157_cast_fp16")];
            bool var_10160 = const()[name = string("op_10160"), val = bool(true)];
            tensor<int32, [1]> s_axes_0 = const()[name = string("s_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> s_cast_fp16 = reduce_max(axes = s_axes_0, keep_dims = var_10160, x = var_10157_cast_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [1, 768]> pooled_cast_fp16 = real_div(x = var_10151_cast_fp16, y = s_cast_fp16)[name = string("pooled_cast_fp16")];
            tensor<fp16, [1, 768]> var_10163_cast_fp16 = mul(x = pooled_cast_fp16, y = pooled_cast_fp16)[name = string("op_10163_cast_fp16")];
            tensor<int32, [1]> var_10168_axes_0 = const()[name = string("op_10168_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10168_keep_dims_0 = const()[name = string("op_10168_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10168_cast_fp16 = reduce_sum(axes = var_10168_axes_0, keep_dims = var_10168_keep_dims_0, x = var_10163_cast_fp16)[name = string("op_10168_cast_fp16")];
            fp32 var_10169_epsilon_0 = const()[name = string("op_10169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_10169_cast_fp16 = rsqrt(epsilon = var_10169_epsilon_0, x = var_10168_cast_fp16)[name = string("op_10169_cast_fp16")];
            tensor<fp16, [1, 768]> var_10170_cast_fp16 = mul(x = pooled_cast_fp16, y = var_10169_cast_fp16)[name = string("op_10170_cast_fp16")];
            string var_10170_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_10170_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 768]> embedding = cast(dtype = var_10170_cast_fp16_to_fp32_dtype_0, x = var_10170_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func embed_512<ios18>(tensor<fp16, [1, 512]> attention_mask, tensor<fp16, [1, 512, 512]> inputs_embeds) {
            tensor<fp16, [12288, 512]> tm_ple_per_layer_model_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [12288, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1728))), scale = tensor<fp16, [12288, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6293248))))[name = string("tm_ple_per_layer_model_projection_weight_to_fp16_quantized")];
            tensor<fp16, [12288]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [12288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6317888)))];
            tensor<fp16, [1, 512, 12288]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = tm_ple_per_layer_model_projection_weight_to_fp16_quantized, x = inputs_embeds)[name = string("linear_0_cast_fp16")];
            fp16 var_2129_to_fp16 = const()[name = string("op_2129_to_fp16"), val = fp16(0x1.6ap-5)];
            tensor<fp16, [1, 512, 12288]> ple_1_cast_fp16 = mul(x = linear_0_cast_fp16, y = var_2129_to_fp16)[name = string("ple_1_cast_fp16")];
            tensor<int32, [4]> var_2135 = const()[name = string("op_2135"), val = tensor<int32, [4]>([1, -1, 24, 512])];
            tensor<fp16, [1, 512, 24, 512]> x_1_cast_fp16 = reshape(shape = var_2135, x = ple_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 512, 24, 512]> var_2137_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_2137_cast_fp16")];
            bool var_2140 = const()[name = string("op_2140"), val = bool(true)];
            tensor<int32, [1]> var_2141_axes_0 = const()[name = string("op_2141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 24, 1]> var_2141_cast_fp16 = reduce_max(axes = var_2141_axes_0, keep_dims = var_2140, x = var_2137_cast_fp16)[name = string("op_2141_cast_fp16")];
            fp16 var_2142_to_fp16 = const()[name = string("op_2142_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 24, 1]> s_1_cast_fp16 = maximum(x = var_2141_cast_fp16, y = var_2142_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 512, 24, 512]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 512, 24, 512]> var_2145_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [1]> var_2150_axes_0 = const()[name = string("op_2150_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2150_keep_dims_0 = const()[name = string("op_2150_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 24, 1]> var_2150_cast_fp16 = reduce_mean(axes = var_2150_axes_0, keep_dims = var_2150_keep_dims_0, x = var_2145_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 512, 24, 1]> var_2151_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 24, 1]> var_2152_cast_fp16 = inverse(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            fp16 var_2153_to_fp16 = const()[name = string("op_2153_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 24, 1]> var_2154_cast_fp16 = mul(x = var_2152_cast_fp16, y = var_2153_to_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 512, 24, 1]> var_2156_cast_fp16 = add(x = var_2150_cast_fp16, y = var_2154_cast_fp16)[name = string("op_2156_cast_fp16")];
            fp32 var_2157_epsilon_0 = const()[name = string("op_2157_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 24, 1]> var_2157_cast_fp16 = rsqrt(epsilon = var_2157_epsilon_0, x = var_2156_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<fp16, [1, 512, 24, 512]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_2157_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [512]> tm_ple_per_layer_projection_norm_weight_to_fp16 = const()[name = string("tm_ple_per_layer_projection_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6342528)))];
            tensor<fp16, [1, 512, 24, 512]> ple_3_cast_fp16 = mul(x = out_1_cast_fp16, y = tm_ple_per_layer_projection_norm_weight_to_fp16)[name = string("ple_3_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 512]> var_2162_cast_fp16 = sub(x = var_2160_to_fp16, y = attention_mask)[name = string("op_2162_cast_fp16")];
            fp16 var_2163_to_fp16 = const()[name = string("op_2163_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 512]> var_2164_cast_fp16 = mul(x = var_2162_cast_fp16, y = var_2163_to_fp16)[name = string("op_2164_cast_fp16")];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 512]> var_2171_cast_fp16 = expand_dims(axes = var_2171_axes_0, x = var_2164_cast_fp16)[name = string("op_2171_cast_fp16")];
            tensor<int32, [1]> var_2173_axes_0 = const()[name = string("op_2173_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 512]> var_2173_cast_fp16 = expand_dims(axes = var_2173_axes_0, x = var_2171_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2179_cast_fp16 = abs(x = inputs_embeds)[name = string("op_2179_cast_fp16")];
            bool var_2182 = const()[name = string("op_2182"), val = bool(true)];
            tensor<int32, [1]> var_2183_axes_0 = const()[name = string("op_2183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2183_cast_fp16 = reduce_max(axes = var_2183_axes_0, keep_dims = var_2182, x = var_2179_cast_fp16)[name = string("op_2183_cast_fp16")];
            fp16 var_2184_to_fp16 = const()[name = string("op_2184_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_3_cast_fp16 = maximum(x = var_2183_cast_fp16, y = var_2184_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_3_cast_fp16 = real_div(x = inputs_embeds, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2187_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> var_2192_axes_0 = const()[name = string("op_2192_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2192_keep_dims_0 = const()[name = string("op_2192_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2192_cast_fp16 = reduce_mean(axes = var_2192_axes_0, keep_dims = var_2192_keep_dims_0, x = var_2187_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2193_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp32 var_2194_epsilon_0 = const()[name = string("op_2194_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2194_cast_fp16 = inverse(epsilon = var_2194_epsilon_0, x = var_2193_cast_fp16)[name = string("op_2194_cast_fp16")];
            fp16 var_2195_to_fp16 = const()[name = string("op_2195_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2196_cast_fp16 = mul(x = var_2194_cast_fp16, y = var_2195_to_fp16)[name = string("op_2196_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2198_cast_fp16 = add(x = var_2192_cast_fp16, y = var_2196_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_2199_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6343616)))];
            tensor<fp16, [1, 512, 512]> input_1_cast_fp16 = mul(x = out_3_cast_fp16, y = tm_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6344704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6869056))))[name = string("tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6871168)))];
            tensor<fp16, [1, 512, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_1_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6873280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7135488))))[name = string("tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7136576)))];
            tensor<fp16, [1, 512, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_2221 = const()[name = string("op_2221"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_2222_cast_fp16 = reshape(shape = var_2221, x = linear_2_cast_fp16)[name = string("op_2222_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7137664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7399872))))[name = string("tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_2234_cast_fp16 = reshape(shape = var_2233, x = linear_3_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<int32, [4]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_2210_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 4, 512, 256]> var_2238_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_2238_cast_fp16")];
            bool var_2241 = const()[name = string("op_2241"), val = bool(true)];
            tensor<int32, [1]> var_2242_axes_0 = const()[name = string("op_2242_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_2242_cast_fp16 = reduce_max(axes = var_2242_axes_0, keep_dims = var_2241, x = var_2238_cast_fp16)[name = string("op_2242_cast_fp16")];
            fp16 var_2243_to_fp16 = const()[name = string("op_2243_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_5_cast_fp16 = maximum(x = var_2242_cast_fp16, y = var_2243_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_2246_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<int32, [1]> var_2251_axes_0 = const()[name = string("op_2251_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2251_keep_dims_0 = const()[name = string("op_2251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_2251_cast_fp16 = reduce_mean(axes = var_2251_axes_0, keep_dims = var_2251_keep_dims_0, x = var_2246_cast_fp16)[name = string("op_2251_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_2252_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_2252_cast_fp16")];
            fp32 var_2253_epsilon_0 = const()[name = string("op_2253_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_2253_cast_fp16 = inverse(epsilon = var_2253_epsilon_0, x = var_2252_cast_fp16)[name = string("op_2253_cast_fp16")];
            fp16 var_2254_to_fp16 = const()[name = string("op_2254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_2255_cast_fp16 = mul(x = var_2253_cast_fp16, y = var_2254_to_fp16)[name = string("op_2255_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_2257_cast_fp16 = add(x = var_2251_cast_fp16, y = var_2255_cast_fp16)[name = string("op_2257_cast_fp16")];
            fp32 var_2258_epsilon_0 = const()[name = string("op_2258_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_2258_cast_fp16 = rsqrt(epsilon = var_2258_epsilon_0, x = var_2257_cast_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_2258_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7400960)))];
            tensor<fp16, [1, 4, 512, 256]> q_1_cast_fp16 = mul(x = out_5_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_2222_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 2, 512, 256]> var_2261_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_2261_cast_fp16")];
            bool var_2264 = const()[name = string("op_2264"), val = bool(true)];
            tensor<int32, [1]> var_2265_axes_0 = const()[name = string("op_2265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_2265_cast_fp16 = reduce_max(axes = var_2265_axes_0, keep_dims = var_2264, x = var_2261_cast_fp16)[name = string("op_2265_cast_fp16")];
            fp16 var_2266_to_fp16 = const()[name = string("op_2266_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_7_cast_fp16 = maximum(x = var_2265_cast_fp16, y = var_2266_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2269_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> var_2274_axes_0 = const()[name = string("op_2274_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2274_keep_dims_0 = const()[name = string("op_2274_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_2274_cast_fp16 = reduce_mean(axes = var_2274_axes_0, keep_dims = var_2274_keep_dims_0, x = var_2269_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2275_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_2276_cast_fp16 = inverse(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            fp16 var_2277_to_fp16 = const()[name = string("op_2277_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_2278_cast_fp16 = mul(x = var_2276_cast_fp16, y = var_2277_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2280_cast_fp16 = add(x = var_2274_cast_fp16, y = var_2278_cast_fp16)[name = string("op_2280_cast_fp16")];
            fp32 var_2281_epsilon_0 = const()[name = string("op_2281_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_2281_cast_fp16 = rsqrt(epsilon = var_2281_epsilon_0, x = var_2280_cast_fp16)[name = string("op_2281_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_7_cast_fp16 = mul(x = y_7_cast_fp16, y = var_2281_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7401536)))];
            tensor<fp16, [1, 2, 512, 256]> k_1_cast_fp16 = mul(x = out_7_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = var_2234_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2, 512, 256]> var_2284_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_2284_cast_fp16")];
            bool var_2287 = const()[name = string("op_2287"), val = bool(true)];
            tensor<int32, [1]> var_2288_axes_0 = const()[name = string("op_2288_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_2288_cast_fp16 = reduce_max(axes = var_2288_axes_0, keep_dims = var_2287, x = var_2284_cast_fp16)[name = string("op_2288_cast_fp16")];
            fp16 var_2289_to_fp16 = const()[name = string("op_2289_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_9_cast_fp16 = maximum(x = var_2288_cast_fp16, y = var_2289_to_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_9_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2292_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_2292_cast_fp16")];
            tensor<int32, [1]> var_2297_axes_0 = const()[name = string("op_2297_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2297_keep_dims_0 = const()[name = string("op_2297_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_2297_cast_fp16 = reduce_mean(axes = var_2297_axes_0, keep_dims = var_2297_keep_dims_0, x = var_2292_cast_fp16)[name = string("op_2297_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2298_cast_fp16 = mul(x = s_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_2298_cast_fp16")];
            fp32 var_2299_epsilon_0 = const()[name = string("op_2299_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_2299_cast_fp16 = inverse(epsilon = var_2299_epsilon_0, x = var_2298_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 var_2300_to_fp16 = const()[name = string("op_2300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_2301_cast_fp16 = mul(x = var_2299_cast_fp16, y = var_2300_to_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2303_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2301_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp32 var_2304_epsilon_0 = const()[name = string("op_2304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_2304_cast_fp16 = rsqrt(epsilon = var_2304_epsilon_0, x = var_2303_cast_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_1_cast_fp16 = mul(x = y_9_cast_fp16, y = var_2304_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<fp16, [512, 256]> cos_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137971072))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138102208))))[name = string("cos_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 512, 256]> var_2306_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2306_cast_fp16")];
            tensor<int32, [4]> var_2311_begin_0 = const()[name = string("op_2311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2311_end_0 = const()[name = string("op_2311_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_2311_end_mask_0 = const()[name = string("op_2311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_2311_cast_fp16 = slice_by_index(begin = var_2311_begin_0, end = var_2311_end_0, end_mask = var_2311_end_mask_0, x = q_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = q_1_cast_fp16)[name = string("op_2317_cast_fp16")];
            int32 var_2319 = const()[name = string("op_2319"), val = int32(-1)];
            bool var_2320_interleave_0 = const()[name = string("op_2320_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_2320_cast_fp16 = concat(axis = var_2319, interleave = var_2320_interleave_0, values = (var_2312_cast_fp16, var_2317_cast_fp16))[name = string("op_2320_cast_fp16")];
            tensor<fp16, [512, 256]> sin_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138103296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138234432))))[name = string("sin_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 512, 256]> var_2321_cast_fp16 = mul(x = var_2320_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_3_cast_fp16 = add(x = var_2306_cast_fp16, y = var_2321_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2324_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2329_begin_0 = const()[name = string("op_2329_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2329_end_0 = const()[name = string("op_2329_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_2329_end_mask_0 = const()[name = string("op_2329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, x = k_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_2330_cast_fp16 = mul(x = var_2329_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_2330_cast_fp16")];
            tensor<int32, [4]> var_2335_begin_0 = const()[name = string("op_2335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2335_end_0 = const()[name = string("op_2335_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_2335_end_mask_0 = const()[name = string("op_2335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_2335_cast_fp16 = slice_by_index(begin = var_2335_begin_0, end = var_2335_end_0, end_mask = var_2335_end_mask_0, x = k_1_cast_fp16)[name = string("op_2335_cast_fp16")];
            int32 var_2337 = const()[name = string("op_2337"), val = int32(-1)];
            bool var_2338_interleave_0 = const()[name = string("op_2338_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_2338_cast_fp16 = concat(axis = var_2337, interleave = var_2338_interleave_0, values = (var_2330_cast_fp16, var_2335_cast_fp16))[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2339_cast_fp16 = mul(x = var_2338_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2339_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_3_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2339_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = k_3_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [4, 1, 512, 256]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
            tensor<int32, [5]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_0_cast_fp16 = reshape(shape = concat_0, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [4, 1, 512, 256]> reshape_1_cast_fp16 = reshape(shape = concat_1, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = v_1_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [4, 1, 512, 256]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
            tensor<int32, [5]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_2_cast_fp16 = reshape(shape = concat_2, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [4, 1, 512, 256]> reshape_3_cast_fp16 = reshape(shape = concat_3, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2353_transpose_x_0 = const()[name = string("op_2353_transpose_x_0"), val = bool(false)];
            bool var_2353_transpose_y_0 = const()[name = string("op_2353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 4, 512, 512]> var_2353_cast_fp16 = matmul(transpose_x = var_2353_transpose_x_0, transpose_y = var_2353_transpose_y_0, x = q_3_cast_fp16, y = transpose_96_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_2355_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2356 = const()[name = string("op_2356"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_9_cast_fp16 = softmax(axis = var_2356, x = var_2355_cast_fp16)[name = string("w_9_cast_fp16")];
            bool var_2359_transpose_x_0 = const()[name = string("op_2359_transpose_x_0"), val = bool(false)];
            bool var_2359_transpose_y_0 = const()[name = string("op_2359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 4, 512, 256]> var_2359_cast_fp16 = matmul(transpose_x = var_2359_transpose_x_0, transpose_y = var_2359_transpose_y_0, x = w_9_cast_fp16, y = v_3_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<int32, [4]> var_2362_perm_0 = const()[name = string("op_2362_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2366 = const()[name = string("op_2366"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_2362_cast_fp16 = transpose(perm = var_2362_perm_0, x = var_2359_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 512, 1024]> input_3_cast_fp16 = reshape(shape = var_2366, x = var_2362_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7402112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7926464))))[name = string("tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2371_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_2371_cast_fp16")];
            bool var_2374 = const()[name = string("op_2374"), val = bool(true)];
            tensor<int32, [1]> var_2375_axes_0 = const()[name = string("op_2375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2375_cast_fp16 = reduce_max(axes = var_2375_axes_0, keep_dims = var_2374, x = var_2371_cast_fp16)[name = string("op_2375_cast_fp16")];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_11_cast_fp16 = maximum(x = var_2375_cast_fp16, y = var_2376_to_fp16)[name = string("s_11_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_11_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_11_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2379_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_2379_cast_fp16")];
            tensor<int32, [1]> var_2384_axes_0 = const()[name = string("op_2384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2384_keep_dims_0 = const()[name = string("op_2384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2384_cast_fp16 = reduce_mean(axes = var_2384_axes_0, keep_dims = var_2384_keep_dims_0, x = var_2379_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2385_cast_fp16 = mul(x = s_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2385_cast_fp16")];
            fp32 var_2386_epsilon_0 = const()[name = string("op_2386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2386_cast_fp16 = inverse(epsilon = var_2386_epsilon_0, x = var_2385_cast_fp16)[name = string("op_2386_cast_fp16")];
            fp16 var_2387_to_fp16 = const()[name = string("op_2387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2388_cast_fp16 = mul(x = var_2386_cast_fp16, y = var_2387_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2390_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2388_cast_fp16)[name = string("op_2390_cast_fp16")];
            fp32 var_2391_epsilon_0 = const()[name = string("op_2391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2391_cast_fp16 = rsqrt(epsilon = var_2391_epsilon_0, x = var_2390_cast_fp16)[name = string("op_2391_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_2391_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7927552)))];
            tensor<fp16, [1, 512, 512]> var_2393_cast_fp16 = mul(x = out_9_cast_fp16, y = tm_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_11_cast_fp16 = add(x = inputs_embeds, y = var_2393_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2396_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_2396_cast_fp16")];
            bool var_2399 = const()[name = string("op_2399"), val = bool(true)];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_2399, x = var_2396_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 var_2401_to_fp16 = const()[name = string("op_2401_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_13_cast_fp16 = maximum(x = var_2400_cast_fp16, y = var_2401_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = s_13_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2404_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [1]> var_2409_axes_0 = const()[name = string("op_2409_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2409_keep_dims_0 = const()[name = string("op_2409_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2409_cast_fp16 = reduce_mean(axes = var_2409_axes_0, keep_dims = var_2409_keep_dims_0, x = var_2404_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2410_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2411_cast_fp16 = inverse(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2413_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("op_2413_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2415_cast_fp16 = add(x = var_2409_cast_fp16, y = var_2413_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp32 var_2416_epsilon_0 = const()[name = string("op_2416_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2416_cast_fp16 = rsqrt(epsilon = var_2416_epsilon_0, x = var_2415_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_2416_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7928640)))];
            tensor<fp16, [1, 512, 512]> input_5_cast_fp16 = mul(x = out_11_cast_fp16, y = tm_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7929728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8978368))))[name = string("tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8982528)))];
            tensor<fp16, [1, 512, 2048]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_2423_mode_0 = const()[name = string("op_2423_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_2423_cast_fp16 = gelu(mode = var_2423_mode_0, x = linear_5_cast_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8986688))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10035328))))[name = string("tm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_7_cast_fp16 = mul(x = var_2423_cast_fp16, y = linear_6_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10039488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11088128))))[name = string("tm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2431_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_2431_cast_fp16")];
            bool var_2434 = const()[name = string("op_2434"), val = bool(true)];
            tensor<int32, [1]> var_2435_axes_0 = const()[name = string("op_2435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2435_cast_fp16 = reduce_max(axes = var_2435_axes_0, keep_dims = var_2434, x = var_2431_cast_fp16)[name = string("op_2435_cast_fp16")];
            fp16 var_2436_to_fp16 = const()[name = string("op_2436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_15_cast_fp16 = maximum(x = var_2435_cast_fp16, y = var_2436_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_15_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_15_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2439_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_2439_cast_fp16")];
            tensor<int32, [1]> var_2444_axes_0 = const()[name = string("op_2444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2444_keep_dims_0 = const()[name = string("op_2444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2444_cast_fp16 = reduce_mean(axes = var_2444_axes_0, keep_dims = var_2444_keep_dims_0, x = var_2439_cast_fp16)[name = string("op_2444_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2445_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_2445_cast_fp16")];
            fp32 var_2446_epsilon_0 = const()[name = string("op_2446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2446_cast_fp16 = inverse(epsilon = var_2446_epsilon_0, x = var_2445_cast_fp16)[name = string("op_2446_cast_fp16")];
            fp16 var_2447_to_fp16 = const()[name = string("op_2447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2448_cast_fp16 = mul(x = var_2446_cast_fp16, y = var_2447_to_fp16)[name = string("op_2448_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2450_cast_fp16 = add(x = var_2444_cast_fp16, y = var_2448_cast_fp16)[name = string("op_2450_cast_fp16")];
            fp32 var_2451_epsilon_0 = const()[name = string("op_2451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2451_cast_fp16 = rsqrt(epsilon = var_2451_epsilon_0, x = var_2450_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_2451_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11089216)))];
            tensor<fp16, [1, 512, 512]> var_2453_cast_fp16 = mul(x = out_13_cast_fp16, y = tm_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_9_cast_fp16 = add(x = x_11_cast_fp16, y = var_2453_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11090304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11352512))))[name = string("tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_8_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_8_cast_fp16")];
            string x_15_mode_0 = const()[name = string("x_15_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_15_cast_fp16 = gelu(mode = x_15_mode_0, x = linear_8_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2473_squeeze_mask_0 = const()[name = string("op_2473_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, squeeze_mask = var_2473_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2473_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_11_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2473_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11353600))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11615808))))[name = string("tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2483_cast_fp16 = abs(x = linear_9_cast_fp16)[name = string("op_2483_cast_fp16")];
            bool var_2486 = const()[name = string("op_2486"), val = bool(true)];
            tensor<int32, [1]> var_2487_axes_0 = const()[name = string("op_2487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2487_cast_fp16 = reduce_max(axes = var_2487_axes_0, keep_dims = var_2486, x = var_2483_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp16 var_2488_to_fp16 = const()[name = string("op_2488_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_17_cast_fp16 = maximum(x = var_2487_cast_fp16, y = var_2488_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_17_cast_fp16 = real_div(x = linear_9_cast_fp16, y = s_17_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2491_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_2491_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2496_keep_dims_0 = const()[name = string("op_2496_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2496_cast_fp16 = reduce_mean(axes = var_2496_axes_0, keep_dims = var_2496_keep_dims_0, x = var_2491_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2497_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_2497_cast_fp16")];
            fp32 var_2498_epsilon_0 = const()[name = string("op_2498_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2498_cast_fp16 = inverse(epsilon = var_2498_epsilon_0, x = var_2497_cast_fp16)[name = string("op_2498_cast_fp16")];
            fp16 var_2499_to_fp16 = const()[name = string("op_2499_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2500_cast_fp16 = mul(x = var_2498_cast_fp16, y = var_2499_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2502_cast_fp16 = add(x = var_2496_cast_fp16, y = var_2500_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp32 var_2503_epsilon_0 = const()[name = string("op_2503_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2503_cast_fp16 = rsqrt(epsilon = var_2503_epsilon_0, x = var_2502_cast_fp16)[name = string("op_2503_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_2503_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11616896)))];
            tensor<fp16, [1, 512, 512]> var_2505_cast_fp16 = mul(x = out_15_cast_fp16, y = tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_1_cast_fp16 = add(x = input_9_cast_fp16, y = var_2505_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<fp16, [1]> tm_layers_0_layer_scalar_to_fp16 = const()[name = string("tm_layers_0_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9p-2])];
            tensor<fp16, [1, 512, 512]> x_19_cast_fp16 = mul(x = h_1_cast_fp16, y = tm_layers_0_layer_scalar_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2509_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool var_2512 = const()[name = string("op_2512"), val = bool(true)];
            tensor<int32, [1]> var_2513_axes_0 = const()[name = string("op_2513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2513_cast_fp16 = reduce_max(axes = var_2513_axes_0, keep_dims = var_2512, x = var_2509_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_19_cast_fp16 = maximum(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_19_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2517_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [1]> var_2522_axes_0 = const()[name = string("op_2522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2522_keep_dims_0 = const()[name = string("op_2522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2522_cast_fp16 = reduce_mean(axes = var_2522_axes_0, keep_dims = var_2522_keep_dims_0, x = var_2517_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2523_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_2523_cast_fp16")];
            fp32 var_2524_epsilon_0 = const()[name = string("op_2524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2524_cast_fp16 = inverse(epsilon = var_2524_epsilon_0, x = var_2523_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp16 var_2525_to_fp16 = const()[name = string("op_2525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2526_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_2525_to_fp16)[name = string("op_2526_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2528_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2526_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_2529_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11617984)))];
            tensor<fp16, [1, 512, 512]> input_13_cast_fp16 = mul(x = out_17_cast_fp16, y = tm_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11619072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12143424))))[name = string("tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_2540_cast_fp16 = reshape(shape = var_2539, x = linear_10_cast_fp16)[name = string("op_2540_cast_fp16")];
            tensor<int32, [4]> x_21_perm_0 = const()[name = string("x_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12145536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12407744))))[name = string("tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<int32, [4]> var_2551 = const()[name = string("op_2551"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_2552_cast_fp16 = reshape(shape = var_2551, x = linear_11_cast_fp16)[name = string("op_2552_cast_fp16")];
            tensor<int32, [4]> x_23_perm_0 = const()[name = string("x_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12408832))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671040))))[name = string("tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [4]> var_2563 = const()[name = string("op_2563"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_2564_cast_fp16 = reshape(shape = var_2563, x = linear_12_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = var_2540_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 4, 512, 256]> var_2568_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_2568_cast_fp16")];
            bool var_2571 = const()[name = string("op_2571"), val = bool(true)];
            tensor<int32, [1]> var_2572_axes_0 = const()[name = string("op_2572_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_2572_cast_fp16 = reduce_max(axes = var_2572_axes_0, keep_dims = var_2571, x = var_2568_cast_fp16)[name = string("op_2572_cast_fp16")];
            fp16 var_2573_to_fp16 = const()[name = string("op_2573_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_21_cast_fp16 = maximum(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_21_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_2576_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> var_2581_axes_0 = const()[name = string("op_2581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2581_keep_dims_0 = const()[name = string("op_2581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_2581_cast_fp16 = reduce_mean(axes = var_2581_axes_0, keep_dims = var_2581_keep_dims_0, x = var_2576_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_2582_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp32 var_2583_epsilon_0 = const()[name = string("op_2583_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_2583_cast_fp16 = inverse(epsilon = var_2583_epsilon_0, x = var_2582_cast_fp16)[name = string("op_2583_cast_fp16")];
            fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_2585_cast_fp16 = mul(x = var_2583_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_2587_cast_fp16 = add(x = var_2581_cast_fp16, y = var_2585_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_19_cast_fp16 = mul(x = y_21_cast_fp16, y = var_2588_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672128)))];
            tensor<fp16, [1, 4, 512, 256]> q_5_cast_fp16 = mul(x = out_19_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_23_cast_fp16 = transpose(perm = x_23_perm_0, x = var_2552_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2, 512, 256]> var_2591_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2594 = const()[name = string("op_2594"), val = bool(true)];
            tensor<int32, [1]> var_2595_axes_0 = const()[name = string("op_2595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_2595_cast_fp16 = reduce_max(axes = var_2595_axes_0, keep_dims = var_2594, x = var_2591_cast_fp16)[name = string("op_2595_cast_fp16")];
            fp16 var_2596_to_fp16 = const()[name = string("op_2596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_23_cast_fp16 = maximum(x = var_2595_cast_fp16, y = var_2596_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_23_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2599_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2599_cast_fp16")];
            tensor<int32, [1]> var_2604_axes_0 = const()[name = string("op_2604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2604_keep_dims_0 = const()[name = string("op_2604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_2604_cast_fp16 = reduce_mean(axes = var_2604_axes_0, keep_dims = var_2604_keep_dims_0, x = var_2599_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2605_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_2605_cast_fp16")];
            fp32 var_2606_epsilon_0 = const()[name = string("op_2606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_2606_cast_fp16 = inverse(epsilon = var_2606_epsilon_0, x = var_2605_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp16 var_2607_to_fp16 = const()[name = string("op_2607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_2608_cast_fp16 = mul(x = var_2606_cast_fp16, y = var_2607_to_fp16)[name = string("op_2608_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2610_cast_fp16 = add(x = var_2604_cast_fp16, y = var_2608_cast_fp16)[name = string("op_2610_cast_fp16")];
            fp32 var_2611_epsilon_0 = const()[name = string("op_2611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_2611_cast_fp16 = rsqrt(epsilon = var_2611_epsilon_0, x = var_2610_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_21_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2611_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672704)))];
            tensor<fp16, [1, 2, 512, 256]> k_7_cast_fp16 = mul(x = out_21_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_2564_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 2, 512, 256]> var_2614_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_2614_cast_fp16")];
            bool var_2617 = const()[name = string("op_2617"), val = bool(true)];
            tensor<int32, [1]> var_2618_axes_0 = const()[name = string("op_2618_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_2618_cast_fp16 = reduce_max(axes = var_2618_axes_0, keep_dims = var_2617, x = var_2614_cast_fp16)[name = string("op_2618_cast_fp16")];
            fp16 var_2619_to_fp16 = const()[name = string("op_2619_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_25_cast_fp16 = maximum(x = var_2618_cast_fp16, y = var_2619_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_25_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2622_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2622_cast_fp16")];
            tensor<int32, [1]> var_2627_axes_0 = const()[name = string("op_2627_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2627_keep_dims_0 = const()[name = string("op_2627_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_2627_cast_fp16 = reduce_mean(axes = var_2627_axes_0, keep_dims = var_2627_keep_dims_0, x = var_2622_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2628_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_2628_cast_fp16")];
            fp32 var_2629_epsilon_0 = const()[name = string("op_2629_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_2629_cast_fp16 = inverse(epsilon = var_2629_epsilon_0, x = var_2628_cast_fp16)[name = string("op_2629_cast_fp16")];
            fp16 var_2630_to_fp16 = const()[name = string("op_2630_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_2631_cast_fp16 = mul(x = var_2629_cast_fp16, y = var_2630_to_fp16)[name = string("op_2631_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2633_cast_fp16 = add(x = var_2627_cast_fp16, y = var_2631_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_2634_cast_fp16 = rsqrt(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_5_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2634_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_2636_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2636_cast_fp16")];
            tensor<int32, [4]> var_2641_begin_0 = const()[name = string("op_2641_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2641_end_0 = const()[name = string("op_2641_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_2641_end_mask_0 = const()[name = string("op_2641_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_2641_cast_fp16 = slice_by_index(begin = var_2641_begin_0, end = var_2641_end_0, end_mask = var_2641_end_mask_0, x = q_5_cast_fp16)[name = string("op_2641_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_2642_cast_fp16 = mul(x = var_2641_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_2642_cast_fp16")];
            tensor<int32, [4]> var_2647_begin_0 = const()[name = string("op_2647_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2647_end_0 = const()[name = string("op_2647_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_2647_end_mask_0 = const()[name = string("op_2647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_2647_cast_fp16 = slice_by_index(begin = var_2647_begin_0, end = var_2647_end_0, end_mask = var_2647_end_mask_0, x = q_5_cast_fp16)[name = string("op_2647_cast_fp16")];
            int32 var_2649 = const()[name = string("op_2649"), val = int32(-1)];
            bool var_2650_interleave_0 = const()[name = string("op_2650_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_2650_cast_fp16 = concat(axis = var_2649, interleave = var_2650_interleave_0, values = (var_2642_cast_fp16, var_2647_cast_fp16))[name = string("op_2650_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2651_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_7_cast_fp16 = add(x = var_2636_cast_fp16, y = var_2651_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2654_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = k_7_cast_fp16)[name = string("op_2659_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_2660_cast_fp16 = mul(x = var_2659_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2660_cast_fp16")];
            tensor<int32, [4]> var_2665_begin_0 = const()[name = string("op_2665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2665_end_0 = const()[name = string("op_2665_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_2665_end_mask_0 = const()[name = string("op_2665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_2665_cast_fp16 = slice_by_index(begin = var_2665_begin_0, end = var_2665_end_0, end_mask = var_2665_end_mask_0, x = k_7_cast_fp16)[name = string("op_2665_cast_fp16")];
            int32 var_2667 = const()[name = string("op_2667"), val = int32(-1)];
            bool var_2668_interleave_0 = const()[name = string("op_2668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_2668_cast_fp16 = concat(axis = var_2667, interleave = var_2668_interleave_0, values = (var_2660_cast_fp16, var_2665_cast_fp16))[name = string("op_2668_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2669_cast_fp16 = mul(x = var_2668_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_9_cast_fp16 = add(x = var_2654_cast_fp16, y = var_2669_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = k_9_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [4, 1, 512, 256]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_4_cast_fp16 = reshape(shape = concat_4, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [4, 1, 512, 256]> reshape_5_cast_fp16 = reshape(shape = concat_5, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = v_5_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [4, 1, 512, 256]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_6_cast_fp16 = reshape(shape = concat_6, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [4, 1, 512, 256]> reshape_7_cast_fp16 = reshape(shape = concat_7, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2683_transpose_x_0 = const()[name = string("op_2683_transpose_x_0"), val = bool(false)];
            bool var_2683_transpose_y_0 = const()[name = string("op_2683_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_97_cast_fp16 = transpose(perm = transpose_97_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 4, 512, 512]> var_2683_cast_fp16 = matmul(transpose_x = var_2683_transpose_x_0, transpose_y = var_2683_transpose_y_0, x = q_7_cast_fp16, y = transpose_97_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_2685_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2685_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_25_cast_fp16 = softmax(axis = var_2686, x = var_2685_cast_fp16)[name = string("w_25_cast_fp16")];
            bool var_2689_transpose_x_0 = const()[name = string("op_2689_transpose_x_0"), val = bool(false)];
            bool var_2689_transpose_y_0 = const()[name = string("op_2689_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 4, 512, 256]> var_2689_cast_fp16 = matmul(transpose_x = var_2689_transpose_x_0, transpose_y = var_2689_transpose_y_0, x = w_25_cast_fp16, y = v_7_cast_fp16)[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> var_2692_perm_0 = const()[name = string("op_2692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2696 = const()[name = string("op_2696"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_2692_cast_fp16 = transpose(perm = var_2692_perm_0, x = var_2689_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 512, 1024]> input_15_cast_fp16 = reshape(shape = var_2696, x = var_2692_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13197632))))[name = string("tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2701_cast_fp16 = abs(x = linear_13_cast_fp16)[name = string("op_2701_cast_fp16")];
            bool var_2704 = const()[name = string("op_2704"), val = bool(true)];
            tensor<int32, [1]> var_2705_axes_0 = const()[name = string("op_2705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2705_cast_fp16 = reduce_max(axes = var_2705_axes_0, keep_dims = var_2704, x = var_2701_cast_fp16)[name = string("op_2705_cast_fp16")];
            fp16 var_2706_to_fp16 = const()[name = string("op_2706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_27_cast_fp16 = maximum(x = var_2705_cast_fp16, y = var_2706_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_27_cast_fp16 = real_div(x = linear_13_cast_fp16, y = s_27_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2709_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2709_cast_fp16")];
            tensor<int32, [1]> var_2714_axes_0 = const()[name = string("op_2714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2714_keep_dims_0 = const()[name = string("op_2714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2714_cast_fp16 = reduce_mean(axes = var_2714_axes_0, keep_dims = var_2714_keep_dims_0, x = var_2709_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2715_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2715_cast_fp16")];
            fp32 var_2716_epsilon_0 = const()[name = string("op_2716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2716_cast_fp16 = inverse(epsilon = var_2716_epsilon_0, x = var_2715_cast_fp16)[name = string("op_2716_cast_fp16")];
            fp16 var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2718_cast_fp16 = mul(x = var_2716_cast_fp16, y = var_2717_to_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2720_cast_fp16 = add(x = var_2714_cast_fp16, y = var_2718_cast_fp16)[name = string("op_2720_cast_fp16")];
            fp32 var_2721_epsilon_0 = const()[name = string("op_2721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2721_cast_fp16 = rsqrt(epsilon = var_2721_epsilon_0, x = var_2720_cast_fp16)[name = string("op_2721_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2721_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13198720)))];
            tensor<fp16, [1, 512, 512]> var_2723_cast_fp16 = mul(x = out_23_cast_fp16, y = tm_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2723_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_29_cast_fp16 = add(x = x_19_cast_fp16, y = var_2723_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2726_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_2726_cast_fp16")];
            bool var_2729 = const()[name = string("op_2729"), val = bool(true)];
            tensor<int32, [1]> var_2730_axes_0 = const()[name = string("op_2730_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2730_cast_fp16 = reduce_max(axes = var_2730_axes_0, keep_dims = var_2729, x = var_2726_cast_fp16)[name = string("op_2730_cast_fp16")];
            fp16 var_2731_to_fp16 = const()[name = string("op_2731_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_29_cast_fp16 = maximum(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_29_cast_fp16 = real_div(x = x_29_cast_fp16, y = s_29_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2734_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2734_cast_fp16")];
            tensor<int32, [1]> var_2739_axes_0 = const()[name = string("op_2739_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2739_keep_dims_0 = const()[name = string("op_2739_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2739_cast_fp16 = reduce_mean(axes = var_2739_axes_0, keep_dims = var_2739_keep_dims_0, x = var_2734_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2740_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2740_cast_fp16")];
            fp32 var_2741_epsilon_0 = const()[name = string("op_2741_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2741_cast_fp16 = inverse(epsilon = var_2741_epsilon_0, x = var_2740_cast_fp16)[name = string("op_2741_cast_fp16")];
            fp16 var_2742_to_fp16 = const()[name = string("op_2742_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2743_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_2742_to_fp16)[name = string("op_2743_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2745_cast_fp16 = add(x = var_2739_cast_fp16, y = var_2743_cast_fp16)[name = string("op_2745_cast_fp16")];
            fp32 var_2746_epsilon_0 = const()[name = string("op_2746_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2746_cast_fp16 = rsqrt(epsilon = var_2746_epsilon_0, x = var_2745_cast_fp16)[name = string("op_2746_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2746_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13199808)))];
            tensor<fp16, [1, 512, 512]> input_17_cast_fp16 = mul(x = out_25_cast_fp16, y = tm_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13200896))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14249536))))[name = string("tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_14_cast_fp16")];
            string var_2753_mode_0 = const()[name = string("op_2753_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_2753_cast_fp16 = gelu(mode = var_2753_mode_0, x = linear_14_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14253696))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15302336))))[name = string("tm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_15_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_19_cast_fp16 = mul(x = var_2753_cast_fp16, y = linear_15_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15306496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16355136))))[name = string("tm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2761_cast_fp16 = abs(x = linear_16_cast_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2764 = const()[name = string("op_2764"), val = bool(true)];
            tensor<int32, [1]> var_2765_axes_0 = const()[name = string("op_2765_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2765_cast_fp16 = reduce_max(axes = var_2765_axes_0, keep_dims = var_2764, x = var_2761_cast_fp16)[name = string("op_2765_cast_fp16")];
            fp16 var_2766_to_fp16 = const()[name = string("op_2766_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_31_cast_fp16 = maximum(x = var_2765_cast_fp16, y = var_2766_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_31_cast_fp16 = real_div(x = linear_16_cast_fp16, y = s_31_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2769_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [1]> var_2774_axes_0 = const()[name = string("op_2774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2774_keep_dims_0 = const()[name = string("op_2774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2774_cast_fp16 = reduce_mean(axes = var_2774_axes_0, keep_dims = var_2774_keep_dims_0, x = var_2769_cast_fp16)[name = string("op_2774_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2775_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp32 var_2776_epsilon_0 = const()[name = string("op_2776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2776_cast_fp16 = inverse(epsilon = var_2776_epsilon_0, x = var_2775_cast_fp16)[name = string("op_2776_cast_fp16")];
            fp16 var_2777_to_fp16 = const()[name = string("op_2777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2778_cast_fp16 = mul(x = var_2776_cast_fp16, y = var_2777_to_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2780_cast_fp16 = add(x = var_2774_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2780_cast_fp16")];
            fp32 var_2781_epsilon_0 = const()[name = string("op_2781_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2781_cast_fp16 = rsqrt(epsilon = var_2781_epsilon_0, x = var_2780_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2781_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16356224)))];
            tensor<fp16, [1, 512, 512]> var_2783_cast_fp16 = mul(x = out_27_cast_fp16, y = tm_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2783_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_21_cast_fp16 = add(x = x_29_cast_fp16, y = var_2783_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16357312))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16619520))))[name = string("tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_17_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            string x_33_mode_0 = const()[name = string("x_33_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_33_cast_fp16 = gelu(mode = x_33_mode_0, x = linear_17_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> var_2803_begin_0 = const()[name = string("op_2803_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2803_end_0 = const()[name = string("op_2803_end_0"), val = tensor<int32, [4]>([1, 512, 2, 512])];
            tensor<bool, [4]> var_2803_end_mask_0 = const()[name = string("op_2803_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2803_squeeze_mask_0 = const()[name = string("op_2803_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_2803_cast_fp16 = slice_by_index(begin = var_2803_begin_0, end = var_2803_end_0, end_mask = var_2803_end_mask_0, squeeze_mask = var_2803_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_23_cast_fp16 = mul(x = x_33_cast_fp16, y = var_2803_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16620608))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16882816))))[name = string("tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2813_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2813_cast_fp16")];
            bool var_2816 = const()[name = string("op_2816"), val = bool(true)];
            tensor<int32, [1]> var_2817_axes_0 = const()[name = string("op_2817_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2817_cast_fp16 = reduce_max(axes = var_2817_axes_0, keep_dims = var_2816, x = var_2813_cast_fp16)[name = string("op_2817_cast_fp16")];
            fp16 var_2818_to_fp16 = const()[name = string("op_2818_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_33_cast_fp16 = maximum(x = var_2817_cast_fp16, y = var_2818_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_33_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_33_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2821_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<int32, [1]> var_2826_axes_0 = const()[name = string("op_2826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2826_keep_dims_0 = const()[name = string("op_2826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2826_cast_fp16 = reduce_mean(axes = var_2826_axes_0, keep_dims = var_2826_keep_dims_0, x = var_2821_cast_fp16)[name = string("op_2826_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2827_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp32 var_2828_epsilon_0 = const()[name = string("op_2828_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2828_cast_fp16 = inverse(epsilon = var_2828_epsilon_0, x = var_2827_cast_fp16)[name = string("op_2828_cast_fp16")];
            fp16 var_2829_to_fp16 = const()[name = string("op_2829_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2830_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_2829_to_fp16)[name = string("op_2830_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2832_cast_fp16 = add(x = var_2826_cast_fp16, y = var_2830_cast_fp16)[name = string("op_2832_cast_fp16")];
            fp32 var_2833_epsilon_0 = const()[name = string("op_2833_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2833_cast_fp16 = rsqrt(epsilon = var_2833_epsilon_0, x = var_2832_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2833_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16883904)))];
            tensor<fp16, [1, 512, 512]> var_2835_cast_fp16 = mul(x = out_29_cast_fp16, y = tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_3_cast_fp16 = add(x = input_21_cast_fp16, y = var_2835_cast_fp16)[name = string("h_3_cast_fp16")];
            tensor<fp16, [1]> tm_layers_1_layer_scalar_to_fp16 = const()[name = string("tm_layers_1_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.5ap-3])];
            tensor<fp16, [1, 512, 512]> x_37_cast_fp16 = mul(x = h_3_cast_fp16, y = tm_layers_1_layer_scalar_to_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2839_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2839_cast_fp16")];
            bool var_2842 = const()[name = string("op_2842"), val = bool(true)];
            tensor<int32, [1]> var_2843_axes_0 = const()[name = string("op_2843_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2843_cast_fp16 = reduce_max(axes = var_2843_axes_0, keep_dims = var_2842, x = var_2839_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp16 var_2844_to_fp16 = const()[name = string("op_2844_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_35_cast_fp16 = maximum(x = var_2843_cast_fp16, y = var_2844_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_35_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_35_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_2847_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2847_cast_fp16")];
            tensor<int32, [1]> var_2852_axes_0 = const()[name = string("op_2852_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2852_keep_dims_0 = const()[name = string("op_2852_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_2852_cast_fp16 = reduce_mean(axes = var_2852_axes_0, keep_dims = var_2852_keep_dims_0, x = var_2847_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2853_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp32 var_2854_epsilon_0 = const()[name = string("op_2854_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2854_cast_fp16 = inverse(epsilon = var_2854_epsilon_0, x = var_2853_cast_fp16)[name = string("op_2854_cast_fp16")];
            fp16 var_2855_to_fp16 = const()[name = string("op_2855_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2856_cast_fp16 = mul(x = var_2854_cast_fp16, y = var_2855_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2858_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2856_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp32 var_2859_epsilon_0 = const()[name = string("op_2859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2859_cast_fp16 = rsqrt(epsilon = var_2859_epsilon_0, x = var_2858_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_31_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2859_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16884992)))];
            tensor<fp16, [1, 512, 512]> input_25_cast_fp16 = mul(x = out_31_cast_fp16, y = tm_layers_2_input_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16886080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17410432))))[name = string("tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [4]> var_2869 = const()[name = string("op_2869"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_2870_cast_fp16 = reshape(shape = var_2869, x = linear_19_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17412544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17674752))))[name = string("tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [4]> var_2881 = const()[name = string("op_2881"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_2882_cast_fp16 = reshape(shape = var_2881, x = linear_20_cast_fp16)[name = string("op_2882_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17675840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17938048))))[name = string("tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_2894_cast_fp16 = reshape(shape = var_2893, x = linear_21_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [4]> x_43_perm_0 = const()[name = string("x_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2870_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 4, 512, 256]> var_2898_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2898_cast_fp16")];
            bool var_2901 = const()[name = string("op_2901"), val = bool(true)];
            tensor<int32, [1]> var_2902_axes_0 = const()[name = string("op_2902_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_2902_cast_fp16 = reduce_max(axes = var_2902_axes_0, keep_dims = var_2901, x = var_2898_cast_fp16)[name = string("op_2902_cast_fp16")];
            fp16 var_2903_to_fp16 = const()[name = string("op_2903_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_37_cast_fp16 = maximum(x = var_2902_cast_fp16, y = var_2903_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_37_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_37_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_2906_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<int32, [1]> var_2911_axes_0 = const()[name = string("op_2911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2911_keep_dims_0 = const()[name = string("op_2911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_2911_cast_fp16 = reduce_mean(axes = var_2911_axes_0, keep_dims = var_2911_keep_dims_0, x = var_2906_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_2912_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2912_cast_fp16")];
            fp32 var_2913_epsilon_0 = const()[name = string("op_2913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_2913_cast_fp16 = inverse(epsilon = var_2913_epsilon_0, x = var_2912_cast_fp16)[name = string("op_2913_cast_fp16")];
            fp16 var_2914_to_fp16 = const()[name = string("op_2914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_2915_cast_fp16 = mul(x = var_2913_cast_fp16, y = var_2914_to_fp16)[name = string("op_2915_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_2917_cast_fp16 = add(x = var_2911_cast_fp16, y = var_2915_cast_fp16)[name = string("op_2917_cast_fp16")];
            fp32 var_2918_epsilon_0 = const()[name = string("op_2918_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_2918_cast_fp16 = rsqrt(epsilon = var_2918_epsilon_0, x = var_2917_cast_fp16)[name = string("op_2918_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_33_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2918_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_9_cast_fp16 = mul(x = out_33_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2882_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 2, 512, 256]> var_2921_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2921_cast_fp16")];
            bool var_2924 = const()[name = string("op_2924"), val = bool(true)];
            tensor<int32, [1]> var_2925_axes_0 = const()[name = string("op_2925_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_2925_cast_fp16 = reduce_max(axes = var_2925_axes_0, keep_dims = var_2924, x = var_2921_cast_fp16)[name = string("op_2925_cast_fp16")];
            fp16 var_2926_to_fp16 = const()[name = string("op_2926_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_39_cast_fp16 = maximum(x = var_2925_cast_fp16, y = var_2926_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_39_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_39_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2929_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2929_cast_fp16")];
            tensor<int32, [1]> var_2934_axes_0 = const()[name = string("op_2934_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2934_keep_dims_0 = const()[name = string("op_2934_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_2934_cast_fp16 = reduce_mean(axes = var_2934_axes_0, keep_dims = var_2934_keep_dims_0, x = var_2929_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2935_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2935_cast_fp16")];
            fp32 var_2936_epsilon_0 = const()[name = string("op_2936_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_2936_cast_fp16 = inverse(epsilon = var_2936_epsilon_0, x = var_2935_cast_fp16)[name = string("op_2936_cast_fp16")];
            fp16 var_2937_to_fp16 = const()[name = string("op_2937_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_2938_cast_fp16 = mul(x = var_2936_cast_fp16, y = var_2937_to_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2940_cast_fp16 = add(x = var_2934_cast_fp16, y = var_2938_cast_fp16)[name = string("op_2940_cast_fp16")];
            fp32 var_2941_epsilon_0 = const()[name = string("op_2941_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_2941_cast_fp16 = rsqrt(epsilon = var_2941_epsilon_0, x = var_2940_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_35_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2941_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_13_cast_fp16 = mul(x = out_35_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_43_cast_fp16 = transpose(perm = x_43_perm_0, x = var_2894_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 2, 512, 256]> var_2944_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2944_cast_fp16")];
            bool var_2947 = const()[name = string("op_2947"), val = bool(true)];
            tensor<int32, [1]> var_2948_axes_0 = const()[name = string("op_2948_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_2948_cast_fp16 = reduce_max(axes = var_2948_axes_0, keep_dims = var_2947, x = var_2944_cast_fp16)[name = string("op_2948_cast_fp16")];
            fp16 var_2949_to_fp16 = const()[name = string("op_2949_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_41_cast_fp16 = maximum(x = var_2948_cast_fp16, y = var_2949_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_41_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_41_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2952_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2952_cast_fp16")];
            tensor<int32, [1]> var_2957_axes_0 = const()[name = string("op_2957_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2957_keep_dims_0 = const()[name = string("op_2957_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_2957_cast_fp16 = reduce_mean(axes = var_2957_axes_0, keep_dims = var_2957_keep_dims_0, x = var_2952_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2958_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2958_cast_fp16")];
            fp32 var_2959_epsilon_0 = const()[name = string("op_2959_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_2959_cast_fp16 = inverse(epsilon = var_2959_epsilon_0, x = var_2958_cast_fp16)[name = string("op_2959_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_2961_cast_fp16 = mul(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_2963_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2961_cast_fp16)[name = string("op_2963_cast_fp16")];
            fp32 var_2964_epsilon_0 = const()[name = string("op_2964_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_2964_cast_fp16 = rsqrt(epsilon = var_2964_epsilon_0, x = var_2963_cast_fp16)[name = string("op_2964_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_9_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2964_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_2966_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2966_cast_fp16")];
            tensor<int32, [4]> var_2971_begin_0 = const()[name = string("op_2971_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2971_end_0 = const()[name = string("op_2971_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_2971_end_mask_0 = const()[name = string("op_2971_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_2971_cast_fp16 = slice_by_index(begin = var_2971_begin_0, end = var_2971_end_0, end_mask = var_2971_end_mask_0, x = q_9_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_2972_cast_fp16 = mul(x = var_2971_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2972_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            int32 var_2979 = const()[name = string("op_2979"), val = int32(-1)];
            bool var_2980_interleave_0 = const()[name = string("op_2980_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_2980_cast_fp16 = concat(axis = var_2979, interleave = var_2980_interleave_0, values = (var_2972_cast_fp16, var_2977_cast_fp16))[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_2981_cast_fp16 = mul(x = var_2980_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2981_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_11_cast_fp16 = add(x = var_2966_cast_fp16, y = var_2981_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2984_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2989_begin_0 = const()[name = string("op_2989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2989_end_0 = const()[name = string("op_2989_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_2989_end_mask_0 = const()[name = string("op_2989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_2989_cast_fp16 = slice_by_index(begin = var_2989_begin_0, end = var_2989_end_0, end_mask = var_2989_end_mask_0, x = k_13_cast_fp16)[name = string("op_2989_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_2990_cast_fp16 = mul(x = var_2989_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2990_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = k_13_cast_fp16)[name = string("op_2995_cast_fp16")];
            int32 var_2997 = const()[name = string("op_2997"), val = int32(-1)];
            bool var_2998_interleave_0 = const()[name = string("op_2998_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_2998_cast_fp16 = concat(axis = var_2997, interleave = var_2998_interleave_0, values = (var_2990_cast_fp16, var_2995_cast_fp16))[name = string("op_2998_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_2999_cast_fp16 = mul(x = var_2998_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2999_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_15_cast_fp16 = add(x = var_2984_cast_fp16, y = var_2999_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = k_15_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [4, 1, 512, 256]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
            tensor<int32, [5]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_8_cast_fp16 = reshape(shape = concat_8, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [4, 1, 512, 256]> reshape_9_cast_fp16 = reshape(shape = concat_9, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = v_9_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [4, 1, 512, 256]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
            tensor<int32, [5]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_10_cast_fp16 = reshape(shape = concat_10, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [4, 1, 512, 256]> reshape_11_cast_fp16 = reshape(shape = concat_11, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3013_transpose_x_0 = const()[name = string("op_3013_transpose_x_0"), val = bool(false)];
            bool var_3013_transpose_y_0 = const()[name = string("op_3013_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 4, 512, 512]> var_3013_cast_fp16 = matmul(transpose_x = var_3013_transpose_x_0, transpose_y = var_3013_transpose_y_0, x = q_11_cast_fp16, y = transpose_98_cast_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_3015_cast_fp16 = add(x = var_3013_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3015_cast_fp16")];
            int32 var_3016 = const()[name = string("op_3016"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_41_cast_fp16 = softmax(axis = var_3016, x = var_3015_cast_fp16)[name = string("w_41_cast_fp16")];
            bool var_3019_transpose_x_0 = const()[name = string("op_3019_transpose_x_0"), val = bool(false)];
            bool var_3019_transpose_y_0 = const()[name = string("op_3019_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 4, 512, 256]> var_3019_cast_fp16 = matmul(transpose_x = var_3019_transpose_x_0, transpose_y = var_3019_transpose_y_0, x = w_41_cast_fp16, y = v_11_cast_fp16)[name = string("op_3019_cast_fp16")];
            tensor<int32, [4]> var_3022_perm_0 = const()[name = string("op_3022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3026 = const()[name = string("op_3026"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_3022_cast_fp16 = transpose(perm = var_3022_perm_0, x = var_3019_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 512, 1024]> input_27_cast_fp16 = reshape(shape = var_3026, x = var_3022_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17939136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18463488))))[name = string("tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3031_cast_fp16 = abs(x = linear_22_cast_fp16)[name = string("op_3031_cast_fp16")];
            bool var_3034 = const()[name = string("op_3034"), val = bool(true)];
            tensor<int32, [1]> var_3035_axes_0 = const()[name = string("op_3035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3035_cast_fp16 = reduce_max(axes = var_3035_axes_0, keep_dims = var_3034, x = var_3031_cast_fp16)[name = string("op_3035_cast_fp16")];
            fp16 var_3036_to_fp16 = const()[name = string("op_3036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_43_cast_fp16 = maximum(x = var_3035_cast_fp16, y = var_3036_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_43_cast_fp16 = real_div(x = linear_22_cast_fp16, y = s_43_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3039_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3039_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3045_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_3045_cast_fp16")];
            fp32 var_3046_epsilon_0 = const()[name = string("op_3046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3046_cast_fp16 = inverse(epsilon = var_3046_epsilon_0, x = var_3045_cast_fp16)[name = string("op_3046_cast_fp16")];
            fp16 var_3047_to_fp16 = const()[name = string("op_3047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3048_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3047_to_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3050_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3048_cast_fp16)[name = string("op_3050_cast_fp16")];
            fp32 var_3051_epsilon_0 = const()[name = string("op_3051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3051_cast_fp16 = rsqrt(epsilon = var_3051_epsilon_0, x = var_3050_cast_fp16)[name = string("op_3051_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_3051_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18464576)))];
            tensor<fp16, [1, 512, 512]> var_3053_cast_fp16 = mul(x = out_37_cast_fp16, y = tm_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = var_3053_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3056_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_3056_cast_fp16")];
            bool var_3059 = const()[name = string("op_3059"), val = bool(true)];
            tensor<int32, [1]> var_3060_axes_0 = const()[name = string("op_3060_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3060_cast_fp16 = reduce_max(axes = var_3060_axes_0, keep_dims = var_3059, x = var_3056_cast_fp16)[name = string("op_3060_cast_fp16")];
            fp16 var_3061_to_fp16 = const()[name = string("op_3061_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_45_cast_fp16 = maximum(x = var_3060_cast_fp16, y = var_3061_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_45_cast_fp16 = real_div(x = x_47_cast_fp16, y = s_45_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3064_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<int32, [1]> var_3069_axes_0 = const()[name = string("op_3069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3069_keep_dims_0 = const()[name = string("op_3069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3069_cast_fp16 = reduce_mean(axes = var_3069_axes_0, keep_dims = var_3069_keep_dims_0, x = var_3064_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3070_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp32 var_3071_epsilon_0 = const()[name = string("op_3071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3071_cast_fp16 = inverse(epsilon = var_3071_epsilon_0, x = var_3070_cast_fp16)[name = string("op_3071_cast_fp16")];
            fp16 var_3072_to_fp16 = const()[name = string("op_3072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3073_cast_fp16 = mul(x = var_3071_cast_fp16, y = var_3072_to_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3075_cast_fp16 = add(x = var_3069_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3075_cast_fp16")];
            fp32 var_3076_epsilon_0 = const()[name = string("op_3076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3076_cast_fp16 = rsqrt(epsilon = var_3076_epsilon_0, x = var_3075_cast_fp16)[name = string("op_3076_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_3076_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18465664)))];
            tensor<fp16, [1, 512, 512]> input_29_cast_fp16 = mul(x = out_39_cast_fp16, y = tm_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18466752))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19515392))))[name = string("tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_23_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_23_cast_fp16")];
            string var_3083_mode_0 = const()[name = string("op_3083_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_3083_cast_fp16 = gelu(mode = var_3083_mode_0, x = linear_23_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19519552))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20568192))))[name = string("tm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_24_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_31_cast_fp16 = mul(x = var_3083_cast_fp16, y = linear_24_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20572352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21620992))))[name = string("tm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3091_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_3091_cast_fp16")];
            bool var_3094 = const()[name = string("op_3094"), val = bool(true)];
            tensor<int32, [1]> var_3095_axes_0 = const()[name = string("op_3095_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3095_cast_fp16 = reduce_max(axes = var_3095_axes_0, keep_dims = var_3094, x = var_3091_cast_fp16)[name = string("op_3095_cast_fp16")];
            fp16 var_3096_to_fp16 = const()[name = string("op_3096_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_47_cast_fp16 = maximum(x = var_3095_cast_fp16, y = var_3096_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_47_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_47_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3099_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3104_keep_dims_0 = const()[name = string("op_3104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3104_cast_fp16 = reduce_mean(axes = var_3104_axes_0, keep_dims = var_3104_keep_dims_0, x = var_3099_cast_fp16)[name = string("op_3104_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3105_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_3105_cast_fp16")];
            fp32 var_3106_epsilon_0 = const()[name = string("op_3106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3106_cast_fp16 = inverse(epsilon = var_3106_epsilon_0, x = var_3105_cast_fp16)[name = string("op_3106_cast_fp16")];
            fp16 var_3107_to_fp16 = const()[name = string("op_3107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3108_cast_fp16 = mul(x = var_3106_cast_fp16, y = var_3107_to_fp16)[name = string("op_3108_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3110_cast_fp16 = add(x = var_3104_cast_fp16, y = var_3108_cast_fp16)[name = string("op_3110_cast_fp16")];
            fp32 var_3111_epsilon_0 = const()[name = string("op_3111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3111_cast_fp16 = rsqrt(epsilon = var_3111_epsilon_0, x = var_3110_cast_fp16)[name = string("op_3111_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_3111_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21622080)))];
            tensor<fp16, [1, 512, 512]> var_3113_cast_fp16 = mul(x = out_41_cast_fp16, y = tm_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_33_cast_fp16 = add(x = x_47_cast_fp16, y = var_3113_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21623168))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21885376))))[name = string("tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_26_cast_fp16")];
            string x_51_mode_0 = const()[name = string("x_51_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_51_cast_fp16 = gelu(mode = x_51_mode_0, x = linear_26_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([1, 512, 3, 512])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3133_squeeze_mask_0 = const()[name = string("op_3133_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, squeeze_mask = var_3133_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_35_cast_fp16 = mul(x = x_51_cast_fp16, y = var_3133_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21886464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22148672))))[name = string("tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3143_cast_fp16 = abs(x = linear_27_cast_fp16)[name = string("op_3143_cast_fp16")];
            bool var_3146 = const()[name = string("op_3146"), val = bool(true)];
            tensor<int32, [1]> var_3147_axes_0 = const()[name = string("op_3147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3147_cast_fp16 = reduce_max(axes = var_3147_axes_0, keep_dims = var_3146, x = var_3143_cast_fp16)[name = string("op_3147_cast_fp16")];
            fp16 var_3148_to_fp16 = const()[name = string("op_3148_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_49_cast_fp16 = maximum(x = var_3147_cast_fp16, y = var_3148_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_49_cast_fp16 = real_div(x = linear_27_cast_fp16, y = s_49_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3151_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_3151_cast_fp16")];
            tensor<int32, [1]> var_3156_axes_0 = const()[name = string("op_3156_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3156_keep_dims_0 = const()[name = string("op_3156_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3156_cast_fp16 = reduce_mean(axes = var_3156_axes_0, keep_dims = var_3156_keep_dims_0, x = var_3151_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3157_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_3157_cast_fp16")];
            fp32 var_3158_epsilon_0 = const()[name = string("op_3158_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3158_cast_fp16 = inverse(epsilon = var_3158_epsilon_0, x = var_3157_cast_fp16)[name = string("op_3158_cast_fp16")];
            fp16 var_3159_to_fp16 = const()[name = string("op_3159_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3160_cast_fp16 = mul(x = var_3158_cast_fp16, y = var_3159_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3162_cast_fp16 = add(x = var_3156_cast_fp16, y = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp32 var_3163_epsilon_0 = const()[name = string("op_3163_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3163_cast_fp16 = rsqrt(epsilon = var_3163_epsilon_0, x = var_3162_cast_fp16)[name = string("op_3163_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_43_cast_fp16 = mul(x = y_49_cast_fp16, y = var_3163_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22149760)))];
            tensor<fp16, [1, 512, 512]> var_3165_cast_fp16 = mul(x = out_43_cast_fp16, y = tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_5_cast_fp16 = add(x = input_33_cast_fp16, y = var_3165_cast_fp16)[name = string("h_5_cast_fp16")];
            tensor<fp16, [1]> tm_layers_2_layer_scalar_to_fp16 = const()[name = string("tm_layers_2_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ep-1])];
            tensor<fp16, [1, 512, 512]> x_55_cast_fp16 = mul(x = h_5_cast_fp16, y = tm_layers_2_layer_scalar_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3169_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_3169_cast_fp16")];
            bool var_3172 = const()[name = string("op_3172"), val = bool(true)];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3173_cast_fp16 = reduce_max(axes = var_3173_axes_0, keep_dims = var_3172, x = var_3169_cast_fp16)[name = string("op_3173_cast_fp16")];
            fp16 var_3174_to_fp16 = const()[name = string("op_3174_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_51_cast_fp16 = maximum(x = var_3173_cast_fp16, y = var_3174_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_51_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_51_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3177_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_3177_cast_fp16")];
            tensor<int32, [1]> var_3182_axes_0 = const()[name = string("op_3182_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3182_keep_dims_0 = const()[name = string("op_3182_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3182_cast_fp16 = reduce_mean(axes = var_3182_axes_0, keep_dims = var_3182_keep_dims_0, x = var_3177_cast_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3183_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_3183_cast_fp16")];
            fp32 var_3184_epsilon_0 = const()[name = string("op_3184_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3184_cast_fp16 = inverse(epsilon = var_3184_epsilon_0, x = var_3183_cast_fp16)[name = string("op_3184_cast_fp16")];
            fp16 var_3185_to_fp16 = const()[name = string("op_3185_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3186_cast_fp16 = mul(x = var_3184_cast_fp16, y = var_3185_to_fp16)[name = string("op_3186_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3188_cast_fp16 = add(x = var_3182_cast_fp16, y = var_3186_cast_fp16)[name = string("op_3188_cast_fp16")];
            fp32 var_3189_epsilon_0 = const()[name = string("op_3189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3189_cast_fp16 = rsqrt(epsilon = var_3189_epsilon_0, x = var_3188_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_45_cast_fp16 = mul(x = y_51_cast_fp16, y = var_3189_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22150848)))];
            tensor<fp16, [1, 512, 512]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = tm_layers_3_input_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22151936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22676288))))[name = string("tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_3200_cast_fp16 = reshape(shape = var_3199, x = linear_28_cast_fp16)[name = string("op_3200_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22678400))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22940608))))[name = string("tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_29_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_59_perm_0 = const()[name = string("x_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22941696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23203904))))[name = string("tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_3223 = const()[name = string("op_3223"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_3224_cast_fp16 = reshape(shape = var_3223, x = linear_30_cast_fp16)[name = string("op_3224_cast_fp16")];
            tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_3200_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 4, 512, 256]> var_3228_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3231 = const()[name = string("op_3231"), val = bool(true)];
            tensor<int32, [1]> var_3232_axes_0 = const()[name = string("op_3232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_3232_cast_fp16 = reduce_max(axes = var_3232_axes_0, keep_dims = var_3231, x = var_3228_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_53_cast_fp16 = maximum(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_53_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_53_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_3236_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_3236_cast_fp16")];
            tensor<int32, [1]> var_3241_axes_0 = const()[name = string("op_3241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3241_keep_dims_0 = const()[name = string("op_3241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_3241_cast_fp16 = reduce_mean(axes = var_3241_axes_0, keep_dims = var_3241_keep_dims_0, x = var_3236_cast_fp16)[name = string("op_3241_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_3242_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_3242_cast_fp16")];
            fp32 var_3243_epsilon_0 = const()[name = string("op_3243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_3243_cast_fp16 = inverse(epsilon = var_3243_epsilon_0, x = var_3242_cast_fp16)[name = string("op_3243_cast_fp16")];
            fp16 var_3244_to_fp16 = const()[name = string("op_3244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_3245_cast_fp16 = mul(x = var_3243_cast_fp16, y = var_3244_to_fp16)[name = string("op_3245_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_3247_cast_fp16 = add(x = var_3241_cast_fp16, y = var_3245_cast_fp16)[name = string("op_3247_cast_fp16")];
            fp32 var_3248_epsilon_0 = const()[name = string("op_3248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_3248_cast_fp16 = rsqrt(epsilon = var_3248_epsilon_0, x = var_3247_cast_fp16)[name = string("op_3248_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_47_cast_fp16 = mul(x = y_53_cast_fp16, y = var_3248_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23204992)))];
            tensor<fp16, [1, 4, 512, 256]> q_13_cast_fp16 = mul(x = out_47_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_59_cast_fp16 = transpose(perm = x_59_perm_0, x = var_3212_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 2, 512, 256]> var_3251_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_3251_cast_fp16")];
            bool var_3254 = const()[name = string("op_3254"), val = bool(true)];
            tensor<int32, [1]> var_3255_axes_0 = const()[name = string("op_3255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_3255_cast_fp16 = reduce_max(axes = var_3255_axes_0, keep_dims = var_3254, x = var_3251_cast_fp16)[name = string("op_3255_cast_fp16")];
            fp16 var_3256_to_fp16 = const()[name = string("op_3256_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_55_cast_fp16 = maximum(x = var_3255_cast_fp16, y = var_3256_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_55_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_55_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_3259_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<int32, [1]> var_3264_axes_0 = const()[name = string("op_3264_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3264_keep_dims_0 = const()[name = string("op_3264_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_3264_cast_fp16 = reduce_mean(axes = var_3264_axes_0, keep_dims = var_3264_keep_dims_0, x = var_3259_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_3265_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_3265_cast_fp16")];
            fp32 var_3266_epsilon_0 = const()[name = string("op_3266_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_3266_cast_fp16 = inverse(epsilon = var_3266_epsilon_0, x = var_3265_cast_fp16)[name = string("op_3266_cast_fp16")];
            fp16 var_3267_to_fp16 = const()[name = string("op_3267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_3268_cast_fp16 = mul(x = var_3266_cast_fp16, y = var_3267_to_fp16)[name = string("op_3268_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_3270_cast_fp16 = add(x = var_3264_cast_fp16, y = var_3268_cast_fp16)[name = string("op_3270_cast_fp16")];
            fp32 var_3271_epsilon_0 = const()[name = string("op_3271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_3271_cast_fp16 = rsqrt(epsilon = var_3271_epsilon_0, x = var_3270_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_49_cast_fp16 = mul(x = y_55_cast_fp16, y = var_3271_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23205568)))];
            tensor<fp16, [1, 2, 512, 256]> k_19_cast_fp16 = mul(x = out_49_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_3224_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2, 512, 256]> var_3274_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_3274_cast_fp16")];
            bool var_3277 = const()[name = string("op_3277"), val = bool(true)];
            tensor<int32, [1]> var_3278_axes_0 = const()[name = string("op_3278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_3278_cast_fp16 = reduce_max(axes = var_3278_axes_0, keep_dims = var_3277, x = var_3274_cast_fp16)[name = string("op_3278_cast_fp16")];
            fp16 var_3279_to_fp16 = const()[name = string("op_3279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_57_cast_fp16 = maximum(x = var_3278_cast_fp16, y = var_3279_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_57_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_57_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_3282_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_3282_cast_fp16")];
            tensor<int32, [1]> var_3287_axes_0 = const()[name = string("op_3287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3287_keep_dims_0 = const()[name = string("op_3287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_3287_cast_fp16 = reduce_mean(axes = var_3287_axes_0, keep_dims = var_3287_keep_dims_0, x = var_3282_cast_fp16)[name = string("op_3287_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_3288_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_3289_cast_fp16 = inverse(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            fp16 var_3290_to_fp16 = const()[name = string("op_3290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_3291_cast_fp16 = mul(x = var_3289_cast_fp16, y = var_3290_to_fp16)[name = string("op_3291_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_3293_cast_fp16 = add(x = var_3287_cast_fp16, y = var_3291_cast_fp16)[name = string("op_3293_cast_fp16")];
            fp32 var_3294_epsilon_0 = const()[name = string("op_3294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_3294_cast_fp16 = rsqrt(epsilon = var_3294_epsilon_0, x = var_3293_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_13_cast_fp16 = mul(x = y_57_cast_fp16, y = var_3294_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_3296_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3296_cast_fp16")];
            tensor<int32, [4]> var_3301_begin_0 = const()[name = string("op_3301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3301_end_0 = const()[name = string("op_3301_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_3301_end_mask_0 = const()[name = string("op_3301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_3301_cast_fp16 = slice_by_index(begin = var_3301_begin_0, end = var_3301_end_0, end_mask = var_3301_end_mask_0, x = q_13_cast_fp16)[name = string("op_3301_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_3302_cast_fp16 = mul(x = var_3301_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3307_begin_0 = const()[name = string("op_3307_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3307_end_0 = const()[name = string("op_3307_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_3307_end_mask_0 = const()[name = string("op_3307_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_3307_cast_fp16 = slice_by_index(begin = var_3307_begin_0, end = var_3307_end_0, end_mask = var_3307_end_mask_0, x = q_13_cast_fp16)[name = string("op_3307_cast_fp16")];
            int32 var_3309 = const()[name = string("op_3309"), val = int32(-1)];
            bool var_3310_interleave_0 = const()[name = string("op_3310_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_3310_cast_fp16 = concat(axis = var_3309, interleave = var_3310_interleave_0, values = (var_3302_cast_fp16, var_3307_cast_fp16))[name = string("op_3310_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_3311_cast_fp16 = mul(x = var_3310_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3311_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_15_cast_fp16 = add(x = var_3296_cast_fp16, y = var_3311_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_3314_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3314_cast_fp16")];
            tensor<int32, [4]> var_3319_begin_0 = const()[name = string("op_3319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3319_end_0 = const()[name = string("op_3319_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_3319_end_mask_0 = const()[name = string("op_3319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_3319_cast_fp16 = slice_by_index(begin = var_3319_begin_0, end = var_3319_end_0, end_mask = var_3319_end_mask_0, x = k_19_cast_fp16)[name = string("op_3319_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_3320_cast_fp16 = mul(x = var_3319_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_3320_cast_fp16")];
            tensor<int32, [4]> var_3325_begin_0 = const()[name = string("op_3325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3325_end_0 = const()[name = string("op_3325_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_3325_end_mask_0 = const()[name = string("op_3325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_3325_cast_fp16 = slice_by_index(begin = var_3325_begin_0, end = var_3325_end_0, end_mask = var_3325_end_mask_0, x = k_19_cast_fp16)[name = string("op_3325_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), val = int32(-1)];
            bool var_3328_interleave_0 = const()[name = string("op_3328_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_3328_cast_fp16 = concat(axis = var_3327, interleave = var_3328_interleave_0, values = (var_3320_cast_fp16, var_3325_cast_fp16))[name = string("op_3328_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_3329_cast_fp16 = mul(x = var_3328_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3329_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_21_cast_fp16 = add(x = var_3314_cast_fp16, y = var_3329_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = k_21_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [4, 1, 512, 256]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_12_cast_fp16 = reshape(shape = concat_12, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [4, 1, 512, 256]> reshape_13_cast_fp16 = reshape(shape = concat_13, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = v_13_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [4, 1, 512, 256]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_14_cast_fp16 = reshape(shape = concat_14, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [4, 1, 512, 256]> reshape_15_cast_fp16 = reshape(shape = concat_15, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3343_transpose_x_0 = const()[name = string("op_3343_transpose_x_0"), val = bool(false)];
            bool var_3343_transpose_y_0 = const()[name = string("op_3343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 4, 512, 512]> var_3343_cast_fp16 = matmul(transpose_x = var_3343_transpose_x_0, transpose_y = var_3343_transpose_y_0, x = q_15_cast_fp16, y = transpose_99_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_3345_cast_fp16 = add(x = var_3343_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3345_cast_fp16")];
            int32 var_3346 = const()[name = string("op_3346"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_57_cast_fp16 = softmax(axis = var_3346, x = var_3345_cast_fp16)[name = string("w_57_cast_fp16")];
            bool var_3349_transpose_x_0 = const()[name = string("op_3349_transpose_x_0"), val = bool(false)];
            bool var_3349_transpose_y_0 = const()[name = string("op_3349_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 4, 512, 256]> var_3349_cast_fp16 = matmul(transpose_x = var_3349_transpose_x_0, transpose_y = var_3349_transpose_y_0, x = w_57_cast_fp16, y = v_15_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<int32, [4]> var_3352_perm_0 = const()[name = string("op_3352_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3356 = const()[name = string("op_3356"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_3352_cast_fp16 = transpose(perm = var_3352_perm_0, x = var_3349_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 512, 1024]> input_39_cast_fp16 = reshape(shape = var_3356, x = var_3352_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23206144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23730496))))[name = string("tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3361_cast_fp16 = abs(x = linear_31_cast_fp16)[name = string("op_3361_cast_fp16")];
            bool var_3364 = const()[name = string("op_3364"), val = bool(true)];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3365_cast_fp16 = reduce_max(axes = var_3365_axes_0, keep_dims = var_3364, x = var_3361_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_59_cast_fp16 = maximum(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_59_cast_fp16 = real_div(x = linear_31_cast_fp16, y = s_59_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3369_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<int32, [1]> var_3374_axes_0 = const()[name = string("op_3374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3374_keep_dims_0 = const()[name = string("op_3374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3374_cast_fp16 = reduce_mean(axes = var_3374_axes_0, keep_dims = var_3374_keep_dims_0, x = var_3369_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3375_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_3375_cast_fp16")];
            fp32 var_3376_epsilon_0 = const()[name = string("op_3376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3376_cast_fp16 = inverse(epsilon = var_3376_epsilon_0, x = var_3375_cast_fp16)[name = string("op_3376_cast_fp16")];
            fp16 var_3377_to_fp16 = const()[name = string("op_3377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3378_cast_fp16 = mul(x = var_3376_cast_fp16, y = var_3377_to_fp16)[name = string("op_3378_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3380_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3380_cast_fp16")];
            fp32 var_3381_epsilon_0 = const()[name = string("op_3381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3381_cast_fp16 = rsqrt(epsilon = var_3381_epsilon_0, x = var_3380_cast_fp16)[name = string("op_3381_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_3381_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23731584)))];
            tensor<fp16, [1, 512, 512]> var_3383_cast_fp16 = mul(x = out_51_cast_fp16, y = tm_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_65_cast_fp16 = add(x = x_55_cast_fp16, y = var_3383_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3386_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_3386_cast_fp16")];
            bool var_3389 = const()[name = string("op_3389"), val = bool(true)];
            tensor<int32, [1]> var_3390_axes_0 = const()[name = string("op_3390_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3390_cast_fp16 = reduce_max(axes = var_3390_axes_0, keep_dims = var_3389, x = var_3386_cast_fp16)[name = string("op_3390_cast_fp16")];
            fp16 var_3391_to_fp16 = const()[name = string("op_3391_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_61_cast_fp16 = maximum(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_61_cast_fp16 = real_div(x = x_65_cast_fp16, y = s_61_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3394_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<int32, [1]> var_3399_axes_0 = const()[name = string("op_3399_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3399_keep_dims_0 = const()[name = string("op_3399_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3399_cast_fp16 = reduce_mean(axes = var_3399_axes_0, keep_dims = var_3399_keep_dims_0, x = var_3394_cast_fp16)[name = string("op_3399_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3400_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_3400_cast_fp16")];
            fp32 var_3401_epsilon_0 = const()[name = string("op_3401_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3401_cast_fp16 = inverse(epsilon = var_3401_epsilon_0, x = var_3400_cast_fp16)[name = string("op_3401_cast_fp16")];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3403_cast_fp16 = mul(x = var_3401_cast_fp16, y = var_3402_to_fp16)[name = string("op_3403_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3405_cast_fp16 = add(x = var_3399_cast_fp16, y = var_3403_cast_fp16)[name = string("op_3405_cast_fp16")];
            fp32 var_3406_epsilon_0 = const()[name = string("op_3406_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3406_cast_fp16 = rsqrt(epsilon = var_3406_epsilon_0, x = var_3405_cast_fp16)[name = string("op_3406_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_3406_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23732672)))];
            tensor<fp16, [1, 512, 512]> input_41_cast_fp16 = mul(x = out_53_cast_fp16, y = tm_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23733760))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24782400))))[name = string("tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_32_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_32_cast_fp16")];
            string var_3413_mode_0 = const()[name = string("op_3413_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_3413_cast_fp16 = gelu(mode = var_3413_mode_0, x = linear_32_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24786560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25835200))))[name = string("tm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_43_cast_fp16 = mul(x = var_3413_cast_fp16, y = linear_33_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25839360))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26888000))))[name = string("tm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3421_cast_fp16 = abs(x = linear_34_cast_fp16)[name = string("op_3421_cast_fp16")];
            bool var_3424 = const()[name = string("op_3424"), val = bool(true)];
            tensor<int32, [1]> var_3425_axes_0 = const()[name = string("op_3425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3425_cast_fp16 = reduce_max(axes = var_3425_axes_0, keep_dims = var_3424, x = var_3421_cast_fp16)[name = string("op_3425_cast_fp16")];
            fp16 var_3426_to_fp16 = const()[name = string("op_3426_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_63_cast_fp16 = maximum(x = var_3425_cast_fp16, y = var_3426_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_63_cast_fp16 = real_div(x = linear_34_cast_fp16, y = s_63_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3429_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_3429_cast_fp16")];
            tensor<int32, [1]> var_3434_axes_0 = const()[name = string("op_3434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3434_keep_dims_0 = const()[name = string("op_3434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3434_cast_fp16 = reduce_mean(axes = var_3434_axes_0, keep_dims = var_3434_keep_dims_0, x = var_3429_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3435_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp32 var_3436_epsilon_0 = const()[name = string("op_3436_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3436_cast_fp16 = inverse(epsilon = var_3436_epsilon_0, x = var_3435_cast_fp16)[name = string("op_3436_cast_fp16")];
            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3438_cast_fp16 = mul(x = var_3436_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3440_cast_fp16 = add(x = var_3434_cast_fp16, y = var_3438_cast_fp16)[name = string("op_3440_cast_fp16")];
            fp32 var_3441_epsilon_0 = const()[name = string("op_3441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3441_cast_fp16 = rsqrt(epsilon = var_3441_epsilon_0, x = var_3440_cast_fp16)[name = string("op_3441_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_55_cast_fp16 = mul(x = y_63_cast_fp16, y = var_3441_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26889088)))];
            tensor<fp16, [1, 512, 512]> var_3443_cast_fp16 = mul(x = out_55_cast_fp16, y = tm_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_45_cast_fp16 = add(x = x_65_cast_fp16, y = var_3443_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26890176))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27152384))))[name = string("tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_35_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_35_cast_fp16")];
            string x_69_mode_0 = const()[name = string("x_69_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_69_cast_fp16 = gelu(mode = x_69_mode_0, x = linear_35_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [4]> var_3463_begin_0 = const()[name = string("op_3463_begin_0"), val = tensor<int32, [4]>([0, 0, 3, 0])];
            tensor<int32, [4]> var_3463_end_0 = const()[name = string("op_3463_end_0"), val = tensor<int32, [4]>([1, 512, 4, 512])];
            tensor<bool, [4]> var_3463_end_mask_0 = const()[name = string("op_3463_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3463_squeeze_mask_0 = const()[name = string("op_3463_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_3463_cast_fp16 = slice_by_index(begin = var_3463_begin_0, end = var_3463_end_0, end_mask = var_3463_end_mask_0, squeeze_mask = var_3463_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3463_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_47_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3463_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27153472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27415680))))[name = string("tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3473_cast_fp16 = abs(x = linear_36_cast_fp16)[name = string("op_3473_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<int32, [1]> var_3477_axes_0 = const()[name = string("op_3477_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3477_cast_fp16 = reduce_max(axes = var_3477_axes_0, keep_dims = var_3476, x = var_3473_cast_fp16)[name = string("op_3477_cast_fp16")];
            fp16 var_3478_to_fp16 = const()[name = string("op_3478_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_65_cast_fp16 = maximum(x = var_3477_cast_fp16, y = var_3478_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_65_cast_fp16 = real_div(x = linear_36_cast_fp16, y = s_65_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3481_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [1]> var_3486_axes_0 = const()[name = string("op_3486_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3486_keep_dims_0 = const()[name = string("op_3486_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3486_cast_fp16 = reduce_mean(axes = var_3486_axes_0, keep_dims = var_3486_keep_dims_0, x = var_3481_cast_fp16)[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3487_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp32 var_3488_epsilon_0 = const()[name = string("op_3488_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3488_cast_fp16 = inverse(epsilon = var_3488_epsilon_0, x = var_3487_cast_fp16)[name = string("op_3488_cast_fp16")];
            fp16 var_3489_to_fp16 = const()[name = string("op_3489_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3490_cast_fp16 = mul(x = var_3488_cast_fp16, y = var_3489_to_fp16)[name = string("op_3490_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3492_cast_fp16 = add(x = var_3486_cast_fp16, y = var_3490_cast_fp16)[name = string("op_3492_cast_fp16")];
            fp32 var_3493_epsilon_0 = const()[name = string("op_3493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3493_cast_fp16 = rsqrt(epsilon = var_3493_epsilon_0, x = var_3492_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_57_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3493_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27416768)))];
            tensor<fp16, [1, 512, 512]> var_3495_cast_fp16 = mul(x = out_57_cast_fp16, y = tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3495_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_7_cast_fp16 = add(x = input_45_cast_fp16, y = var_3495_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<fp16, [1]> tm_layers_3_layer_scalar_to_fp16 = const()[name = string("tm_layers_3_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 512, 512]> x_73_cast_fp16 = mul(x = h_7_cast_fp16, y = tm_layers_3_layer_scalar_to_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3499_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_3499_cast_fp16")];
            bool var_3502 = const()[name = string("op_3502"), val = bool(true)];
            tensor<int32, [1]> var_3503_axes_0 = const()[name = string("op_3503_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3503_cast_fp16 = reduce_max(axes = var_3503_axes_0, keep_dims = var_3502, x = var_3499_cast_fp16)[name = string("op_3503_cast_fp16")];
            fp16 var_3504_to_fp16 = const()[name = string("op_3504_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_67_cast_fp16 = maximum(x = var_3503_cast_fp16, y = var_3504_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_67_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_67_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3507_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<int32, [1]> var_3512_axes_0 = const()[name = string("op_3512_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3512_keep_dims_0 = const()[name = string("op_3512_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3512_cast_fp16 = reduce_mean(axes = var_3512_axes_0, keep_dims = var_3512_keep_dims_0, x = var_3507_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3513_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_3513_cast_fp16")];
            fp32 var_3514_epsilon_0 = const()[name = string("op_3514_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3514_cast_fp16 = inverse(epsilon = var_3514_epsilon_0, x = var_3513_cast_fp16)[name = string("op_3514_cast_fp16")];
            fp16 var_3515_to_fp16 = const()[name = string("op_3515_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3516_cast_fp16 = mul(x = var_3514_cast_fp16, y = var_3515_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3518_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp32 var_3519_epsilon_0 = const()[name = string("op_3519_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3519_cast_fp16 = rsqrt(epsilon = var_3519_epsilon_0, x = var_3518_cast_fp16)[name = string("op_3519_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_59_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3519_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27417856)))];
            tensor<fp16, [1, 512, 512]> input_49_cast_fp16 = mul(x = out_59_cast_fp16, y = tm_layers_4_input_layernorm_weight_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27418944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27943296))))[name = string("tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3529 = const()[name = string("op_3529"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_3530_cast_fp16 = reshape(shape = var_3529, x = linear_37_cast_fp16)[name = string("op_3530_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27945408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28207616))))[name = string("tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3541 = const()[name = string("op_3541"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_3542_cast_fp16 = reshape(shape = var_3541, x = linear_38_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28208704))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28470912))))[name = string("tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_3553 = const()[name = string("op_3553"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_3554_cast_fp16 = reshape(shape = var_3553, x = linear_39_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<int32, [4]> x_79_perm_0 = const()[name = string("x_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3530_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 4, 512, 256]> var_3558_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3558_cast_fp16")];
            bool var_3561 = const()[name = string("op_3561"), val = bool(true)];
            tensor<int32, [1]> var_3562_axes_0 = const()[name = string("op_3562_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_3562_cast_fp16 = reduce_max(axes = var_3562_axes_0, keep_dims = var_3561, x = var_3558_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp16 var_3563_to_fp16 = const()[name = string("op_3563_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_69_cast_fp16 = maximum(x = var_3562_cast_fp16, y = var_3563_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_69_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_69_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_3566_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3566_cast_fp16")];
            tensor<int32, [1]> var_3571_axes_0 = const()[name = string("op_3571_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3571_keep_dims_0 = const()[name = string("op_3571_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_3571_cast_fp16 = reduce_mean(axes = var_3571_axes_0, keep_dims = var_3571_keep_dims_0, x = var_3566_cast_fp16)[name = string("op_3571_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_3572_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3572_cast_fp16")];
            fp32 var_3573_epsilon_0 = const()[name = string("op_3573_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_3573_cast_fp16 = inverse(epsilon = var_3573_epsilon_0, x = var_3572_cast_fp16)[name = string("op_3573_cast_fp16")];
            fp16 var_3574_to_fp16 = const()[name = string("op_3574_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_3575_cast_fp16 = mul(x = var_3573_cast_fp16, y = var_3574_to_fp16)[name = string("op_3575_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_3577_cast_fp16 = add(x = var_3571_cast_fp16, y = var_3575_cast_fp16)[name = string("op_3577_cast_fp16")];
            fp32 var_3578_epsilon_0 = const()[name = string("op_3578_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_3578_cast_fp16 = rsqrt(epsilon = var_3578_epsilon_0, x = var_3577_cast_fp16)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_61_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3578_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [256]> tm_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472000)))];
            tensor<fp16, [1, 4, 512, 256]> q_17_cast_fp16 = mul(x = out_61_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3542_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2, 512, 256]> var_3581_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3581_cast_fp16")];
            bool var_3584 = const()[name = string("op_3584"), val = bool(true)];
            tensor<int32, [1]> var_3585_axes_0 = const()[name = string("op_3585_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_3585_cast_fp16 = reduce_max(axes = var_3585_axes_0, keep_dims = var_3584, x = var_3581_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp16 var_3586_to_fp16 = const()[name = string("op_3586_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_71_cast_fp16 = maximum(x = var_3585_cast_fp16, y = var_3586_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_71_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_71_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_3589_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3589_cast_fp16")];
            tensor<int32, [1]> var_3594_axes_0 = const()[name = string("op_3594_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3594_keep_dims_0 = const()[name = string("op_3594_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_3594_cast_fp16 = reduce_mean(axes = var_3594_axes_0, keep_dims = var_3594_keep_dims_0, x = var_3589_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_3595_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp32 var_3596_epsilon_0 = const()[name = string("op_3596_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_3596_cast_fp16 = inverse(epsilon = var_3596_epsilon_0, x = var_3595_cast_fp16)[name = string("op_3596_cast_fp16")];
            fp16 var_3597_to_fp16 = const()[name = string("op_3597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_3598_cast_fp16 = mul(x = var_3596_cast_fp16, y = var_3597_to_fp16)[name = string("op_3598_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_3600_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3598_cast_fp16)[name = string("op_3600_cast_fp16")];
            fp32 var_3601_epsilon_0 = const()[name = string("op_3601_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_3601_cast_fp16 = rsqrt(epsilon = var_3601_epsilon_0, x = var_3600_cast_fp16)[name = string("op_3601_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_63_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3601_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472576)))];
            tensor<fp16, [1, 2, 512, 256]> k_25_cast_fp16 = mul(x = out_63_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_79_cast_fp16 = transpose(perm = x_79_perm_0, x = var_3554_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 2, 512, 256]> var_3604_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3604_cast_fp16")];
            bool var_3607 = const()[name = string("op_3607"), val = bool(true)];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_3607, x = var_3604_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_73_cast_fp16 = maximum(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_73_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_73_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_3612_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<int32, [1]> var_3617_axes_0 = const()[name = string("op_3617_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3617_keep_dims_0 = const()[name = string("op_3617_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_3617_cast_fp16 = reduce_mean(axes = var_3617_axes_0, keep_dims = var_3617_keep_dims_0, x = var_3612_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_3618_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_3619_cast_fp16 = inverse(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            fp16 var_3620_to_fp16 = const()[name = string("op_3620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_3621_cast_fp16 = mul(x = var_3619_cast_fp16, y = var_3620_to_fp16)[name = string("op_3621_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_3623_cast_fp16 = add(x = var_3617_cast_fp16, y = var_3621_cast_fp16)[name = string("op_3623_cast_fp16")];
            fp32 var_3624_epsilon_0 = const()[name = string("op_3624_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_3624_cast_fp16 = rsqrt(epsilon = var_3624_epsilon_0, x = var_3623_cast_fp16)[name = string("op_3624_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_17_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3624_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_3626_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3626_cast_fp16")];
            tensor<int32, [4]> var_3631_begin_0 = const()[name = string("op_3631_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3631_end_0 = const()[name = string("op_3631_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_3631_end_mask_0 = const()[name = string("op_3631_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_3631_cast_fp16 = slice_by_index(begin = var_3631_begin_0, end = var_3631_end_0, end_mask = var_3631_end_mask_0, x = q_17_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_3632_cast_fp16 = mul(x = var_3631_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> var_3637_begin_0 = const()[name = string("op_3637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3637_end_0 = const()[name = string("op_3637_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_3637_end_mask_0 = const()[name = string("op_3637_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_3637_cast_fp16 = slice_by_index(begin = var_3637_begin_0, end = var_3637_end_0, end_mask = var_3637_end_mask_0, x = q_17_cast_fp16)[name = string("op_3637_cast_fp16")];
            int32 var_3639 = const()[name = string("op_3639"), val = int32(-1)];
            bool var_3640_interleave_0 = const()[name = string("op_3640_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_3640_cast_fp16 = concat(axis = var_3639, interleave = var_3640_interleave_0, values = (var_3632_cast_fp16, var_3637_cast_fp16))[name = string("op_3640_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_3641_cast_fp16 = mul(x = var_3640_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3641_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_19_cast_fp16 = add(x = var_3626_cast_fp16, y = var_3641_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_3644_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = k_25_cast_fp16)[name = string("op_3649_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_3650_cast_fp16 = mul(x = var_3649_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            tensor<int32, [4]> var_3655_begin_0 = const()[name = string("op_3655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3655_end_0 = const()[name = string("op_3655_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_3655_end_mask_0 = const()[name = string("op_3655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_3655_cast_fp16 = slice_by_index(begin = var_3655_begin_0, end = var_3655_end_0, end_mask = var_3655_end_mask_0, x = k_25_cast_fp16)[name = string("op_3655_cast_fp16")];
            int32 var_3657 = const()[name = string("op_3657"), val = int32(-1)];
            bool var_3658_interleave_0 = const()[name = string("op_3658_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_3658_cast_fp16 = concat(axis = var_3657, interleave = var_3658_interleave_0, values = (var_3650_cast_fp16, var_3655_cast_fp16))[name = string("op_3658_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_3659_cast_fp16 = mul(x = var_3658_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3659_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_27_cast_fp16 = add(x = var_3644_cast_fp16, y = var_3659_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = k_27_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [4, 1, 512, 256]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
            tensor<int32, [5]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_16_cast_fp16 = reshape(shape = concat_16, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [4, 1, 512, 256]> reshape_17_cast_fp16 = reshape(shape = concat_17, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = v_17_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [4, 1, 512, 256]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
            tensor<int32, [5]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_18_cast_fp16 = reshape(shape = concat_18, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [4, 1, 512, 256]> reshape_19_cast_fp16 = reshape(shape = concat_19, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3673_transpose_x_0 = const()[name = string("op_3673_transpose_x_0"), val = bool(false)];
            bool var_3673_transpose_y_0 = const()[name = string("op_3673_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 4, 512, 512]> var_3673_cast_fp16 = matmul(transpose_x = var_3673_transpose_x_0, transpose_y = var_3673_transpose_y_0, x = q_19_cast_fp16, y = transpose_100_cast_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_3675_cast_fp16 = add(x = var_3673_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3675_cast_fp16")];
            int32 var_3676 = const()[name = string("op_3676"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_73_cast_fp16 = softmax(axis = var_3676, x = var_3675_cast_fp16)[name = string("w_73_cast_fp16")];
            bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)];
            bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 4, 512, 256]> var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = w_73_cast_fp16, y = v_19_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<int32, [4]> var_3682_perm_0 = const()[name = string("op_3682_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3686 = const()[name = string("op_3686"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_3682_cast_fp16 = transpose(perm = var_3682_perm_0, x = var_3679_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 512, 1024]> input_51_cast_fp16 = reshape(shape = var_3686, x = var_3682_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28473152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28997504))))[name = string("tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3691_cast_fp16 = abs(x = linear_40_cast_fp16)[name = string("op_3691_cast_fp16")];
            bool var_3694 = const()[name = string("op_3694"), val = bool(true)];
            tensor<int32, [1]> var_3695_axes_0 = const()[name = string("op_3695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3695_cast_fp16 = reduce_max(axes = var_3695_axes_0, keep_dims = var_3694, x = var_3691_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 var_3696_to_fp16 = const()[name = string("op_3696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_75_cast_fp16 = maximum(x = var_3695_cast_fp16, y = var_3696_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_75_cast_fp16 = real_div(x = linear_40_cast_fp16, y = s_75_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3699_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3699_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3704_keep_dims_0 = const()[name = string("op_3704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3704_cast_fp16 = reduce_mean(axes = var_3704_axes_0, keep_dims = var_3704_keep_dims_0, x = var_3699_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3705_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3705_cast_fp16")];
            fp32 var_3706_epsilon_0 = const()[name = string("op_3706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3706_cast_fp16 = inverse(epsilon = var_3706_epsilon_0, x = var_3705_cast_fp16)[name = string("op_3706_cast_fp16")];
            fp16 var_3707_to_fp16 = const()[name = string("op_3707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3708_cast_fp16 = mul(x = var_3706_cast_fp16, y = var_3707_to_fp16)[name = string("op_3708_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3710_cast_fp16 = add(x = var_3704_cast_fp16, y = var_3708_cast_fp16)[name = string("op_3710_cast_fp16")];
            fp32 var_3711_epsilon_0 = const()[name = string("op_3711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3711_cast_fp16 = rsqrt(epsilon = var_3711_epsilon_0, x = var_3710_cast_fp16)[name = string("op_3711_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3711_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28998592)))];
            tensor<fp16, [1, 512, 512]> var_3713_cast_fp16 = mul(x = out_65_cast_fp16, y = tm_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = var_3713_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3716_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3716_cast_fp16")];
            bool var_3719 = const()[name = string("op_3719"), val = bool(true)];
            tensor<int32, [1]> var_3720_axes_0 = const()[name = string("op_3720_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3720_cast_fp16 = reduce_max(axes = var_3720_axes_0, keep_dims = var_3719, x = var_3716_cast_fp16)[name = string("op_3720_cast_fp16")];
            fp16 var_3721_to_fp16 = const()[name = string("op_3721_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_77_cast_fp16 = maximum(x = var_3720_cast_fp16, y = var_3721_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_77_cast_fp16 = real_div(x = x_83_cast_fp16, y = s_77_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3724_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3729_axes_0 = const()[name = string("op_3729_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3729_keep_dims_0 = const()[name = string("op_3729_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3729_cast_fp16 = reduce_mean(axes = var_3729_axes_0, keep_dims = var_3729_keep_dims_0, x = var_3724_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3730_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3730_cast_fp16")];
            fp32 var_3731_epsilon_0 = const()[name = string("op_3731_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3731_cast_fp16 = inverse(epsilon = var_3731_epsilon_0, x = var_3730_cast_fp16)[name = string("op_3731_cast_fp16")];
            fp16 var_3732_to_fp16 = const()[name = string("op_3732_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3733_cast_fp16 = mul(x = var_3731_cast_fp16, y = var_3732_to_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3735_cast_fp16 = add(x = var_3729_cast_fp16, y = var_3733_cast_fp16)[name = string("op_3735_cast_fp16")];
            fp32 var_3736_epsilon_0 = const()[name = string("op_3736_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3736_cast_fp16 = rsqrt(epsilon = var_3736_epsilon_0, x = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_67_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3736_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28999680)))];
            tensor<fp16, [1, 512, 512]> input_53_cast_fp16 = mul(x = out_67_cast_fp16, y = tm_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29000768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30049408))))[name = string("tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_41_cast_fp16")];
            string var_3743_mode_0 = const()[name = string("op_3743_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_3743_cast_fp16 = gelu(mode = var_3743_mode_0, x = linear_41_cast_fp16)[name = string("op_3743_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30053568))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31102208))))[name = string("tm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_42_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_55_cast_fp16 = mul(x = var_3743_cast_fp16, y = linear_42_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31106368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32155008))))[name = string("tm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3751_cast_fp16 = abs(x = linear_43_cast_fp16)[name = string("op_3751_cast_fp16")];
            bool var_3754 = const()[name = string("op_3754"), val = bool(true)];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_3754, x = var_3751_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 var_3756_to_fp16 = const()[name = string("op_3756_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_79_cast_fp16 = maximum(x = var_3755_cast_fp16, y = var_3756_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_79_cast_fp16 = real_div(x = linear_43_cast_fp16, y = s_79_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3759_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<int32, [1]> var_3764_axes_0 = const()[name = string("op_3764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3764_keep_dims_0 = const()[name = string("op_3764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3764_cast_fp16 = reduce_mean(axes = var_3764_axes_0, keep_dims = var_3764_keep_dims_0, x = var_3759_cast_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3765_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3766_cast_fp16 = inverse(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3768_cast_fp16 = mul(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("op_3768_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3770_cast_fp16 = add(x = var_3764_cast_fp16, y = var_3768_cast_fp16)[name = string("op_3770_cast_fp16")];
            fp32 var_3771_epsilon_0 = const()[name = string("op_3771_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3771_cast_fp16 = rsqrt(epsilon = var_3771_epsilon_0, x = var_3770_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_69_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3771_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32156096)))];
            tensor<fp16, [1, 512, 512]> var_3773_cast_fp16 = mul(x = out_69_cast_fp16, y = tm_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3773_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_57_cast_fp16 = add(x = x_83_cast_fp16, y = var_3773_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32157184))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32419392))))[name = string("tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_44_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_44_cast_fp16")];
            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_87_cast_fp16 = gelu(mode = x_87_mode_0, x = linear_44_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [4]> var_3793_begin_0 = const()[name = string("op_3793_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_3793_end_0 = const()[name = string("op_3793_end_0"), val = tensor<int32, [4]>([1, 512, 5, 512])];
            tensor<bool, [4]> var_3793_end_mask_0 = const()[name = string("op_3793_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3793_squeeze_mask_0 = const()[name = string("op_3793_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_3793_cast_fp16 = slice_by_index(begin = var_3793_begin_0, end = var_3793_end_0, end_mask = var_3793_end_mask_0, squeeze_mask = var_3793_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_59_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3793_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32420480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32682688))))[name = string("tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3803_cast_fp16 = abs(x = linear_45_cast_fp16)[name = string("op_3803_cast_fp16")];
            bool var_3806 = const()[name = string("op_3806"), val = bool(true)];
            tensor<int32, [1]> var_3807_axes_0 = const()[name = string("op_3807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3807_cast_fp16 = reduce_max(axes = var_3807_axes_0, keep_dims = var_3806, x = var_3803_cast_fp16)[name = string("op_3807_cast_fp16")];
            fp16 var_3808_to_fp16 = const()[name = string("op_3808_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_81_cast_fp16 = maximum(x = var_3807_cast_fp16, y = var_3808_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_81_cast_fp16 = real_div(x = linear_45_cast_fp16, y = s_81_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3811_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [1]> var_3816_axes_0 = const()[name = string("op_3816_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3816_keep_dims_0 = const()[name = string("op_3816_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3816_cast_fp16 = reduce_mean(axes = var_3816_axes_0, keep_dims = var_3816_keep_dims_0, x = var_3811_cast_fp16)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3817_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3817_cast_fp16")];
            fp32 var_3818_epsilon_0 = const()[name = string("op_3818_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3818_cast_fp16 = inverse(epsilon = var_3818_epsilon_0, x = var_3817_cast_fp16)[name = string("op_3818_cast_fp16")];
            fp16 var_3819_to_fp16 = const()[name = string("op_3819_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3820_cast_fp16 = mul(x = var_3818_cast_fp16, y = var_3819_to_fp16)[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3822_cast_fp16 = add(x = var_3816_cast_fp16, y = var_3820_cast_fp16)[name = string("op_3822_cast_fp16")];
            fp32 var_3823_epsilon_0 = const()[name = string("op_3823_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3823_cast_fp16 = rsqrt(epsilon = var_3823_epsilon_0, x = var_3822_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_71_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3823_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32683776)))];
            tensor<fp16, [1, 512, 512]> var_3825_cast_fp16 = mul(x = out_71_cast_fp16, y = tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3825_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_9_cast_fp16 = add(x = input_57_cast_fp16, y = var_3825_cast_fp16)[name = string("h_9_cast_fp16")];
            tensor<fp16, [1]> tm_layers_4_layer_scalar_to_fp16 = const()[name = string("tm_layers_4_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.62p-1])];
            tensor<fp16, [1, 512, 512]> x_91_cast_fp16 = mul(x = h_9_cast_fp16, y = tm_layers_4_layer_scalar_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3829_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3829_cast_fp16")];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<int32, [1]> var_3833_axes_0 = const()[name = string("op_3833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3833_cast_fp16 = reduce_max(axes = var_3833_axes_0, keep_dims = var_3832, x = var_3829_cast_fp16)[name = string("op_3833_cast_fp16")];
            fp16 var_3834_to_fp16 = const()[name = string("op_3834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_83_cast_fp16 = maximum(x = var_3833_cast_fp16, y = var_3834_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_83_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_83_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_3837_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3837_cast_fp16")];
            tensor<int32, [1]> var_3842_axes_0 = const()[name = string("op_3842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3842_keep_dims_0 = const()[name = string("op_3842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3842_cast_fp16 = reduce_mean(axes = var_3842_axes_0, keep_dims = var_3842_keep_dims_0, x = var_3837_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3843_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp32 var_3844_epsilon_0 = const()[name = string("op_3844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3844_cast_fp16 = inverse(epsilon = var_3844_epsilon_0, x = var_3843_cast_fp16)[name = string("op_3844_cast_fp16")];
            fp16 var_3845_to_fp16 = const()[name = string("op_3845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3846_cast_fp16 = mul(x = var_3844_cast_fp16, y = var_3845_to_fp16)[name = string("op_3846_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3848_cast_fp16 = add(x = var_3842_cast_fp16, y = var_3846_cast_fp16)[name = string("op_3848_cast_fp16")];
            fp32 var_3849_epsilon_0 = const()[name = string("op_3849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3849_cast_fp16 = rsqrt(epsilon = var_3849_epsilon_0, x = var_3848_cast_fp16)[name = string("op_3849_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_73_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3849_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32684864)))];
            tensor<fp16, [1, 512, 512]> input_61_cast_fp16 = mul(x = out_73_cast_fp16, y = tm_layers_5_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32685952))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33734592))))[name = string("tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_46_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [4]> var_3859 = const()[name = string("op_3859"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 512, 4, 512]> var_3860_cast_fp16 = reshape(shape = var_3859, x = linear_46_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<int32, [4]> x_93_perm_0 = const()[name = string("x_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33738752))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34000960))))[name = string("tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<int32, [4]> var_3871 = const()[name = string("op_3871"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 512, 1, 512]> var_3872_cast_fp16 = reshape(shape = var_3871, x = linear_47_cast_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [4]> x_95_perm_0 = const()[name = string("x_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34002048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34264256))))[name = string("tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [4]> var_3883 = const()[name = string("op_3883"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 512, 1, 512]> var_3884_cast_fp16 = reshape(shape = var_3883, x = linear_48_cast_fp16)[name = string("op_3884_cast_fp16")];
            tensor<int32, [4]> x_97_perm_0 = const()[name = string("x_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 512]> x_93_cast_fp16 = transpose(perm = x_93_perm_0, x = var_3860_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 4, 512, 512]> var_3888_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3888_cast_fp16")];
            bool var_3891 = const()[name = string("op_3891"), val = bool(true)];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_3892_cast_fp16 = reduce_max(axes = var_3892_axes_0, keep_dims = var_3891, x = var_3888_cast_fp16)[name = string("op_3892_cast_fp16")];
            fp16 var_3893_to_fp16 = const()[name = string("op_3893_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_85_cast_fp16 = maximum(x = var_3892_cast_fp16, y = var_3893_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> y_85_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_85_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_3896_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3896_cast_fp16")];
            tensor<int32, [1]> var_3901_axes_0 = const()[name = string("op_3901_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3901_keep_dims_0 = const()[name = string("op_3901_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_3901_cast_fp16 = reduce_mean(axes = var_3901_axes_0, keep_dims = var_3901_keep_dims_0, x = var_3896_cast_fp16)[name = string("op_3901_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_3902_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3902_cast_fp16")];
            fp32 var_3903_epsilon_0 = const()[name = string("op_3903_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_3903_cast_fp16 = inverse(epsilon = var_3903_epsilon_0, x = var_3902_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_3905_cast_fp16 = mul(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_3907_cast_fp16 = add(x = var_3901_cast_fp16, y = var_3905_cast_fp16)[name = string("op_3907_cast_fp16")];
            fp32 var_3908_epsilon_0 = const()[name = string("op_3908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_3908_cast_fp16 = rsqrt(epsilon = var_3908_epsilon_0, x = var_3907_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> out_75_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3908_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34265344)))];
            tensor<fp16, [1, 4, 512, 512]> q_21_cast_fp16 = mul(x = out_75_cast_fp16, y = tm_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> x_95_cast_fp16 = transpose(perm = x_95_perm_0, x = var_3872_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1, 512, 512]> var_3911_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 512, 1]> s_87_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> y_87_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_87_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_3919_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_3925_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 512, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> out_77_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3931_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34266432)))];
            tensor<fp16, [1, 1, 512, 512]> k_31_cast_fp16 = mul(x = out_77_cast_fp16, y = tm_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = var_3884_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 1, 512, 512]> var_3934_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3934_cast_fp16")];
            bool var_3937 = const()[name = string("op_3937"), val = bool(true)];
            tensor<int32, [1]> var_3938_axes_0 = const()[name = string("op_3938_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_3938_cast_fp16 = reduce_max(axes = var_3938_axes_0, keep_dims = var_3937, x = var_3934_cast_fp16)[name = string("op_3938_cast_fp16")];
            fp16 var_3939_to_fp16 = const()[name = string("op_3939_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 512, 1]> s_89_cast_fp16 = maximum(x = var_3938_cast_fp16, y = var_3939_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> y_89_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_89_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_3942_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3942_cast_fp16")];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3947_keep_dims_0 = const()[name = string("op_3947_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> var_3947_cast_fp16 = reduce_mean(axes = var_3947_axes_0, keep_dims = var_3947_keep_dims_0, x = var_3942_cast_fp16)[name = string("op_3947_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_3948_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3948_cast_fp16")];
            fp32 var_3949_epsilon_0 = const()[name = string("op_3949_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 512, 1]> var_3949_cast_fp16 = inverse(epsilon = var_3949_epsilon_0, x = var_3948_cast_fp16)[name = string("op_3949_cast_fp16")];
            fp16 var_3950_to_fp16 = const()[name = string("op_3950_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> var_3951_cast_fp16 = mul(x = var_3949_cast_fp16, y = var_3950_to_fp16)[name = string("op_3951_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_3953_cast_fp16 = add(x = var_3947_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3953_cast_fp16")];
            fp32 var_3954_epsilon_0 = const()[name = string("op_3954_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> var_3954_cast_fp16 = rsqrt(epsilon = var_3954_epsilon_0, x = var_3953_cast_fp16)[name = string("op_3954_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> v_21_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3954_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<fp16, [512, 512]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138235520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138497728))))[name = string("cos_to_fp16_quantized")];
            tensor<fp16, [1, 4, 512, 512]> var_3956_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3956_cast_fp16")];
            tensor<int32, [4]> var_3961_begin_0 = const()[name = string("op_3961_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3961_end_0 = const()[name = string("op_3961_end_0"), val = tensor<int32, [4]>([1, 4, 512, 512])];
            tensor<bool, [4]> var_3961_end_mask_0 = const()[name = string("op_3961_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 256]> var_3961_cast_fp16 = slice_by_index(begin = var_3961_begin_0, end = var_3961_end_0, end_mask = var_3961_end_mask_0, x = q_21_cast_fp16)[name = string("op_3961_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 256]> var_3962_cast_fp16 = mul(x = var_3961_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_3962_cast_fp16")];
            tensor<int32, [4]> var_3967_begin_0 = const()[name = string("op_3967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3967_end_0 = const()[name = string("op_3967_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_3967_end_mask_0 = const()[name = string("op_3967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 256]> var_3967_cast_fp16 = slice_by_index(begin = var_3967_begin_0, end = var_3967_end_0, end_mask = var_3967_end_mask_0, x = q_21_cast_fp16)[name = string("op_3967_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(-1)];
            bool var_3970_interleave_0 = const()[name = string("op_3970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> var_3970_cast_fp16 = concat(axis = var_3969, interleave = var_3970_interleave_0, values = (var_3962_cast_fp16, var_3967_cast_fp16))[name = string("op_3970_cast_fp16")];
            tensor<fp16, [512, 512]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138498816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138761024))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [1, 4, 512, 512]> var_3971_cast_fp16 = mul(x = var_3970_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3971_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> q_23_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3971_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_3974_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3974_cast_fp16")];
            tensor<int32, [4]> var_3979_begin_0 = const()[name = string("op_3979_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3979_end_0 = const()[name = string("op_3979_end_0"), val = tensor<int32, [4]>([1, 1, 512, 512])];
            tensor<bool, [4]> var_3979_end_mask_0 = const()[name = string("op_3979_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 256]> var_3979_cast_fp16 = slice_by_index(begin = var_3979_begin_0, end = var_3979_end_0, end_mask = var_3979_end_mask_0, x = k_31_cast_fp16)[name = string("op_3979_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 256]> var_3980_cast_fp16 = mul(x = var_3979_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_3980_cast_fp16")];
            tensor<int32, [4]> var_3985_begin_0 = const()[name = string("op_3985_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3985_end_0 = const()[name = string("op_3985_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
            tensor<bool, [4]> var_3985_end_mask_0 = const()[name = string("op_3985_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 256]> var_3985_cast_fp16 = slice_by_index(begin = var_3985_begin_0, end = var_3985_end_0, end_mask = var_3985_end_mask_0, x = k_31_cast_fp16)[name = string("op_3985_cast_fp16")];
            int32 var_3987 = const()[name = string("op_3987"), val = int32(-1)];
            bool var_3988_interleave_0 = const()[name = string("op_3988_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 512]> var_3988_cast_fp16 = concat(axis = var_3987, interleave = var_3988_interleave_0, values = (var_3980_cast_fp16, var_3985_cast_fp16))[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_3989_cast_fp16 = mul(x = var_3988_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> k_33_cast_fp16 = add(x = var_3974_cast_fp16, y = var_3989_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 512, 512]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = k_33_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [4, 1, 512, 512]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([4, 1, 1, 512, 512])];
            tensor<fp16, [4, 1, 1, 512, 512]> reshape_20_cast_fp16 = reshape(shape = concat_20, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
            tensor<fp16, [1, 4, 1, 512, 512]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [4, 1, 512, 512]> reshape_21_cast_fp16 = reshape(shape = concat_21, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 512, 512]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = v_21_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [4, 1, 512, 512]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([4, 1, 1, 512, 512])];
            tensor<fp16, [4, 1, 1, 512, 512]> reshape_22_cast_fp16 = reshape(shape = concat_22, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
            tensor<fp16, [1, 4, 1, 512, 512]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [4, 1, 512, 512]> reshape_23_cast_fp16 = reshape(shape = concat_23, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4003_transpose_x_0 = const()[name = string("op_4003_transpose_x_0"), val = bool(false)];
            bool var_4003_transpose_y_0 = const()[name = string("op_4003_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 4, 512, 512]> var_4003_cast_fp16 = matmul(transpose_x = var_4003_transpose_x_0, transpose_y = var_4003_transpose_y_0, x = q_23_cast_fp16, y = transpose_101_cast_fp16)[name = string("op_4003_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_4005_cast_fp16 = add(x = var_4003_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4005_cast_fp16")];
            int32 var_4006 = const()[name = string("op_4006"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_89_cast_fp16 = softmax(axis = var_4006, x = var_4005_cast_fp16)[name = string("w_89_cast_fp16")];
            bool var_4009_transpose_x_0 = const()[name = string("op_4009_transpose_x_0"), val = bool(false)];
            bool var_4009_transpose_y_0 = const()[name = string("op_4009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 4, 512, 512]> var_4009_cast_fp16 = matmul(transpose_x = var_4009_transpose_x_0, transpose_y = var_4009_transpose_y_0, x = w_89_cast_fp16, y = v_23_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4012_perm_0 = const()[name = string("op_4012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4016 = const()[name = string("op_4016"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 512, 4, 512]> var_4012_cast_fp16 = transpose(perm = var_4012_perm_0, x = var_4009_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 512, 2048]> input_63_cast_fp16 = reshape(shape = var_4016, x = var_4012_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34267520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35316160))))[name = string("tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4021_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_4021_cast_fp16")];
            bool var_4024 = const()[name = string("op_4024"), val = bool(true)];
            tensor<int32, [1]> var_4025_axes_0 = const()[name = string("op_4025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4025_cast_fp16 = reduce_max(axes = var_4025_axes_0, keep_dims = var_4024, x = var_4021_cast_fp16)[name = string("op_4025_cast_fp16")];
            fp16 var_4026_to_fp16 = const()[name = string("op_4026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_91_cast_fp16 = maximum(x = var_4025_cast_fp16, y = var_4026_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_91_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_91_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4029_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<int32, [1]> var_4034_axes_0 = const()[name = string("op_4034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4034_keep_dims_0 = const()[name = string("op_4034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4034_cast_fp16 = reduce_mean(axes = var_4034_axes_0, keep_dims = var_4034_keep_dims_0, x = var_4029_cast_fp16)[name = string("op_4034_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4035_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_4035_cast_fp16")];
            fp32 var_4036_epsilon_0 = const()[name = string("op_4036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4036_cast_fp16 = inverse(epsilon = var_4036_epsilon_0, x = var_4035_cast_fp16)[name = string("op_4036_cast_fp16")];
            fp16 var_4037_to_fp16 = const()[name = string("op_4037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4038_cast_fp16 = mul(x = var_4036_cast_fp16, y = var_4037_to_fp16)[name = string("op_4038_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4040_cast_fp16 = add(x = var_4034_cast_fp16, y = var_4038_cast_fp16)[name = string("op_4040_cast_fp16")];
            fp32 var_4041_epsilon_0 = const()[name = string("op_4041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4041_cast_fp16 = rsqrt(epsilon = var_4041_epsilon_0, x = var_4040_cast_fp16)[name = string("op_4041_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_79_cast_fp16 = mul(x = y_91_cast_fp16, y = var_4041_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35317248)))];
            tensor<fp16, [1, 512, 512]> var_4043_cast_fp16 = mul(x = out_79_cast_fp16, y = tm_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_4043_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_101_cast_fp16 = add(x = x_91_cast_fp16, y = var_4043_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4046_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_4046_cast_fp16")];
            bool var_4049 = const()[name = string("op_4049"), val = bool(true)];
            tensor<int32, [1]> var_4050_axes_0 = const()[name = string("op_4050_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4050_cast_fp16 = reduce_max(axes = var_4050_axes_0, keep_dims = var_4049, x = var_4046_cast_fp16)[name = string("op_4050_cast_fp16")];
            fp16 var_4051_to_fp16 = const()[name = string("op_4051_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_93_cast_fp16 = maximum(x = var_4050_cast_fp16, y = var_4051_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_93_cast_fp16 = real_div(x = x_101_cast_fp16, y = s_93_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4054_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [1]> var_4059_axes_0 = const()[name = string("op_4059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4059_keep_dims_0 = const()[name = string("op_4059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4059_cast_fp16 = reduce_mean(axes = var_4059_axes_0, keep_dims = var_4059_keep_dims_0, x = var_4054_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4060_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_4060_cast_fp16")];
            fp32 var_4061_epsilon_0 = const()[name = string("op_4061_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4061_cast_fp16 = inverse(epsilon = var_4061_epsilon_0, x = var_4060_cast_fp16)[name = string("op_4061_cast_fp16")];
            fp16 var_4062_to_fp16 = const()[name = string("op_4062_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4063_cast_fp16 = mul(x = var_4061_cast_fp16, y = var_4062_to_fp16)[name = string("op_4063_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4065_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4063_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp32 var_4066_epsilon_0 = const()[name = string("op_4066_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4066_cast_fp16 = rsqrt(epsilon = var_4066_epsilon_0, x = var_4065_cast_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_81_cast_fp16 = mul(x = y_93_cast_fp16, y = var_4066_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35318336)))];
            tensor<fp16, [1, 512, 512]> input_65_cast_fp16 = mul(x = out_81_cast_fp16, y = tm_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35319424))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36368064))))[name = string("tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_50_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_50_cast_fp16")];
            string var_4073_mode_0 = const()[name = string("op_4073_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_4073_cast_fp16 = gelu(mode = var_4073_mode_0, x = linear_50_cast_fp16)[name = string("op_4073_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36372224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37420864))))[name = string("tm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_51_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_67_cast_fp16 = mul(x = var_4073_cast_fp16, y = linear_51_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37425024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38473664))))[name = string("tm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4081_cast_fp16 = abs(x = linear_52_cast_fp16)[name = string("op_4081_cast_fp16")];
            bool var_4084 = const()[name = string("op_4084"), val = bool(true)];
            tensor<int32, [1]> var_4085_axes_0 = const()[name = string("op_4085_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4085_cast_fp16 = reduce_max(axes = var_4085_axes_0, keep_dims = var_4084, x = var_4081_cast_fp16)[name = string("op_4085_cast_fp16")];
            fp16 var_4086_to_fp16 = const()[name = string("op_4086_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_95_cast_fp16 = maximum(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_95_cast_fp16 = real_div(x = linear_52_cast_fp16, y = s_95_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4089_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_4089_cast_fp16")];
            tensor<int32, [1]> var_4094_axes_0 = const()[name = string("op_4094_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4094_keep_dims_0 = const()[name = string("op_4094_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4094_cast_fp16 = reduce_mean(axes = var_4094_axes_0, keep_dims = var_4094_keep_dims_0, x = var_4089_cast_fp16)[name = string("op_4094_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4095_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_4095_cast_fp16")];
            fp32 var_4096_epsilon_0 = const()[name = string("op_4096_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4096_cast_fp16 = inverse(epsilon = var_4096_epsilon_0, x = var_4095_cast_fp16)[name = string("op_4096_cast_fp16")];
            fp16 var_4097_to_fp16 = const()[name = string("op_4097_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4098_cast_fp16 = mul(x = var_4096_cast_fp16, y = var_4097_to_fp16)[name = string("op_4098_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4100_cast_fp16 = add(x = var_4094_cast_fp16, y = var_4098_cast_fp16)[name = string("op_4100_cast_fp16")];
            fp32 var_4101_epsilon_0 = const()[name = string("op_4101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4101_cast_fp16 = rsqrt(epsilon = var_4101_epsilon_0, x = var_4100_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_83_cast_fp16 = mul(x = y_95_cast_fp16, y = var_4101_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38474752)))];
            tensor<fp16, [1, 512, 512]> var_4103_cast_fp16 = mul(x = out_83_cast_fp16, y = tm_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_69_cast_fp16 = add(x = x_101_cast_fp16, y = var_4103_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38475840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38738048))))[name = string("tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_53_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_53_cast_fp16")];
            string x_105_mode_0 = const()[name = string("x_105_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_105_cast_fp16 = gelu(mode = x_105_mode_0, x = linear_53_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [4]> var_4123_begin_0 = const()[name = string("op_4123_begin_0"), val = tensor<int32, [4]>([0, 0, 5, 0])];
            tensor<int32, [4]> var_4123_end_0 = const()[name = string("op_4123_end_0"), val = tensor<int32, [4]>([1, 512, 6, 512])];
            tensor<bool, [4]> var_4123_end_mask_0 = const()[name = string("op_4123_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4123_squeeze_mask_0 = const()[name = string("op_4123_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_4123_cast_fp16 = slice_by_index(begin = var_4123_begin_0, end = var_4123_end_0, end_mask = var_4123_end_mask_0, squeeze_mask = var_4123_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_71_cast_fp16 = mul(x = x_105_cast_fp16, y = var_4123_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38739136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39001344))))[name = string("tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4133_cast_fp16 = abs(x = linear_54_cast_fp16)[name = string("op_4133_cast_fp16")];
            bool var_4136 = const()[name = string("op_4136"), val = bool(true)];
            tensor<int32, [1]> var_4137_axes_0 = const()[name = string("op_4137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4137_cast_fp16 = reduce_max(axes = var_4137_axes_0, keep_dims = var_4136, x = var_4133_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 var_4138_to_fp16 = const()[name = string("op_4138_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_97_cast_fp16 = maximum(x = var_4137_cast_fp16, y = var_4138_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_97_cast_fp16 = real_div(x = linear_54_cast_fp16, y = s_97_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4141_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_4141_cast_fp16")];
            tensor<int32, [1]> var_4146_axes_0 = const()[name = string("op_4146_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4146_keep_dims_0 = const()[name = string("op_4146_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4146_cast_fp16 = reduce_mean(axes = var_4146_axes_0, keep_dims = var_4146_keep_dims_0, x = var_4141_cast_fp16)[name = string("op_4146_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4147_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_4147_cast_fp16")];
            fp32 var_4148_epsilon_0 = const()[name = string("op_4148_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4148_cast_fp16 = inverse(epsilon = var_4148_epsilon_0, x = var_4147_cast_fp16)[name = string("op_4148_cast_fp16")];
            fp16 var_4149_to_fp16 = const()[name = string("op_4149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4150_cast_fp16 = mul(x = var_4148_cast_fp16, y = var_4149_to_fp16)[name = string("op_4150_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4152_cast_fp16 = add(x = var_4146_cast_fp16, y = var_4150_cast_fp16)[name = string("op_4152_cast_fp16")];
            fp32 var_4153_epsilon_0 = const()[name = string("op_4153_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4153_cast_fp16 = rsqrt(epsilon = var_4153_epsilon_0, x = var_4152_cast_fp16)[name = string("op_4153_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_85_cast_fp16 = mul(x = y_97_cast_fp16, y = var_4153_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39002432)))];
            tensor<fp16, [1, 512, 512]> var_4155_cast_fp16 = mul(x = out_85_cast_fp16, y = tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4155_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_11_cast_fp16 = add(x = input_69_cast_fp16, y = var_4155_cast_fp16)[name = string("h_11_cast_fp16")];
            tensor<fp16, [1]> tm_layers_5_layer_scalar_to_fp16 = const()[name = string("tm_layers_5_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-2])];
            tensor<fp16, [1, 512, 512]> x_109_cast_fp16 = mul(x = h_11_cast_fp16, y = tm_layers_5_layer_scalar_to_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4159_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_4159_cast_fp16")];
            bool var_4162 = const()[name = string("op_4162"), val = bool(true)];
            tensor<int32, [1]> var_4163_axes_0 = const()[name = string("op_4163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4163_cast_fp16 = reduce_max(axes = var_4163_axes_0, keep_dims = var_4162, x = var_4159_cast_fp16)[name = string("op_4163_cast_fp16")];
            fp16 var_4164_to_fp16 = const()[name = string("op_4164_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_99_cast_fp16 = maximum(x = var_4163_cast_fp16, y = var_4164_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_99_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_99_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4167_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_4167_cast_fp16")];
            tensor<int32, [1]> var_4172_axes_0 = const()[name = string("op_4172_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4172_keep_dims_0 = const()[name = string("op_4172_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4172_cast_fp16 = reduce_mean(axes = var_4172_axes_0, keep_dims = var_4172_keep_dims_0, x = var_4167_cast_fp16)[name = string("op_4172_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4173_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4174_cast_fp16 = inverse(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            fp16 var_4175_to_fp16 = const()[name = string("op_4175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4176_cast_fp16 = mul(x = var_4174_cast_fp16, y = var_4175_to_fp16)[name = string("op_4176_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4178_cast_fp16 = add(x = var_4172_cast_fp16, y = var_4176_cast_fp16)[name = string("op_4178_cast_fp16")];
            fp32 var_4179_epsilon_0 = const()[name = string("op_4179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4179_cast_fp16 = rsqrt(epsilon = var_4179_epsilon_0, x = var_4178_cast_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_87_cast_fp16 = mul(x = y_99_cast_fp16, y = var_4179_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39003520)))];
            tensor<fp16, [1, 512, 512]> input_73_cast_fp16 = mul(x = out_87_cast_fp16, y = tm_layers_6_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39004608))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39528960))))[name = string("tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [4]> var_4189 = const()[name = string("op_4189"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_4190_cast_fp16 = reshape(shape = var_4189, x = linear_55_cast_fp16)[name = string("op_4190_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39531072))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39793280))))[name = string("tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_4201 = const()[name = string("op_4201"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_4202_cast_fp16 = reshape(shape = var_4201, x = linear_56_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39794368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40056576))))[name = string("tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_57_cast_fp16)[name = string("op_4214_cast_fp16")];
            tensor<int32, [4]> x_115_perm_0 = const()[name = string("x_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4190_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 4, 512, 256]> var_4218_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_4218_cast_fp16")];
            bool var_4221 = const()[name = string("op_4221"), val = bool(true)];
            tensor<int32, [1]> var_4222_axes_0 = const()[name = string("op_4222_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_4222_cast_fp16 = reduce_max(axes = var_4222_axes_0, keep_dims = var_4221, x = var_4218_cast_fp16)[name = string("op_4222_cast_fp16")];
            fp16 var_4223_to_fp16 = const()[name = string("op_4223_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_101_cast_fp16 = maximum(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_101_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_101_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4226_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [1]> var_4231_axes_0 = const()[name = string("op_4231_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4231_keep_dims_0 = const()[name = string("op_4231_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_4231_cast_fp16 = reduce_mean(axes = var_4231_axes_0, keep_dims = var_4231_keep_dims_0, x = var_4226_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_4232_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_4232_cast_fp16")];
            fp32 var_4233_epsilon_0 = const()[name = string("op_4233_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_4233_cast_fp16 = inverse(epsilon = var_4233_epsilon_0, x = var_4232_cast_fp16)[name = string("op_4233_cast_fp16")];
            fp16 var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_4235_cast_fp16 = mul(x = var_4233_cast_fp16, y = var_4234_to_fp16)[name = string("op_4235_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_4237_cast_fp16 = add(x = var_4231_cast_fp16, y = var_4235_cast_fp16)[name = string("op_4237_cast_fp16")];
            fp32 var_4238_epsilon_0 = const()[name = string("op_4238_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_4238_cast_fp16 = rsqrt(epsilon = var_4238_epsilon_0, x = var_4237_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_89_cast_fp16 = mul(x = y_101_cast_fp16, y = var_4238_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40057664)))];
            tensor<fp16, [1, 4, 512, 256]> q_25_cast_fp16 = mul(x = out_89_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4202_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 2, 512, 256]> var_4241_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_4241_cast_fp16")];
            bool var_4244 = const()[name = string("op_4244"), val = bool(true)];
            tensor<int32, [1]> var_4245_axes_0 = const()[name = string("op_4245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_4245_cast_fp16 = reduce_max(axes = var_4245_axes_0, keep_dims = var_4244, x = var_4241_cast_fp16)[name = string("op_4245_cast_fp16")];
            fp16 var_4246_to_fp16 = const()[name = string("op_4246_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_103_cast_fp16 = maximum(x = var_4245_cast_fp16, y = var_4246_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_103_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_103_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4249_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<int32, [1]> var_4254_axes_0 = const()[name = string("op_4254_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4254_keep_dims_0 = const()[name = string("op_4254_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_4254_cast_fp16 = reduce_mean(axes = var_4254_axes_0, keep_dims = var_4254_keep_dims_0, x = var_4249_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4255_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_4255_cast_fp16")];
            fp32 var_4256_epsilon_0 = const()[name = string("op_4256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_4256_cast_fp16 = inverse(epsilon = var_4256_epsilon_0, x = var_4255_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp16 var_4257_to_fp16 = const()[name = string("op_4257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_4258_cast_fp16 = mul(x = var_4256_cast_fp16, y = var_4257_to_fp16)[name = string("op_4258_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4260_cast_fp16 = add(x = var_4254_cast_fp16, y = var_4258_cast_fp16)[name = string("op_4260_cast_fp16")];
            fp32 var_4261_epsilon_0 = const()[name = string("op_4261_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_4261_cast_fp16 = rsqrt(epsilon = var_4261_epsilon_0, x = var_4260_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_91_cast_fp16 = mul(x = y_103_cast_fp16, y = var_4261_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058240)))];
            tensor<fp16, [1, 2, 512, 256]> k_37_cast_fp16 = mul(x = out_91_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_37_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_115_cast_fp16 = transpose(perm = x_115_perm_0, x = var_4214_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 2, 512, 256]> var_4264_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_4264_cast_fp16")];
            bool var_4267 = const()[name = string("op_4267"), val = bool(true)];
            tensor<int32, [1]> var_4268_axes_0 = const()[name = string("op_4268_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_4268_cast_fp16 = reduce_max(axes = var_4268_axes_0, keep_dims = var_4267, x = var_4264_cast_fp16)[name = string("op_4268_cast_fp16")];
            fp16 var_4269_to_fp16 = const()[name = string("op_4269_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_105_cast_fp16 = maximum(x = var_4268_cast_fp16, y = var_4269_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_105_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_105_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4272_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_4272_cast_fp16")];
            tensor<int32, [1]> var_4277_axes_0 = const()[name = string("op_4277_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4277_keep_dims_0 = const()[name = string("op_4277_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_4277_cast_fp16 = reduce_mean(axes = var_4277_axes_0, keep_dims = var_4277_keep_dims_0, x = var_4272_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4278_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp32 var_4279_epsilon_0 = const()[name = string("op_4279_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_4279_cast_fp16 = inverse(epsilon = var_4279_epsilon_0, x = var_4278_cast_fp16)[name = string("op_4279_cast_fp16")];
            fp16 var_4280_to_fp16 = const()[name = string("op_4280_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_4281_cast_fp16 = mul(x = var_4279_cast_fp16, y = var_4280_to_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4283_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4281_cast_fp16)[name = string("op_4283_cast_fp16")];
            fp32 var_4284_epsilon_0 = const()[name = string("op_4284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_4284_cast_fp16 = rsqrt(epsilon = var_4284_epsilon_0, x = var_4283_cast_fp16)[name = string("op_4284_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_25_cast_fp16 = mul(x = y_105_cast_fp16, y = var_4284_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4286_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4286_cast_fp16")];
            tensor<int32, [4]> var_4291_begin_0 = const()[name = string("op_4291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4291_end_0 = const()[name = string("op_4291_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_4291_end_mask_0 = const()[name = string("op_4291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_4291_cast_fp16 = slice_by_index(begin = var_4291_begin_0, end = var_4291_end_0, end_mask = var_4291_end_mask_0, x = q_25_cast_fp16)[name = string("op_4291_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_4292_cast_fp16 = mul(x = var_4291_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = q_25_cast_fp16)[name = string("op_4297_cast_fp16")];
            int32 var_4299 = const()[name = string("op_4299"), val = int32(-1)];
            bool var_4300_interleave_0 = const()[name = string("op_4300_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_4300_cast_fp16 = concat(axis = var_4299, interleave = var_4300_interleave_0, values = (var_4292_cast_fp16, var_4297_cast_fp16))[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4301_cast_fp16 = mul(x = var_4300_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4301_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_27_cast_fp16 = add(x = var_4286_cast_fp16, y = var_4301_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4304_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4304_cast_fp16")];
            tensor<int32, [4]> var_4309_begin_0 = const()[name = string("op_4309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4309_end_0 = const()[name = string("op_4309_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_4309_end_mask_0 = const()[name = string("op_4309_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_4309_cast_fp16 = slice_by_index(begin = var_4309_begin_0, end = var_4309_end_0, end_mask = var_4309_end_mask_0, x = k_37_cast_fp16)[name = string("op_4309_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_4310_cast_fp16 = mul(x = var_4309_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_4310_cast_fp16")];
            tensor<int32, [4]> var_4315_begin_0 = const()[name = string("op_4315_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4315_end_0 = const()[name = string("op_4315_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_4315_end_mask_0 = const()[name = string("op_4315_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_4315_cast_fp16 = slice_by_index(begin = var_4315_begin_0, end = var_4315_end_0, end_mask = var_4315_end_mask_0, x = k_37_cast_fp16)[name = string("op_4315_cast_fp16")];
            int32 var_4317 = const()[name = string("op_4317"), val = int32(-1)];
            bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_4318_cast_fp16 = concat(axis = var_4317, interleave = var_4318_interleave_0, values = (var_4310_cast_fp16, var_4315_cast_fp16))[name = string("op_4318_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4319_cast_fp16 = mul(x = var_4318_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_39_cast_fp16 = add(x = var_4304_cast_fp16, y = var_4319_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = k_39_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [4, 1, 512, 256]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
            tensor<int32, [5]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_24_cast_fp16 = reshape(shape = concat_24, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [4, 1, 512, 256]> reshape_25_cast_fp16 = reshape(shape = concat_25, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = v_25_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [4, 1, 512, 256]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
            tensor<int32, [5]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_26_cast_fp16 = reshape(shape = concat_26, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [4, 1, 512, 256]> reshape_27_cast_fp16 = reshape(shape = concat_27, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4333_transpose_x_0 = const()[name = string("op_4333_transpose_x_0"), val = bool(false)];
            bool var_4333_transpose_y_0 = const()[name = string("op_4333_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_102_cast_fp16 = transpose(perm = transpose_102_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 4, 512, 512]> var_4333_cast_fp16 = matmul(transpose_x = var_4333_transpose_x_0, transpose_y = var_4333_transpose_y_0, x = q_27_cast_fp16, y = transpose_102_cast_fp16)[name = string("op_4333_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_4335_cast_fp16 = add(x = var_4333_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4335_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_105_cast_fp16 = softmax(axis = var_4336, x = var_4335_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4339_transpose_x_0 = const()[name = string("op_4339_transpose_x_0"), val = bool(false)];
            bool var_4339_transpose_y_0 = const()[name = string("op_4339_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 4, 512, 256]> var_4339_cast_fp16 = matmul(transpose_x = var_4339_transpose_x_0, transpose_y = var_4339_transpose_y_0, x = w_105_cast_fp16, y = v_27_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<int32, [4]> var_4342_perm_0 = const()[name = string("op_4342_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4346 = const()[name = string("op_4346"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_4342_cast_fp16 = transpose(perm = var_4342_perm_0, x = var_4339_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 512, 1024]> input_75_cast_fp16 = reshape(shape = var_4346, x = var_4342_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40583168))))[name = string("tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4351_cast_fp16 = abs(x = linear_58_cast_fp16)[name = string("op_4351_cast_fp16")];
            bool var_4354 = const()[name = string("op_4354"), val = bool(true)];
            tensor<int32, [1]> var_4355_axes_0 = const()[name = string("op_4355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4355_cast_fp16 = reduce_max(axes = var_4355_axes_0, keep_dims = var_4354, x = var_4351_cast_fp16)[name = string("op_4355_cast_fp16")];
            fp16 var_4356_to_fp16 = const()[name = string("op_4356_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_107_cast_fp16 = maximum(x = var_4355_cast_fp16, y = var_4356_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_107_cast_fp16 = real_div(x = linear_58_cast_fp16, y = s_107_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4359_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<int32, [1]> var_4364_axes_0 = const()[name = string("op_4364_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4364_keep_dims_0 = const()[name = string("op_4364_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4364_cast_fp16 = reduce_mean(axes = var_4364_axes_0, keep_dims = var_4364_keep_dims_0, x = var_4359_cast_fp16)[name = string("op_4364_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4365_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_4365_cast_fp16")];
            fp32 var_4366_epsilon_0 = const()[name = string("op_4366_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4366_cast_fp16 = inverse(epsilon = var_4366_epsilon_0, x = var_4365_cast_fp16)[name = string("op_4366_cast_fp16")];
            fp16 var_4367_to_fp16 = const()[name = string("op_4367_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4368_cast_fp16 = mul(x = var_4366_cast_fp16, y = var_4367_to_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4370_cast_fp16 = add(x = var_4364_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            fp32 var_4371_epsilon_0 = const()[name = string("op_4371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4371_cast_fp16 = rsqrt(epsilon = var_4371_epsilon_0, x = var_4370_cast_fp16)[name = string("op_4371_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_93_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4371_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40584256)))];
            tensor<fp16, [1, 512, 512]> var_4373_cast_fp16 = mul(x = out_93_cast_fp16, y = tm_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_4373_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = var_4373_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4376_cast_fp16 = abs(x = x_119_cast_fp16)[name = string("op_4376_cast_fp16")];
            bool var_4379 = const()[name = string("op_4379"), val = bool(true)];
            tensor<int32, [1]> var_4380_axes_0 = const()[name = string("op_4380_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4380_cast_fp16 = reduce_max(axes = var_4380_axes_0, keep_dims = var_4379, x = var_4376_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 var_4381_to_fp16 = const()[name = string("op_4381_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_109_cast_fp16 = maximum(x = var_4380_cast_fp16, y = var_4381_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_109_cast_fp16 = real_div(x = x_119_cast_fp16, y = s_109_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4384_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4384_cast_fp16")];
            tensor<int32, [1]> var_4389_axes_0 = const()[name = string("op_4389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4389_keep_dims_0 = const()[name = string("op_4389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4389_cast_fp16 = reduce_mean(axes = var_4389_axes_0, keep_dims = var_4389_keep_dims_0, x = var_4384_cast_fp16)[name = string("op_4389_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4390_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_4390_cast_fp16")];
            fp32 var_4391_epsilon_0 = const()[name = string("op_4391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4391_cast_fp16 = inverse(epsilon = var_4391_epsilon_0, x = var_4390_cast_fp16)[name = string("op_4391_cast_fp16")];
            fp16 var_4392_to_fp16 = const()[name = string("op_4392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = string("op_4393_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4395_cast_fp16 = add(x = var_4389_cast_fp16, y = var_4393_cast_fp16)[name = string("op_4395_cast_fp16")];
            fp32 var_4396_epsilon_0 = const()[name = string("op_4396_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4396_cast_fp16 = rsqrt(epsilon = var_4396_epsilon_0, x = var_4395_cast_fp16)[name = string("op_4396_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_95_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4396_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40585344)))];
            tensor<fp16, [1, 512, 512]> input_77_cast_fp16 = mul(x = out_95_cast_fp16, y = tm_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40586432))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41635072))))[name = string("tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_59_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_59_cast_fp16")];
            string var_4403_mode_0 = const()[name = string("op_4403_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_4403_cast_fp16 = gelu(mode = var_4403_mode_0, x = linear_59_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41639232))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42687872))))[name = string("tm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_60_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_79_cast_fp16 = mul(x = var_4403_cast_fp16, y = linear_60_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42692032))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43740672))))[name = string("tm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4411_cast_fp16 = abs(x = linear_61_cast_fp16)[name = string("op_4411_cast_fp16")];
            bool var_4414 = const()[name = string("op_4414"), val = bool(true)];
            tensor<int32, [1]> var_4415_axes_0 = const()[name = string("op_4415_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4415_cast_fp16 = reduce_max(axes = var_4415_axes_0, keep_dims = var_4414, x = var_4411_cast_fp16)[name = string("op_4415_cast_fp16")];
            fp16 var_4416_to_fp16 = const()[name = string("op_4416_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_111_cast_fp16 = maximum(x = var_4415_cast_fp16, y = var_4416_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_111_cast_fp16 = real_div(x = linear_61_cast_fp16, y = s_111_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4419_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4419_cast_fp16")];
            tensor<int32, [1]> var_4424_axes_0 = const()[name = string("op_4424_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4424_keep_dims_0 = const()[name = string("op_4424_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4424_cast_fp16 = reduce_mean(axes = var_4424_axes_0, keep_dims = var_4424_keep_dims_0, x = var_4419_cast_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4425_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4426_cast_fp16 = inverse(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            fp16 var_4427_to_fp16 = const()[name = string("op_4427_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4428_cast_fp16 = mul(x = var_4426_cast_fp16, y = var_4427_to_fp16)[name = string("op_4428_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4430_cast_fp16 = add(x = var_4424_cast_fp16, y = var_4428_cast_fp16)[name = string("op_4430_cast_fp16")];
            fp32 var_4431_epsilon_0 = const()[name = string("op_4431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4431_cast_fp16 = rsqrt(epsilon = var_4431_epsilon_0, x = var_4430_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_97_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4431_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43741760)))];
            tensor<fp16, [1, 512, 512]> var_4433_cast_fp16 = mul(x = out_97_cast_fp16, y = tm_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4433_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_81_cast_fp16 = add(x = x_119_cast_fp16, y = var_4433_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43742848))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44005056))))[name = string("tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_62_cast_fp16")];
            string x_123_mode_0 = const()[name = string("x_123_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_123_cast_fp16 = gelu(mode = x_123_mode_0, x = linear_62_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [4]> var_4453_begin_0 = const()[name = string("op_4453_begin_0"), val = tensor<int32, [4]>([0, 0, 6, 0])];
            tensor<int32, [4]> var_4453_end_0 = const()[name = string("op_4453_end_0"), val = tensor<int32, [4]>([1, 512, 7, 512])];
            tensor<bool, [4]> var_4453_end_mask_0 = const()[name = string("op_4453_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4453_squeeze_mask_0 = const()[name = string("op_4453_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_4453_cast_fp16 = slice_by_index(begin = var_4453_begin_0, end = var_4453_end_0, end_mask = var_4453_end_mask_0, squeeze_mask = var_4453_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_83_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4453_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44006144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44268352))))[name = string("tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4463_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4463_cast_fp16")];
            bool var_4466 = const()[name = string("op_4466"), val = bool(true)];
            tensor<int32, [1]> var_4467_axes_0 = const()[name = string("op_4467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4467_cast_fp16 = reduce_max(axes = var_4467_axes_0, keep_dims = var_4466, x = var_4463_cast_fp16)[name = string("op_4467_cast_fp16")];
            fp16 var_4468_to_fp16 = const()[name = string("op_4468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_113_cast_fp16 = maximum(x = var_4467_cast_fp16, y = var_4468_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_113_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_113_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4471_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [1]> var_4476_axes_0 = const()[name = string("op_4476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4476_keep_dims_0 = const()[name = string("op_4476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4476_cast_fp16 = reduce_mean(axes = var_4476_axes_0, keep_dims = var_4476_keep_dims_0, x = var_4471_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4477_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_fp16)[name = string("op_4477_cast_fp16")];
            fp32 var_4478_epsilon_0 = const()[name = string("op_4478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4478_cast_fp16 = inverse(epsilon = var_4478_epsilon_0, x = var_4477_cast_fp16)[name = string("op_4478_cast_fp16")];
            fp16 var_4479_to_fp16 = const()[name = string("op_4479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4480_cast_fp16 = mul(x = var_4478_cast_fp16, y = var_4479_to_fp16)[name = string("op_4480_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4482_cast_fp16 = add(x = var_4476_cast_fp16, y = var_4480_cast_fp16)[name = string("op_4482_cast_fp16")];
            fp32 var_4483_epsilon_0 = const()[name = string("op_4483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4483_cast_fp16 = rsqrt(epsilon = var_4483_epsilon_0, x = var_4482_cast_fp16)[name = string("op_4483_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_99_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4483_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44269440)))];
            tensor<fp16, [1, 512, 512]> var_4485_cast_fp16 = mul(x = out_99_cast_fp16, y = tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_13_cast_fp16 = add(x = input_81_cast_fp16, y = var_4485_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<fp16, [1]> tm_layers_6_layer_scalar_to_fp16 = const()[name = string("tm_layers_6_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.cep-3])];
            tensor<fp16, [1, 512, 512]> x_127_cast_fp16 = mul(x = h_13_cast_fp16, y = tm_layers_6_layer_scalar_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4489_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_4489_cast_fp16")];
            bool var_4492 = const()[name = string("op_4492"), val = bool(true)];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4493_cast_fp16 = reduce_max(axes = var_4493_axes_0, keep_dims = var_4492, x = var_4489_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_115_cast_fp16 = maximum(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_115_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_115_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4497_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<int32, [1]> var_4502_axes_0 = const()[name = string("op_4502_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4502_keep_dims_0 = const()[name = string("op_4502_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4502_cast_fp16 = reduce_mean(axes = var_4502_axes_0, keep_dims = var_4502_keep_dims_0, x = var_4497_cast_fp16)[name = string("op_4502_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4503_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4504_cast_fp16 = inverse(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            fp16 var_4505_to_fp16 = const()[name = string("op_4505_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4506_cast_fp16 = mul(x = var_4504_cast_fp16, y = var_4505_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4508_cast_fp16 = add(x = var_4502_cast_fp16, y = var_4506_cast_fp16)[name = string("op_4508_cast_fp16")];
            fp32 var_4509_epsilon_0 = const()[name = string("op_4509_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4509_cast_fp16 = rsqrt(epsilon = var_4509_epsilon_0, x = var_4508_cast_fp16)[name = string("op_4509_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_101_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4509_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44270528)))];
            tensor<fp16, [1, 512, 512]> input_85_cast_fp16 = mul(x = out_101_cast_fp16, y = tm_layers_7_input_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44271616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44795968))))[name = string("tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4519 = const()[name = string("op_4519"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_4520_cast_fp16 = reshape(shape = var_4519, x = linear_64_cast_fp16)[name = string("op_4520_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44798080))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45060288))))[name = string("tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4531 = const()[name = string("op_4531"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_4532_cast_fp16 = reshape(shape = var_4531, x = linear_65_cast_fp16)[name = string("op_4532_cast_fp16")];
            tensor<int32, [4]> x_131_perm_0 = const()[name = string("x_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45061376))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45323584))))[name = string("tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4543 = const()[name = string("op_4543"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_4544_cast_fp16 = reshape(shape = var_4543, x = linear_66_cast_fp16)[name = string("op_4544_cast_fp16")];
            tensor<int32, [4]> x_133_perm_0 = const()[name = string("x_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_4520_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 4, 512, 256]> var_4548_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_4548_cast_fp16")];
            bool var_4551 = const()[name = string("op_4551"), val = bool(true)];
            tensor<int32, [1]> var_4552_axes_0 = const()[name = string("op_4552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_4552_cast_fp16 = reduce_max(axes = var_4552_axes_0, keep_dims = var_4551, x = var_4548_cast_fp16)[name = string("op_4552_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_117_cast_fp16 = maximum(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_117_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_117_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4556_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<int32, [1]> var_4561_axes_0 = const()[name = string("op_4561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4561_keep_dims_0 = const()[name = string("op_4561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_4561_cast_fp16 = reduce_mean(axes = var_4561_axes_0, keep_dims = var_4561_keep_dims_0, x = var_4556_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_4562_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4562_cast_fp16")];
            fp32 var_4563_epsilon_0 = const()[name = string("op_4563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_4563_cast_fp16 = inverse(epsilon = var_4563_epsilon_0, x = var_4562_cast_fp16)[name = string("op_4563_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_4565_cast_fp16 = mul(x = var_4563_cast_fp16, y = var_4564_to_fp16)[name = string("op_4565_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_4567_cast_fp16 = add(x = var_4561_cast_fp16, y = var_4565_cast_fp16)[name = string("op_4567_cast_fp16")];
            fp32 var_4568_epsilon_0 = const()[name = string("op_4568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_4568_cast_fp16 = rsqrt(epsilon = var_4568_epsilon_0, x = var_4567_cast_fp16)[name = string("op_4568_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_103_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4568_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_29_cast_fp16 = mul(x = out_103_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_131_cast_fp16 = transpose(perm = x_131_perm_0, x = var_4532_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2, 512, 256]> var_4571_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_4571_cast_fp16")];
            bool var_4574 = const()[name = string("op_4574"), val = bool(true)];
            tensor<int32, [1]> var_4575_axes_0 = const()[name = string("op_4575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_4575_cast_fp16 = reduce_max(axes = var_4575_axes_0, keep_dims = var_4574, x = var_4571_cast_fp16)[name = string("op_4575_cast_fp16")];
            fp16 var_4576_to_fp16 = const()[name = string("op_4576_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_119_cast_fp16 = maximum(x = var_4575_cast_fp16, y = var_4576_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_119_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_119_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4579_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4579_cast_fp16")];
            tensor<int32, [1]> var_4584_axes_0 = const()[name = string("op_4584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4584_keep_dims_0 = const()[name = string("op_4584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_4584_cast_fp16 = reduce_mean(axes = var_4584_axes_0, keep_dims = var_4584_keep_dims_0, x = var_4579_cast_fp16)[name = string("op_4584_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4585_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4585_cast_fp16")];
            fp32 var_4586_epsilon_0 = const()[name = string("op_4586_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_4586_cast_fp16 = inverse(epsilon = var_4586_epsilon_0, x = var_4585_cast_fp16)[name = string("op_4586_cast_fp16")];
            fp16 var_4587_to_fp16 = const()[name = string("op_4587_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_4588_cast_fp16 = mul(x = var_4586_cast_fp16, y = var_4587_to_fp16)[name = string("op_4588_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4590_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4588_cast_fp16)[name = string("op_4590_cast_fp16")];
            fp32 var_4591_epsilon_0 = const()[name = string("op_4591_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_4591_cast_fp16 = rsqrt(epsilon = var_4591_epsilon_0, x = var_4590_cast_fp16)[name = string("op_4591_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_105_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4591_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [256]> tm_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45324672)))];
            tensor<fp16, [1, 2, 512, 256]> k_43_cast_fp16 = mul(x = out_105_cast_fp16, y = tm_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_133_cast_fp16 = transpose(perm = x_133_perm_0, x = var_4544_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 2, 512, 256]> var_4594_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_4594_cast_fp16")];
            bool var_4597 = const()[name = string("op_4597"), val = bool(true)];
            tensor<int32, [1]> var_4598_axes_0 = const()[name = string("op_4598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_4598_cast_fp16 = reduce_max(axes = var_4598_axes_0, keep_dims = var_4597, x = var_4594_cast_fp16)[name = string("op_4598_cast_fp16")];
            fp16 var_4599_to_fp16 = const()[name = string("op_4599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_121_cast_fp16 = maximum(x = var_4598_cast_fp16, y = var_4599_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_121_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_121_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4602_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4602_cast_fp16")];
            tensor<int32, [1]> var_4607_axes_0 = const()[name = string("op_4607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4607_keep_dims_0 = const()[name = string("op_4607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_4607_cast_fp16 = reduce_mean(axes = var_4607_axes_0, keep_dims = var_4607_keep_dims_0, x = var_4602_cast_fp16)[name = string("op_4607_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4608_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4608_cast_fp16")];
            fp32 var_4609_epsilon_0 = const()[name = string("op_4609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_4609_cast_fp16 = inverse(epsilon = var_4609_epsilon_0, x = var_4608_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_4611_cast_fp16 = mul(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("op_4611_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4613_cast_fp16 = add(x = var_4607_cast_fp16, y = var_4611_cast_fp16)[name = string("op_4613_cast_fp16")];
            fp32 var_4614_epsilon_0 = const()[name = string("op_4614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_4614_cast_fp16 = rsqrt(epsilon = var_4614_epsilon_0, x = var_4613_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_29_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4614_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4616_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4616_cast_fp16")];
            tensor<int32, [4]> var_4621_begin_0 = const()[name = string("op_4621_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4621_end_0 = const()[name = string("op_4621_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_4621_end_mask_0 = const()[name = string("op_4621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_4621_cast_fp16 = slice_by_index(begin = var_4621_begin_0, end = var_4621_end_0, end_mask = var_4621_end_mask_0, x = q_29_cast_fp16)[name = string("op_4621_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_4622_cast_fp16 = mul(x = var_4621_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<int32, [4]> var_4627_begin_0 = const()[name = string("op_4627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4627_end_0 = const()[name = string("op_4627_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_4627_end_mask_0 = const()[name = string("op_4627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_4627_cast_fp16 = slice_by_index(begin = var_4627_begin_0, end = var_4627_end_0, end_mask = var_4627_end_mask_0, x = q_29_cast_fp16)[name = string("op_4627_cast_fp16")];
            int32 var_4629 = const()[name = string("op_4629"), val = int32(-1)];
            bool var_4630_interleave_0 = const()[name = string("op_4630_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_4630_cast_fp16 = concat(axis = var_4629, interleave = var_4630_interleave_0, values = (var_4622_cast_fp16, var_4627_cast_fp16))[name = string("op_4630_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4631_cast_fp16 = mul(x = var_4630_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4631_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_31_cast_fp16 = add(x = var_4616_cast_fp16, y = var_4631_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4634_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4634_cast_fp16")];
            tensor<int32, [4]> var_4639_begin_0 = const()[name = string("op_4639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4639_end_0 = const()[name = string("op_4639_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_4639_end_mask_0 = const()[name = string("op_4639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_4639_cast_fp16 = slice_by_index(begin = var_4639_begin_0, end = var_4639_end_0, end_mask = var_4639_end_mask_0, x = k_43_cast_fp16)[name = string("op_4639_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_4640_cast_fp16 = mul(x = var_4639_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4640_cast_fp16")];
            tensor<int32, [4]> var_4645_begin_0 = const()[name = string("op_4645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4645_end_0 = const()[name = string("op_4645_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_4645_end_mask_0 = const()[name = string("op_4645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_4645_cast_fp16 = slice_by_index(begin = var_4645_begin_0, end = var_4645_end_0, end_mask = var_4645_end_mask_0, x = k_43_cast_fp16)[name = string("op_4645_cast_fp16")];
            int32 var_4647 = const()[name = string("op_4647"), val = int32(-1)];
            bool var_4648_interleave_0 = const()[name = string("op_4648_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_4648_cast_fp16 = concat(axis = var_4647, interleave = var_4648_interleave_0, values = (var_4640_cast_fp16, var_4645_cast_fp16))[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4649_cast_fp16 = mul(x = var_4648_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4649_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_45_cast_fp16 = add(x = var_4634_cast_fp16, y = var_4649_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_14_reps_0 = const()[name = string("tile_14_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = k_45_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [4, 1, 512, 256]> tile_14_cast_fp16 = tile(reps = tile_14_reps_0, x = transpose_28_cast_fp16)[name = string("tile_14_cast_fp16")];
            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_28_cast_fp16 = reshape(shape = concat_28, x = tile_14_cast_fp16)[name = string("reshape_28_cast_fp16")];
            tensor<int32, [5]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_28_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [4, 1, 512, 256]> reshape_29_cast_fp16 = reshape(shape = concat_29, x = transpose_29_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_15_reps_0 = const()[name = string("tile_15_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = v_29_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [4, 1, 512, 256]> tile_15_cast_fp16 = tile(reps = tile_15_reps_0, x = transpose_30_cast_fp16)[name = string("tile_15_cast_fp16")];
            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_30_cast_fp16 = reshape(shape = concat_30, x = tile_15_cast_fp16)[name = string("reshape_30_cast_fp16")];
            tensor<int32, [5]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_30_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [4, 1, 512, 256]> reshape_31_cast_fp16 = reshape(shape = concat_31, x = transpose_31_cast_fp16)[name = string("reshape_31_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4663_transpose_x_0 = const()[name = string("op_4663_transpose_x_0"), val = bool(false)];
            bool var_4663_transpose_y_0 = const()[name = string("op_4663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 4, 512, 512]> var_4663_cast_fp16 = matmul(transpose_x = var_4663_transpose_x_0, transpose_y = var_4663_transpose_y_0, x = q_31_cast_fp16, y = transpose_103_cast_fp16)[name = string("op_4663_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_4665_cast_fp16 = add(x = var_4663_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4666 = const()[name = string("op_4666"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_121_cast_fp16 = softmax(axis = var_4666, x = var_4665_cast_fp16)[name = string("w_121_cast_fp16")];
            bool var_4669_transpose_x_0 = const()[name = string("op_4669_transpose_x_0"), val = bool(false)];
            bool var_4669_transpose_y_0 = const()[name = string("op_4669_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = reshape_31_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 4, 512, 256]> var_4669_cast_fp16 = matmul(transpose_x = var_4669_transpose_x_0, transpose_y = var_4669_transpose_y_0, x = w_121_cast_fp16, y = v_31_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<int32, [4]> var_4672_perm_0 = const()[name = string("op_4672_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4676 = const()[name = string("op_4676"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_4672_cast_fp16 = transpose(perm = var_4672_perm_0, x = var_4669_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 512, 1024]> input_87_cast_fp16 = reshape(shape = var_4676, x = var_4672_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45325248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45849600))))[name = string("tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4681_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4684 = const()[name = string("op_4684"), val = bool(true)];
            tensor<int32, [1]> var_4685_axes_0 = const()[name = string("op_4685_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4685_cast_fp16 = reduce_max(axes = var_4685_axes_0, keep_dims = var_4684, x = var_4681_cast_fp16)[name = string("op_4685_cast_fp16")];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_123_cast_fp16 = maximum(x = var_4685_cast_fp16, y = var_4686_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_123_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_123_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4689_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [1]> var_4694_axes_0 = const()[name = string("op_4694_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4694_keep_dims_0 = const()[name = string("op_4694_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4694_cast_fp16 = reduce_mean(axes = var_4694_axes_0, keep_dims = var_4694_keep_dims_0, x = var_4689_cast_fp16)[name = string("op_4694_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4695_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4695_cast_fp16")];
            fp32 var_4696_epsilon_0 = const()[name = string("op_4696_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4696_cast_fp16 = inverse(epsilon = var_4696_epsilon_0, x = var_4695_cast_fp16)[name = string("op_4696_cast_fp16")];
            fp16 var_4697_to_fp16 = const()[name = string("op_4697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4698_cast_fp16 = mul(x = var_4696_cast_fp16, y = var_4697_to_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4700_cast_fp16 = add(x = var_4694_cast_fp16, y = var_4698_cast_fp16)[name = string("op_4700_cast_fp16")];
            fp32 var_4701_epsilon_0 = const()[name = string("op_4701_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4701_cast_fp16 = rsqrt(epsilon = var_4701_epsilon_0, x = var_4700_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_107_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4701_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45850688)))];
            tensor<fp16, [1, 512, 512]> var_4703_cast_fp16 = mul(x = out_107_cast_fp16, y = tm_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4703_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_137_cast_fp16 = add(x = x_127_cast_fp16, y = var_4703_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4706_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_4706_cast_fp16")];
            bool var_4709 = const()[name = string("op_4709"), val = bool(true)];
            tensor<int32, [1]> var_4710_axes_0 = const()[name = string("op_4710_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4710_cast_fp16 = reduce_max(axes = var_4710_axes_0, keep_dims = var_4709, x = var_4706_cast_fp16)[name = string("op_4710_cast_fp16")];
            fp16 var_4711_to_fp16 = const()[name = string("op_4711_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_125_cast_fp16 = maximum(x = var_4710_cast_fp16, y = var_4711_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_125_cast_fp16 = real_div(x = x_137_cast_fp16, y = s_125_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4714_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4714_cast_fp16")];
            tensor<int32, [1]> var_4719_axes_0 = const()[name = string("op_4719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4719_keep_dims_0 = const()[name = string("op_4719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4719_cast_fp16 = reduce_mean(axes = var_4719_axes_0, keep_dims = var_4719_keep_dims_0, x = var_4714_cast_fp16)[name = string("op_4719_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4720_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4720_cast_fp16")];
            fp32 var_4721_epsilon_0 = const()[name = string("op_4721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4721_cast_fp16 = inverse(epsilon = var_4721_epsilon_0, x = var_4720_cast_fp16)[name = string("op_4721_cast_fp16")];
            fp16 var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4723_cast_fp16 = mul(x = var_4721_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4725_cast_fp16 = add(x = var_4719_cast_fp16, y = var_4723_cast_fp16)[name = string("op_4725_cast_fp16")];
            fp32 var_4726_epsilon_0 = const()[name = string("op_4726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4726_cast_fp16 = rsqrt(epsilon = var_4726_epsilon_0, x = var_4725_cast_fp16)[name = string("op_4726_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_109_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4726_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45851776)))];
            tensor<fp16, [1, 512, 512]> input_89_cast_fp16 = mul(x = out_109_cast_fp16, y = tm_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45852864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46901504))))[name = string("tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4733_mode_0 = const()[name = string("op_4733_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_4733_cast_fp16 = gelu(mode = var_4733_mode_0, x = linear_68_cast_fp16)[name = string("op_4733_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46905664))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47954304))))[name = string("tm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_91_cast_fp16 = mul(x = var_4733_cast_fp16, y = linear_69_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47958464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49007104))))[name = string("tm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4741_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4741_cast_fp16")];
            bool var_4744 = const()[name = string("op_4744"), val = bool(true)];
            tensor<int32, [1]> var_4745_axes_0 = const()[name = string("op_4745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4745_cast_fp16 = reduce_max(axes = var_4745_axes_0, keep_dims = var_4744, x = var_4741_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_127_cast_fp16 = maximum(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_127_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_127_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4749_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<int32, [1]> var_4754_axes_0 = const()[name = string("op_4754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4754_keep_dims_0 = const()[name = string("op_4754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4754_cast_fp16 = reduce_mean(axes = var_4754_axes_0, keep_dims = var_4754_keep_dims_0, x = var_4749_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4755_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4755_cast_fp16")];
            fp32 var_4756_epsilon_0 = const()[name = string("op_4756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4756_cast_fp16 = inverse(epsilon = var_4756_epsilon_0, x = var_4755_cast_fp16)[name = string("op_4756_cast_fp16")];
            fp16 var_4757_to_fp16 = const()[name = string("op_4757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4758_cast_fp16 = mul(x = var_4756_cast_fp16, y = var_4757_to_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4760_cast_fp16 = add(x = var_4754_cast_fp16, y = var_4758_cast_fp16)[name = string("op_4760_cast_fp16")];
            fp32 var_4761_epsilon_0 = const()[name = string("op_4761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4761_cast_fp16 = rsqrt(epsilon = var_4761_epsilon_0, x = var_4760_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_111_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4761_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49008192)))];
            tensor<fp16, [1, 512, 512]> var_4763_cast_fp16 = mul(x = out_111_cast_fp16, y = tm_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4763_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_93_cast_fp16 = add(x = x_137_cast_fp16, y = var_4763_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49009280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49271488))))[name = string("tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_71_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_71_cast_fp16")];
            string x_141_mode_0 = const()[name = string("x_141_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_141_cast_fp16 = gelu(mode = x_141_mode_0, x = linear_71_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_4783_begin_0 = const()[name = string("op_4783_begin_0"), val = tensor<int32, [4]>([0, 0, 7, 0])];
            tensor<int32, [4]> var_4783_end_0 = const()[name = string("op_4783_end_0"), val = tensor<int32, [4]>([1, 512, 8, 512])];
            tensor<bool, [4]> var_4783_end_mask_0 = const()[name = string("op_4783_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4783_squeeze_mask_0 = const()[name = string("op_4783_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_4783_cast_fp16 = slice_by_index(begin = var_4783_begin_0, end = var_4783_end_0, end_mask = var_4783_end_mask_0, squeeze_mask = var_4783_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_95_cast_fp16 = mul(x = x_141_cast_fp16, y = var_4783_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49272576))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49534784))))[name = string("tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4793_cast_fp16 = abs(x = linear_72_cast_fp16)[name = string("op_4793_cast_fp16")];
            bool var_4796 = const()[name = string("op_4796"), val = bool(true)];
            tensor<int32, [1]> var_4797_axes_0 = const()[name = string("op_4797_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4797_cast_fp16 = reduce_max(axes = var_4797_axes_0, keep_dims = var_4796, x = var_4793_cast_fp16)[name = string("op_4797_cast_fp16")];
            fp16 var_4798_to_fp16 = const()[name = string("op_4798_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_129_cast_fp16 = maximum(x = var_4797_cast_fp16, y = var_4798_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_129_cast_fp16 = real_div(x = linear_72_cast_fp16, y = s_129_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4801_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4801_cast_fp16")];
            tensor<int32, [1]> var_4806_axes_0 = const()[name = string("op_4806_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4806_keep_dims_0 = const()[name = string("op_4806_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4806_cast_fp16 = reduce_mean(axes = var_4806_axes_0, keep_dims = var_4806_keep_dims_0, x = var_4801_cast_fp16)[name = string("op_4806_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4807_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4807_cast_fp16")];
            fp32 var_4808_epsilon_0 = const()[name = string("op_4808_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4808_cast_fp16 = inverse(epsilon = var_4808_epsilon_0, x = var_4807_cast_fp16)[name = string("op_4808_cast_fp16")];
            fp16 var_4809_to_fp16 = const()[name = string("op_4809_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4810_cast_fp16 = mul(x = var_4808_cast_fp16, y = var_4809_to_fp16)[name = string("op_4810_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4812_cast_fp16 = add(x = var_4806_cast_fp16, y = var_4810_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp32 var_4813_epsilon_0 = const()[name = string("op_4813_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4813_cast_fp16 = rsqrt(epsilon = var_4813_epsilon_0, x = var_4812_cast_fp16)[name = string("op_4813_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_113_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4813_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49535872)))];
            tensor<fp16, [1, 512, 512]> var_4815_cast_fp16 = mul(x = out_113_cast_fp16, y = tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4815_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_15_cast_fp16 = add(x = input_93_cast_fp16, y = var_4815_cast_fp16)[name = string("h_15_cast_fp16")];
            tensor<fp16, [1]> tm_layers_7_layer_scalar_to_fp16 = const()[name = string("tm_layers_7_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-1])];
            tensor<fp16, [1, 512, 512]> x_145_cast_fp16 = mul(x = h_15_cast_fp16, y = tm_layers_7_layer_scalar_to_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4819_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4819_cast_fp16")];
            bool var_4822 = const()[name = string("op_4822"), val = bool(true)];
            tensor<int32, [1]> var_4823_axes_0 = const()[name = string("op_4823_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_4823_cast_fp16 = reduce_max(axes = var_4823_axes_0, keep_dims = var_4822, x = var_4819_cast_fp16)[name = string("op_4823_cast_fp16")];
            fp16 var_4824_to_fp16 = const()[name = string("op_4824_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_131_cast_fp16 = maximum(x = var_4823_cast_fp16, y = var_4824_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_131_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_131_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_4827_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [1]> var_4832_axes_0 = const()[name = string("op_4832_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4832_keep_dims_0 = const()[name = string("op_4832_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_4832_cast_fp16 = reduce_mean(axes = var_4832_axes_0, keep_dims = var_4832_keep_dims_0, x = var_4827_cast_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4833_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4833_cast_fp16")];
            fp32 var_4834_epsilon_0 = const()[name = string("op_4834_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_4834_cast_fp16 = inverse(epsilon = var_4834_epsilon_0, x = var_4833_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp16 var_4835_to_fp16 = const()[name = string("op_4835_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4836_cast_fp16 = mul(x = var_4834_cast_fp16, y = var_4835_to_fp16)[name = string("op_4836_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_4838_cast_fp16 = add(x = var_4832_cast_fp16, y = var_4836_cast_fp16)[name = string("op_4838_cast_fp16")];
            fp32 var_4839_epsilon_0 = const()[name = string("op_4839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4839_cast_fp16 = rsqrt(epsilon = var_4839_epsilon_0, x = var_4838_cast_fp16)[name = string("op_4839_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_115_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4839_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49536960)))];
            tensor<fp16, [1, 512, 512]> input_97_cast_fp16 = mul(x = out_115_cast_fp16, y = tm_layers_8_input_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49538048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50062400))))[name = string("tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4849 = const()[name = string("op_4849"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_4850_cast_fp16 = reshape(shape = var_4849, x = linear_73_cast_fp16)[name = string("op_4850_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50064512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50326720))))[name = string("tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [4]> var_4861 = const()[name = string("op_4861"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_4862_cast_fp16 = reshape(shape = var_4861, x = linear_74_cast_fp16)[name = string("op_4862_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50327808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50590016))))[name = string("tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [4]> var_4873 = const()[name = string("op_4873"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_4874_cast_fp16 = reshape(shape = var_4873, x = linear_75_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<int32, [4]> x_151_perm_0 = const()[name = string("x_151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_4850_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 4, 512, 256]> var_4878_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4878_cast_fp16")];
            bool var_4881 = const()[name = string("op_4881"), val = bool(true)];
            tensor<int32, [1]> var_4882_axes_0 = const()[name = string("op_4882_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_4882_cast_fp16 = reduce_max(axes = var_4882_axes_0, keep_dims = var_4881, x = var_4878_cast_fp16)[name = string("op_4882_cast_fp16")];
            fp16 var_4883_to_fp16 = const()[name = string("op_4883_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_133_cast_fp16 = maximum(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_133_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_133_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4886_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4886_cast_fp16")];
            tensor<int32, [1]> var_4891_axes_0 = const()[name = string("op_4891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4891_keep_dims_0 = const()[name = string("op_4891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_4891_cast_fp16 = reduce_mean(axes = var_4891_axes_0, keep_dims = var_4891_keep_dims_0, x = var_4886_cast_fp16)[name = string("op_4891_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_4892_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4892_cast_fp16")];
            fp32 var_4893_epsilon_0 = const()[name = string("op_4893_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_4893_cast_fp16 = inverse(epsilon = var_4893_epsilon_0, x = var_4892_cast_fp16)[name = string("op_4893_cast_fp16")];
            fp16 var_4894_to_fp16 = const()[name = string("op_4894_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_4895_cast_fp16 = mul(x = var_4893_cast_fp16, y = var_4894_to_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_4897_cast_fp16 = add(x = var_4891_cast_fp16, y = var_4895_cast_fp16)[name = string("op_4897_cast_fp16")];
            fp32 var_4898_epsilon_0 = const()[name = string("op_4898_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_4898_cast_fp16 = rsqrt(epsilon = var_4898_epsilon_0, x = var_4897_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_117_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4898_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591104)))];
            tensor<fp16, [1, 4, 512, 256]> q_33_cast_fp16 = mul(x = out_117_cast_fp16, y = tm_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_4862_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 2, 512, 256]> var_4901_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4901_cast_fp16")];
            bool var_4904 = const()[name = string("op_4904"), val = bool(true)];
            tensor<int32, [1]> var_4905_axes_0 = const()[name = string("op_4905_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_4905_cast_fp16 = reduce_max(axes = var_4905_axes_0, keep_dims = var_4904, x = var_4901_cast_fp16)[name = string("op_4905_cast_fp16")];
            fp16 var_4906_to_fp16 = const()[name = string("op_4906_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_135_cast_fp16 = maximum(x = var_4905_cast_fp16, y = var_4906_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_135_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_135_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4909_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [1]> var_4914_axes_0 = const()[name = string("op_4914_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4914_keep_dims_0 = const()[name = string("op_4914_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_4914_cast_fp16 = reduce_mean(axes = var_4914_axes_0, keep_dims = var_4914_keep_dims_0, x = var_4909_cast_fp16)[name = string("op_4914_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4915_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4915_cast_fp16")];
            fp32 var_4916_epsilon_0 = const()[name = string("op_4916_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_4916_cast_fp16 = inverse(epsilon = var_4916_epsilon_0, x = var_4915_cast_fp16)[name = string("op_4916_cast_fp16")];
            fp16 var_4917_to_fp16 = const()[name = string("op_4917_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_4918_cast_fp16 = mul(x = var_4916_cast_fp16, y = var_4917_to_fp16)[name = string("op_4918_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4920_cast_fp16 = add(x = var_4914_cast_fp16, y = var_4918_cast_fp16)[name = string("op_4920_cast_fp16")];
            fp32 var_4921_epsilon_0 = const()[name = string("op_4921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_4921_cast_fp16 = rsqrt(epsilon = var_4921_epsilon_0, x = var_4920_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_119_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4921_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591680)))];
            tensor<fp16, [1, 2, 512, 256]> k_49_cast_fp16 = mul(x = out_119_cast_fp16, y = tm_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_151_cast_fp16 = transpose(perm = x_151_perm_0, x = var_4874_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 2, 512, 256]> var_4924_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4924_cast_fp16")];
            bool var_4927 = const()[name = string("op_4927"), val = bool(true)];
            tensor<int32, [1]> var_4928_axes_0 = const()[name = string("op_4928_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_4928_cast_fp16 = reduce_max(axes = var_4928_axes_0, keep_dims = var_4927, x = var_4924_cast_fp16)[name = string("op_4928_cast_fp16")];
            fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_137_cast_fp16 = maximum(x = var_4928_cast_fp16, y = var_4929_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_137_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_137_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4932_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4932_cast_fp16")];
            tensor<int32, [1]> var_4937_axes_0 = const()[name = string("op_4937_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4937_keep_dims_0 = const()[name = string("op_4937_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_4937_cast_fp16 = reduce_mean(axes = var_4937_axes_0, keep_dims = var_4937_keep_dims_0, x = var_4932_cast_fp16)[name = string("op_4937_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4938_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4938_cast_fp16")];
            fp32 var_4939_epsilon_0 = const()[name = string("op_4939_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_4939_cast_fp16 = inverse(epsilon = var_4939_epsilon_0, x = var_4938_cast_fp16)[name = string("op_4939_cast_fp16")];
            fp16 var_4940_to_fp16 = const()[name = string("op_4940_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_4941_cast_fp16 = mul(x = var_4939_cast_fp16, y = var_4940_to_fp16)[name = string("op_4941_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_4943_cast_fp16 = add(x = var_4937_cast_fp16, y = var_4941_cast_fp16)[name = string("op_4943_cast_fp16")];
            fp32 var_4944_epsilon_0 = const()[name = string("op_4944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_4944_cast_fp16 = rsqrt(epsilon = var_4944_epsilon_0, x = var_4943_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_33_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4944_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4946_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = q_33_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_4952_cast_fp16 = mul(x = var_4951_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4957_begin_0 = const()[name = string("op_4957_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4957_end_0 = const()[name = string("op_4957_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_4957_end_mask_0 = const()[name = string("op_4957_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_4957_cast_fp16 = slice_by_index(begin = var_4957_begin_0, end = var_4957_end_0, end_mask = var_4957_end_mask_0, x = q_33_cast_fp16)[name = string("op_4957_cast_fp16")];
            int32 var_4959 = const()[name = string("op_4959"), val = int32(-1)];
            bool var_4960_interleave_0 = const()[name = string("op_4960_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_4960_cast_fp16 = concat(axis = var_4959, interleave = var_4960_interleave_0, values = (var_4952_cast_fp16, var_4957_cast_fp16))[name = string("op_4960_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_4961_cast_fp16 = mul(x = var_4960_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4961_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_35_cast_fp16 = add(x = var_4946_cast_fp16, y = var_4961_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4964_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4964_cast_fp16")];
            tensor<int32, [4]> var_4969_begin_0 = const()[name = string("op_4969_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4969_end_0 = const()[name = string("op_4969_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_4969_end_mask_0 = const()[name = string("op_4969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_4969_cast_fp16 = slice_by_index(begin = var_4969_begin_0, end = var_4969_end_0, end_mask = var_4969_end_mask_0, x = k_49_cast_fp16)[name = string("op_4969_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_4970_cast_fp16 = mul(x = var_4969_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_4970_cast_fp16")];
            tensor<int32, [4]> var_4975_begin_0 = const()[name = string("op_4975_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4975_end_0 = const()[name = string("op_4975_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_4975_end_mask_0 = const()[name = string("op_4975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_4975_cast_fp16 = slice_by_index(begin = var_4975_begin_0, end = var_4975_end_0, end_mask = var_4975_end_mask_0, x = k_49_cast_fp16)[name = string("op_4975_cast_fp16")];
            int32 var_4977 = const()[name = string("op_4977"), val = int32(-1)];
            bool var_4978_interleave_0 = const()[name = string("op_4978_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_4978_cast_fp16 = concat(axis = var_4977, interleave = var_4978_interleave_0, values = (var_4970_cast_fp16, var_4975_cast_fp16))[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_4979_cast_fp16 = mul(x = var_4978_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_51_cast_fp16 = add(x = var_4964_cast_fp16, y = var_4979_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_16_reps_0 = const()[name = string("tile_16_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = k_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [4, 1, 512, 256]> tile_16_cast_fp16 = tile(reps = tile_16_reps_0, x = transpose_32_cast_fp16)[name = string("tile_16_cast_fp16")];
            tensor<int32, [5]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_32_cast_fp16 = reshape(shape = concat_32, x = tile_16_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [5]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [4, 1, 512, 256]> reshape_33_cast_fp16 = reshape(shape = concat_33, x = transpose_33_cast_fp16)[name = string("reshape_33_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_17_reps_0 = const()[name = string("tile_17_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = v_33_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [4, 1, 512, 256]> tile_17_cast_fp16 = tile(reps = tile_17_reps_0, x = transpose_34_cast_fp16)[name = string("tile_17_cast_fp16")];
            tensor<int32, [5]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_34_cast_fp16 = reshape(shape = concat_34, x = tile_17_cast_fp16)[name = string("reshape_34_cast_fp16")];
            tensor<int32, [5]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = reshape_34_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [4, 1, 512, 256]> reshape_35_cast_fp16 = reshape(shape = concat_35, x = transpose_35_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = reshape_33_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 4, 512, 512]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = q_35_cast_fp16, y = transpose_104_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_4995_cast_fp16 = add(x = var_4993_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4995_cast_fp16")];
            int32 var_4996 = const()[name = string("op_4996"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_137_cast_fp16 = softmax(axis = var_4996, x = var_4995_cast_fp16)[name = string("w_137_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 4, 512, 256]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = w_137_cast_fp16, y = v_35_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5002_perm_0 = const()[name = string("op_5002_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5006 = const()[name = string("op_5006"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_5002_cast_fp16 = transpose(perm = var_5002_perm_0, x = var_4999_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 512, 1024]> input_99_cast_fp16 = reshape(shape = var_5006, x = var_5002_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50592256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51116608))))[name = string("tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5011_cast_fp16 = abs(x = linear_76_cast_fp16)[name = string("op_5011_cast_fp16")];
            bool var_5014 = const()[name = string("op_5014"), val = bool(true)];
            tensor<int32, [1]> var_5015_axes_0 = const()[name = string("op_5015_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5015_cast_fp16 = reduce_max(axes = var_5015_axes_0, keep_dims = var_5014, x = var_5011_cast_fp16)[name = string("op_5015_cast_fp16")];
            fp16 var_5016_to_fp16 = const()[name = string("op_5016_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_139_cast_fp16 = maximum(x = var_5015_cast_fp16, y = var_5016_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_139_cast_fp16 = real_div(x = linear_76_cast_fp16, y = s_139_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5019_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_5019_cast_fp16")];
            tensor<int32, [1]> var_5024_axes_0 = const()[name = string("op_5024_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5024_keep_dims_0 = const()[name = string("op_5024_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5024_cast_fp16 = reduce_mean(axes = var_5024_axes_0, keep_dims = var_5024_keep_dims_0, x = var_5019_cast_fp16)[name = string("op_5024_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5025_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_5025_cast_fp16")];
            fp32 var_5026_epsilon_0 = const()[name = string("op_5026_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5026_cast_fp16 = inverse(epsilon = var_5026_epsilon_0, x = var_5025_cast_fp16)[name = string("op_5026_cast_fp16")];
            fp16 var_5027_to_fp16 = const()[name = string("op_5027_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5028_cast_fp16 = mul(x = var_5026_cast_fp16, y = var_5027_to_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5030_cast_fp16 = add(x = var_5024_cast_fp16, y = var_5028_cast_fp16)[name = string("op_5030_cast_fp16")];
            fp32 var_5031_epsilon_0 = const()[name = string("op_5031_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5031_cast_fp16 = rsqrt(epsilon = var_5031_epsilon_0, x = var_5030_cast_fp16)[name = string("op_5031_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_121_cast_fp16 = mul(x = y_139_cast_fp16, y = var_5031_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51117696)))];
            tensor<fp16, [1, 512, 512]> var_5033_cast_fp16 = mul(x = out_121_cast_fp16, y = tm_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_5033_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = var_5033_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5036_cast_fp16 = abs(x = x_155_cast_fp16)[name = string("op_5036_cast_fp16")];
            bool var_5039 = const()[name = string("op_5039"), val = bool(true)];
            tensor<int32, [1]> var_5040_axes_0 = const()[name = string("op_5040_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5040_cast_fp16 = reduce_max(axes = var_5040_axes_0, keep_dims = var_5039, x = var_5036_cast_fp16)[name = string("op_5040_cast_fp16")];
            fp16 var_5041_to_fp16 = const()[name = string("op_5041_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_141_cast_fp16 = maximum(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_141_cast_fp16 = real_div(x = x_155_cast_fp16, y = s_141_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5044_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<int32, [1]> var_5049_axes_0 = const()[name = string("op_5049_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5049_keep_dims_0 = const()[name = string("op_5049_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5049_cast_fp16 = reduce_mean(axes = var_5049_axes_0, keep_dims = var_5049_keep_dims_0, x = var_5044_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5050_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_5050_cast_fp16")];
            fp32 var_5051_epsilon_0 = const()[name = string("op_5051_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5051_cast_fp16 = inverse(epsilon = var_5051_epsilon_0, x = var_5050_cast_fp16)[name = string("op_5051_cast_fp16")];
            fp16 var_5052_to_fp16 = const()[name = string("op_5052_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5053_cast_fp16 = mul(x = var_5051_cast_fp16, y = var_5052_to_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5055_cast_fp16 = add(x = var_5049_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            fp32 var_5056_epsilon_0 = const()[name = string("op_5056_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5056_cast_fp16 = rsqrt(epsilon = var_5056_epsilon_0, x = var_5055_cast_fp16)[name = string("op_5056_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_123_cast_fp16 = mul(x = y_141_cast_fp16, y = var_5056_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51118784)))];
            tensor<fp16, [1, 512, 512]> input_101_cast_fp16 = mul(x = out_123_cast_fp16, y = tm_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51119872))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52168512))))[name = string("tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_77_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_77_cast_fp16")];
            string var_5063_mode_0 = const()[name = string("op_5063_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_5063_cast_fp16 = gelu(mode = var_5063_mode_0, x = linear_77_cast_fp16)[name = string("op_5063_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52172672))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53221312))))[name = string("tm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_78_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_103_cast_fp16 = mul(x = var_5063_cast_fp16, y = linear_78_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53225472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54274112))))[name = string("tm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5071_cast_fp16 = abs(x = linear_79_cast_fp16)[name = string("op_5071_cast_fp16")];
            bool var_5074 = const()[name = string("op_5074"), val = bool(true)];
            tensor<int32, [1]> var_5075_axes_0 = const()[name = string("op_5075_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5075_cast_fp16 = reduce_max(axes = var_5075_axes_0, keep_dims = var_5074, x = var_5071_cast_fp16)[name = string("op_5075_cast_fp16")];
            fp16 var_5076_to_fp16 = const()[name = string("op_5076_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_143_cast_fp16 = maximum(x = var_5075_cast_fp16, y = var_5076_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_143_cast_fp16 = real_div(x = linear_79_cast_fp16, y = s_143_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5079_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_5079_cast_fp16")];
            tensor<int32, [1]> var_5084_axes_0 = const()[name = string("op_5084_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5084_keep_dims_0 = const()[name = string("op_5084_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5084_cast_fp16 = reduce_mean(axes = var_5084_axes_0, keep_dims = var_5084_keep_dims_0, x = var_5079_cast_fp16)[name = string("op_5084_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5085_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5086_cast_fp16 = inverse(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            fp16 var_5087_to_fp16 = const()[name = string("op_5087_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5088_cast_fp16 = mul(x = var_5086_cast_fp16, y = var_5087_to_fp16)[name = string("op_5088_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5090_cast_fp16 = add(x = var_5084_cast_fp16, y = var_5088_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp32 var_5091_epsilon_0 = const()[name = string("op_5091_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5091_cast_fp16 = rsqrt(epsilon = var_5091_epsilon_0, x = var_5090_cast_fp16)[name = string("op_5091_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_125_cast_fp16 = mul(x = y_143_cast_fp16, y = var_5091_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54275200)))];
            tensor<fp16, [1, 512, 512]> var_5093_cast_fp16 = mul(x = out_125_cast_fp16, y = tm_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5093_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_105_cast_fp16 = add(x = x_155_cast_fp16, y = var_5093_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54276288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54538496))))[name = string("tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_80_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_80_cast_fp16")];
            string x_159_mode_0 = const()[name = string("x_159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_159_cast_fp16 = gelu(mode = x_159_mode_0, x = linear_80_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [4]> var_5113_begin_0 = const()[name = string("op_5113_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_5113_end_0 = const()[name = string("op_5113_end_0"), val = tensor<int32, [4]>([1, 512, 9, 512])];
            tensor<bool, [4]> var_5113_end_mask_0 = const()[name = string("op_5113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5113_squeeze_mask_0 = const()[name = string("op_5113_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_5113_cast_fp16 = slice_by_index(begin = var_5113_begin_0, end = var_5113_end_0, end_mask = var_5113_end_mask_0, squeeze_mask = var_5113_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_107_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5113_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54539584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54801792))))[name = string("tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5123_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5123_cast_fp16")];
            bool var_5126 = const()[name = string("op_5126"), val = bool(true)];
            tensor<int32, [1]> var_5127_axes_0 = const()[name = string("op_5127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5127_cast_fp16 = reduce_max(axes = var_5127_axes_0, keep_dims = var_5126, x = var_5123_cast_fp16)[name = string("op_5127_cast_fp16")];
            fp16 var_5128_to_fp16 = const()[name = string("op_5128_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_145_cast_fp16 = maximum(x = var_5127_cast_fp16, y = var_5128_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_145_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_145_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5131_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5136_axes_0 = const()[name = string("op_5136_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5136_keep_dims_0 = const()[name = string("op_5136_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5136_cast_fp16 = reduce_mean(axes = var_5136_axes_0, keep_dims = var_5136_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5136_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5137_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5138_cast_fp16 = inverse(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp16 var_5139_to_fp16 = const()[name = string("op_5139_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5140_cast_fp16 = mul(x = var_5138_cast_fp16, y = var_5139_to_fp16)[name = string("op_5140_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5142_cast_fp16 = add(x = var_5136_cast_fp16, y = var_5140_cast_fp16)[name = string("op_5142_cast_fp16")];
            fp32 var_5143_epsilon_0 = const()[name = string("op_5143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5143_cast_fp16 = rsqrt(epsilon = var_5143_epsilon_0, x = var_5142_cast_fp16)[name = string("op_5143_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_127_cast_fp16 = mul(x = y_145_cast_fp16, y = var_5143_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54802880)))];
            tensor<fp16, [1, 512, 512]> var_5145_cast_fp16 = mul(x = out_127_cast_fp16, y = tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5145_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_17_cast_fp16 = add(x = input_105_cast_fp16, y = var_5145_cast_fp16)[name = string("h_17_cast_fp16")];
            tensor<fp16, [1]> tm_layers_8_layer_scalar_to_fp16 = const()[name = string("tm_layers_8_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bcp-1])];
            tensor<fp16, [1, 512, 512]> x_163_cast_fp16 = mul(x = h_17_cast_fp16, y = tm_layers_8_layer_scalar_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5149_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_5149_cast_fp16")];
            bool var_5152 = const()[name = string("op_5152"), val = bool(true)];
            tensor<int32, [1]> var_5153_axes_0 = const()[name = string("op_5153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5153_cast_fp16 = reduce_max(axes = var_5153_axes_0, keep_dims = var_5152, x = var_5149_cast_fp16)[name = string("op_5153_cast_fp16")];
            fp16 var_5154_to_fp16 = const()[name = string("op_5154_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_147_cast_fp16 = maximum(x = var_5153_cast_fp16, y = var_5154_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_147_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_147_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5157_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_5157_cast_fp16")];
            tensor<int32, [1]> var_5162_axes_0 = const()[name = string("op_5162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5162_keep_dims_0 = const()[name = string("op_5162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5162_cast_fp16 = reduce_mean(axes = var_5162_axes_0, keep_dims = var_5162_keep_dims_0, x = var_5157_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5163_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_5163_cast_fp16")];
            fp32 var_5164_epsilon_0 = const()[name = string("op_5164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5164_cast_fp16 = inverse(epsilon = var_5164_epsilon_0, x = var_5163_cast_fp16)[name = string("op_5164_cast_fp16")];
            fp16 var_5165_to_fp16 = const()[name = string("op_5165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5166_cast_fp16 = mul(x = var_5164_cast_fp16, y = var_5165_to_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5168_cast_fp16 = add(x = var_5162_cast_fp16, y = var_5166_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp32 var_5169_epsilon_0 = const()[name = string("op_5169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5169_cast_fp16 = rsqrt(epsilon = var_5169_epsilon_0, x = var_5168_cast_fp16)[name = string("op_5169_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_129_cast_fp16 = mul(x = y_147_cast_fp16, y = var_5169_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54803968)))];
            tensor<fp16, [1, 512, 512]> input_109_cast_fp16 = mul(x = out_129_cast_fp16, y = tm_layers_9_input_layernorm_weight_to_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54805056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55329408))))[name = string("tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [4]> var_5179 = const()[name = string("op_5179"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_5180_cast_fp16 = reshape(shape = var_5179, x = linear_82_cast_fp16)[name = string("op_5180_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55331520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55593728))))[name = string("tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [4]> var_5191 = const()[name = string("op_5191"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_5192_cast_fp16 = reshape(shape = var_5191, x = linear_83_cast_fp16)[name = string("op_5192_cast_fp16")];
            tensor<int32, [4]> x_167_perm_0 = const()[name = string("x_167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55594816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55857024))))[name = string("tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5203 = const()[name = string("op_5203"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_5204_cast_fp16 = reshape(shape = var_5203, x = linear_84_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_5180_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 4, 512, 256]> var_5208_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_5208_cast_fp16")];
            bool var_5211 = const()[name = string("op_5211"), val = bool(true)];
            tensor<int32, [1]> var_5212_axes_0 = const()[name = string("op_5212_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_5212_cast_fp16 = reduce_max(axes = var_5212_axes_0, keep_dims = var_5211, x = var_5208_cast_fp16)[name = string("op_5212_cast_fp16")];
            fp16 var_5213_to_fp16 = const()[name = string("op_5213_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_149_cast_fp16 = maximum(x = var_5212_cast_fp16, y = var_5213_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_149_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_149_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_5216_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<int32, [1]> var_5221_axes_0 = const()[name = string("op_5221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5221_keep_dims_0 = const()[name = string("op_5221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_5221_cast_fp16 = reduce_mean(axes = var_5221_axes_0, keep_dims = var_5221_keep_dims_0, x = var_5216_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_5222_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_5223_cast_fp16 = inverse(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            fp16 var_5224_to_fp16 = const()[name = string("op_5224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_5225_cast_fp16 = mul(x = var_5223_cast_fp16, y = var_5224_to_fp16)[name = string("op_5225_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_5227_cast_fp16 = add(x = var_5221_cast_fp16, y = var_5225_cast_fp16)[name = string("op_5227_cast_fp16")];
            fp32 var_5228_epsilon_0 = const()[name = string("op_5228_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_5228_cast_fp16 = rsqrt(epsilon = var_5228_epsilon_0, x = var_5227_cast_fp16)[name = string("op_5228_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_131_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5228_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = var_5192_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 2, 512, 256]> var_5231_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_5231_cast_fp16")];
            bool var_5234 = const()[name = string("op_5234"), val = bool(true)];
            tensor<int32, [1]> var_5235_axes_0 = const()[name = string("op_5235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_5235_cast_fp16 = reduce_max(axes = var_5235_axes_0, keep_dims = var_5234, x = var_5231_cast_fp16)[name = string("op_5235_cast_fp16")];
            fp16 var_5236_to_fp16 = const()[name = string("op_5236_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_151_cast_fp16 = maximum(x = var_5235_cast_fp16, y = var_5236_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_151_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_151_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_5239_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5239_cast_fp16")];
            tensor<int32, [1]> var_5244_axes_0 = const()[name = string("op_5244_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5244_keep_dims_0 = const()[name = string("op_5244_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_5244_cast_fp16 = reduce_mean(axes = var_5244_axes_0, keep_dims = var_5244_keep_dims_0, x = var_5239_cast_fp16)[name = string("op_5244_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_5245_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_5245_cast_fp16")];
            fp32 var_5246_epsilon_0 = const()[name = string("op_5246_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_5246_cast_fp16 = inverse(epsilon = var_5246_epsilon_0, x = var_5245_cast_fp16)[name = string("op_5246_cast_fp16")];
            fp16 var_5247_to_fp16 = const()[name = string("op_5247_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_5248_cast_fp16 = mul(x = var_5246_cast_fp16, y = var_5247_to_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_5250_cast_fp16 = add(x = var_5244_cast_fp16, y = var_5248_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp32 var_5251_epsilon_0 = const()[name = string("op_5251_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_5251_cast_fp16 = rsqrt(epsilon = var_5251_epsilon_0, x = var_5250_cast_fp16)[name = string("op_5251_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_133_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5251_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [256]> tm_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858112)))];
            tensor<fp16, [1, 2, 512, 256]> k_55_cast_fp16 = mul(x = out_133_cast_fp16, y = tm_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_5204_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2, 512, 256]> var_5254_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_5254_cast_fp16")];
            bool var_5257 = const()[name = string("op_5257"), val = bool(true)];
            tensor<int32, [1]> var_5258_axes_0 = const()[name = string("op_5258_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_5258_cast_fp16 = reduce_max(axes = var_5258_axes_0, keep_dims = var_5257, x = var_5254_cast_fp16)[name = string("op_5258_cast_fp16")];
            fp16 var_5259_to_fp16 = const()[name = string("op_5259_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_153_cast_fp16 = maximum(x = var_5258_cast_fp16, y = var_5259_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_153_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_153_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_5262_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5262_cast_fp16")];
            tensor<int32, [1]> var_5267_axes_0 = const()[name = string("op_5267_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5267_keep_dims_0 = const()[name = string("op_5267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_5267_cast_fp16 = reduce_mean(axes = var_5267_axes_0, keep_dims = var_5267_keep_dims_0, x = var_5262_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_5268_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5268_cast_fp16")];
            fp32 var_5269_epsilon_0 = const()[name = string("op_5269_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_5269_cast_fp16 = inverse(epsilon = var_5269_epsilon_0, x = var_5268_cast_fp16)[name = string("op_5269_cast_fp16")];
            fp16 var_5270_to_fp16 = const()[name = string("op_5270_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_5271_cast_fp16 = mul(x = var_5269_cast_fp16, y = var_5270_to_fp16)[name = string("op_5271_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_5273_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5271_cast_fp16)[name = string("op_5273_cast_fp16")];
            fp32 var_5274_epsilon_0 = const()[name = string("op_5274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_5274_cast_fp16 = rsqrt(epsilon = var_5274_epsilon_0, x = var_5273_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_37_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5274_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_5276_cast_fp16 = mul(x = out_131_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5276_cast_fp16")];
            tensor<int32, [4]> var_5281_begin_0 = const()[name = string("op_5281_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5281_end_0 = const()[name = string("op_5281_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_5281_end_mask_0 = const()[name = string("op_5281_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_5281_cast_fp16 = slice_by_index(begin = var_5281_begin_0, end = var_5281_end_0, end_mask = var_5281_end_mask_0, x = out_131_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_5282_cast_fp16 = mul(x = var_5281_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_5282_cast_fp16")];
            tensor<int32, [4]> var_5287_begin_0 = const()[name = string("op_5287_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5287_end_0 = const()[name = string("op_5287_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_5287_end_mask_0 = const()[name = string("op_5287_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_5287_cast_fp16 = slice_by_index(begin = var_5287_begin_0, end = var_5287_end_0, end_mask = var_5287_end_mask_0, x = out_131_cast_fp16)[name = string("op_5287_cast_fp16")];
            int32 var_5289 = const()[name = string("op_5289"), val = int32(-1)];
            bool var_5290_interleave_0 = const()[name = string("op_5290_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_5290_cast_fp16 = concat(axis = var_5289, interleave = var_5290_interleave_0, values = (var_5282_cast_fp16, var_5287_cast_fp16))[name = string("op_5290_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_5291_cast_fp16 = mul(x = var_5290_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5291_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_39_cast_fp16 = add(x = var_5276_cast_fp16, y = var_5291_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_5294_cast_fp16 = mul(x = k_55_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5294_cast_fp16")];
            tensor<int32, [4]> var_5299_begin_0 = const()[name = string("op_5299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5299_end_0 = const()[name = string("op_5299_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_5299_end_mask_0 = const()[name = string("op_5299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_5299_cast_fp16 = slice_by_index(begin = var_5299_begin_0, end = var_5299_end_0, end_mask = var_5299_end_mask_0, x = k_55_cast_fp16)[name = string("op_5299_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_5300_cast_fp16 = mul(x = var_5299_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_5300_cast_fp16")];
            tensor<int32, [4]> var_5305_begin_0 = const()[name = string("op_5305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5305_end_0 = const()[name = string("op_5305_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_5305_end_mask_0 = const()[name = string("op_5305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_5305_cast_fp16 = slice_by_index(begin = var_5305_begin_0, end = var_5305_end_0, end_mask = var_5305_end_mask_0, x = k_55_cast_fp16)[name = string("op_5305_cast_fp16")];
            int32 var_5307 = const()[name = string("op_5307"), val = int32(-1)];
            bool var_5308_interleave_0 = const()[name = string("op_5308_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_5308_cast_fp16 = concat(axis = var_5307, interleave = var_5308_interleave_0, values = (var_5300_cast_fp16, var_5305_cast_fp16))[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_5309_cast_fp16 = mul(x = var_5308_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_57_cast_fp16 = add(x = var_5294_cast_fp16, y = var_5309_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_18_reps_0 = const()[name = string("tile_18_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = k_57_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [4, 1, 512, 256]> tile_18_cast_fp16 = tile(reps = tile_18_reps_0, x = transpose_36_cast_fp16)[name = string("tile_18_cast_fp16")];
            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_36_cast_fp16 = reshape(shape = concat_36, x = tile_18_cast_fp16)[name = string("reshape_36_cast_fp16")];
            tensor<int32, [5]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = reshape_36_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [4, 1, 512, 256]> reshape_37_cast_fp16 = reshape(shape = concat_37, x = transpose_37_cast_fp16)[name = string("reshape_37_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_19_reps_0 = const()[name = string("tile_19_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = v_37_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [4, 1, 512, 256]> tile_19_cast_fp16 = tile(reps = tile_19_reps_0, x = transpose_38_cast_fp16)[name = string("tile_19_cast_fp16")];
            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_38_cast_fp16 = reshape(shape = concat_38, x = tile_19_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [5]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [4, 1, 512, 256]> reshape_39_cast_fp16 = reshape(shape = concat_39, x = transpose_39_cast_fp16)[name = string("reshape_39_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5323_transpose_x_0 = const()[name = string("op_5323_transpose_x_0"), val = bool(false)];
            bool var_5323_transpose_y_0 = const()[name = string("op_5323_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = reshape_37_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 4, 512, 512]> var_5323_cast_fp16 = matmul(transpose_x = var_5323_transpose_x_0, transpose_y = var_5323_transpose_y_0, x = q_39_cast_fp16, y = transpose_105_cast_fp16)[name = string("op_5323_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_5325_cast_fp16 = add(x = var_5323_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5325_cast_fp16")];
            int32 var_5326 = const()[name = string("op_5326"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_153_cast_fp16 = softmax(axis = var_5326, x = var_5325_cast_fp16)[name = string("w_153_cast_fp16")];
            bool var_5329_transpose_x_0 = const()[name = string("op_5329_transpose_x_0"), val = bool(false)];
            bool var_5329_transpose_y_0 = const()[name = string("op_5329_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = reshape_39_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 4, 512, 256]> var_5329_cast_fp16 = matmul(transpose_x = var_5329_transpose_x_0, transpose_y = var_5329_transpose_y_0, x = w_153_cast_fp16, y = v_39_cast_fp16)[name = string("op_5329_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = var_5329_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 512, 1024]> input_111_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56383040))))[name = string("tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5341_cast_fp16 = abs(x = linear_85_cast_fp16)[name = string("op_5341_cast_fp16")];
            bool var_5344 = const()[name = string("op_5344"), val = bool(true)];
            tensor<int32, [1]> var_5345_axes_0 = const()[name = string("op_5345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5345_cast_fp16 = reduce_max(axes = var_5345_axes_0, keep_dims = var_5344, x = var_5341_cast_fp16)[name = string("op_5345_cast_fp16")];
            fp16 var_5346_to_fp16 = const()[name = string("op_5346_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_155_cast_fp16 = maximum(x = var_5345_cast_fp16, y = var_5346_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_155_cast_fp16 = real_div(x = linear_85_cast_fp16, y = s_155_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5349_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5354_axes_0 = const()[name = string("op_5354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5354_keep_dims_0 = const()[name = string("op_5354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5354_cast_fp16 = reduce_mean(axes = var_5354_axes_0, keep_dims = var_5354_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5355_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5355_cast_fp16")];
            fp32 var_5356_epsilon_0 = const()[name = string("op_5356_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5356_cast_fp16 = inverse(epsilon = var_5356_epsilon_0, x = var_5355_cast_fp16)[name = string("op_5356_cast_fp16")];
            fp16 var_5357_to_fp16 = const()[name = string("op_5357_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5358_cast_fp16 = mul(x = var_5356_cast_fp16, y = var_5357_to_fp16)[name = string("op_5358_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5360_cast_fp16 = add(x = var_5354_cast_fp16, y = var_5358_cast_fp16)[name = string("op_5360_cast_fp16")];
            fp32 var_5361_epsilon_0 = const()[name = string("op_5361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5361_cast_fp16 = rsqrt(epsilon = var_5361_epsilon_0, x = var_5360_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_135_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5361_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56384128)))];
            tensor<fp16, [1, 512, 512]> var_5363_cast_fp16 = mul(x = out_135_cast_fp16, y = tm_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_5363_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_173_cast_fp16 = add(x = x_163_cast_fp16, y = var_5363_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5366_cast_fp16 = abs(x = x_173_cast_fp16)[name = string("op_5366_cast_fp16")];
            bool var_5369 = const()[name = string("op_5369"), val = bool(true)];
            tensor<int32, [1]> var_5370_axes_0 = const()[name = string("op_5370_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5370_cast_fp16 = reduce_max(axes = var_5370_axes_0, keep_dims = var_5369, x = var_5366_cast_fp16)[name = string("op_5370_cast_fp16")];
            fp16 var_5371_to_fp16 = const()[name = string("op_5371_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_157_cast_fp16 = maximum(x = var_5370_cast_fp16, y = var_5371_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_157_cast_fp16 = real_div(x = x_173_cast_fp16, y = s_157_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5374_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [1]> var_5379_axes_0 = const()[name = string("op_5379_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5379_keep_dims_0 = const()[name = string("op_5379_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5379_cast_fp16 = reduce_mean(axes = var_5379_axes_0, keep_dims = var_5379_keep_dims_0, x = var_5374_cast_fp16)[name = string("op_5379_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5380_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5381_cast_fp16 = inverse(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            fp16 var_5382_to_fp16 = const()[name = string("op_5382_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5383_cast_fp16 = mul(x = var_5381_cast_fp16, y = var_5382_to_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5385_cast_fp16 = add(x = var_5379_cast_fp16, y = var_5383_cast_fp16)[name = string("op_5385_cast_fp16")];
            fp32 var_5386_epsilon_0 = const()[name = string("op_5386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5386_cast_fp16 = rsqrt(epsilon = var_5386_epsilon_0, x = var_5385_cast_fp16)[name = string("op_5386_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_137_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5386_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56385216)))];
            tensor<fp16, [1, 512, 512]> input_113_cast_fp16 = mul(x = out_137_cast_fp16, y = tm_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56386304))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57434944))))[name = string("tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_86_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_86_cast_fp16")];
            string var_5393_mode_0 = const()[name = string("op_5393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_5393_cast_fp16 = gelu(mode = var_5393_mode_0, x = linear_86_cast_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57439104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58487744))))[name = string("tm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_87_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_115_cast_fp16 = mul(x = var_5393_cast_fp16, y = linear_87_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58491904))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59540544))))[name = string("tm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5401_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5401_cast_fp16")];
            bool var_5404 = const()[name = string("op_5404"), val = bool(true)];
            tensor<int32, [1]> var_5405_axes_0 = const()[name = string("op_5405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5405_cast_fp16 = reduce_max(axes = var_5405_axes_0, keep_dims = var_5404, x = var_5401_cast_fp16)[name = string("op_5405_cast_fp16")];
            fp16 var_5406_to_fp16 = const()[name = string("op_5406_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_159_cast_fp16 = maximum(x = var_5405_cast_fp16, y = var_5406_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_159_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_159_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5409_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5409_cast_fp16")];
            tensor<int32, [1]> var_5414_axes_0 = const()[name = string("op_5414_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5414_keep_dims_0 = const()[name = string("op_5414_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5414_cast_fp16 = reduce_mean(axes = var_5414_axes_0, keep_dims = var_5414_keep_dims_0, x = var_5409_cast_fp16)[name = string("op_5414_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5415_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5415_cast_fp16")];
            fp32 var_5416_epsilon_0 = const()[name = string("op_5416_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5416_cast_fp16 = inverse(epsilon = var_5416_epsilon_0, x = var_5415_cast_fp16)[name = string("op_5416_cast_fp16")];
            fp16 var_5417_to_fp16 = const()[name = string("op_5417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5418_cast_fp16 = mul(x = var_5416_cast_fp16, y = var_5417_to_fp16)[name = string("op_5418_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5420_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5418_cast_fp16)[name = string("op_5420_cast_fp16")];
            fp32 var_5421_epsilon_0 = const()[name = string("op_5421_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5421_cast_fp16 = rsqrt(epsilon = var_5421_epsilon_0, x = var_5420_cast_fp16)[name = string("op_5421_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_139_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5421_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59541632)))];
            tensor<fp16, [1, 512, 512]> var_5423_cast_fp16 = mul(x = out_139_cast_fp16, y = tm_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5423_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_117_cast_fp16 = add(x = x_173_cast_fp16, y = var_5423_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59542720))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59804928))))[name = string("tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_89_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_89_cast_fp16")];
            string x_177_mode_0 = const()[name = string("x_177_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_177_cast_fp16 = gelu(mode = x_177_mode_0, x = linear_89_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [4]> var_5443_begin_0 = const()[name = string("op_5443_begin_0"), val = tensor<int32, [4]>([0, 0, 9, 0])];
            tensor<int32, [4]> var_5443_end_0 = const()[name = string("op_5443_end_0"), val = tensor<int32, [4]>([1, 512, 10, 512])];
            tensor<bool, [4]> var_5443_end_mask_0 = const()[name = string("op_5443_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5443_squeeze_mask_0 = const()[name = string("op_5443_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_5443_cast_fp16 = slice_by_index(begin = var_5443_begin_0, end = var_5443_end_0, end_mask = var_5443_end_mask_0, squeeze_mask = var_5443_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5443_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_119_cast_fp16 = mul(x = x_177_cast_fp16, y = var_5443_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59806016))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60068224))))[name = string("tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5453_cast_fp16 = abs(x = linear_90_cast_fp16)[name = string("op_5453_cast_fp16")];
            bool var_5456 = const()[name = string("op_5456"), val = bool(true)];
            tensor<int32, [1]> var_5457_axes_0 = const()[name = string("op_5457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5457_cast_fp16 = reduce_max(axes = var_5457_axes_0, keep_dims = var_5456, x = var_5453_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_161_cast_fp16 = maximum(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_161_cast_fp16 = real_div(x = linear_90_cast_fp16, y = s_161_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5461_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<int32, [1]> var_5466_axes_0 = const()[name = string("op_5466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5466_keep_dims_0 = const()[name = string("op_5466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5466_cast_fp16 = reduce_mean(axes = var_5466_axes_0, keep_dims = var_5466_keep_dims_0, x = var_5461_cast_fp16)[name = string("op_5466_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5467_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5467_cast_fp16")];
            fp32 var_5468_epsilon_0 = const()[name = string("op_5468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5468_cast_fp16 = inverse(epsilon = var_5468_epsilon_0, x = var_5467_cast_fp16)[name = string("op_5468_cast_fp16")];
            fp16 var_5469_to_fp16 = const()[name = string("op_5469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5470_cast_fp16 = mul(x = var_5468_cast_fp16, y = var_5469_to_fp16)[name = string("op_5470_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5472_cast_fp16 = add(x = var_5466_cast_fp16, y = var_5470_cast_fp16)[name = string("op_5472_cast_fp16")];
            fp32 var_5473_epsilon_0 = const()[name = string("op_5473_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5473_cast_fp16 = rsqrt(epsilon = var_5473_epsilon_0, x = var_5472_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_141_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5473_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60069312)))];
            tensor<fp16, [1, 512, 512]> var_5475_cast_fp16 = mul(x = out_141_cast_fp16, y = tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5475_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_19_cast_fp16 = add(x = input_117_cast_fp16, y = var_5475_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<fp16, [1]> tm_layers_9_layer_scalar_to_fp16 = const()[name = string("tm_layers_9_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9ep-1])];
            tensor<fp16, [1, 512, 512]> x_181_cast_fp16 = mul(x = h_19_cast_fp16, y = tm_layers_9_layer_scalar_to_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5479_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_5479_cast_fp16")];
            bool var_5482 = const()[name = string("op_5482"), val = bool(true)];
            tensor<int32, [1]> var_5483_axes_0 = const()[name = string("op_5483_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5483_cast_fp16 = reduce_max(axes = var_5483_axes_0, keep_dims = var_5482, x = var_5479_cast_fp16)[name = string("op_5483_cast_fp16")];
            fp16 var_5484_to_fp16 = const()[name = string("op_5484_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_163_cast_fp16 = maximum(x = var_5483_cast_fp16, y = var_5484_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_163_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_163_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5487_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5492_axes_0 = const()[name = string("op_5492_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5492_keep_dims_0 = const()[name = string("op_5492_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5492_cast_fp16 = reduce_mean(axes = var_5492_axes_0, keep_dims = var_5492_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5493_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5493_cast_fp16")];
            fp32 var_5494_epsilon_0 = const()[name = string("op_5494_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5494_cast_fp16 = inverse(epsilon = var_5494_epsilon_0, x = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp16 var_5495_to_fp16 = const()[name = string("op_5495_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5496_cast_fp16 = mul(x = var_5494_cast_fp16, y = var_5495_to_fp16)[name = string("op_5496_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5498_cast_fp16 = add(x = var_5492_cast_fp16, y = var_5496_cast_fp16)[name = string("op_5498_cast_fp16")];
            fp32 var_5499_epsilon_0 = const()[name = string("op_5499_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5499_cast_fp16 = rsqrt(epsilon = var_5499_epsilon_0, x = var_5498_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_143_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5499_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60070400)))];
            tensor<fp16, [1, 512, 512]> input_121_cast_fp16 = mul(x = out_143_cast_fp16, y = tm_layers_10_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60071488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60595840))))[name = string("tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5509 = const()[name = string("op_5509"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_5510_cast_fp16 = reshape(shape = var_5509, x = linear_91_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60597952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60860160))))[name = string("tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5521 = const()[name = string("op_5521"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_5522_cast_fp16 = reshape(shape = var_5521, x = linear_92_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60861248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61123456))))[name = string("tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5533 = const()[name = string("op_5533"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_5534_cast_fp16 = reshape(shape = var_5533, x = linear_93_cast_fp16)[name = string("op_5534_cast_fp16")];
            tensor<int32, [4]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5510_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 4, 512, 256]> var_5538_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_5538_cast_fp16")];
            bool var_5541 = const()[name = string("op_5541"), val = bool(true)];
            tensor<int32, [1]> var_5542_axes_0 = const()[name = string("op_5542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_5542_cast_fp16 = reduce_max(axes = var_5542_axes_0, keep_dims = var_5541, x = var_5538_cast_fp16)[name = string("op_5542_cast_fp16")];
            fp16 var_5543_to_fp16 = const()[name = string("op_5543_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_165_cast_fp16 = maximum(x = var_5542_cast_fp16, y = var_5543_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_165_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_165_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_5546_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<int32, [1]> var_5551_axes_0 = const()[name = string("op_5551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5551_keep_dims_0 = const()[name = string("op_5551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_5551_cast_fp16 = reduce_mean(axes = var_5551_axes_0, keep_dims = var_5551_keep_dims_0, x = var_5546_cast_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_5552_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5552_cast_fp16")];
            fp32 var_5553_epsilon_0 = const()[name = string("op_5553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_5553_cast_fp16 = inverse(epsilon = var_5553_epsilon_0, x = var_5552_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp16 var_5554_to_fp16 = const()[name = string("op_5554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_5555_cast_fp16 = mul(x = var_5553_cast_fp16, y = var_5554_to_fp16)[name = string("op_5555_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_5557_cast_fp16 = add(x = var_5551_cast_fp16, y = var_5555_cast_fp16)[name = string("op_5557_cast_fp16")];
            fp32 var_5558_epsilon_0 = const()[name = string("op_5558_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_5558_cast_fp16 = rsqrt(epsilon = var_5558_epsilon_0, x = var_5557_cast_fp16)[name = string("op_5558_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_145_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5558_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61124544)))];
            tensor<fp16, [1, 4, 512, 256]> q_41_cast_fp16 = mul(x = out_145_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_41_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_5522_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2, 512, 256]> var_5561_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_5561_cast_fp16")];
            bool var_5564 = const()[name = string("op_5564"), val = bool(true)];
            tensor<int32, [1]> var_5565_axes_0 = const()[name = string("op_5565_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_5565_cast_fp16 = reduce_max(axes = var_5565_axes_0, keep_dims = var_5564, x = var_5561_cast_fp16)[name = string("op_5565_cast_fp16")];
            fp16 var_5566_to_fp16 = const()[name = string("op_5566_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_167_cast_fp16 = maximum(x = var_5565_cast_fp16, y = var_5566_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_167_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_167_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_5569_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5569_cast_fp16")];
            tensor<int32, [1]> var_5574_axes_0 = const()[name = string("op_5574_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5574_keep_dims_0 = const()[name = string("op_5574_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_5574_cast_fp16 = reduce_mean(axes = var_5574_axes_0, keep_dims = var_5574_keep_dims_0, x = var_5569_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_5575_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5575_cast_fp16")];
            fp32 var_5576_epsilon_0 = const()[name = string("op_5576_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_5576_cast_fp16 = inverse(epsilon = var_5576_epsilon_0, x = var_5575_cast_fp16)[name = string("op_5576_cast_fp16")];
            fp16 var_5577_to_fp16 = const()[name = string("op_5577_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_5578_cast_fp16 = mul(x = var_5576_cast_fp16, y = var_5577_to_fp16)[name = string("op_5578_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_5580_cast_fp16 = add(x = var_5574_cast_fp16, y = var_5578_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_5581_cast_fp16 = rsqrt(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_147_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5581_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [256]> tm_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125120)))];
            tensor<fp16, [1, 2, 512, 256]> k_61_cast_fp16 = mul(x = out_147_cast_fp16, y = tm_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = var_5534_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 2, 512, 256]> var_5584_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_5584_cast_fp16")];
            bool var_5587 = const()[name = string("op_5587"), val = bool(true)];
            tensor<int32, [1]> var_5588_axes_0 = const()[name = string("op_5588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_5588_cast_fp16 = reduce_max(axes = var_5588_axes_0, keep_dims = var_5587, x = var_5584_cast_fp16)[name = string("op_5588_cast_fp16")];
            fp16 var_5589_to_fp16 = const()[name = string("op_5589_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_169_cast_fp16 = maximum(x = var_5588_cast_fp16, y = var_5589_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_169_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_169_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_5592_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5592_cast_fp16")];
            tensor<int32, [1]> var_5597_axes_0 = const()[name = string("op_5597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5597_keep_dims_0 = const()[name = string("op_5597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_5597_cast_fp16 = reduce_mean(axes = var_5597_axes_0, keep_dims = var_5597_keep_dims_0, x = var_5592_cast_fp16)[name = string("op_5597_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_5598_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5598_cast_fp16")];
            fp32 var_5599_epsilon_0 = const()[name = string("op_5599_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_5599_cast_fp16 = inverse(epsilon = var_5599_epsilon_0, x = var_5598_cast_fp16)[name = string("op_5599_cast_fp16")];
            fp16 var_5600_to_fp16 = const()[name = string("op_5600_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_5601_cast_fp16 = mul(x = var_5599_cast_fp16, y = var_5600_to_fp16)[name = string("op_5601_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_5603_cast_fp16 = add(x = var_5597_cast_fp16, y = var_5601_cast_fp16)[name = string("op_5603_cast_fp16")];
            fp32 var_5604_epsilon_0 = const()[name = string("op_5604_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_5604_cast_fp16 = rsqrt(epsilon = var_5604_epsilon_0, x = var_5603_cast_fp16)[name = string("op_5604_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_41_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5604_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_5606_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5611_begin_0 = const()[name = string("op_5611_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5611_end_0 = const()[name = string("op_5611_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_5611_end_mask_0 = const()[name = string("op_5611_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_5611_cast_fp16 = slice_by_index(begin = var_5611_begin_0, end = var_5611_end_0, end_mask = var_5611_end_mask_0, x = q_41_cast_fp16)[name = string("op_5611_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_5612_cast_fp16 = mul(x = var_5611_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_5612_cast_fp16")];
            tensor<int32, [4]> var_5617_begin_0 = const()[name = string("op_5617_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5617_end_0 = const()[name = string("op_5617_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_5617_end_mask_0 = const()[name = string("op_5617_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_5617_cast_fp16 = slice_by_index(begin = var_5617_begin_0, end = var_5617_end_0, end_mask = var_5617_end_mask_0, x = q_41_cast_fp16)[name = string("op_5617_cast_fp16")];
            int32 var_5619 = const()[name = string("op_5619"), val = int32(-1)];
            bool var_5620_interleave_0 = const()[name = string("op_5620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_5620_cast_fp16 = concat(axis = var_5619, interleave = var_5620_interleave_0, values = (var_5612_cast_fp16, var_5617_cast_fp16))[name = string("op_5620_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_5621_cast_fp16 = mul(x = var_5620_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5621_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_43_cast_fp16 = add(x = var_5606_cast_fp16, y = var_5621_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_5624_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = k_61_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = k_61_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool var_5638_interleave_0 = const()[name = string("op_5638_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_5638_cast_fp16 = concat(axis = var_5637, interleave = var_5638_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("op_5638_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_5639_cast_fp16 = mul(x = var_5638_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_63_cast_fp16 = add(x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_20_reps_0 = const()[name = string("tile_20_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = k_63_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [4, 1, 512, 256]> tile_20_cast_fp16 = tile(reps = tile_20_reps_0, x = transpose_40_cast_fp16)[name = string("tile_20_cast_fp16")];
            tensor<int32, [5]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_40_cast_fp16 = reshape(shape = concat_40, x = tile_20_cast_fp16)[name = string("reshape_40_cast_fp16")];
            tensor<int32, [5]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = reshape_40_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [4, 1, 512, 256]> reshape_41_cast_fp16 = reshape(shape = concat_41, x = transpose_41_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_21_reps_0 = const()[name = string("tile_21_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = v_41_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [4, 1, 512, 256]> tile_21_cast_fp16 = tile(reps = tile_21_reps_0, x = transpose_42_cast_fp16)[name = string("tile_21_cast_fp16")];
            tensor<int32, [5]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_42_cast_fp16 = reshape(shape = concat_42, x = tile_21_cast_fp16)[name = string("reshape_42_cast_fp16")];
            tensor<int32, [5]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = reshape_42_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [4, 1, 512, 256]> reshape_43_cast_fp16 = reshape(shape = concat_43, x = transpose_43_cast_fp16)[name = string("reshape_43_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5653_transpose_x_0 = const()[name = string("op_5653_transpose_x_0"), val = bool(false)];
            bool var_5653_transpose_y_0 = const()[name = string("op_5653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 4, 512, 512]> var_5653_cast_fp16 = matmul(transpose_x = var_5653_transpose_x_0, transpose_y = var_5653_transpose_y_0, x = q_43_cast_fp16, y = transpose_106_cast_fp16)[name = string("op_5653_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_5655_cast_fp16 = add(x = var_5653_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5655_cast_fp16")];
            int32 var_5656 = const()[name = string("op_5656"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_169_cast_fp16 = softmax(axis = var_5656, x = var_5655_cast_fp16)[name = string("w_169_cast_fp16")];
            bool var_5659_transpose_x_0 = const()[name = string("op_5659_transpose_x_0"), val = bool(false)];
            bool var_5659_transpose_y_0 = const()[name = string("op_5659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = reshape_43_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 4, 512, 256]> var_5659_cast_fp16 = matmul(transpose_x = var_5659_transpose_x_0, transpose_y = var_5659_transpose_y_0, x = w_169_cast_fp16, y = v_43_cast_fp16)[name = string("op_5659_cast_fp16")];
            tensor<int32, [4]> var_5662_perm_0 = const()[name = string("op_5662_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5666 = const()[name = string("op_5666"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_5662_cast_fp16 = transpose(perm = var_5662_perm_0, x = var_5659_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 512, 1024]> input_123_cast_fp16 = reshape(shape = var_5666, x = var_5662_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61650048))))[name = string("tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5671_cast_fp16 = abs(x = linear_94_cast_fp16)[name = string("op_5671_cast_fp16")];
            bool var_5674 = const()[name = string("op_5674"), val = bool(true)];
            tensor<int32, [1]> var_5675_axes_0 = const()[name = string("op_5675_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5675_cast_fp16 = reduce_max(axes = var_5675_axes_0, keep_dims = var_5674, x = var_5671_cast_fp16)[name = string("op_5675_cast_fp16")];
            fp16 var_5676_to_fp16 = const()[name = string("op_5676_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_171_cast_fp16 = maximum(x = var_5675_cast_fp16, y = var_5676_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_171_cast_fp16 = real_div(x = linear_94_cast_fp16, y = s_171_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5679_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5679_cast_fp16")];
            tensor<int32, [1]> var_5684_axes_0 = const()[name = string("op_5684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5684_keep_dims_0 = const()[name = string("op_5684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5684_cast_fp16 = reduce_mean(axes = var_5684_axes_0, keep_dims = var_5684_keep_dims_0, x = var_5679_cast_fp16)[name = string("op_5684_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5685_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5685_cast_fp16")];
            fp32 var_5686_epsilon_0 = const()[name = string("op_5686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5686_cast_fp16 = inverse(epsilon = var_5686_epsilon_0, x = var_5685_cast_fp16)[name = string("op_5686_cast_fp16")];
            fp16 var_5687_to_fp16 = const()[name = string("op_5687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5688_cast_fp16 = mul(x = var_5686_cast_fp16, y = var_5687_to_fp16)[name = string("op_5688_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5690_cast_fp16 = add(x = var_5684_cast_fp16, y = var_5688_cast_fp16)[name = string("op_5690_cast_fp16")];
            fp32 var_5691_epsilon_0 = const()[name = string("op_5691_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5691_cast_fp16 = rsqrt(epsilon = var_5691_epsilon_0, x = var_5690_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_149_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5691_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61651136)))];
            tensor<fp16, [1, 512, 512]> var_5693_cast_fp16 = mul(x = out_149_cast_fp16, y = tm_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5693_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = var_5693_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5696_cast_fp16 = abs(x = x_191_cast_fp16)[name = string("op_5696_cast_fp16")];
            bool var_5699 = const()[name = string("op_5699"), val = bool(true)];
            tensor<int32, [1]> var_5700_axes_0 = const()[name = string("op_5700_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5700_cast_fp16 = reduce_max(axes = var_5700_axes_0, keep_dims = var_5699, x = var_5696_cast_fp16)[name = string("op_5700_cast_fp16")];
            fp16 var_5701_to_fp16 = const()[name = string("op_5701_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_173_cast_fp16 = maximum(x = var_5700_cast_fp16, y = var_5701_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_173_cast_fp16 = real_div(x = x_191_cast_fp16, y = s_173_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5704_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5704_cast_fp16")];
            tensor<int32, [1]> var_5709_axes_0 = const()[name = string("op_5709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5709_keep_dims_0 = const()[name = string("op_5709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5709_cast_fp16 = reduce_mean(axes = var_5709_axes_0, keep_dims = var_5709_keep_dims_0, x = var_5704_cast_fp16)[name = string("op_5709_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5710_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5710_cast_fp16")];
            fp32 var_5711_epsilon_0 = const()[name = string("op_5711_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5711_cast_fp16 = inverse(epsilon = var_5711_epsilon_0, x = var_5710_cast_fp16)[name = string("op_5711_cast_fp16")];
            fp16 var_5712_to_fp16 = const()[name = string("op_5712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = string("op_5713_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5715_cast_fp16 = add(x = var_5709_cast_fp16, y = var_5713_cast_fp16)[name = string("op_5715_cast_fp16")];
            fp32 var_5716_epsilon_0 = const()[name = string("op_5716_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5716_cast_fp16 = rsqrt(epsilon = var_5716_epsilon_0, x = var_5715_cast_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_151_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5716_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61652224)))];
            tensor<fp16, [1, 512, 512]> input_125_cast_fp16 = mul(x = out_151_cast_fp16, y = tm_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61653312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62701952))))[name = string("tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_95_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_95_cast_fp16")];
            string var_5723_mode_0 = const()[name = string("op_5723_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_5723_cast_fp16 = gelu(mode = var_5723_mode_0, x = linear_95_cast_fp16)[name = string("op_5723_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62706112))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63754752))))[name = string("tm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_127_cast_fp16 = mul(x = var_5723_cast_fp16, y = linear_96_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63758912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64807552))))[name = string("tm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5731_cast_fp16 = abs(x = linear_97_cast_fp16)[name = string("op_5731_cast_fp16")];
            bool var_5734 = const()[name = string("op_5734"), val = bool(true)];
            tensor<int32, [1]> var_5735_axes_0 = const()[name = string("op_5735_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5735_cast_fp16 = reduce_max(axes = var_5735_axes_0, keep_dims = var_5734, x = var_5731_cast_fp16)[name = string("op_5735_cast_fp16")];
            fp16 var_5736_to_fp16 = const()[name = string("op_5736_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_175_cast_fp16 = maximum(x = var_5735_cast_fp16, y = var_5736_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_175_cast_fp16 = real_div(x = linear_97_cast_fp16, y = s_175_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5739_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [1]> var_5744_axes_0 = const()[name = string("op_5744_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5744_keep_dims_0 = const()[name = string("op_5744_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5744_cast_fp16 = reduce_mean(axes = var_5744_axes_0, keep_dims = var_5744_keep_dims_0, x = var_5739_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5745_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5745_cast_fp16")];
            fp32 var_5746_epsilon_0 = const()[name = string("op_5746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5746_cast_fp16 = inverse(epsilon = var_5746_epsilon_0, x = var_5745_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5747_to_fp16 = const()[name = string("op_5747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5748_cast_fp16 = mul(x = var_5746_cast_fp16, y = var_5747_to_fp16)[name = string("op_5748_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5750_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5748_cast_fp16)[name = string("op_5750_cast_fp16")];
            fp32 var_5751_epsilon_0 = const()[name = string("op_5751_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5751_cast_fp16 = rsqrt(epsilon = var_5751_epsilon_0, x = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_153_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5751_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64808640)))];
            tensor<fp16, [1, 512, 512]> var_5753_cast_fp16 = mul(x = out_153_cast_fp16, y = tm_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_129_cast_fp16 = add(x = x_191_cast_fp16, y = var_5753_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64809728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65071936))))[name = string("tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_98_cast_fp16")];
            string x_195_mode_0 = const()[name = string("x_195_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_195_cast_fp16 = gelu(mode = x_195_mode_0, x = linear_98_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [4]> var_5773_begin_0 = const()[name = string("op_5773_begin_0"), val = tensor<int32, [4]>([0, 0, 10, 0])];
            tensor<int32, [4]> var_5773_end_0 = const()[name = string("op_5773_end_0"), val = tensor<int32, [4]>([1, 512, 11, 512])];
            tensor<bool, [4]> var_5773_end_mask_0 = const()[name = string("op_5773_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5773_squeeze_mask_0 = const()[name = string("op_5773_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_5773_cast_fp16 = slice_by_index(begin = var_5773_begin_0, end = var_5773_end_0, end_mask = var_5773_end_mask_0, squeeze_mask = var_5773_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_131_cast_fp16 = mul(x = x_195_cast_fp16, y = var_5773_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65073024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65335232))))[name = string("tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5783_cast_fp16 = abs(x = linear_99_cast_fp16)[name = string("op_5783_cast_fp16")];
            bool var_5786 = const()[name = string("op_5786"), val = bool(true)];
            tensor<int32, [1]> var_5787_axes_0 = const()[name = string("op_5787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5787_cast_fp16 = reduce_max(axes = var_5787_axes_0, keep_dims = var_5786, x = var_5783_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp16 var_5788_to_fp16 = const()[name = string("op_5788_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_177_cast_fp16 = maximum(x = var_5787_cast_fp16, y = var_5788_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_177_cast_fp16 = real_div(x = linear_99_cast_fp16, y = s_177_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5791_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5791_cast_fp16")];
            tensor<int32, [1]> var_5796_axes_0 = const()[name = string("op_5796_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5796_keep_dims_0 = const()[name = string("op_5796_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5796_cast_fp16 = reduce_mean(axes = var_5796_axes_0, keep_dims = var_5796_keep_dims_0, x = var_5791_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5797_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5797_cast_fp16")];
            fp32 var_5798_epsilon_0 = const()[name = string("op_5798_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5798_cast_fp16 = inverse(epsilon = var_5798_epsilon_0, x = var_5797_cast_fp16)[name = string("op_5798_cast_fp16")];
            fp16 var_5799_to_fp16 = const()[name = string("op_5799_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5800_cast_fp16 = mul(x = var_5798_cast_fp16, y = var_5799_to_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5802_cast_fp16 = add(x = var_5796_cast_fp16, y = var_5800_cast_fp16)[name = string("op_5802_cast_fp16")];
            fp32 var_5803_epsilon_0 = const()[name = string("op_5803_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5803_cast_fp16 = rsqrt(epsilon = var_5803_epsilon_0, x = var_5802_cast_fp16)[name = string("op_5803_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_155_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5803_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65336320)))];
            tensor<fp16, [1, 512, 512]> var_5805_cast_fp16 = mul(x = out_155_cast_fp16, y = tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5805_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_21_cast_fp16 = add(x = input_129_cast_fp16, y = var_5805_cast_fp16)[name = string("h_21_cast_fp16")];
            tensor<fp16, [1]> tm_layers_10_layer_scalar_to_fp16 = const()[name = string("tm_layers_10_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.1cp-1])];
            tensor<fp16, [1, 512, 512]> x_199_cast_fp16 = mul(x = h_21_cast_fp16, y = tm_layers_10_layer_scalar_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5809_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5809_cast_fp16")];
            bool var_5812 = const()[name = string("op_5812"), val = bool(true)];
            tensor<int32, [1]> var_5813_axes_0 = const()[name = string("op_5813_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_5813_cast_fp16 = reduce_max(axes = var_5813_axes_0, keep_dims = var_5812, x = var_5809_cast_fp16)[name = string("op_5813_cast_fp16")];
            fp16 var_5814_to_fp16 = const()[name = string("op_5814_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_179_cast_fp16 = maximum(x = var_5813_cast_fp16, y = var_5814_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_179_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_179_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_5817_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5817_cast_fp16")];
            tensor<int32, [1]> var_5822_axes_0 = const()[name = string("op_5822_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5822_keep_dims_0 = const()[name = string("op_5822_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5822_cast_fp16 = reduce_mean(axes = var_5822_axes_0, keep_dims = var_5822_keep_dims_0, x = var_5817_cast_fp16)[name = string("op_5822_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5823_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp32 var_5824_epsilon_0 = const()[name = string("op_5824_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_5824_cast_fp16 = inverse(epsilon = var_5824_epsilon_0, x = var_5823_cast_fp16)[name = string("op_5824_cast_fp16")];
            fp16 var_5825_to_fp16 = const()[name = string("op_5825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5826_cast_fp16 = mul(x = var_5824_cast_fp16, y = var_5825_to_fp16)[name = string("op_5826_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_5828_cast_fp16 = add(x = var_5822_cast_fp16, y = var_5826_cast_fp16)[name = string("op_5828_cast_fp16")];
            fp32 var_5829_epsilon_0 = const()[name = string("op_5829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5829_cast_fp16 = rsqrt(epsilon = var_5829_epsilon_0, x = var_5828_cast_fp16)[name = string("op_5829_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_157_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5829_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65337408)))];
            tensor<fp16, [1, 512, 512]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = tm_layers_11_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65338496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66387136))))[name = string("tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_100_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5839 = const()[name = string("op_5839"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 512, 4, 512]> var_5840_cast_fp16 = reshape(shape = var_5839, x = linear_100_cast_fp16)[name = string("op_5840_cast_fp16")];
            tensor<int32, [4]> x_201_perm_0 = const()[name = string("x_201_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66391296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66653504))))[name = string("tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_5851 = const()[name = string("op_5851"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 512, 1, 512]> var_5852_cast_fp16 = reshape(shape = var_5851, x = linear_101_cast_fp16)[name = string("op_5852_cast_fp16")];
            tensor<int32, [4]> x_203_perm_0 = const()[name = string("x_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66654592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66916800))))[name = string("tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 512, 1, 512]> var_5864_cast_fp16 = reshape(shape = var_5863, x = linear_102_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 512]> x_201_cast_fp16 = transpose(perm = x_201_perm_0, x = var_5840_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 4, 512, 512]> var_5868_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5868_cast_fp16")];
            bool var_5871 = const()[name = string("op_5871"), val = bool(true)];
            tensor<int32, [1]> var_5872_axes_0 = const()[name = string("op_5872_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_5872_cast_fp16 = reduce_max(axes = var_5872_axes_0, keep_dims = var_5871, x = var_5868_cast_fp16)[name = string("op_5872_cast_fp16")];
            fp16 var_5873_to_fp16 = const()[name = string("op_5873_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_181_cast_fp16 = maximum(x = var_5872_cast_fp16, y = var_5873_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> y_181_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_181_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_5876_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5876_cast_fp16")];
            tensor<int32, [1]> var_5881_axes_0 = const()[name = string("op_5881_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5881_keep_dims_0 = const()[name = string("op_5881_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_5881_cast_fp16 = reduce_mean(axes = var_5881_axes_0, keep_dims = var_5881_keep_dims_0, x = var_5876_cast_fp16)[name = string("op_5881_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_5882_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5882_cast_fp16")];
            fp32 var_5883_epsilon_0 = const()[name = string("op_5883_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_5883_cast_fp16 = inverse(epsilon = var_5883_epsilon_0, x = var_5882_cast_fp16)[name = string("op_5883_cast_fp16")];
            fp16 var_5884_to_fp16 = const()[name = string("op_5884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_5885_cast_fp16 = mul(x = var_5883_cast_fp16, y = var_5884_to_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_5887_cast_fp16 = add(x = var_5881_cast_fp16, y = var_5885_cast_fp16)[name = string("op_5887_cast_fp16")];
            fp32 var_5888_epsilon_0 = const()[name = string("op_5888_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_5888_cast_fp16 = rsqrt(epsilon = var_5888_epsilon_0, x = var_5887_cast_fp16)[name = string("op_5888_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> out_159_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5888_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66917888)))];
            tensor<fp16, [1, 4, 512, 512]> q_45_cast_fp16 = mul(x = out_159_cast_fp16, y = tm_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> x_203_cast_fp16 = transpose(perm = x_203_perm_0, x = var_5852_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1, 512, 512]> var_5891_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5891_cast_fp16")];
            bool var_5894 = const()[name = string("op_5894"), val = bool(true)];
            tensor<int32, [1]> var_5895_axes_0 = const()[name = string("op_5895_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_5895_cast_fp16 = reduce_max(axes = var_5895_axes_0, keep_dims = var_5894, x = var_5891_cast_fp16)[name = string("op_5895_cast_fp16")];
            fp16 var_5896_to_fp16 = const()[name = string("op_5896_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 512, 1]> s_183_cast_fp16 = maximum(x = var_5895_cast_fp16, y = var_5896_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> y_183_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_183_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_5899_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5899_cast_fp16")];
            tensor<int32, [1]> var_5904_axes_0 = const()[name = string("op_5904_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5904_keep_dims_0 = const()[name = string("op_5904_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> var_5904_cast_fp16 = reduce_mean(axes = var_5904_axes_0, keep_dims = var_5904_keep_dims_0, x = var_5899_cast_fp16)[name = string("op_5904_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_5905_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5905_cast_fp16")];
            fp32 var_5906_epsilon_0 = const()[name = string("op_5906_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 512, 1]> var_5906_cast_fp16 = inverse(epsilon = var_5906_epsilon_0, x = var_5905_cast_fp16)[name = string("op_5906_cast_fp16")];
            fp16 var_5907_to_fp16 = const()[name = string("op_5907_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> var_5908_cast_fp16 = mul(x = var_5906_cast_fp16, y = var_5907_to_fp16)[name = string("op_5908_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_5910_cast_fp16 = add(x = var_5904_cast_fp16, y = var_5908_cast_fp16)[name = string("op_5910_cast_fp16")];
            fp32 var_5911_epsilon_0 = const()[name = string("op_5911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> var_5911_cast_fp16 = rsqrt(epsilon = var_5911_epsilon_0, x = var_5910_cast_fp16)[name = string("op_5911_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> out_161_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5911_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66918976)))];
            tensor<fp16, [1, 1, 512, 512]> k_67_cast_fp16 = mul(x = out_161_cast_fp16, y = tm_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_5864_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1, 512, 512]> var_5914_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5914_cast_fp16")];
            bool var_5917 = const()[name = string("op_5917"), val = bool(true)];
            tensor<int32, [1]> var_5918_axes_0 = const()[name = string("op_5918_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_5918_cast_fp16 = reduce_max(axes = var_5918_axes_0, keep_dims = var_5917, x = var_5914_cast_fp16)[name = string("op_5918_cast_fp16")];
            fp16 var_5919_to_fp16 = const()[name = string("op_5919_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 512, 1]> s_185_cast_fp16 = maximum(x = var_5918_cast_fp16, y = var_5919_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> y_185_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_185_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_5922_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5922_cast_fp16")];
            tensor<int32, [1]> var_5927_axes_0 = const()[name = string("op_5927_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5927_keep_dims_0 = const()[name = string("op_5927_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> var_5927_cast_fp16 = reduce_mean(axes = var_5927_axes_0, keep_dims = var_5927_keep_dims_0, x = var_5922_cast_fp16)[name = string("op_5927_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_5928_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5928_cast_fp16")];
            fp32 var_5929_epsilon_0 = const()[name = string("op_5929_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 512, 1]> var_5929_cast_fp16 = inverse(epsilon = var_5929_epsilon_0, x = var_5928_cast_fp16)[name = string("op_5929_cast_fp16")];
            fp16 var_5930_to_fp16 = const()[name = string("op_5930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> var_5931_cast_fp16 = mul(x = var_5929_cast_fp16, y = var_5930_to_fp16)[name = string("op_5931_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_5933_cast_fp16 = add(x = var_5927_cast_fp16, y = var_5931_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp32 var_5934_epsilon_0 = const()[name = string("op_5934_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> var_5934_cast_fp16 = rsqrt(epsilon = var_5934_epsilon_0, x = var_5933_cast_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> v_45_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5934_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_5936_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5936_cast_fp16")];
            tensor<int32, [4]> var_5941_begin_0 = const()[name = string("op_5941_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5941_end_0 = const()[name = string("op_5941_end_0"), val = tensor<int32, [4]>([1, 4, 512, 512])];
            tensor<bool, [4]> var_5941_end_mask_0 = const()[name = string("op_5941_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 256]> var_5941_cast_fp16 = slice_by_index(begin = var_5941_begin_0, end = var_5941_end_0, end_mask = var_5941_end_mask_0, x = q_45_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 256]> var_5942_cast_fp16 = mul(x = var_5941_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_5942_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 256]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = q_45_cast_fp16)[name = string("op_5947_cast_fp16")];
            int32 var_5949 = const()[name = string("op_5949"), val = int32(-1)];
            bool var_5950_interleave_0 = const()[name = string("op_5950_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> var_5950_cast_fp16 = concat(axis = var_5949, interleave = var_5950_interleave_0, values = (var_5942_cast_fp16, var_5947_cast_fp16))[name = string("op_5950_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_5951_cast_fp16 = mul(x = var_5950_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5951_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> q_47_cast_fp16 = add(x = var_5936_cast_fp16, y = var_5951_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_5954_cast_fp16 = mul(x = k_67_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5954_cast_fp16")];
            tensor<int32, [4]> var_5959_begin_0 = const()[name = string("op_5959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5959_end_0 = const()[name = string("op_5959_end_0"), val = tensor<int32, [4]>([1, 1, 512, 512])];
            tensor<bool, [4]> var_5959_end_mask_0 = const()[name = string("op_5959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 256]> var_5959_cast_fp16 = slice_by_index(begin = var_5959_begin_0, end = var_5959_end_0, end_mask = var_5959_end_mask_0, x = k_67_cast_fp16)[name = string("op_5959_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 256]> var_5960_cast_fp16 = mul(x = var_5959_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_5960_cast_fp16")];
            tensor<int32, [4]> var_5965_begin_0 = const()[name = string("op_5965_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5965_end_0 = const()[name = string("op_5965_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
            tensor<bool, [4]> var_5965_end_mask_0 = const()[name = string("op_5965_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 256]> var_5965_cast_fp16 = slice_by_index(begin = var_5965_begin_0, end = var_5965_end_0, end_mask = var_5965_end_mask_0, x = k_67_cast_fp16)[name = string("op_5965_cast_fp16")];
            int32 var_5967 = const()[name = string("op_5967"), val = int32(-1)];
            bool var_5968_interleave_0 = const()[name = string("op_5968_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 512]> var_5968_cast_fp16 = concat(axis = var_5967, interleave = var_5968_interleave_0, values = (var_5960_cast_fp16, var_5965_cast_fp16))[name = string("op_5968_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_5969_cast_fp16 = mul(x = var_5968_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5969_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> k_69_cast_fp16 = add(x = var_5954_cast_fp16, y = var_5969_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_22_reps_0 = const()[name = string("tile_22_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 512, 512]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = k_69_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [4, 1, 512, 512]> tile_22_cast_fp16 = tile(reps = tile_22_reps_0, x = transpose_44_cast_fp16)[name = string("tile_22_cast_fp16")];
            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([4, 1, 1, 512, 512])];
            tensor<fp16, [4, 1, 1, 512, 512]> reshape_44_cast_fp16 = reshape(shape = concat_44, x = tile_22_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [5]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
            tensor<fp16, [1, 4, 1, 512, 512]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [4, 1, 512, 512]> reshape_45_cast_fp16 = reshape(shape = concat_45, x = transpose_45_cast_fp16)[name = string("reshape_45_cast_fp16")];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_23_reps_0 = const()[name = string("tile_23_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 512, 512]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = v_45_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [4, 1, 512, 512]> tile_23_cast_fp16 = tile(reps = tile_23_reps_0, x = transpose_46_cast_fp16)[name = string("tile_23_cast_fp16")];
            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([4, 1, 1, 512, 512])];
            tensor<fp16, [4, 1, 1, 512, 512]> reshape_46_cast_fp16 = reshape(shape = concat_46, x = tile_23_cast_fp16)[name = string("reshape_46_cast_fp16")];
            tensor<int32, [5]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
            tensor<fp16, [1, 4, 1, 512, 512]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = reshape_46_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [4, 1, 512, 512]> reshape_47_cast_fp16 = reshape(shape = concat_47, x = transpose_47_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5983_transpose_x_0 = const()[name = string("op_5983_transpose_x_0"), val = bool(false)];
            bool var_5983_transpose_y_0 = const()[name = string("op_5983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = reshape_45_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 4, 512, 512]> var_5983_cast_fp16 = matmul(transpose_x = var_5983_transpose_x_0, transpose_y = var_5983_transpose_y_0, x = q_47_cast_fp16, y = transpose_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_5985_cast_fp16 = add(x = var_5983_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5985_cast_fp16")];
            int32 var_5986 = const()[name = string("op_5986"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_185_cast_fp16 = softmax(axis = var_5986, x = var_5985_cast_fp16)[name = string("w_185_cast_fp16")];
            bool var_5989_transpose_x_0 = const()[name = string("op_5989_transpose_x_0"), val = bool(false)];
            bool var_5989_transpose_y_0 = const()[name = string("op_5989_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 4, 512, 512]> var_5989_cast_fp16 = matmul(transpose_x = var_5989_transpose_x_0, transpose_y = var_5989_transpose_y_0, x = w_185_cast_fp16, y = v_47_cast_fp16)[name = string("op_5989_cast_fp16")];
            tensor<int32, [4]> var_5992_perm_0 = const()[name = string("op_5992_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5996 = const()[name = string("op_5996"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 512, 4, 512]> var_5992_cast_fp16 = transpose(perm = var_5992_perm_0, x = var_5989_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 512, 2048]> input_135_cast_fp16 = reshape(shape = var_5996, x = var_5992_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66920064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67968704))))[name = string("tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6001_cast_fp16 = abs(x = linear_103_cast_fp16)[name = string("op_6001_cast_fp16")];
            bool var_6004 = const()[name = string("op_6004"), val = bool(true)];
            tensor<int32, [1]> var_6005_axes_0 = const()[name = string("op_6005_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6005_cast_fp16 = reduce_max(axes = var_6005_axes_0, keep_dims = var_6004, x = var_6001_cast_fp16)[name = string("op_6005_cast_fp16")];
            fp16 var_6006_to_fp16 = const()[name = string("op_6006_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_187_cast_fp16 = maximum(x = var_6005_cast_fp16, y = var_6006_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_187_cast_fp16 = real_div(x = linear_103_cast_fp16, y = s_187_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6009_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_6009_cast_fp16")];
            tensor<int32, [1]> var_6014_axes_0 = const()[name = string("op_6014_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6014_keep_dims_0 = const()[name = string("op_6014_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6014_cast_fp16 = reduce_mean(axes = var_6014_axes_0, keep_dims = var_6014_keep_dims_0, x = var_6009_cast_fp16)[name = string("op_6014_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6015_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_6015_cast_fp16")];
            fp32 var_6016_epsilon_0 = const()[name = string("op_6016_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6016_cast_fp16 = inverse(epsilon = var_6016_epsilon_0, x = var_6015_cast_fp16)[name = string("op_6016_cast_fp16")];
            fp16 var_6017_to_fp16 = const()[name = string("op_6017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6018_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6020_cast_fp16 = add(x = var_6014_cast_fp16, y = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp32 var_6021_epsilon_0 = const()[name = string("op_6021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6021_cast_fp16 = rsqrt(epsilon = var_6021_epsilon_0, x = var_6020_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_163_cast_fp16 = mul(x = y_187_cast_fp16, y = var_6021_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969792)))];
            tensor<fp16, [1, 512, 512]> var_6023_cast_fp16 = mul(x = out_163_cast_fp16, y = tm_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_209_cast_fp16 = add(x = x_199_cast_fp16, y = var_6023_cast_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6026_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_6026_cast_fp16")];
            bool var_6029 = const()[name = string("op_6029"), val = bool(true)];
            tensor<int32, [1]> var_6030_axes_0 = const()[name = string("op_6030_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6030_cast_fp16 = reduce_max(axes = var_6030_axes_0, keep_dims = var_6029, x = var_6026_cast_fp16)[name = string("op_6030_cast_fp16")];
            fp16 var_6031_to_fp16 = const()[name = string("op_6031_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_189_cast_fp16 = maximum(x = var_6030_cast_fp16, y = var_6031_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_189_cast_fp16 = real_div(x = x_209_cast_fp16, y = s_189_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6034_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [1]> var_6039_axes_0 = const()[name = string("op_6039_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6039_keep_dims_0 = const()[name = string("op_6039_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6039_cast_fp16 = reduce_mean(axes = var_6039_axes_0, keep_dims = var_6039_keep_dims_0, x = var_6034_cast_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6040_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6041_cast_fp16 = inverse(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            fp16 var_6042_to_fp16 = const()[name = string("op_6042_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6043_cast_fp16 = mul(x = var_6041_cast_fp16, y = var_6042_to_fp16)[name = string("op_6043_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6045_cast_fp16 = add(x = var_6039_cast_fp16, y = var_6043_cast_fp16)[name = string("op_6045_cast_fp16")];
            fp32 var_6046_epsilon_0 = const()[name = string("op_6046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6046_cast_fp16 = rsqrt(epsilon = var_6046_epsilon_0, x = var_6045_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_165_cast_fp16 = mul(x = y_189_cast_fp16, y = var_6046_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67970880)))];
            tensor<fp16, [1, 512, 512]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = tm_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971968))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69020608))))[name = string("tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_104_cast_fp16")];
            string var_6053_mode_0 = const()[name = string("op_6053_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_6053_cast_fp16 = gelu(mode = var_6053_mode_0, x = linear_104_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70073408))))[name = string("tm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_105_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_139_cast_fp16 = mul(x = var_6053_cast_fp16, y = linear_105_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70077568))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71126208))))[name = string("tm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6061_cast_fp16 = abs(x = linear_106_cast_fp16)[name = string("op_6061_cast_fp16")];
            bool var_6064 = const()[name = string("op_6064"), val = bool(true)];
            tensor<int32, [1]> var_6065_axes_0 = const()[name = string("op_6065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6065_cast_fp16 = reduce_max(axes = var_6065_axes_0, keep_dims = var_6064, x = var_6061_cast_fp16)[name = string("op_6065_cast_fp16")];
            fp16 var_6066_to_fp16 = const()[name = string("op_6066_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_191_cast_fp16 = maximum(x = var_6065_cast_fp16, y = var_6066_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_191_cast_fp16 = real_div(x = linear_106_cast_fp16, y = s_191_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6069_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<int32, [1]> var_6074_axes_0 = const()[name = string("op_6074_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6074_keep_dims_0 = const()[name = string("op_6074_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6074_cast_fp16 = reduce_mean(axes = var_6074_axes_0, keep_dims = var_6074_keep_dims_0, x = var_6069_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6075_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_6075_cast_fp16")];
            fp32 var_6076_epsilon_0 = const()[name = string("op_6076_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6076_cast_fp16 = inverse(epsilon = var_6076_epsilon_0, x = var_6075_cast_fp16)[name = string("op_6076_cast_fp16")];
            fp16 var_6077_to_fp16 = const()[name = string("op_6077_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6078_cast_fp16 = mul(x = var_6076_cast_fp16, y = var_6077_to_fp16)[name = string("op_6078_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6080_cast_fp16 = add(x = var_6074_cast_fp16, y = var_6078_cast_fp16)[name = string("op_6080_cast_fp16")];
            fp32 var_6081_epsilon_0 = const()[name = string("op_6081_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6081_cast_fp16 = rsqrt(epsilon = var_6081_epsilon_0, x = var_6080_cast_fp16)[name = string("op_6081_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_167_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6081_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71127296)))];
            tensor<fp16, [1, 512, 512]> var_6083_cast_fp16 = mul(x = out_167_cast_fp16, y = tm_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6083_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_141_cast_fp16 = add(x = x_209_cast_fp16, y = var_6083_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71128384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71390592))))[name = string("tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_107_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_107_cast_fp16")];
            string x_213_mode_0 = const()[name = string("x_213_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_213_cast_fp16 = gelu(mode = x_213_mode_0, x = linear_107_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [4]> var_6103_begin_0 = const()[name = string("op_6103_begin_0"), val = tensor<int32, [4]>([0, 0, 11, 0])];
            tensor<int32, [4]> var_6103_end_0 = const()[name = string("op_6103_end_0"), val = tensor<int32, [4]>([1, 512, 12, 512])];
            tensor<bool, [4]> var_6103_end_mask_0 = const()[name = string("op_6103_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6103_squeeze_mask_0 = const()[name = string("op_6103_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_6103_cast_fp16 = slice_by_index(begin = var_6103_begin_0, end = var_6103_end_0, end_mask = var_6103_end_mask_0, squeeze_mask = var_6103_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6103_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_143_cast_fp16 = mul(x = x_213_cast_fp16, y = var_6103_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71391680))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71653888))))[name = string("tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6113_cast_fp16 = abs(x = linear_108_cast_fp16)[name = string("op_6113_cast_fp16")];
            bool var_6116 = const()[name = string("op_6116"), val = bool(true)];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6117_cast_fp16 = reduce_max(axes = var_6117_axes_0, keep_dims = var_6116, x = var_6113_cast_fp16)[name = string("op_6117_cast_fp16")];
            fp16 var_6118_to_fp16 = const()[name = string("op_6118_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_193_cast_fp16 = maximum(x = var_6117_cast_fp16, y = var_6118_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_193_cast_fp16 = real_div(x = linear_108_cast_fp16, y = s_193_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6121_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6121_cast_fp16")];
            tensor<int32, [1]> var_6126_axes_0 = const()[name = string("op_6126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6126_keep_dims_0 = const()[name = string("op_6126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6126_cast_fp16 = reduce_mean(axes = var_6126_axes_0, keep_dims = var_6126_keep_dims_0, x = var_6121_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6127_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_6127_cast_fp16")];
            fp32 var_6128_epsilon_0 = const()[name = string("op_6128_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6128_cast_fp16 = inverse(epsilon = var_6128_epsilon_0, x = var_6127_cast_fp16)[name = string("op_6128_cast_fp16")];
            fp16 var_6129_to_fp16 = const()[name = string("op_6129_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6130_cast_fp16 = mul(x = var_6128_cast_fp16, y = var_6129_to_fp16)[name = string("op_6130_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6132_cast_fp16 = add(x = var_6126_cast_fp16, y = var_6130_cast_fp16)[name = string("op_6132_cast_fp16")];
            fp32 var_6133_epsilon_0 = const()[name = string("op_6133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6133_cast_fp16 = rsqrt(epsilon = var_6133_epsilon_0, x = var_6132_cast_fp16)[name = string("op_6133_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_169_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6133_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71654976)))];
            tensor<fp16, [1, 512, 512]> var_6135_cast_fp16 = mul(x = out_169_cast_fp16, y = tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6135_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_23_cast_fp16 = add(x = input_141_cast_fp16, y = var_6135_cast_fp16)[name = string("h_23_cast_fp16")];
            tensor<fp16, [1]> tm_layers_11_layer_scalar_to_fp16 = const()[name = string("tm_layers_11_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.86p-2])];
            tensor<fp16, [1, 512, 512]> x_217_cast_fp16 = mul(x = h_23_cast_fp16, y = tm_layers_11_layer_scalar_to_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6139_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_195_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_195_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_195_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6153_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_171_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71656064)))];
            tensor<fp16, [1, 512, 512]> input_145_cast_fp16 = mul(x = out_171_cast_fp16, y = tm_layers_12_input_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71657152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72181504))))[name = string("tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<int32, [4]> var_6169 = const()[name = string("op_6169"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_6170_cast_fp16 = reshape(shape = var_6169, x = linear_109_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72183616))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72445824))))[name = string("tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [4]> var_6181 = const()[name = string("op_6181"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_6182_cast_fp16 = reshape(shape = var_6181, x = linear_110_cast_fp16)[name = string("op_6182_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72446912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72709120))))[name = string("tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<int32, [4]> var_6193 = const()[name = string("op_6193"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_6194_cast_fp16 = reshape(shape = var_6193, x = linear_111_cast_fp16)[name = string("op_6194_cast_fp16")];
            tensor<int32, [4]> x_223_perm_0 = const()[name = string("x_223_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6170_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 4, 512, 256]> var_6198_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_6198_cast_fp16")];
            bool var_6201 = const()[name = string("op_6201"), val = bool(true)];
            tensor<int32, [1]> var_6202_axes_0 = const()[name = string("op_6202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_6202_cast_fp16 = reduce_max(axes = var_6202_axes_0, keep_dims = var_6201, x = var_6198_cast_fp16)[name = string("op_6202_cast_fp16")];
            fp16 var_6203_to_fp16 = const()[name = string("op_6203_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_197_cast_fp16 = maximum(x = var_6202_cast_fp16, y = var_6203_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_197_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_197_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6206_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6206_cast_fp16")];
            tensor<int32, [1]> var_6211_axes_0 = const()[name = string("op_6211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6211_keep_dims_0 = const()[name = string("op_6211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_6211_cast_fp16 = reduce_mean(axes = var_6211_axes_0, keep_dims = var_6211_keep_dims_0, x = var_6206_cast_fp16)[name = string("op_6211_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_6212_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6212_cast_fp16")];
            fp32 var_6213_epsilon_0 = const()[name = string("op_6213_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_6213_cast_fp16 = inverse(epsilon = var_6213_epsilon_0, x = var_6212_cast_fp16)[name = string("op_6213_cast_fp16")];
            fp16 var_6214_to_fp16 = const()[name = string("op_6214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_6215_cast_fp16 = mul(x = var_6213_cast_fp16, y = var_6214_to_fp16)[name = string("op_6215_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_6217_cast_fp16 = add(x = var_6211_cast_fp16, y = var_6215_cast_fp16)[name = string("op_6217_cast_fp16")];
            fp32 var_6218_epsilon_0 = const()[name = string("op_6218_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_6218_cast_fp16 = rsqrt(epsilon = var_6218_epsilon_0, x = var_6217_cast_fp16)[name = string("op_6218_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_173_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6218_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_49_cast_fp16 = mul(x = out_173_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_49_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6182_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 2, 512, 256]> var_6221_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_6221_cast_fp16")];
            bool var_6224 = const()[name = string("op_6224"), val = bool(true)];
            tensor<int32, [1]> var_6225_axes_0 = const()[name = string("op_6225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_6225_cast_fp16 = reduce_max(axes = var_6225_axes_0, keep_dims = var_6224, x = var_6221_cast_fp16)[name = string("op_6225_cast_fp16")];
            fp16 var_6226_to_fp16 = const()[name = string("op_6226_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_199_cast_fp16 = maximum(x = var_6225_cast_fp16, y = var_6226_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_199_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_199_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6229_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6229_cast_fp16")];
            tensor<int32, [1]> var_6234_axes_0 = const()[name = string("op_6234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6234_keep_dims_0 = const()[name = string("op_6234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_6234_cast_fp16 = reduce_mean(axes = var_6234_axes_0, keep_dims = var_6234_keep_dims_0, x = var_6229_cast_fp16)[name = string("op_6234_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6235_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6235_cast_fp16")];
            fp32 var_6236_epsilon_0 = const()[name = string("op_6236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_6236_cast_fp16 = inverse(epsilon = var_6236_epsilon_0, x = var_6235_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_6238_cast_fp16 = mul(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6240_cast_fp16 = add(x = var_6234_cast_fp16, y = var_6238_cast_fp16)[name = string("op_6240_cast_fp16")];
            fp32 var_6241_epsilon_0 = const()[name = string("op_6241_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_6241_cast_fp16 = rsqrt(epsilon = var_6241_epsilon_0, x = var_6240_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_175_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6241_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_73_cast_fp16 = mul(x = out_175_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_223_cast_fp16 = transpose(perm = x_223_perm_0, x = var_6194_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2, 512, 256]> var_6244_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_6244_cast_fp16")];
            bool var_6247 = const()[name = string("op_6247"), val = bool(true)];
            tensor<int32, [1]> var_6248_axes_0 = const()[name = string("op_6248_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_6248_cast_fp16 = reduce_max(axes = var_6248_axes_0, keep_dims = var_6247, x = var_6244_cast_fp16)[name = string("op_6248_cast_fp16")];
            fp16 var_6249_to_fp16 = const()[name = string("op_6249_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_201_cast_fp16 = maximum(x = var_6248_cast_fp16, y = var_6249_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_201_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_201_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6252_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6252_cast_fp16")];
            tensor<int32, [1]> var_6257_axes_0 = const()[name = string("op_6257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6257_keep_dims_0 = const()[name = string("op_6257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_6257_cast_fp16 = reduce_mean(axes = var_6257_axes_0, keep_dims = var_6257_keep_dims_0, x = var_6252_cast_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6258_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6258_cast_fp16")];
            fp32 var_6259_epsilon_0 = const()[name = string("op_6259_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_6259_cast_fp16 = inverse(epsilon = var_6259_epsilon_0, x = var_6258_cast_fp16)[name = string("op_6259_cast_fp16")];
            fp16 var_6260_to_fp16 = const()[name = string("op_6260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_6261_cast_fp16 = mul(x = var_6259_cast_fp16, y = var_6260_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6263_cast_fp16 = add(x = var_6257_cast_fp16, y = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp32 var_6264_epsilon_0 = const()[name = string("op_6264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_6264_cast_fp16 = rsqrt(epsilon = var_6264_epsilon_0, x = var_6263_cast_fp16)[name = string("op_6264_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_49_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6264_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6266_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6266_cast_fp16")];
            tensor<int32, [4]> var_6271_begin_0 = const()[name = string("op_6271_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6271_end_0 = const()[name = string("op_6271_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_6271_end_mask_0 = const()[name = string("op_6271_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_6271_cast_fp16 = slice_by_index(begin = var_6271_begin_0, end = var_6271_end_0, end_mask = var_6271_end_mask_0, x = q_49_cast_fp16)[name = string("op_6271_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_6272_cast_fp16 = mul(x = var_6271_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_6272_cast_fp16")];
            tensor<int32, [4]> var_6277_begin_0 = const()[name = string("op_6277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6277_end_0 = const()[name = string("op_6277_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_6277_end_mask_0 = const()[name = string("op_6277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_6277_cast_fp16 = slice_by_index(begin = var_6277_begin_0, end = var_6277_end_0, end_mask = var_6277_end_mask_0, x = q_49_cast_fp16)[name = string("op_6277_cast_fp16")];
            int32 var_6279 = const()[name = string("op_6279"), val = int32(-1)];
            bool var_6280_interleave_0 = const()[name = string("op_6280_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_6280_cast_fp16 = concat(axis = var_6279, interleave = var_6280_interleave_0, values = (var_6272_cast_fp16, var_6277_cast_fp16))[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6281_cast_fp16 = mul(x = var_6280_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6281_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_51_cast_fp16 = add(x = var_6266_cast_fp16, y = var_6281_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6284_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6284_cast_fp16")];
            tensor<int32, [4]> var_6289_begin_0 = const()[name = string("op_6289_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6289_end_0 = const()[name = string("op_6289_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_6289_end_mask_0 = const()[name = string("op_6289_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_6289_cast_fp16 = slice_by_index(begin = var_6289_begin_0, end = var_6289_end_0, end_mask = var_6289_end_mask_0, x = k_73_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_6290_cast_fp16 = mul(x = var_6289_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_6290_cast_fp16")];
            tensor<int32, [4]> var_6295_begin_0 = const()[name = string("op_6295_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6295_end_0 = const()[name = string("op_6295_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_6295_end_mask_0 = const()[name = string("op_6295_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_6295_cast_fp16 = slice_by_index(begin = var_6295_begin_0, end = var_6295_end_0, end_mask = var_6295_end_mask_0, x = k_73_cast_fp16)[name = string("op_6295_cast_fp16")];
            int32 var_6297 = const()[name = string("op_6297"), val = int32(-1)];
            bool var_6298_interleave_0 = const()[name = string("op_6298_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_6298_cast_fp16 = concat(axis = var_6297, interleave = var_6298_interleave_0, values = (var_6290_cast_fp16, var_6295_cast_fp16))[name = string("op_6298_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6299_cast_fp16 = mul(x = var_6298_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6299_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_75_cast_fp16 = add(x = var_6284_cast_fp16, y = var_6299_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_24_reps_0 = const()[name = string("tile_24_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = k_75_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [4, 1, 512, 256]> tile_24_cast_fp16 = tile(reps = tile_24_reps_0, x = transpose_48_cast_fp16)[name = string("tile_24_cast_fp16")];
            tensor<int32, [5]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_48_cast_fp16 = reshape(shape = concat_48, x = tile_24_cast_fp16)[name = string("reshape_48_cast_fp16")];
            tensor<int32, [5]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = reshape_48_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [4, 1, 512, 256]> reshape_49_cast_fp16 = reshape(shape = concat_49, x = transpose_49_cast_fp16)[name = string("reshape_49_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_25_reps_0 = const()[name = string("tile_25_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = v_49_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [4, 1, 512, 256]> tile_25_cast_fp16 = tile(reps = tile_25_reps_0, x = transpose_50_cast_fp16)[name = string("tile_25_cast_fp16")];
            tensor<int32, [5]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_50_cast_fp16 = reshape(shape = concat_50, x = tile_25_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [5]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [4, 1, 512, 256]> reshape_51_cast_fp16 = reshape(shape = concat_51, x = transpose_51_cast_fp16)[name = string("reshape_51_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6313_transpose_x_0 = const()[name = string("op_6313_transpose_x_0"), val = bool(false)];
            bool var_6313_transpose_y_0 = const()[name = string("op_6313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = reshape_49_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 4, 512, 512]> var_6313_cast_fp16 = matmul(transpose_x = var_6313_transpose_x_0, transpose_y = var_6313_transpose_y_0, x = q_51_cast_fp16, y = transpose_108_cast_fp16)[name = string("op_6313_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_6315_cast_fp16 = add(x = var_6313_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6315_cast_fp16")];
            int32 var_6316 = const()[name = string("op_6316"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_201_cast_fp16 = softmax(axis = var_6316, x = var_6315_cast_fp16)[name = string("w_201_cast_fp16")];
            bool var_6319_transpose_x_0 = const()[name = string("op_6319_transpose_x_0"), val = bool(false)];
            bool var_6319_transpose_y_0 = const()[name = string("op_6319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = reshape_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 4, 512, 256]> var_6319_cast_fp16 = matmul(transpose_x = var_6319_transpose_x_0, transpose_y = var_6319_transpose_y_0, x = w_201_cast_fp16, y = v_51_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<int32, [4]> var_6322_perm_0 = const()[name = string("op_6322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6326 = const()[name = string("op_6326"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_6322_cast_fp16 = transpose(perm = var_6322_perm_0, x = var_6319_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 512, 1024]> input_147_cast_fp16 = reshape(shape = var_6326, x = var_6322_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72710208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73234560))))[name = string("tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6331_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6331_cast_fp16")];
            bool var_6334 = const()[name = string("op_6334"), val = bool(true)];
            tensor<int32, [1]> var_6335_axes_0 = const()[name = string("op_6335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6335_cast_fp16 = reduce_max(axes = var_6335_axes_0, keep_dims = var_6334, x = var_6331_cast_fp16)[name = string("op_6335_cast_fp16")];
            fp16 var_6336_to_fp16 = const()[name = string("op_6336_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_203_cast_fp16 = maximum(x = var_6335_cast_fp16, y = var_6336_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_203_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_203_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6339_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6339_cast_fp16")];
            tensor<int32, [1]> var_6344_axes_0 = const()[name = string("op_6344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6344_keep_dims_0 = const()[name = string("op_6344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6344_cast_fp16 = reduce_mean(axes = var_6344_axes_0, keep_dims = var_6344_keep_dims_0, x = var_6339_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6345_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6345_cast_fp16")];
            fp32 var_6346_epsilon_0 = const()[name = string("op_6346_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6346_cast_fp16 = inverse(epsilon = var_6346_epsilon_0, x = var_6345_cast_fp16)[name = string("op_6346_cast_fp16")];
            fp16 var_6347_to_fp16 = const()[name = string("op_6347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6348_cast_fp16 = mul(x = var_6346_cast_fp16, y = var_6347_to_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6350_cast_fp16 = add(x = var_6344_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            fp32 var_6351_epsilon_0 = const()[name = string("op_6351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6351_cast_fp16 = rsqrt(epsilon = var_6351_epsilon_0, x = var_6350_cast_fp16)[name = string("op_6351_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_177_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6351_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73235648)))];
            tensor<fp16, [1, 512, 512]> var_6353_cast_fp16 = mul(x = out_177_cast_fp16, y = tm_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = var_6353_cast_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6356_cast_fp16 = abs(x = x_227_cast_fp16)[name = string("op_6356_cast_fp16")];
            bool var_6359 = const()[name = string("op_6359"), val = bool(true)];
            tensor<int32, [1]> var_6360_axes_0 = const()[name = string("op_6360_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6360_cast_fp16 = reduce_max(axes = var_6360_axes_0, keep_dims = var_6359, x = var_6356_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_205_cast_fp16 = maximum(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_205_cast_fp16 = real_div(x = x_227_cast_fp16, y = s_205_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6364_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<int32, [1]> var_6369_axes_0 = const()[name = string("op_6369_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6369_keep_dims_0 = const()[name = string("op_6369_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6369_cast_fp16 = reduce_mean(axes = var_6369_axes_0, keep_dims = var_6369_keep_dims_0, x = var_6364_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6370_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6370_cast_fp16")];
            fp32 var_6371_epsilon_0 = const()[name = string("op_6371_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6371_cast_fp16 = inverse(epsilon = var_6371_epsilon_0, x = var_6370_cast_fp16)[name = string("op_6371_cast_fp16")];
            fp16 var_6372_to_fp16 = const()[name = string("op_6372_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6373_cast_fp16 = mul(x = var_6371_cast_fp16, y = var_6372_to_fp16)[name = string("op_6373_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6375_cast_fp16 = add(x = var_6369_cast_fp16, y = var_6373_cast_fp16)[name = string("op_6375_cast_fp16")];
            fp32 var_6376_epsilon_0 = const()[name = string("op_6376_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6376_cast_fp16 = rsqrt(epsilon = var_6376_epsilon_0, x = var_6375_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_179_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6376_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73236736)))];
            tensor<fp16, [1, 512, 512]> input_149_cast_fp16 = mul(x = out_179_cast_fp16, y = tm_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73237824))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74286464))))[name = string("tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_113_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_113_cast_fp16")];
            string var_6383_mode_0 = const()[name = string("op_6383_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_6383_cast_fp16 = gelu(mode = var_6383_mode_0, x = linear_113_cast_fp16)[name = string("op_6383_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74290624))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339264))))[name = string("tm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_114_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_151_cast_fp16 = mul(x = var_6383_cast_fp16, y = linear_114_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75343424))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76392064))))[name = string("tm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6391_cast_fp16 = abs(x = linear_115_cast_fp16)[name = string("op_6391_cast_fp16")];
            bool var_6394 = const()[name = string("op_6394"), val = bool(true)];
            tensor<int32, [1]> var_6395_axes_0 = const()[name = string("op_6395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6395_cast_fp16 = reduce_max(axes = var_6395_axes_0, keep_dims = var_6394, x = var_6391_cast_fp16)[name = string("op_6395_cast_fp16")];
            fp16 var_6396_to_fp16 = const()[name = string("op_6396_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_207_cast_fp16 = maximum(x = var_6395_cast_fp16, y = var_6396_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_207_cast_fp16 = real_div(x = linear_115_cast_fp16, y = s_207_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6399_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6399_cast_fp16")];
            tensor<int32, [1]> var_6404_axes_0 = const()[name = string("op_6404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6404_keep_dims_0 = const()[name = string("op_6404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6404_cast_fp16 = reduce_mean(axes = var_6404_axes_0, keep_dims = var_6404_keep_dims_0, x = var_6399_cast_fp16)[name = string("op_6404_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6405_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6405_cast_fp16")];
            fp32 var_6406_epsilon_0 = const()[name = string("op_6406_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6406_cast_fp16 = inverse(epsilon = var_6406_epsilon_0, x = var_6405_cast_fp16)[name = string("op_6406_cast_fp16")];
            fp16 var_6407_to_fp16 = const()[name = string("op_6407_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6408_cast_fp16 = mul(x = var_6406_cast_fp16, y = var_6407_to_fp16)[name = string("op_6408_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6410_cast_fp16 = add(x = var_6404_cast_fp16, y = var_6408_cast_fp16)[name = string("op_6410_cast_fp16")];
            fp32 var_6411_epsilon_0 = const()[name = string("op_6411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6411_cast_fp16 = rsqrt(epsilon = var_6411_epsilon_0, x = var_6410_cast_fp16)[name = string("op_6411_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_181_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6411_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76393152)))];
            tensor<fp16, [1, 512, 512]> var_6413_cast_fp16 = mul(x = out_181_cast_fp16, y = tm_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6413_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_153_cast_fp16 = add(x = x_227_cast_fp16, y = var_6413_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76394240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76656448))))[name = string("tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_116_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_116_cast_fp16")];
            string x_231_mode_0 = const()[name = string("x_231_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_231_cast_fp16 = gelu(mode = x_231_mode_0, x = linear_116_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_6433_begin_0 = const()[name = string("op_6433_begin_0"), val = tensor<int32, [4]>([0, 0, 12, 0])];
            tensor<int32, [4]> var_6433_end_0 = const()[name = string("op_6433_end_0"), val = tensor<int32, [4]>([1, 512, 13, 512])];
            tensor<bool, [4]> var_6433_end_mask_0 = const()[name = string("op_6433_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6433_squeeze_mask_0 = const()[name = string("op_6433_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_6433_cast_fp16 = slice_by_index(begin = var_6433_begin_0, end = var_6433_end_0, end_mask = var_6433_end_mask_0, squeeze_mask = var_6433_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6433_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_155_cast_fp16 = mul(x = x_231_cast_fp16, y = var_6433_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76657536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76919744))))[name = string("tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6443_cast_fp16 = abs(x = linear_117_cast_fp16)[name = string("op_6443_cast_fp16")];
            bool var_6446 = const()[name = string("op_6446"), val = bool(true)];
            tensor<int32, [1]> var_6447_axes_0 = const()[name = string("op_6447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6447_cast_fp16 = reduce_max(axes = var_6447_axes_0, keep_dims = var_6446, x = var_6443_cast_fp16)[name = string("op_6447_cast_fp16")];
            fp16 var_6448_to_fp16 = const()[name = string("op_6448_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_209_cast_fp16 = maximum(x = var_6447_cast_fp16, y = var_6448_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_209_cast_fp16 = real_div(x = linear_117_cast_fp16, y = s_209_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6451_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6451_cast_fp16")];
            tensor<int32, [1]> var_6456_axes_0 = const()[name = string("op_6456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6456_keep_dims_0 = const()[name = string("op_6456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6456_cast_fp16 = reduce_mean(axes = var_6456_axes_0, keep_dims = var_6456_keep_dims_0, x = var_6451_cast_fp16)[name = string("op_6456_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6457_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6457_cast_fp16")];
            fp32 var_6458_epsilon_0 = const()[name = string("op_6458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6458_cast_fp16 = inverse(epsilon = var_6458_epsilon_0, x = var_6457_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6459_to_fp16 = const()[name = string("op_6459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6460_cast_fp16 = mul(x = var_6458_cast_fp16, y = var_6459_to_fp16)[name = string("op_6460_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6462_cast_fp16 = add(x = var_6456_cast_fp16, y = var_6460_cast_fp16)[name = string("op_6462_cast_fp16")];
            fp32 var_6463_epsilon_0 = const()[name = string("op_6463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6463_cast_fp16 = rsqrt(epsilon = var_6463_epsilon_0, x = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_183_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6463_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76920832)))];
            tensor<fp16, [1, 512, 512]> var_6465_cast_fp16 = mul(x = out_183_cast_fp16, y = tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6465_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_25_cast_fp16 = add(x = input_153_cast_fp16, y = var_6465_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<fp16, [1]> tm_layers_12_layer_scalar_to_fp16 = const()[name = string("tm_layers_12_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.8cp-1])];
            tensor<fp16, [1, 512, 512]> x_235_cast_fp16 = mul(x = h_25_cast_fp16, y = tm_layers_12_layer_scalar_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6469_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_6469_cast_fp16")];
            bool var_6472 = const()[name = string("op_6472"), val = bool(true)];
            tensor<int32, [1]> var_6473_axes_0 = const()[name = string("op_6473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6473_cast_fp16 = reduce_max(axes = var_6473_axes_0, keep_dims = var_6472, x = var_6469_cast_fp16)[name = string("op_6473_cast_fp16")];
            fp16 var_6474_to_fp16 = const()[name = string("op_6474_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_211_cast_fp16 = maximum(x = var_6473_cast_fp16, y = var_6474_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_211_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_211_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6477_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6477_cast_fp16")];
            tensor<int32, [1]> var_6482_axes_0 = const()[name = string("op_6482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6482_keep_dims_0 = const()[name = string("op_6482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6482_cast_fp16 = reduce_mean(axes = var_6482_axes_0, keep_dims = var_6482_keep_dims_0, x = var_6477_cast_fp16)[name = string("op_6482_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6483_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6483_cast_fp16")];
            fp32 var_6484_epsilon_0 = const()[name = string("op_6484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6484_cast_fp16 = inverse(epsilon = var_6484_epsilon_0, x = var_6483_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp16 var_6485_to_fp16 = const()[name = string("op_6485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6486_cast_fp16 = mul(x = var_6484_cast_fp16, y = var_6485_to_fp16)[name = string("op_6486_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6488_cast_fp16 = add(x = var_6482_cast_fp16, y = var_6486_cast_fp16)[name = string("op_6488_cast_fp16")];
            fp32 var_6489_epsilon_0 = const()[name = string("op_6489_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6489_cast_fp16 = rsqrt(epsilon = var_6489_epsilon_0, x = var_6488_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_185_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6489_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76921920)))];
            tensor<fp16, [1, 512, 512]> input_157_cast_fp16 = mul(x = out_185_cast_fp16, y = tm_layers_13_input_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76923008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77447360))))[name = string("tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<int32, [4]> var_6499 = const()[name = string("op_6499"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_6500_cast_fp16 = reshape(shape = var_6499, x = linear_118_cast_fp16)[name = string("op_6500_cast_fp16")];
            tensor<int32, [4]> x_237_perm_0 = const()[name = string("x_237_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77449472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77711680))))[name = string("tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6511 = const()[name = string("op_6511"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_6512_cast_fp16 = reshape(shape = var_6511, x = linear_119_cast_fp16)[name = string("op_6512_cast_fp16")];
            tensor<int32, [4]> x_239_perm_0 = const()[name = string("x_239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77712768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77974976))))[name = string("tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6523 = const()[name = string("op_6523"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_6524_cast_fp16 = reshape(shape = var_6523, x = linear_120_cast_fp16)[name = string("op_6524_cast_fp16")];
            tensor<int32, [4]> x_241_perm_0 = const()[name = string("x_241_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_237_cast_fp16 = transpose(perm = x_237_perm_0, x = var_6500_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 4, 512, 256]> var_6528_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_6528_cast_fp16")];
            bool var_6531 = const()[name = string("op_6531"), val = bool(true)];
            tensor<int32, [1]> var_6532_axes_0 = const()[name = string("op_6532_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_6532_cast_fp16 = reduce_max(axes = var_6532_axes_0, keep_dims = var_6531, x = var_6528_cast_fp16)[name = string("op_6532_cast_fp16")];
            fp16 var_6533_to_fp16 = const()[name = string("op_6533_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_213_cast_fp16 = maximum(x = var_6532_cast_fp16, y = var_6533_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_213_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_213_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6536_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6536_cast_fp16")];
            tensor<int32, [1]> var_6541_axes_0 = const()[name = string("op_6541_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6541_keep_dims_0 = const()[name = string("op_6541_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_6541_cast_fp16 = reduce_mean(axes = var_6541_axes_0, keep_dims = var_6541_keep_dims_0, x = var_6536_cast_fp16)[name = string("op_6541_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_6542_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp32 var_6543_epsilon_0 = const()[name = string("op_6543_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_6543_cast_fp16 = inverse(epsilon = var_6543_epsilon_0, x = var_6542_cast_fp16)[name = string("op_6543_cast_fp16")];
            fp16 var_6544_to_fp16 = const()[name = string("op_6544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_6545_cast_fp16 = mul(x = var_6543_cast_fp16, y = var_6544_to_fp16)[name = string("op_6545_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_6547_cast_fp16 = add(x = var_6541_cast_fp16, y = var_6545_cast_fp16)[name = string("op_6547_cast_fp16")];
            fp32 var_6548_epsilon_0 = const()[name = string("op_6548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_6548_cast_fp16 = rsqrt(epsilon = var_6548_epsilon_0, x = var_6547_cast_fp16)[name = string("op_6548_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_187_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6548_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976064)))];
            tensor<fp16, [1, 4, 512, 256]> q_53_cast_fp16 = mul(x = out_187_cast_fp16, y = tm_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("q_53_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_239_cast_fp16 = transpose(perm = x_239_perm_0, x = var_6512_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2, 512, 256]> var_6551_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_6551_cast_fp16")];
            bool var_6554 = const()[name = string("op_6554"), val = bool(true)];
            tensor<int32, [1]> var_6555_axes_0 = const()[name = string("op_6555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_6555_cast_fp16 = reduce_max(axes = var_6555_axes_0, keep_dims = var_6554, x = var_6551_cast_fp16)[name = string("op_6555_cast_fp16")];
            fp16 var_6556_to_fp16 = const()[name = string("op_6556_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_215_cast_fp16 = maximum(x = var_6555_cast_fp16, y = var_6556_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_215_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_215_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6559_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6559_cast_fp16")];
            tensor<int32, [1]> var_6564_axes_0 = const()[name = string("op_6564_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6564_keep_dims_0 = const()[name = string("op_6564_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_6564_cast_fp16 = reduce_mean(axes = var_6564_axes_0, keep_dims = var_6564_keep_dims_0, x = var_6559_cast_fp16)[name = string("op_6564_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6565_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6565_cast_fp16")];
            fp32 var_6566_epsilon_0 = const()[name = string("op_6566_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_6566_cast_fp16 = inverse(epsilon = var_6566_epsilon_0, x = var_6565_cast_fp16)[name = string("op_6566_cast_fp16")];
            fp16 var_6567_to_fp16 = const()[name = string("op_6567_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_6568_cast_fp16 = mul(x = var_6566_cast_fp16, y = var_6567_to_fp16)[name = string("op_6568_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6570_cast_fp16 = add(x = var_6564_cast_fp16, y = var_6568_cast_fp16)[name = string("op_6570_cast_fp16")];
            fp32 var_6571_epsilon_0 = const()[name = string("op_6571_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_6571_cast_fp16 = rsqrt(epsilon = var_6571_epsilon_0, x = var_6570_cast_fp16)[name = string("op_6571_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_189_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6571_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976640)))];
            tensor<fp16, [1, 2, 512, 256]> k_79_cast_fp16 = mul(x = out_189_cast_fp16, y = tm_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("k_79_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_241_cast_fp16 = transpose(perm = x_241_perm_0, x = var_6524_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 2, 512, 256]> var_6574_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_6574_cast_fp16")];
            bool var_6577 = const()[name = string("op_6577"), val = bool(true)];
            tensor<int32, [1]> var_6578_axes_0 = const()[name = string("op_6578_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_6578_cast_fp16 = reduce_max(axes = var_6578_axes_0, keep_dims = var_6577, x = var_6574_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_217_cast_fp16 = maximum(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_217_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_217_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6582_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6587_axes_0 = const()[name = string("op_6587_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6587_keep_dims_0 = const()[name = string("op_6587_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_6587_cast_fp16 = reduce_mean(axes = var_6587_axes_0, keep_dims = var_6587_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6588_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6588_cast_fp16")];
            fp32 var_6589_epsilon_0 = const()[name = string("op_6589_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_6589_cast_fp16 = inverse(epsilon = var_6589_epsilon_0, x = var_6588_cast_fp16)[name = string("op_6589_cast_fp16")];
            fp16 var_6590_to_fp16 = const()[name = string("op_6590_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_6591_cast_fp16 = mul(x = var_6589_cast_fp16, y = var_6590_to_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6593_cast_fp16 = add(x = var_6587_cast_fp16, y = var_6591_cast_fp16)[name = string("op_6593_cast_fp16")];
            fp32 var_6594_epsilon_0 = const()[name = string("op_6594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_6594_cast_fp16 = rsqrt(epsilon = var_6594_epsilon_0, x = var_6593_cast_fp16)[name = string("op_6594_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_53_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6594_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6596_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6596_cast_fp16")];
            tensor<int32, [4]> var_6601_begin_0 = const()[name = string("op_6601_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6601_end_0 = const()[name = string("op_6601_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_6601_end_mask_0 = const()[name = string("op_6601_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_6601_cast_fp16 = slice_by_index(begin = var_6601_begin_0, end = var_6601_end_0, end_mask = var_6601_end_mask_0, x = q_53_cast_fp16)[name = string("op_6601_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_6602_cast_fp16 = mul(x = var_6601_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_6602_cast_fp16")];
            tensor<int32, [4]> var_6607_begin_0 = const()[name = string("op_6607_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6607_end_0 = const()[name = string("op_6607_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_6607_end_mask_0 = const()[name = string("op_6607_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_6607_cast_fp16 = slice_by_index(begin = var_6607_begin_0, end = var_6607_end_0, end_mask = var_6607_end_mask_0, x = q_53_cast_fp16)[name = string("op_6607_cast_fp16")];
            int32 var_6609 = const()[name = string("op_6609"), val = int32(-1)];
            bool var_6610_interleave_0 = const()[name = string("op_6610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_6610_cast_fp16 = concat(axis = var_6609, interleave = var_6610_interleave_0, values = (var_6602_cast_fp16, var_6607_cast_fp16))[name = string("op_6610_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6611_cast_fp16 = mul(x = var_6610_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_55_cast_fp16 = add(x = var_6596_cast_fp16, y = var_6611_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6614_cast_fp16 = mul(x = k_79_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = k_79_cast_fp16)[name = string("op_6619_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_6620_cast_fp16 = mul(x = var_6619_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_6620_cast_fp16")];
            tensor<int32, [4]> var_6625_begin_0 = const()[name = string("op_6625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6625_end_0 = const()[name = string("op_6625_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_6625_end_mask_0 = const()[name = string("op_6625_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_6625_cast_fp16 = slice_by_index(begin = var_6625_begin_0, end = var_6625_end_0, end_mask = var_6625_end_mask_0, x = k_79_cast_fp16)[name = string("op_6625_cast_fp16")];
            int32 var_6627 = const()[name = string("op_6627"), val = int32(-1)];
            bool var_6628_interleave_0 = const()[name = string("op_6628_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_6628_cast_fp16 = concat(axis = var_6627, interleave = var_6628_interleave_0, values = (var_6620_cast_fp16, var_6625_cast_fp16))[name = string("op_6628_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6629_cast_fp16 = mul(x = var_6628_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6629_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_81_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6629_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_26_reps_0 = const()[name = string("tile_26_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = k_81_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [4, 1, 512, 256]> tile_26_cast_fp16 = tile(reps = tile_26_reps_0, x = transpose_52_cast_fp16)[name = string("tile_26_cast_fp16")];
            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_52_cast_fp16 = reshape(shape = concat_52, x = tile_26_cast_fp16)[name = string("reshape_52_cast_fp16")];
            tensor<int32, [5]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = reshape_52_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [4, 1, 512, 256]> reshape_53_cast_fp16 = reshape(shape = concat_53, x = transpose_53_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_27_reps_0 = const()[name = string("tile_27_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = v_53_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [4, 1, 512, 256]> tile_27_cast_fp16 = tile(reps = tile_27_reps_0, x = transpose_54_cast_fp16)[name = string("tile_27_cast_fp16")];
            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_54_cast_fp16 = reshape(shape = concat_54, x = tile_27_cast_fp16)[name = string("reshape_54_cast_fp16")];
            tensor<int32, [5]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = reshape_54_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [4, 1, 512, 256]> reshape_55_cast_fp16 = reshape(shape = concat_55, x = transpose_55_cast_fp16)[name = string("reshape_55_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6643_transpose_x_0 = const()[name = string("op_6643_transpose_x_0"), val = bool(false)];
            bool var_6643_transpose_y_0 = const()[name = string("op_6643_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 4, 512, 512]> var_6643_cast_fp16 = matmul(transpose_x = var_6643_transpose_x_0, transpose_y = var_6643_transpose_y_0, x = q_55_cast_fp16, y = transpose_109_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_6645_cast_fp16 = add(x = var_6643_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6646 = const()[name = string("op_6646"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_217_cast_fp16 = softmax(axis = var_6646, x = var_6645_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6649_transpose_x_0 = const()[name = string("op_6649_transpose_x_0"), val = bool(false)];
            bool var_6649_transpose_y_0 = const()[name = string("op_6649_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = reshape_55_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 4, 512, 256]> var_6649_cast_fp16 = matmul(transpose_x = var_6649_transpose_x_0, transpose_y = var_6649_transpose_y_0, x = w_217_cast_fp16, y = v_55_cast_fp16)[name = string("op_6649_cast_fp16")];
            tensor<int32, [4]> var_6652_perm_0 = const()[name = string("op_6652_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6656 = const()[name = string("op_6656"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_6652_cast_fp16 = transpose(perm = var_6652_perm_0, x = var_6649_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 512, 1024]> input_159_cast_fp16 = reshape(shape = var_6656, x = var_6652_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77977216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78501568))))[name = string("tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6661_cast_fp16 = abs(x = linear_121_cast_fp16)[name = string("op_6661_cast_fp16")];
            bool var_6664 = const()[name = string("op_6664"), val = bool(true)];
            tensor<int32, [1]> var_6665_axes_0 = const()[name = string("op_6665_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6665_cast_fp16 = reduce_max(axes = var_6665_axes_0, keep_dims = var_6664, x = var_6661_cast_fp16)[name = string("op_6665_cast_fp16")];
            fp16 var_6666_to_fp16 = const()[name = string("op_6666_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_219_cast_fp16 = maximum(x = var_6665_cast_fp16, y = var_6666_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_219_cast_fp16 = real_div(x = linear_121_cast_fp16, y = s_219_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6669_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6669_cast_fp16")];
            tensor<int32, [1]> var_6674_axes_0 = const()[name = string("op_6674_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6674_keep_dims_0 = const()[name = string("op_6674_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6674_cast_fp16 = reduce_mean(axes = var_6674_axes_0, keep_dims = var_6674_keep_dims_0, x = var_6669_cast_fp16)[name = string("op_6674_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6675_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6675_cast_fp16")];
            fp32 var_6676_epsilon_0 = const()[name = string("op_6676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6676_cast_fp16 = inverse(epsilon = var_6676_epsilon_0, x = var_6675_cast_fp16)[name = string("op_6676_cast_fp16")];
            fp16 var_6677_to_fp16 = const()[name = string("op_6677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6678_cast_fp16 = mul(x = var_6676_cast_fp16, y = var_6677_to_fp16)[name = string("op_6678_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6680_cast_fp16 = add(x = var_6674_cast_fp16, y = var_6678_cast_fp16)[name = string("op_6680_cast_fp16")];
            fp32 var_6681_epsilon_0 = const()[name = string("op_6681_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6681_cast_fp16 = rsqrt(epsilon = var_6681_epsilon_0, x = var_6680_cast_fp16)[name = string("op_6681_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_191_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6681_cast_fp16)[name = string("out_191_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78502656)))];
            tensor<fp16, [1, 512, 512]> var_6683_cast_fp16 = mul(x = out_191_cast_fp16, y = tm_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6683_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_245_cast_fp16 = add(x = x_235_cast_fp16, y = var_6683_cast_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6686_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_6686_cast_fp16")];
            bool var_6689 = const()[name = string("op_6689"), val = bool(true)];
            tensor<int32, [1]> var_6690_axes_0 = const()[name = string("op_6690_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6690_cast_fp16 = reduce_max(axes = var_6690_axes_0, keep_dims = var_6689, x = var_6686_cast_fp16)[name = string("op_6690_cast_fp16")];
            fp16 var_6691_to_fp16 = const()[name = string("op_6691_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_221_cast_fp16 = maximum(x = var_6690_cast_fp16, y = var_6691_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_221_cast_fp16 = real_div(x = x_245_cast_fp16, y = s_221_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6694_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6694_cast_fp16")];
            tensor<int32, [1]> var_6699_axes_0 = const()[name = string("op_6699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6699_keep_dims_0 = const()[name = string("op_6699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6699_cast_fp16 = reduce_mean(axes = var_6699_axes_0, keep_dims = var_6699_keep_dims_0, x = var_6694_cast_fp16)[name = string("op_6699_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6700_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6700_cast_fp16")];
            fp32 var_6701_epsilon_0 = const()[name = string("op_6701_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6701_cast_fp16 = inverse(epsilon = var_6701_epsilon_0, x = var_6700_cast_fp16)[name = string("op_6701_cast_fp16")];
            fp16 var_6702_to_fp16 = const()[name = string("op_6702_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6703_cast_fp16 = mul(x = var_6701_cast_fp16, y = var_6702_to_fp16)[name = string("op_6703_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6705_cast_fp16 = add(x = var_6699_cast_fp16, y = var_6703_cast_fp16)[name = string("op_6705_cast_fp16")];
            fp32 var_6706_epsilon_0 = const()[name = string("op_6706_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6706_cast_fp16 = rsqrt(epsilon = var_6706_epsilon_0, x = var_6705_cast_fp16)[name = string("op_6706_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_193_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6706_cast_fp16)[name = string("out_193_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78503744)))];
            tensor<fp16, [1, 512, 512]> input_161_cast_fp16 = mul(x = out_193_cast_fp16, y = tm_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78504832))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79553472))))[name = string("tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_122_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_122_cast_fp16")];
            string var_6713_mode_0 = const()[name = string("op_6713_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_6713_cast_fp16 = gelu(mode = var_6713_mode_0, x = linear_122_cast_fp16)[name = string("op_6713_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79557632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80606272))))[name = string("tm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_123_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_163_cast_fp16 = mul(x = var_6713_cast_fp16, y = linear_123_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80610432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81659072))))[name = string("tm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6721_cast_fp16 = abs(x = linear_124_cast_fp16)[name = string("op_6721_cast_fp16")];
            bool var_6724 = const()[name = string("op_6724"), val = bool(true)];
            tensor<int32, [1]> var_6725_axes_0 = const()[name = string("op_6725_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6725_cast_fp16 = reduce_max(axes = var_6725_axes_0, keep_dims = var_6724, x = var_6721_cast_fp16)[name = string("op_6725_cast_fp16")];
            fp16 var_6726_to_fp16 = const()[name = string("op_6726_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_223_cast_fp16 = maximum(x = var_6725_cast_fp16, y = var_6726_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_223_cast_fp16 = real_div(x = linear_124_cast_fp16, y = s_223_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6729_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [1]> var_6734_axes_0 = const()[name = string("op_6734_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6734_keep_dims_0 = const()[name = string("op_6734_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6734_cast_fp16 = reduce_mean(axes = var_6734_axes_0, keep_dims = var_6734_keep_dims_0, x = var_6729_cast_fp16)[name = string("op_6734_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6735_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6735_cast_fp16")];
            fp32 var_6736_epsilon_0 = const()[name = string("op_6736_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6736_cast_fp16 = inverse(epsilon = var_6736_epsilon_0, x = var_6735_cast_fp16)[name = string("op_6736_cast_fp16")];
            fp16 var_6737_to_fp16 = const()[name = string("op_6737_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6738_cast_fp16 = mul(x = var_6736_cast_fp16, y = var_6737_to_fp16)[name = string("op_6738_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6740_cast_fp16 = add(x = var_6734_cast_fp16, y = var_6738_cast_fp16)[name = string("op_6740_cast_fp16")];
            fp32 var_6741_epsilon_0 = const()[name = string("op_6741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6741_cast_fp16 = rsqrt(epsilon = var_6741_epsilon_0, x = var_6740_cast_fp16)[name = string("op_6741_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_195_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6741_cast_fp16)[name = string("out_195_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81660160)))];
            tensor<fp16, [1, 512, 512]> var_6743_cast_fp16 = mul(x = out_195_cast_fp16, y = tm_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6743_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_165_cast_fp16 = add(x = x_245_cast_fp16, y = var_6743_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81661248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81923456))))[name = string("tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_125_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_125_cast_fp16")];
            string x_249_mode_0 = const()[name = string("x_249_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_249_cast_fp16 = gelu(mode = x_249_mode_0, x = linear_125_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<int32, [4]> var_6763_begin_0 = const()[name = string("op_6763_begin_0"), val = tensor<int32, [4]>([0, 0, 13, 0])];
            tensor<int32, [4]> var_6763_end_0 = const()[name = string("op_6763_end_0"), val = tensor<int32, [4]>([1, 512, 14, 512])];
            tensor<bool, [4]> var_6763_end_mask_0 = const()[name = string("op_6763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6763_squeeze_mask_0 = const()[name = string("op_6763_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_6763_cast_fp16 = slice_by_index(begin = var_6763_begin_0, end = var_6763_end_0, end_mask = var_6763_end_mask_0, squeeze_mask = var_6763_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_167_cast_fp16 = mul(x = x_249_cast_fp16, y = var_6763_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81924544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82186752))))[name = string("tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6773_cast_fp16 = abs(x = linear_126_cast_fp16)[name = string("op_6773_cast_fp16")];
            bool var_6776 = const()[name = string("op_6776"), val = bool(true)];
            tensor<int32, [1]> var_6777_axes_0 = const()[name = string("op_6777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6777_cast_fp16 = reduce_max(axes = var_6777_axes_0, keep_dims = var_6776, x = var_6773_cast_fp16)[name = string("op_6777_cast_fp16")];
            fp16 var_6778_to_fp16 = const()[name = string("op_6778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_225_cast_fp16 = maximum(x = var_6777_cast_fp16, y = var_6778_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_225_cast_fp16 = real_div(x = linear_126_cast_fp16, y = s_225_cast_fp16)[name = string("y_225_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6781_cast_fp16 = mul(x = y_225_cast_fp16, y = y_225_cast_fp16)[name = string("op_6781_cast_fp16")];
            tensor<int32, [1]> var_6786_axes_0 = const()[name = string("op_6786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6786_keep_dims_0 = const()[name = string("op_6786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6786_cast_fp16 = reduce_mean(axes = var_6786_axes_0, keep_dims = var_6786_keep_dims_0, x = var_6781_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6787_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6787_cast_fp16")];
            fp32 var_6788_epsilon_0 = const()[name = string("op_6788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6788_cast_fp16 = inverse(epsilon = var_6788_epsilon_0, x = var_6787_cast_fp16)[name = string("op_6788_cast_fp16")];
            fp16 var_6789_to_fp16 = const()[name = string("op_6789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6790_cast_fp16 = mul(x = var_6788_cast_fp16, y = var_6789_to_fp16)[name = string("op_6790_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6792_cast_fp16 = add(x = var_6786_cast_fp16, y = var_6790_cast_fp16)[name = string("op_6792_cast_fp16")];
            fp32 var_6793_epsilon_0 = const()[name = string("op_6793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6793_cast_fp16 = rsqrt(epsilon = var_6793_epsilon_0, x = var_6792_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_197_cast_fp16 = mul(x = y_225_cast_fp16, y = var_6793_cast_fp16)[name = string("out_197_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82187840)))];
            tensor<fp16, [1, 512, 512]> var_6795_cast_fp16 = mul(x = out_197_cast_fp16, y = tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6795_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_27_cast_fp16 = add(x = input_165_cast_fp16, y = var_6795_cast_fp16)[name = string("h_27_cast_fp16")];
            tensor<fp16, [1]> tm_layers_13_layer_scalar_to_fp16 = const()[name = string("tm_layers_13_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9cp-1])];
            tensor<fp16, [1, 512, 512]> x_253_cast_fp16 = mul(x = h_27_cast_fp16, y = tm_layers_13_layer_scalar_to_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6799_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6799_cast_fp16")];
            bool var_6802 = const()[name = string("op_6802"), val = bool(true)];
            tensor<int32, [1]> var_6803_axes_0 = const()[name = string("op_6803_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6803_cast_fp16 = reduce_max(axes = var_6803_axes_0, keep_dims = var_6802, x = var_6799_cast_fp16)[name = string("op_6803_cast_fp16")];
            fp16 var_6804_to_fp16 = const()[name = string("op_6804_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_227_cast_fp16 = maximum(x = var_6803_cast_fp16, y = var_6804_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_227_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_227_cast_fp16)[name = string("y_227_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6807_cast_fp16 = mul(x = y_227_cast_fp16, y = y_227_cast_fp16)[name = string("op_6807_cast_fp16")];
            tensor<int32, [1]> var_6812_axes_0 = const()[name = string("op_6812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6812_keep_dims_0 = const()[name = string("op_6812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_6812_cast_fp16 = reduce_mean(axes = var_6812_axes_0, keep_dims = var_6812_keep_dims_0, x = var_6807_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6813_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6813_cast_fp16")];
            fp32 var_6814_epsilon_0 = const()[name = string("op_6814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_6814_cast_fp16 = inverse(epsilon = var_6814_epsilon_0, x = var_6813_cast_fp16)[name = string("op_6814_cast_fp16")];
            fp16 var_6815_to_fp16 = const()[name = string("op_6815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_6816_cast_fp16 = mul(x = var_6814_cast_fp16, y = var_6815_to_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_6818_cast_fp16 = add(x = var_6812_cast_fp16, y = var_6816_cast_fp16)[name = string("op_6818_cast_fp16")];
            fp32 var_6819_epsilon_0 = const()[name = string("op_6819_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_6819_cast_fp16 = rsqrt(epsilon = var_6819_epsilon_0, x = var_6818_cast_fp16)[name = string("op_6819_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_199_cast_fp16 = mul(x = y_227_cast_fp16, y = var_6819_cast_fp16)[name = string("out_199_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82188928)))];
            tensor<fp16, [1, 512, 512]> input_169_cast_fp16 = mul(x = out_199_cast_fp16, y = tm_layers_14_input_layernorm_weight_to_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82190016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82714368))))[name = string("tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6829 = const()[name = string("op_6829"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_6830_cast_fp16 = reshape(shape = var_6829, x = linear_127_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82716480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82978688))))[name = string("tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6841 = const()[name = string("op_6841"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_6842_cast_fp16 = reshape(shape = var_6841, x = linear_128_cast_fp16)[name = string("op_6842_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82979776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83241984))))[name = string("tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_6853 = const()[name = string("op_6853"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_6854_cast_fp16 = reshape(shape = var_6853, x = linear_129_cast_fp16)[name = string("op_6854_cast_fp16")];
            tensor<int32, [4]> x_259_perm_0 = const()[name = string("x_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_6830_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 4, 512, 256]> var_6858_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6858_cast_fp16")];
            bool var_6861 = const()[name = string("op_6861"), val = bool(true)];
            tensor<int32, [1]> var_6862_axes_0 = const()[name = string("op_6862_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_6862_cast_fp16 = reduce_max(axes = var_6862_axes_0, keep_dims = var_6861, x = var_6858_cast_fp16)[name = string("op_6862_cast_fp16")];
            fp16 var_6863_to_fp16 = const()[name = string("op_6863_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_229_cast_fp16 = maximum(x = var_6862_cast_fp16, y = var_6863_to_fp16)[name = string("s_229_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_229_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_229_cast_fp16)[name = string("y_229_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6866_cast_fp16 = mul(x = y_229_cast_fp16, y = y_229_cast_fp16)[name = string("op_6866_cast_fp16")];
            tensor<int32, [1]> var_6871_axes_0 = const()[name = string("op_6871_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6871_keep_dims_0 = const()[name = string("op_6871_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_6871_cast_fp16 = reduce_mean(axes = var_6871_axes_0, keep_dims = var_6871_keep_dims_0, x = var_6866_cast_fp16)[name = string("op_6871_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_6872_cast_fp16 = mul(x = s_229_cast_fp16, y = s_229_cast_fp16)[name = string("op_6872_cast_fp16")];
            fp32 var_6873_epsilon_0 = const()[name = string("op_6873_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_6873_cast_fp16 = inverse(epsilon = var_6873_epsilon_0, x = var_6872_cast_fp16)[name = string("op_6873_cast_fp16")];
            fp16 var_6874_to_fp16 = const()[name = string("op_6874_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_6875_cast_fp16 = mul(x = var_6873_cast_fp16, y = var_6874_to_fp16)[name = string("op_6875_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_6877_cast_fp16 = add(x = var_6871_cast_fp16, y = var_6875_cast_fp16)[name = string("op_6877_cast_fp16")];
            fp32 var_6878_epsilon_0 = const()[name = string("op_6878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_6878_cast_fp16 = rsqrt(epsilon = var_6878_epsilon_0, x = var_6877_cast_fp16)[name = string("op_6878_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_201_cast_fp16 = mul(x = y_229_cast_fp16, y = var_6878_cast_fp16)[name = string("out_201_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243072)))];
            tensor<fp16, [1, 4, 512, 256]> q_57_cast_fp16 = mul(x = out_201_cast_fp16, y = tm_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("q_57_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_6842_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, 512, 256]> var_6881_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6884 = const()[name = string("op_6884"), val = bool(true)];
            tensor<int32, [1]> var_6885_axes_0 = const()[name = string("op_6885_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_6885_cast_fp16 = reduce_max(axes = var_6885_axes_0, keep_dims = var_6884, x = var_6881_cast_fp16)[name = string("op_6885_cast_fp16")];
            fp16 var_6886_to_fp16 = const()[name = string("op_6886_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_231_cast_fp16 = maximum(x = var_6885_cast_fp16, y = var_6886_to_fp16)[name = string("s_231_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_231_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_231_cast_fp16)[name = string("y_231_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6889_cast_fp16 = mul(x = y_231_cast_fp16, y = y_231_cast_fp16)[name = string("op_6889_cast_fp16")];
            tensor<int32, [1]> var_6894_axes_0 = const()[name = string("op_6894_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6894_keep_dims_0 = const()[name = string("op_6894_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_6894_cast_fp16 = reduce_mean(axes = var_6894_axes_0, keep_dims = var_6894_keep_dims_0, x = var_6889_cast_fp16)[name = string("op_6894_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6895_cast_fp16 = mul(x = s_231_cast_fp16, y = s_231_cast_fp16)[name = string("op_6895_cast_fp16")];
            fp32 var_6896_epsilon_0 = const()[name = string("op_6896_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_6896_cast_fp16 = inverse(epsilon = var_6896_epsilon_0, x = var_6895_cast_fp16)[name = string("op_6896_cast_fp16")];
            fp16 var_6897_to_fp16 = const()[name = string("op_6897_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_6898_cast_fp16 = mul(x = var_6896_cast_fp16, y = var_6897_to_fp16)[name = string("op_6898_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6900_cast_fp16 = add(x = var_6894_cast_fp16, y = var_6898_cast_fp16)[name = string("op_6900_cast_fp16")];
            fp32 var_6901_epsilon_0 = const()[name = string("op_6901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_6901_cast_fp16 = rsqrt(epsilon = var_6901_epsilon_0, x = var_6900_cast_fp16)[name = string("op_6901_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_203_cast_fp16 = mul(x = y_231_cast_fp16, y = var_6901_cast_fp16)[name = string("out_203_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243648)))];
            tensor<fp16, [1, 2, 512, 256]> k_85_cast_fp16 = mul(x = out_203_cast_fp16, y = tm_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("k_85_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_259_cast_fp16 = transpose(perm = x_259_perm_0, x = var_6854_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 2, 512, 256]> var_6904_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6904_cast_fp16")];
            bool var_6907 = const()[name = string("op_6907"), val = bool(true)];
            tensor<int32, [1]> var_6908_axes_0 = const()[name = string("op_6908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_6908_cast_fp16 = reduce_max(axes = var_6908_axes_0, keep_dims = var_6907, x = var_6904_cast_fp16)[name = string("op_6908_cast_fp16")];
            fp16 var_6909_to_fp16 = const()[name = string("op_6909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_233_cast_fp16 = maximum(x = var_6908_cast_fp16, y = var_6909_to_fp16)[name = string("s_233_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_233_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_233_cast_fp16)[name = string("y_233_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6912_cast_fp16 = mul(x = y_233_cast_fp16, y = y_233_cast_fp16)[name = string("op_6912_cast_fp16")];
            tensor<int32, [1]> var_6917_axes_0 = const()[name = string("op_6917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6917_keep_dims_0 = const()[name = string("op_6917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_6917_cast_fp16 = reduce_mean(axes = var_6917_axes_0, keep_dims = var_6917_keep_dims_0, x = var_6912_cast_fp16)[name = string("op_6917_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6918_cast_fp16 = mul(x = s_233_cast_fp16, y = s_233_cast_fp16)[name = string("op_6918_cast_fp16")];
            fp32 var_6919_epsilon_0 = const()[name = string("op_6919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_6919_cast_fp16 = inverse(epsilon = var_6919_epsilon_0, x = var_6918_cast_fp16)[name = string("op_6919_cast_fp16")];
            fp16 var_6920_to_fp16 = const()[name = string("op_6920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_6921_cast_fp16 = mul(x = var_6919_cast_fp16, y = var_6920_to_fp16)[name = string("op_6921_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_6923_cast_fp16 = add(x = var_6917_cast_fp16, y = var_6921_cast_fp16)[name = string("op_6923_cast_fp16")];
            fp32 var_6924_epsilon_0 = const()[name = string("op_6924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_6924_cast_fp16 = rsqrt(epsilon = var_6924_epsilon_0, x = var_6923_cast_fp16)[name = string("op_6924_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_57_cast_fp16 = mul(x = y_233_cast_fp16, y = var_6924_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6926_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6926_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = q_57_cast_fp16)[name = string("op_6931_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_6932_cast_fp16 = mul(x = var_6931_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_6932_cast_fp16")];
            tensor<int32, [4]> var_6937_begin_0 = const()[name = string("op_6937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6937_end_0 = const()[name = string("op_6937_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_6937_end_mask_0 = const()[name = string("op_6937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_6937_cast_fp16 = slice_by_index(begin = var_6937_begin_0, end = var_6937_end_0, end_mask = var_6937_end_mask_0, x = q_57_cast_fp16)[name = string("op_6937_cast_fp16")];
            int32 var_6939 = const()[name = string("op_6939"), val = int32(-1)];
            bool var_6940_interleave_0 = const()[name = string("op_6940_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_6940_cast_fp16 = concat(axis = var_6939, interleave = var_6940_interleave_0, values = (var_6932_cast_fp16, var_6937_cast_fp16))[name = string("op_6940_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_6941_cast_fp16 = mul(x = var_6940_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6941_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_59_cast_fp16 = add(x = var_6926_cast_fp16, y = var_6941_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6944_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6944_cast_fp16")];
            tensor<int32, [4]> var_6949_begin_0 = const()[name = string("op_6949_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6949_end_0 = const()[name = string("op_6949_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_6949_end_mask_0 = const()[name = string("op_6949_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_6949_cast_fp16 = slice_by_index(begin = var_6949_begin_0, end = var_6949_end_0, end_mask = var_6949_end_mask_0, x = k_85_cast_fp16)[name = string("op_6949_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_6950_cast_fp16 = mul(x = var_6949_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_6950_cast_fp16")];
            tensor<int32, [4]> var_6955_begin_0 = const()[name = string("op_6955_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6955_end_0 = const()[name = string("op_6955_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_6955_end_mask_0 = const()[name = string("op_6955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_6955_cast_fp16 = slice_by_index(begin = var_6955_begin_0, end = var_6955_end_0, end_mask = var_6955_end_mask_0, x = k_85_cast_fp16)[name = string("op_6955_cast_fp16")];
            int32 var_6957 = const()[name = string("op_6957"), val = int32(-1)];
            bool var_6958_interleave_0 = const()[name = string("op_6958_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_6958_cast_fp16 = concat(axis = var_6957, interleave = var_6958_interleave_0, values = (var_6950_cast_fp16, var_6955_cast_fp16))[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_6959_cast_fp16 = mul(x = var_6958_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_87_cast_fp16 = add(x = var_6944_cast_fp16, y = var_6959_cast_fp16)[name = string("k_87_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_28_reps_0 = const()[name = string("tile_28_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = k_87_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 1, 512, 256]> tile_28_cast_fp16 = tile(reps = tile_28_reps_0, x = transpose_56_cast_fp16)[name = string("tile_28_cast_fp16")];
            tensor<int32, [5]> concat_56 = const()[name = string("concat_56"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_56_cast_fp16 = reshape(shape = concat_56, x = tile_28_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [5]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_57 = const()[name = string("concat_57"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 1, 512, 256]> reshape_57_cast_fp16 = reshape(shape = concat_57, x = transpose_57_cast_fp16)[name = string("reshape_57_cast_fp16")];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_29_reps_0 = const()[name = string("tile_29_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = v_57_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 1, 512, 256]> tile_29_cast_fp16 = tile(reps = tile_29_reps_0, x = transpose_58_cast_fp16)[name = string("tile_29_cast_fp16")];
            tensor<int32, [5]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_58_cast_fp16 = reshape(shape = concat_58, x = tile_29_cast_fp16)[name = string("reshape_58_cast_fp16")];
            tensor<int32, [5]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = reshape_58_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 1, 512, 256]> reshape_59_cast_fp16 = reshape(shape = concat_59, x = transpose_59_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6973_transpose_x_0 = const()[name = string("op_6973_transpose_x_0"), val = bool(false)];
            bool var_6973_transpose_y_0 = const()[name = string("op_6973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_110_cast_fp16 = transpose(perm = transpose_110_perm_0, x = reshape_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 4, 512, 512]> var_6973_cast_fp16 = matmul(transpose_x = var_6973_transpose_x_0, transpose_y = var_6973_transpose_y_0, x = q_59_cast_fp16, y = transpose_110_cast_fp16)[name = string("op_6973_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_6975_cast_fp16 = add(x = var_6973_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6976 = const()[name = string("op_6976"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_233_cast_fp16 = softmax(axis = var_6976, x = var_6975_cast_fp16)[name = string("w_233_cast_fp16")];
            bool var_6979_transpose_x_0 = const()[name = string("op_6979_transpose_x_0"), val = bool(false)];
            bool var_6979_transpose_y_0 = const()[name = string("op_6979_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 4, 512, 256]> var_6979_cast_fp16 = matmul(transpose_x = var_6979_transpose_x_0, transpose_y = var_6979_transpose_y_0, x = w_233_cast_fp16, y = v_59_cast_fp16)[name = string("op_6979_cast_fp16")];
            tensor<int32, [4]> var_6982_perm_0 = const()[name = string("op_6982_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6986 = const()[name = string("op_6986"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_6982_cast_fp16 = transpose(perm = var_6982_perm_0, x = var_6979_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 512, 1024]> input_171_cast_fp16 = reshape(shape = var_6986, x = var_6982_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83244224))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83768576))))[name = string("tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6991_cast_fp16 = abs(x = linear_130_cast_fp16)[name = string("op_6991_cast_fp16")];
            bool var_6994 = const()[name = string("op_6994"), val = bool(true)];
            tensor<int32, [1]> var_6995_axes_0 = const()[name = string("op_6995_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_6995_cast_fp16 = reduce_max(axes = var_6995_axes_0, keep_dims = var_6994, x = var_6991_cast_fp16)[name = string("op_6995_cast_fp16")];
            fp16 var_6996_to_fp16 = const()[name = string("op_6996_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_235_cast_fp16 = maximum(x = var_6995_cast_fp16, y = var_6996_to_fp16)[name = string("s_235_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_235_cast_fp16 = real_div(x = linear_130_cast_fp16, y = s_235_cast_fp16)[name = string("y_235_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_6999_cast_fp16 = mul(x = y_235_cast_fp16, y = y_235_cast_fp16)[name = string("op_6999_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_6999_cast_fp16)[name = string("op_7004_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7005_cast_fp16 = mul(x = s_235_cast_fp16, y = s_235_cast_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7010_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7010_cast_fp16")];
            fp32 var_7011_epsilon_0 = const()[name = string("op_7011_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7011_cast_fp16 = rsqrt(epsilon = var_7011_epsilon_0, x = var_7010_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_205_cast_fp16 = mul(x = y_235_cast_fp16, y = var_7011_cast_fp16)[name = string("out_205_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83769664)))];
            tensor<fp16, [1, 512, 512]> var_7013_cast_fp16 = mul(x = out_205_cast_fp16, y = tm_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_7013_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = var_7013_cast_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7016_cast_fp16 = abs(x = x_263_cast_fp16)[name = string("op_7016_cast_fp16")];
            bool var_7019 = const()[name = string("op_7019"), val = bool(true)];
            tensor<int32, [1]> var_7020_axes_0 = const()[name = string("op_7020_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7020_cast_fp16 = reduce_max(axes = var_7020_axes_0, keep_dims = var_7019, x = var_7016_cast_fp16)[name = string("op_7020_cast_fp16")];
            fp16 var_7021_to_fp16 = const()[name = string("op_7021_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_237_cast_fp16 = maximum(x = var_7020_cast_fp16, y = var_7021_to_fp16)[name = string("s_237_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_237_cast_fp16 = real_div(x = x_263_cast_fp16, y = s_237_cast_fp16)[name = string("y_237_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7024_cast_fp16 = mul(x = y_237_cast_fp16, y = y_237_cast_fp16)[name = string("op_7024_cast_fp16")];
            tensor<int32, [1]> var_7029_axes_0 = const()[name = string("op_7029_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7029_keep_dims_0 = const()[name = string("op_7029_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7029_cast_fp16 = reduce_mean(axes = var_7029_axes_0, keep_dims = var_7029_keep_dims_0, x = var_7024_cast_fp16)[name = string("op_7029_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7030_cast_fp16 = mul(x = s_237_cast_fp16, y = s_237_cast_fp16)[name = string("op_7030_cast_fp16")];
            fp32 var_7031_epsilon_0 = const()[name = string("op_7031_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7031_cast_fp16 = inverse(epsilon = var_7031_epsilon_0, x = var_7030_cast_fp16)[name = string("op_7031_cast_fp16")];
            fp16 var_7032_to_fp16 = const()[name = string("op_7032_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7033_cast_fp16 = mul(x = var_7031_cast_fp16, y = var_7032_to_fp16)[name = string("op_7033_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7035_cast_fp16 = add(x = var_7029_cast_fp16, y = var_7033_cast_fp16)[name = string("op_7035_cast_fp16")];
            fp32 var_7036_epsilon_0 = const()[name = string("op_7036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7036_cast_fp16 = rsqrt(epsilon = var_7036_epsilon_0, x = var_7035_cast_fp16)[name = string("op_7036_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_207_cast_fp16 = mul(x = y_237_cast_fp16, y = var_7036_cast_fp16)[name = string("out_207_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83770752)))];
            tensor<fp16, [1, 512, 512]> input_173_cast_fp16 = mul(x = out_207_cast_fp16, y = tm_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83771840))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84820480))))[name = string("tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_131_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_131_cast_fp16")];
            string var_7043_mode_0 = const()[name = string("op_7043_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_7043_cast_fp16 = gelu(mode = var_7043_mode_0, x = linear_131_cast_fp16)[name = string("op_7043_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84824640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85873280))))[name = string("tm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_132_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_175_cast_fp16 = mul(x = var_7043_cast_fp16, y = linear_132_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85877440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86926080))))[name = string("tm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7051_cast_fp16 = abs(x = linear_133_cast_fp16)[name = string("op_7051_cast_fp16")];
            bool var_7054 = const()[name = string("op_7054"), val = bool(true)];
            tensor<int32, [1]> var_7055_axes_0 = const()[name = string("op_7055_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7055_cast_fp16 = reduce_max(axes = var_7055_axes_0, keep_dims = var_7054, x = var_7051_cast_fp16)[name = string("op_7055_cast_fp16")];
            fp16 var_7056_to_fp16 = const()[name = string("op_7056_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_239_cast_fp16 = maximum(x = var_7055_cast_fp16, y = var_7056_to_fp16)[name = string("s_239_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_239_cast_fp16 = real_div(x = linear_133_cast_fp16, y = s_239_cast_fp16)[name = string("y_239_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7059_cast_fp16 = mul(x = y_239_cast_fp16, y = y_239_cast_fp16)[name = string("op_7059_cast_fp16")];
            tensor<int32, [1]> var_7064_axes_0 = const()[name = string("op_7064_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7064_keep_dims_0 = const()[name = string("op_7064_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7064_cast_fp16 = reduce_mean(axes = var_7064_axes_0, keep_dims = var_7064_keep_dims_0, x = var_7059_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7065_cast_fp16 = mul(x = s_239_cast_fp16, y = s_239_cast_fp16)[name = string("op_7065_cast_fp16")];
            fp32 var_7066_epsilon_0 = const()[name = string("op_7066_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7066_cast_fp16 = inverse(epsilon = var_7066_epsilon_0, x = var_7065_cast_fp16)[name = string("op_7066_cast_fp16")];
            fp16 var_7067_to_fp16 = const()[name = string("op_7067_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7068_cast_fp16 = mul(x = var_7066_cast_fp16, y = var_7067_to_fp16)[name = string("op_7068_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7070_cast_fp16 = add(x = var_7064_cast_fp16, y = var_7068_cast_fp16)[name = string("op_7070_cast_fp16")];
            fp32 var_7071_epsilon_0 = const()[name = string("op_7071_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7071_cast_fp16 = rsqrt(epsilon = var_7071_epsilon_0, x = var_7070_cast_fp16)[name = string("op_7071_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_209_cast_fp16 = mul(x = y_239_cast_fp16, y = var_7071_cast_fp16)[name = string("out_209_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86927168)))];
            tensor<fp16, [1, 512, 512]> var_7073_cast_fp16 = mul(x = out_209_cast_fp16, y = tm_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7073_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_177_cast_fp16 = add(x = x_263_cast_fp16, y = var_7073_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86928256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87190464))))[name = string("tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_134_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_134_cast_fp16")];
            string x_267_mode_0 = const()[name = string("x_267_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_267_cast_fp16 = gelu(mode = x_267_mode_0, x = linear_134_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<int32, [4]> var_7093_begin_0 = const()[name = string("op_7093_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
            tensor<int32, [4]> var_7093_end_0 = const()[name = string("op_7093_end_0"), val = tensor<int32, [4]>([1, 512, 15, 512])];
            tensor<bool, [4]> var_7093_end_mask_0 = const()[name = string("op_7093_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7093_squeeze_mask_0 = const()[name = string("op_7093_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_7093_cast_fp16 = slice_by_index(begin = var_7093_begin_0, end = var_7093_end_0, end_mask = var_7093_end_mask_0, squeeze_mask = var_7093_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7093_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_179_cast_fp16 = mul(x = x_267_cast_fp16, y = var_7093_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87191552))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87453760))))[name = string("tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7103_cast_fp16 = abs(x = linear_135_cast_fp16)[name = string("op_7103_cast_fp16")];
            bool var_7106 = const()[name = string("op_7106"), val = bool(true)];
            tensor<int32, [1]> var_7107_axes_0 = const()[name = string("op_7107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7107_cast_fp16 = reduce_max(axes = var_7107_axes_0, keep_dims = var_7106, x = var_7103_cast_fp16)[name = string("op_7107_cast_fp16")];
            fp16 var_7108_to_fp16 = const()[name = string("op_7108_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_241_cast_fp16 = maximum(x = var_7107_cast_fp16, y = var_7108_to_fp16)[name = string("s_241_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_241_cast_fp16 = real_div(x = linear_135_cast_fp16, y = s_241_cast_fp16)[name = string("y_241_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7111_cast_fp16 = mul(x = y_241_cast_fp16, y = y_241_cast_fp16)[name = string("op_7111_cast_fp16")];
            tensor<int32, [1]> var_7116_axes_0 = const()[name = string("op_7116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7116_keep_dims_0 = const()[name = string("op_7116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7116_cast_fp16 = reduce_mean(axes = var_7116_axes_0, keep_dims = var_7116_keep_dims_0, x = var_7111_cast_fp16)[name = string("op_7116_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7117_cast_fp16 = mul(x = s_241_cast_fp16, y = s_241_cast_fp16)[name = string("op_7117_cast_fp16")];
            fp32 var_7118_epsilon_0 = const()[name = string("op_7118_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7118_cast_fp16 = inverse(epsilon = var_7118_epsilon_0, x = var_7117_cast_fp16)[name = string("op_7118_cast_fp16")];
            fp16 var_7119_to_fp16 = const()[name = string("op_7119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7120_cast_fp16 = mul(x = var_7118_cast_fp16, y = var_7119_to_fp16)[name = string("op_7120_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7122_cast_fp16 = add(x = var_7116_cast_fp16, y = var_7120_cast_fp16)[name = string("op_7122_cast_fp16")];
            fp32 var_7123_epsilon_0 = const()[name = string("op_7123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7123_cast_fp16 = rsqrt(epsilon = var_7123_epsilon_0, x = var_7122_cast_fp16)[name = string("op_7123_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_211_cast_fp16 = mul(x = y_241_cast_fp16, y = var_7123_cast_fp16)[name = string("out_211_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87454848)))];
            tensor<fp16, [1, 512, 512]> var_7125_cast_fp16 = mul(x = out_211_cast_fp16, y = tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_29_cast_fp16 = add(x = input_177_cast_fp16, y = var_7125_cast_fp16)[name = string("h_29_cast_fp16")];
            tensor<fp16, [1]> tm_layers_14_layer_scalar_to_fp16 = const()[name = string("tm_layers_14_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1p-1])];
            tensor<fp16, [1, 512, 512]> x_271_cast_fp16 = mul(x = h_29_cast_fp16, y = tm_layers_14_layer_scalar_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7129_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_7129_cast_fp16")];
            bool var_7132 = const()[name = string("op_7132"), val = bool(true)];
            tensor<int32, [1]> var_7133_axes_0 = const()[name = string("op_7133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7133_cast_fp16 = reduce_max(axes = var_7133_axes_0, keep_dims = var_7132, x = var_7129_cast_fp16)[name = string("op_7133_cast_fp16")];
            fp16 var_7134_to_fp16 = const()[name = string("op_7134_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_243_cast_fp16 = maximum(x = var_7133_cast_fp16, y = var_7134_to_fp16)[name = string("s_243_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_243_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_243_cast_fp16)[name = string("y_243_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7137_cast_fp16 = mul(x = y_243_cast_fp16, y = y_243_cast_fp16)[name = string("op_7137_cast_fp16")];
            tensor<int32, [1]> var_7142_axes_0 = const()[name = string("op_7142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7142_keep_dims_0 = const()[name = string("op_7142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7142_cast_fp16 = reduce_mean(axes = var_7142_axes_0, keep_dims = var_7142_keep_dims_0, x = var_7137_cast_fp16)[name = string("op_7142_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7143_cast_fp16 = mul(x = s_243_cast_fp16, y = s_243_cast_fp16)[name = string("op_7143_cast_fp16")];
            fp32 var_7144_epsilon_0 = const()[name = string("op_7144_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7144_cast_fp16 = inverse(epsilon = var_7144_epsilon_0, x = var_7143_cast_fp16)[name = string("op_7144_cast_fp16")];
            fp16 var_7145_to_fp16 = const()[name = string("op_7145_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7146_cast_fp16 = mul(x = var_7144_cast_fp16, y = var_7145_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7148_cast_fp16 = add(x = var_7142_cast_fp16, y = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp32 var_7149_epsilon_0 = const()[name = string("op_7149_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7149_cast_fp16 = rsqrt(epsilon = var_7149_epsilon_0, x = var_7148_cast_fp16)[name = string("op_7149_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_213_cast_fp16 = mul(x = y_243_cast_fp16, y = var_7149_cast_fp16)[name = string("out_213_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87455936)))];
            tensor<fp16, [1, 512, 512]> input_181_cast_fp16 = mul(x = out_213_cast_fp16, y = tm_layers_15_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87457024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87981376))))[name = string("tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<int32, [4]> var_7159 = const()[name = string("op_7159"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_7160_cast_fp16 = reshape(shape = var_7159, x = linear_136_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<int32, [4]> x_273_perm_0 = const()[name = string("x_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87983488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88245696))))[name = string("tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<int32, [4]> var_7171 = const()[name = string("op_7171"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_7172_cast_fp16 = reshape(shape = var_7171, x = linear_137_cast_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [4]> x_275_perm_0 = const()[name = string("x_275_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88246784))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88508992))))[name = string("tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_7183 = const()[name = string("op_7183"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_7184_cast_fp16 = reshape(shape = var_7183, x = linear_138_cast_fp16)[name = string("op_7184_cast_fp16")];
            tensor<int32, [4]> x_277_perm_0 = const()[name = string("x_277_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_273_cast_fp16 = transpose(perm = x_273_perm_0, x = var_7160_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 4, 512, 256]> var_7188_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_7188_cast_fp16")];
            bool var_7191 = const()[name = string("op_7191"), val = bool(true)];
            tensor<int32, [1]> var_7192_axes_0 = const()[name = string("op_7192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_7192_cast_fp16 = reduce_max(axes = var_7192_axes_0, keep_dims = var_7191, x = var_7188_cast_fp16)[name = string("op_7192_cast_fp16")];
            fp16 var_7193_to_fp16 = const()[name = string("op_7193_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_245_cast_fp16 = maximum(x = var_7192_cast_fp16, y = var_7193_to_fp16)[name = string("s_245_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_245_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_245_cast_fp16)[name = string("y_245_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_7196_cast_fp16 = mul(x = y_245_cast_fp16, y = y_245_cast_fp16)[name = string("op_7196_cast_fp16")];
            tensor<int32, [1]> var_7201_axes_0 = const()[name = string("op_7201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7201_keep_dims_0 = const()[name = string("op_7201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_7201_cast_fp16 = reduce_mean(axes = var_7201_axes_0, keep_dims = var_7201_keep_dims_0, x = var_7196_cast_fp16)[name = string("op_7201_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_7202_cast_fp16 = mul(x = s_245_cast_fp16, y = s_245_cast_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_7203_cast_fp16 = inverse(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            fp16 var_7204_to_fp16 = const()[name = string("op_7204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_7205_cast_fp16 = mul(x = var_7203_cast_fp16, y = var_7204_to_fp16)[name = string("op_7205_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_7207_cast_fp16 = add(x = var_7201_cast_fp16, y = var_7205_cast_fp16)[name = string("op_7207_cast_fp16")];
            fp32 var_7208_epsilon_0 = const()[name = string("op_7208_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_7208_cast_fp16 = rsqrt(epsilon = var_7208_epsilon_0, x = var_7207_cast_fp16)[name = string("op_7208_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_215_cast_fp16 = mul(x = y_245_cast_fp16, y = var_7208_cast_fp16)[name = string("out_215_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_61_cast_fp16 = mul(x = out_215_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_61_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_275_cast_fp16 = transpose(perm = x_275_perm_0, x = var_7172_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 2, 512, 256]> var_7211_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_7211_cast_fp16")];
            bool var_7214 = const()[name = string("op_7214"), val = bool(true)];
            tensor<int32, [1]> var_7215_axes_0 = const()[name = string("op_7215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_7215_cast_fp16 = reduce_max(axes = var_7215_axes_0, keep_dims = var_7214, x = var_7211_cast_fp16)[name = string("op_7215_cast_fp16")];
            fp16 var_7216_to_fp16 = const()[name = string("op_7216_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_247_cast_fp16 = maximum(x = var_7215_cast_fp16, y = var_7216_to_fp16)[name = string("s_247_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_247_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_247_cast_fp16)[name = string("y_247_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_7219_cast_fp16 = mul(x = y_247_cast_fp16, y = y_247_cast_fp16)[name = string("op_7219_cast_fp16")];
            tensor<int32, [1]> var_7224_axes_0 = const()[name = string("op_7224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7224_keep_dims_0 = const()[name = string("op_7224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_7224_cast_fp16 = reduce_mean(axes = var_7224_axes_0, keep_dims = var_7224_keep_dims_0, x = var_7219_cast_fp16)[name = string("op_7224_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_7225_cast_fp16 = mul(x = s_247_cast_fp16, y = s_247_cast_fp16)[name = string("op_7225_cast_fp16")];
            fp32 var_7226_epsilon_0 = const()[name = string("op_7226_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_7226_cast_fp16 = inverse(epsilon = var_7226_epsilon_0, x = var_7225_cast_fp16)[name = string("op_7226_cast_fp16")];
            fp16 var_7227_to_fp16 = const()[name = string("op_7227_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_7228_cast_fp16 = mul(x = var_7226_cast_fp16, y = var_7227_to_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_7230_cast_fp16 = add(x = var_7224_cast_fp16, y = var_7228_cast_fp16)[name = string("op_7230_cast_fp16")];
            fp32 var_7231_epsilon_0 = const()[name = string("op_7231_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_7231_cast_fp16 = rsqrt(epsilon = var_7231_epsilon_0, x = var_7230_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_217_cast_fp16 = mul(x = y_247_cast_fp16, y = var_7231_cast_fp16)[name = string("out_217_cast_fp16")];
            tensor<fp16, [256]> tm_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510080)))];
            tensor<fp16, [1, 2, 512, 256]> k_91_cast_fp16 = mul(x = out_217_cast_fp16, y = tm_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_277_cast_fp16 = transpose(perm = x_277_perm_0, x = var_7184_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2, 512, 256]> var_7234_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_7234_cast_fp16")];
            bool var_7237 = const()[name = string("op_7237"), val = bool(true)];
            tensor<int32, [1]> var_7238_axes_0 = const()[name = string("op_7238_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_7238_cast_fp16 = reduce_max(axes = var_7238_axes_0, keep_dims = var_7237, x = var_7234_cast_fp16)[name = string("op_7238_cast_fp16")];
            fp16 var_7239_to_fp16 = const()[name = string("op_7239_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_249_cast_fp16 = maximum(x = var_7238_cast_fp16, y = var_7239_to_fp16)[name = string("s_249_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_249_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_249_cast_fp16)[name = string("y_249_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_7242_cast_fp16 = mul(x = y_249_cast_fp16, y = y_249_cast_fp16)[name = string("op_7242_cast_fp16")];
            tensor<int32, [1]> var_7247_axes_0 = const()[name = string("op_7247_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7247_keep_dims_0 = const()[name = string("op_7247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_7247_cast_fp16 = reduce_mean(axes = var_7247_axes_0, keep_dims = var_7247_keep_dims_0, x = var_7242_cast_fp16)[name = string("op_7247_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_7248_cast_fp16 = mul(x = s_249_cast_fp16, y = s_249_cast_fp16)[name = string("op_7248_cast_fp16")];
            fp32 var_7249_epsilon_0 = const()[name = string("op_7249_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_7249_cast_fp16 = inverse(epsilon = var_7249_epsilon_0, x = var_7248_cast_fp16)[name = string("op_7249_cast_fp16")];
            fp16 var_7250_to_fp16 = const()[name = string("op_7250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_7251_cast_fp16 = mul(x = var_7249_cast_fp16, y = var_7250_to_fp16)[name = string("op_7251_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_7253_cast_fp16 = add(x = var_7247_cast_fp16, y = var_7251_cast_fp16)[name = string("op_7253_cast_fp16")];
            fp32 var_7254_epsilon_0 = const()[name = string("op_7254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_7254_cast_fp16 = rsqrt(epsilon = var_7254_epsilon_0, x = var_7253_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_61_cast_fp16 = mul(x = y_249_cast_fp16, y = var_7254_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_7256_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7256_cast_fp16")];
            tensor<int32, [4]> var_7261_begin_0 = const()[name = string("op_7261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7261_end_0 = const()[name = string("op_7261_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_7261_end_mask_0 = const()[name = string("op_7261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_7261_cast_fp16 = slice_by_index(begin = var_7261_begin_0, end = var_7261_end_0, end_mask = var_7261_end_mask_0, x = q_61_cast_fp16)[name = string("op_7261_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_7262_cast_fp16 = mul(x = var_7261_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_7262_cast_fp16")];
            tensor<int32, [4]> var_7267_begin_0 = const()[name = string("op_7267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7267_end_0 = const()[name = string("op_7267_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_7267_end_mask_0 = const()[name = string("op_7267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_7267_cast_fp16 = slice_by_index(begin = var_7267_begin_0, end = var_7267_end_0, end_mask = var_7267_end_mask_0, x = q_61_cast_fp16)[name = string("op_7267_cast_fp16")];
            int32 var_7269 = const()[name = string("op_7269"), val = int32(-1)];
            bool var_7270_interleave_0 = const()[name = string("op_7270_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_7270_cast_fp16 = concat(axis = var_7269, interleave = var_7270_interleave_0, values = (var_7262_cast_fp16, var_7267_cast_fp16))[name = string("op_7270_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_7271_cast_fp16 = mul(x = var_7270_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7271_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_63_cast_fp16 = add(x = var_7256_cast_fp16, y = var_7271_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_7274_cast_fp16 = mul(x = k_91_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7274_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, x = k_91_cast_fp16)[name = string("op_7279_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_7280_cast_fp16 = mul(x = var_7279_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_7280_cast_fp16")];
            tensor<int32, [4]> var_7285_begin_0 = const()[name = string("op_7285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7285_end_0 = const()[name = string("op_7285_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_7285_end_mask_0 = const()[name = string("op_7285_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, x = k_91_cast_fp16)[name = string("op_7285_cast_fp16")];
            int32 var_7287 = const()[name = string("op_7287"), val = int32(-1)];
            bool var_7288_interleave_0 = const()[name = string("op_7288_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_7288_cast_fp16 = concat(axis = var_7287, interleave = var_7288_interleave_0, values = (var_7280_cast_fp16, var_7285_cast_fp16))[name = string("op_7288_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_7289_cast_fp16 = mul(x = var_7288_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7289_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_93_cast_fp16 = add(x = var_7274_cast_fp16, y = var_7289_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_30_reps_0 = const()[name = string("tile_30_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = k_93_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 1, 512, 256]> tile_30_cast_fp16 = tile(reps = tile_30_reps_0, x = transpose_60_cast_fp16)[name = string("tile_30_cast_fp16")];
            tensor<int32, [5]> concat_60 = const()[name = string("concat_60"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_60_cast_fp16 = reshape(shape = concat_60, x = tile_30_cast_fp16)[name = string("reshape_60_cast_fp16")];
            tensor<int32, [5]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = reshape_60_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 1, 512, 256]> reshape_61_cast_fp16 = reshape(shape = concat_61, x = transpose_61_cast_fp16)[name = string("reshape_61_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_31_reps_0 = const()[name = string("tile_31_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = v_61_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 1, 512, 256]> tile_31_cast_fp16 = tile(reps = tile_31_reps_0, x = transpose_62_cast_fp16)[name = string("tile_31_cast_fp16")];
            tensor<int32, [5]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_62_cast_fp16 = reshape(shape = concat_62, x = tile_31_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [5]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 1, 512, 256]> reshape_63_cast_fp16 = reshape(shape = concat_63, x = transpose_63_cast_fp16)[name = string("reshape_63_cast_fp16")];
            tensor<int32, [4]> v_63_perm_0 = const()[name = string("v_63_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7303_transpose_x_0 = const()[name = string("op_7303_transpose_x_0"), val = bool(false)];
            bool var_7303_transpose_y_0 = const()[name = string("op_7303_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = reshape_61_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 4, 512, 512]> var_7303_cast_fp16 = matmul(transpose_x = var_7303_transpose_x_0, transpose_y = var_7303_transpose_y_0, x = q_63_cast_fp16, y = transpose_111_cast_fp16)[name = string("op_7303_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_7305_cast_fp16 = add(x = var_7303_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7305_cast_fp16")];
            int32 var_7306 = const()[name = string("op_7306"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_249_cast_fp16 = softmax(axis = var_7306, x = var_7305_cast_fp16)[name = string("w_249_cast_fp16")];
            bool var_7309_transpose_x_0 = const()[name = string("op_7309_transpose_x_0"), val = bool(false)];
            bool var_7309_transpose_y_0 = const()[name = string("op_7309_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = reshape_63_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 4, 512, 256]> var_7309_cast_fp16 = matmul(transpose_x = var_7309_transpose_x_0, transpose_y = var_7309_transpose_y_0, x = w_249_cast_fp16, y = v_63_cast_fp16)[name = string("op_7309_cast_fp16")];
            tensor<int32, [4]> var_7312_perm_0 = const()[name = string("op_7312_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7316 = const()[name = string("op_7316"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_7312_cast_fp16 = transpose(perm = var_7312_perm_0, x = var_7309_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 512, 1024]> input_183_cast_fp16 = reshape(shape = var_7316, x = var_7312_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510656))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89035008))))[name = string("tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7321_cast_fp16 = abs(x = linear_139_cast_fp16)[name = string("op_7321_cast_fp16")];
            bool var_7324 = const()[name = string("op_7324"), val = bool(true)];
            tensor<int32, [1]> var_7325_axes_0 = const()[name = string("op_7325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7325_cast_fp16 = reduce_max(axes = var_7325_axes_0, keep_dims = var_7324, x = var_7321_cast_fp16)[name = string("op_7325_cast_fp16")];
            fp16 var_7326_to_fp16 = const()[name = string("op_7326_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_251_cast_fp16 = maximum(x = var_7325_cast_fp16, y = var_7326_to_fp16)[name = string("s_251_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_251_cast_fp16 = real_div(x = linear_139_cast_fp16, y = s_251_cast_fp16)[name = string("y_251_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7329_cast_fp16 = mul(x = y_251_cast_fp16, y = y_251_cast_fp16)[name = string("op_7329_cast_fp16")];
            tensor<int32, [1]> var_7334_axes_0 = const()[name = string("op_7334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7334_keep_dims_0 = const()[name = string("op_7334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7334_cast_fp16 = reduce_mean(axes = var_7334_axes_0, keep_dims = var_7334_keep_dims_0, x = var_7329_cast_fp16)[name = string("op_7334_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7335_cast_fp16 = mul(x = s_251_cast_fp16, y = s_251_cast_fp16)[name = string("op_7335_cast_fp16")];
            fp32 var_7336_epsilon_0 = const()[name = string("op_7336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7336_cast_fp16 = inverse(epsilon = var_7336_epsilon_0, x = var_7335_cast_fp16)[name = string("op_7336_cast_fp16")];
            fp16 var_7337_to_fp16 = const()[name = string("op_7337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7338_cast_fp16 = mul(x = var_7336_cast_fp16, y = var_7337_to_fp16)[name = string("op_7338_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7340_cast_fp16 = add(x = var_7334_cast_fp16, y = var_7338_cast_fp16)[name = string("op_7340_cast_fp16")];
            fp32 var_7341_epsilon_0 = const()[name = string("op_7341_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7341_cast_fp16 = rsqrt(epsilon = var_7341_epsilon_0, x = var_7340_cast_fp16)[name = string("op_7341_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_219_cast_fp16 = mul(x = y_251_cast_fp16, y = var_7341_cast_fp16)[name = string("out_219_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89036096)))];
            tensor<fp16, [1, 512, 512]> var_7343_cast_fp16 = mul(x = out_219_cast_fp16, y = tm_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_7343_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_281_cast_fp16 = add(x = x_271_cast_fp16, y = var_7343_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7346_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_7346_cast_fp16")];
            bool var_7349 = const()[name = string("op_7349"), val = bool(true)];
            tensor<int32, [1]> var_7350_axes_0 = const()[name = string("op_7350_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7350_cast_fp16 = reduce_max(axes = var_7350_axes_0, keep_dims = var_7349, x = var_7346_cast_fp16)[name = string("op_7350_cast_fp16")];
            fp16 var_7351_to_fp16 = const()[name = string("op_7351_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_253_cast_fp16 = maximum(x = var_7350_cast_fp16, y = var_7351_to_fp16)[name = string("s_253_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_253_cast_fp16 = real_div(x = x_281_cast_fp16, y = s_253_cast_fp16)[name = string("y_253_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7354_cast_fp16 = mul(x = y_253_cast_fp16, y = y_253_cast_fp16)[name = string("op_7354_cast_fp16")];
            tensor<int32, [1]> var_7359_axes_0 = const()[name = string("op_7359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7359_keep_dims_0 = const()[name = string("op_7359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7359_cast_fp16 = reduce_mean(axes = var_7359_axes_0, keep_dims = var_7359_keep_dims_0, x = var_7354_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7360_cast_fp16 = mul(x = s_253_cast_fp16, y = s_253_cast_fp16)[name = string("op_7360_cast_fp16")];
            fp32 var_7361_epsilon_0 = const()[name = string("op_7361_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7361_cast_fp16 = inverse(epsilon = var_7361_epsilon_0, x = var_7360_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7362_to_fp16 = const()[name = string("op_7362_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7363_cast_fp16 = mul(x = var_7361_cast_fp16, y = var_7362_to_fp16)[name = string("op_7363_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7365_cast_fp16 = add(x = var_7359_cast_fp16, y = var_7363_cast_fp16)[name = string("op_7365_cast_fp16")];
            fp32 var_7366_epsilon_0 = const()[name = string("op_7366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7366_cast_fp16 = rsqrt(epsilon = var_7366_epsilon_0, x = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_221_cast_fp16 = mul(x = y_253_cast_fp16, y = var_7366_cast_fp16)[name = string("out_221_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89037184)))];
            tensor<fp16, [1, 512, 512]> input_185_cast_fp16 = mul(x = out_221_cast_fp16, y = tm_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89038272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90086912))))[name = string("tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_140_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_140_cast_fp16")];
            string var_7373_mode_0 = const()[name = string("op_7373_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_7373_cast_fp16 = gelu(mode = var_7373_mode_0, x = linear_140_cast_fp16)[name = string("op_7373_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90091072))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91139712))))[name = string("tm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_141_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_187_cast_fp16 = mul(x = var_7373_cast_fp16, y = linear_141_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91143872))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92192512))))[name = string("tm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7381_cast_fp16 = abs(x = linear_142_cast_fp16)[name = string("op_7381_cast_fp16")];
            bool var_7384 = const()[name = string("op_7384"), val = bool(true)];
            tensor<int32, [1]> var_7385_axes_0 = const()[name = string("op_7385_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7385_cast_fp16 = reduce_max(axes = var_7385_axes_0, keep_dims = var_7384, x = var_7381_cast_fp16)[name = string("op_7385_cast_fp16")];
            fp16 var_7386_to_fp16 = const()[name = string("op_7386_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_255_cast_fp16 = maximum(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = string("s_255_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_255_cast_fp16 = real_div(x = linear_142_cast_fp16, y = s_255_cast_fp16)[name = string("y_255_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7389_cast_fp16 = mul(x = y_255_cast_fp16, y = y_255_cast_fp16)[name = string("op_7389_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7389_cast_fp16)[name = string("op_7394_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7395_cast_fp16 = mul(x = s_255_cast_fp16, y = s_255_cast_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7400_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7400_cast_fp16")];
            fp32 var_7401_epsilon_0 = const()[name = string("op_7401_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7401_cast_fp16 = rsqrt(epsilon = var_7401_epsilon_0, x = var_7400_cast_fp16)[name = string("op_7401_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_223_cast_fp16 = mul(x = y_255_cast_fp16, y = var_7401_cast_fp16)[name = string("out_223_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92193600)))];
            tensor<fp16, [1, 512, 512]> var_7403_cast_fp16 = mul(x = out_223_cast_fp16, y = tm_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7403_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_189_cast_fp16 = add(x = x_281_cast_fp16, y = var_7403_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92194688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92456896))))[name = string("tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_143_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_143_cast_fp16")];
            string x_285_mode_0 = const()[name = string("x_285_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_285_cast_fp16 = gelu(mode = x_285_mode_0, x = linear_143_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [4]> var_7423_begin_0 = const()[name = string("op_7423_begin_0"), val = tensor<int32, [4]>([0, 0, 15, 0])];
            tensor<int32, [4]> var_7423_end_0 = const()[name = string("op_7423_end_0"), val = tensor<int32, [4]>([1, 512, 16, 512])];
            tensor<bool, [4]> var_7423_end_mask_0 = const()[name = string("op_7423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7423_squeeze_mask_0 = const()[name = string("op_7423_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_7423_cast_fp16 = slice_by_index(begin = var_7423_begin_0, end = var_7423_end_0, end_mask = var_7423_end_mask_0, squeeze_mask = var_7423_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7423_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_191_cast_fp16 = mul(x = x_285_cast_fp16, y = var_7423_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92457984))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92720192))))[name = string("tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7433_cast_fp16 = abs(x = linear_144_cast_fp16)[name = string("op_7433_cast_fp16")];
            bool var_7436 = const()[name = string("op_7436"), val = bool(true)];
            tensor<int32, [1]> var_7437_axes_0 = const()[name = string("op_7437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7437_cast_fp16 = reduce_max(axes = var_7437_axes_0, keep_dims = var_7436, x = var_7433_cast_fp16)[name = string("op_7437_cast_fp16")];
            fp16 var_7438_to_fp16 = const()[name = string("op_7438_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_257_cast_fp16 = maximum(x = var_7437_cast_fp16, y = var_7438_to_fp16)[name = string("s_257_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_257_cast_fp16 = real_div(x = linear_144_cast_fp16, y = s_257_cast_fp16)[name = string("y_257_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7441_cast_fp16 = mul(x = y_257_cast_fp16, y = y_257_cast_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7446_axes_0 = const()[name = string("op_7446_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7446_keep_dims_0 = const()[name = string("op_7446_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7446_cast_fp16 = reduce_mean(axes = var_7446_axes_0, keep_dims = var_7446_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7447_cast_fp16 = mul(x = s_257_cast_fp16, y = s_257_cast_fp16)[name = string("op_7447_cast_fp16")];
            fp32 var_7448_epsilon_0 = const()[name = string("op_7448_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7448_cast_fp16 = inverse(epsilon = var_7448_epsilon_0, x = var_7447_cast_fp16)[name = string("op_7448_cast_fp16")];
            fp16 var_7449_to_fp16 = const()[name = string("op_7449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7450_cast_fp16 = mul(x = var_7448_cast_fp16, y = var_7449_to_fp16)[name = string("op_7450_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7452_cast_fp16 = add(x = var_7446_cast_fp16, y = var_7450_cast_fp16)[name = string("op_7452_cast_fp16")];
            fp32 var_7453_epsilon_0 = const()[name = string("op_7453_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7453_cast_fp16 = rsqrt(epsilon = var_7453_epsilon_0, x = var_7452_cast_fp16)[name = string("op_7453_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_225_cast_fp16 = mul(x = y_257_cast_fp16, y = var_7453_cast_fp16)[name = string("out_225_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92721280)))];
            tensor<fp16, [1, 512, 512]> var_7455_cast_fp16 = mul(x = out_225_cast_fp16, y = tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7455_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_31_cast_fp16 = add(x = input_189_cast_fp16, y = var_7455_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<fp16, [1]> tm_layers_15_layer_scalar_to_fp16 = const()[name = string("tm_layers_15_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7ap-1])];
            tensor<fp16, [1, 512, 512]> x_289_cast_fp16 = mul(x = h_31_cast_fp16, y = tm_layers_15_layer_scalar_to_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7459_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_7459_cast_fp16")];
            bool var_7462 = const()[name = string("op_7462"), val = bool(true)];
            tensor<int32, [1]> var_7463_axes_0 = const()[name = string("op_7463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7463_cast_fp16 = reduce_max(axes = var_7463_axes_0, keep_dims = var_7462, x = var_7459_cast_fp16)[name = string("op_7463_cast_fp16")];
            fp16 var_7464_to_fp16 = const()[name = string("op_7464_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_259_cast_fp16 = maximum(x = var_7463_cast_fp16, y = var_7464_to_fp16)[name = string("s_259_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_259_cast_fp16 = real_div(x = x_289_cast_fp16, y = s_259_cast_fp16)[name = string("y_259_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7467_cast_fp16 = mul(x = y_259_cast_fp16, y = y_259_cast_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7472_axes_0 = const()[name = string("op_7472_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7472_keep_dims_0 = const()[name = string("op_7472_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7472_cast_fp16 = reduce_mean(axes = var_7472_axes_0, keep_dims = var_7472_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7473_cast_fp16 = mul(x = s_259_cast_fp16, y = s_259_cast_fp16)[name = string("op_7473_cast_fp16")];
            fp32 var_7474_epsilon_0 = const()[name = string("op_7474_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7474_cast_fp16 = inverse(epsilon = var_7474_epsilon_0, x = var_7473_cast_fp16)[name = string("op_7474_cast_fp16")];
            fp16 var_7475_to_fp16 = const()[name = string("op_7475_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7476_cast_fp16 = mul(x = var_7474_cast_fp16, y = var_7475_to_fp16)[name = string("op_7476_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7478_cast_fp16 = add(x = var_7472_cast_fp16, y = var_7476_cast_fp16)[name = string("op_7478_cast_fp16")];
            fp32 var_7479_epsilon_0 = const()[name = string("op_7479_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7479_cast_fp16 = rsqrt(epsilon = var_7479_epsilon_0, x = var_7478_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_227_cast_fp16 = mul(x = y_259_cast_fp16, y = var_7479_cast_fp16)[name = string("out_227_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92722368)))];
            tensor<fp16, [1, 512, 512]> input_193_cast_fp16 = mul(x = out_227_cast_fp16, y = tm_layers_16_input_layernorm_weight_to_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92723456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93247808))))[name = string("tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<int32, [4]> var_7489 = const()[name = string("op_7489"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_7490_cast_fp16 = reshape(shape = var_7489, x = linear_145_cast_fp16)[name = string("op_7490_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93249920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93512128))))[name = string("tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<int32, [4]> var_7501 = const()[name = string("op_7501"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_7502_cast_fp16 = reshape(shape = var_7501, x = linear_146_cast_fp16)[name = string("op_7502_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93513216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93775424))))[name = string("tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7513 = const()[name = string("op_7513"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_7514_cast_fp16 = reshape(shape = var_7513, x = linear_147_cast_fp16)[name = string("op_7514_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_7490_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 4, 512, 256]> var_7518_cast_fp16 = abs(x = x_291_cast_fp16)[name = string("op_7518_cast_fp16")];
            bool var_7521 = const()[name = string("op_7521"), val = bool(true)];
            tensor<int32, [1]> var_7522_axes_0 = const()[name = string("op_7522_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_7522_cast_fp16 = reduce_max(axes = var_7522_axes_0, keep_dims = var_7521, x = var_7518_cast_fp16)[name = string("op_7522_cast_fp16")];
            fp16 var_7523_to_fp16 = const()[name = string("op_7523_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_261_cast_fp16 = maximum(x = var_7522_cast_fp16, y = var_7523_to_fp16)[name = string("s_261_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_261_cast_fp16 = real_div(x = x_291_cast_fp16, y = s_261_cast_fp16)[name = string("y_261_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_7526_cast_fp16 = mul(x = y_261_cast_fp16, y = y_261_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<int32, [1]> var_7531_axes_0 = const()[name = string("op_7531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7531_keep_dims_0 = const()[name = string("op_7531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_7531_cast_fp16 = reduce_mean(axes = var_7531_axes_0, keep_dims = var_7531_keep_dims_0, x = var_7526_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_7532_cast_fp16 = mul(x = s_261_cast_fp16, y = s_261_cast_fp16)[name = string("op_7532_cast_fp16")];
            fp32 var_7533_epsilon_0 = const()[name = string("op_7533_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_7533_cast_fp16 = inverse(epsilon = var_7533_epsilon_0, x = var_7532_cast_fp16)[name = string("op_7533_cast_fp16")];
            fp16 var_7534_to_fp16 = const()[name = string("op_7534_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_7535_cast_fp16 = mul(x = var_7533_cast_fp16, y = var_7534_to_fp16)[name = string("op_7535_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_7537_cast_fp16 = add(x = var_7531_cast_fp16, y = var_7535_cast_fp16)[name = string("op_7537_cast_fp16")];
            fp32 var_7538_epsilon_0 = const()[name = string("op_7538_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_7538_cast_fp16 = rsqrt(epsilon = var_7538_epsilon_0, x = var_7537_cast_fp16)[name = string("op_7538_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_229_cast_fp16 = mul(x = y_261_cast_fp16, y = var_7538_cast_fp16)[name = string("out_229_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_65_cast_fp16 = mul(x = out_229_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_65_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_7502_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 512, 256]> var_7541_cast_fp16 = abs(x = x_293_cast_fp16)[name = string("op_7541_cast_fp16")];
            bool var_7544 = const()[name = string("op_7544"), val = bool(true)];
            tensor<int32, [1]> var_7545_axes_0 = const()[name = string("op_7545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_7545_cast_fp16 = reduce_max(axes = var_7545_axes_0, keep_dims = var_7544, x = var_7541_cast_fp16)[name = string("op_7545_cast_fp16")];
            fp16 var_7546_to_fp16 = const()[name = string("op_7546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_263_cast_fp16 = maximum(x = var_7545_cast_fp16, y = var_7546_to_fp16)[name = string("s_263_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_263_cast_fp16 = real_div(x = x_293_cast_fp16, y = s_263_cast_fp16)[name = string("y_263_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_7549_cast_fp16 = mul(x = y_263_cast_fp16, y = y_263_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<int32, [1]> var_7554_axes_0 = const()[name = string("op_7554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7554_keep_dims_0 = const()[name = string("op_7554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_7554_cast_fp16 = reduce_mean(axes = var_7554_axes_0, keep_dims = var_7554_keep_dims_0, x = var_7549_cast_fp16)[name = string("op_7554_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_7555_cast_fp16 = mul(x = s_263_cast_fp16, y = s_263_cast_fp16)[name = string("op_7555_cast_fp16")];
            fp32 var_7556_epsilon_0 = const()[name = string("op_7556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_7556_cast_fp16 = inverse(epsilon = var_7556_epsilon_0, x = var_7555_cast_fp16)[name = string("op_7556_cast_fp16")];
            fp16 var_7557_to_fp16 = const()[name = string("op_7557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_7558_cast_fp16 = mul(x = var_7556_cast_fp16, y = var_7557_to_fp16)[name = string("op_7558_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_7560_cast_fp16 = add(x = var_7554_cast_fp16, y = var_7558_cast_fp16)[name = string("op_7560_cast_fp16")];
            fp32 var_7561_epsilon_0 = const()[name = string("op_7561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_7561_cast_fp16 = rsqrt(epsilon = var_7561_epsilon_0, x = var_7560_cast_fp16)[name = string("op_7561_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_231_cast_fp16 = mul(x = y_263_cast_fp16, y = var_7561_cast_fp16)[name = string("out_231_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_97_cast_fp16 = mul(x = out_231_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_97_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_7514_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 512, 256]> var_7564_cast_fp16 = abs(x = x_295_cast_fp16)[name = string("op_7564_cast_fp16")];
            bool var_7567 = const()[name = string("op_7567"), val = bool(true)];
            tensor<int32, [1]> var_7568_axes_0 = const()[name = string("op_7568_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_7568_cast_fp16 = reduce_max(axes = var_7568_axes_0, keep_dims = var_7567, x = var_7564_cast_fp16)[name = string("op_7568_cast_fp16")];
            fp16 var_7569_to_fp16 = const()[name = string("op_7569_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_265_cast_fp16 = maximum(x = var_7568_cast_fp16, y = var_7569_to_fp16)[name = string("s_265_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_265_cast_fp16 = real_div(x = x_295_cast_fp16, y = s_265_cast_fp16)[name = string("y_265_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_7572_cast_fp16 = mul(x = y_265_cast_fp16, y = y_265_cast_fp16)[name = string("op_7572_cast_fp16")];
            tensor<int32, [1]> var_7577_axes_0 = const()[name = string("op_7577_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7577_keep_dims_0 = const()[name = string("op_7577_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_7577_cast_fp16 = reduce_mean(axes = var_7577_axes_0, keep_dims = var_7577_keep_dims_0, x = var_7572_cast_fp16)[name = string("op_7577_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_7578_cast_fp16 = mul(x = s_265_cast_fp16, y = s_265_cast_fp16)[name = string("op_7578_cast_fp16")];
            fp32 var_7579_epsilon_0 = const()[name = string("op_7579_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_7579_cast_fp16 = inverse(epsilon = var_7579_epsilon_0, x = var_7578_cast_fp16)[name = string("op_7579_cast_fp16")];
            fp16 var_7580_to_fp16 = const()[name = string("op_7580_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_7581_cast_fp16 = mul(x = var_7579_cast_fp16, y = var_7580_to_fp16)[name = string("op_7581_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_7583_cast_fp16 = add(x = var_7577_cast_fp16, y = var_7581_cast_fp16)[name = string("op_7583_cast_fp16")];
            fp32 var_7584_epsilon_0 = const()[name = string("op_7584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_7584_cast_fp16 = rsqrt(epsilon = var_7584_epsilon_0, x = var_7583_cast_fp16)[name = string("op_7584_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_65_cast_fp16 = mul(x = y_265_cast_fp16, y = var_7584_cast_fp16)[name = string("v_65_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_7586_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7586_cast_fp16")];
            tensor<int32, [4]> var_7591_begin_0 = const()[name = string("op_7591_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7591_end_0 = const()[name = string("op_7591_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_7591_end_mask_0 = const()[name = string("op_7591_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_7591_cast_fp16 = slice_by_index(begin = var_7591_begin_0, end = var_7591_end_0, end_mask = var_7591_end_mask_0, x = q_65_cast_fp16)[name = string("op_7591_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_7592_cast_fp16 = mul(x = var_7591_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_7592_cast_fp16")];
            tensor<int32, [4]> var_7597_begin_0 = const()[name = string("op_7597_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7597_end_0 = const()[name = string("op_7597_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_7597_end_mask_0 = const()[name = string("op_7597_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, x = q_65_cast_fp16)[name = string("op_7597_cast_fp16")];
            int32 var_7599 = const()[name = string("op_7599"), val = int32(-1)];
            bool var_7600_interleave_0 = const()[name = string("op_7600_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_7600_cast_fp16 = concat(axis = var_7599, interleave = var_7600_interleave_0, values = (var_7592_cast_fp16, var_7597_cast_fp16))[name = string("op_7600_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_7601_cast_fp16 = mul(x = var_7600_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7601_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_67_cast_fp16 = add(x = var_7586_cast_fp16, y = var_7601_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_7604_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7604_cast_fp16")];
            tensor<int32, [4]> var_7609_begin_0 = const()[name = string("op_7609_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7609_end_0 = const()[name = string("op_7609_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_7609_end_mask_0 = const()[name = string("op_7609_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_7609_cast_fp16 = slice_by_index(begin = var_7609_begin_0, end = var_7609_end_0, end_mask = var_7609_end_mask_0, x = k_97_cast_fp16)[name = string("op_7609_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_7610_cast_fp16 = mul(x = var_7609_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_7610_cast_fp16")];
            tensor<int32, [4]> var_7615_begin_0 = const()[name = string("op_7615_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7615_end_0 = const()[name = string("op_7615_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_7615_end_mask_0 = const()[name = string("op_7615_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_7615_cast_fp16 = slice_by_index(begin = var_7615_begin_0, end = var_7615_end_0, end_mask = var_7615_end_mask_0, x = k_97_cast_fp16)[name = string("op_7615_cast_fp16")];
            int32 var_7617 = const()[name = string("op_7617"), val = int32(-1)];
            bool var_7618_interleave_0 = const()[name = string("op_7618_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_7618_cast_fp16 = concat(axis = var_7617, interleave = var_7618_interleave_0, values = (var_7610_cast_fp16, var_7615_cast_fp16))[name = string("op_7618_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_7619_cast_fp16 = mul(x = var_7618_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7619_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_99_cast_fp16 = add(x = var_7604_cast_fp16, y = var_7619_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_32_reps_0 = const()[name = string("tile_32_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = k_99_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 1, 512, 256]> tile_32_cast_fp16 = tile(reps = tile_32_reps_0, x = transpose_64_cast_fp16)[name = string("tile_32_cast_fp16")];
            tensor<int32, [5]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_64_cast_fp16 = reshape(shape = concat_64, x = tile_32_cast_fp16)[name = string("reshape_64_cast_fp16")];
            tensor<int32, [5]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = reshape_64_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 1, 512, 256]> reshape_65_cast_fp16 = reshape(shape = concat_65, x = transpose_65_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_33_reps_0 = const()[name = string("tile_33_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = v_65_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 1, 512, 256]> tile_33_cast_fp16 = tile(reps = tile_33_reps_0, x = transpose_66_cast_fp16)[name = string("tile_33_cast_fp16")];
            tensor<int32, [5]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_66_cast_fp16 = reshape(shape = concat_66, x = tile_33_cast_fp16)[name = string("reshape_66_cast_fp16")];
            tensor<int32, [5]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = reshape_66_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 1, 512, 256]> reshape_67_cast_fp16 = reshape(shape = concat_67, x = transpose_67_cast_fp16)[name = string("reshape_67_cast_fp16")];
            tensor<int32, [4]> v_67_perm_0 = const()[name = string("v_67_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7633_transpose_x_0 = const()[name = string("op_7633_transpose_x_0"), val = bool(false)];
            bool var_7633_transpose_y_0 = const()[name = string("op_7633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_112_cast_fp16 = transpose(perm = transpose_112_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 4, 512, 512]> var_7633_cast_fp16 = matmul(transpose_x = var_7633_transpose_x_0, transpose_y = var_7633_transpose_y_0, x = q_67_cast_fp16, y = transpose_112_cast_fp16)[name = string("op_7633_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_7635_cast_fp16 = add(x = var_7633_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7635_cast_fp16")];
            int32 var_7636 = const()[name = string("op_7636"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_265_cast_fp16 = softmax(axis = var_7636, x = var_7635_cast_fp16)[name = string("w_265_cast_fp16")];
            bool var_7639_transpose_x_0 = const()[name = string("op_7639_transpose_x_0"), val = bool(false)];
            bool var_7639_transpose_y_0 = const()[name = string("op_7639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = reshape_67_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 4, 512, 256]> var_7639_cast_fp16 = matmul(transpose_x = var_7639_transpose_x_0, transpose_y = var_7639_transpose_y_0, x = w_265_cast_fp16, y = v_67_cast_fp16)[name = string("op_7639_cast_fp16")];
            tensor<int32, [4]> var_7642_perm_0 = const()[name = string("op_7642_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7646 = const()[name = string("op_7646"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_7642_cast_fp16 = transpose(perm = var_7642_perm_0, x = var_7639_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 512, 1024]> input_195_cast_fp16 = reshape(shape = var_7646, x = var_7642_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93776512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94300864))))[name = string("tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7651_cast_fp16 = abs(x = linear_148_cast_fp16)[name = string("op_7651_cast_fp16")];
            bool var_7654 = const()[name = string("op_7654"), val = bool(true)];
            tensor<int32, [1]> var_7655_axes_0 = const()[name = string("op_7655_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7655_cast_fp16 = reduce_max(axes = var_7655_axes_0, keep_dims = var_7654, x = var_7651_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp16 var_7656_to_fp16 = const()[name = string("op_7656_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_267_cast_fp16 = maximum(x = var_7655_cast_fp16, y = var_7656_to_fp16)[name = string("s_267_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_267_cast_fp16 = real_div(x = linear_148_cast_fp16, y = s_267_cast_fp16)[name = string("y_267_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7659_cast_fp16 = mul(x = y_267_cast_fp16, y = y_267_cast_fp16)[name = string("op_7659_cast_fp16")];
            tensor<int32, [1]> var_7664_axes_0 = const()[name = string("op_7664_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7664_keep_dims_0 = const()[name = string("op_7664_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7664_cast_fp16 = reduce_mean(axes = var_7664_axes_0, keep_dims = var_7664_keep_dims_0, x = var_7659_cast_fp16)[name = string("op_7664_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7665_cast_fp16 = mul(x = s_267_cast_fp16, y = s_267_cast_fp16)[name = string("op_7665_cast_fp16")];
            fp32 var_7666_epsilon_0 = const()[name = string("op_7666_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7666_cast_fp16 = inverse(epsilon = var_7666_epsilon_0, x = var_7665_cast_fp16)[name = string("op_7666_cast_fp16")];
            fp16 var_7667_to_fp16 = const()[name = string("op_7667_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7668_cast_fp16 = mul(x = var_7666_cast_fp16, y = var_7667_to_fp16)[name = string("op_7668_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7670_cast_fp16 = add(x = var_7664_cast_fp16, y = var_7668_cast_fp16)[name = string("op_7670_cast_fp16")];
            fp32 var_7671_epsilon_0 = const()[name = string("op_7671_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7671_cast_fp16 = rsqrt(epsilon = var_7671_epsilon_0, x = var_7670_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_233_cast_fp16 = mul(x = y_267_cast_fp16, y = var_7671_cast_fp16)[name = string("out_233_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94301952)))];
            tensor<fp16, [1, 512, 512]> var_7673_cast_fp16 = mul(x = out_233_cast_fp16, y = tm_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("op_7673_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = var_7673_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7676_cast_fp16 = abs(x = x_299_cast_fp16)[name = string("op_7676_cast_fp16")];
            bool var_7679 = const()[name = string("op_7679"), val = bool(true)];
            tensor<int32, [1]> var_7680_axes_0 = const()[name = string("op_7680_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7680_cast_fp16 = reduce_max(axes = var_7680_axes_0, keep_dims = var_7679, x = var_7676_cast_fp16)[name = string("op_7680_cast_fp16")];
            fp16 var_7681_to_fp16 = const()[name = string("op_7681_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_269_cast_fp16 = maximum(x = var_7680_cast_fp16, y = var_7681_to_fp16)[name = string("s_269_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_269_cast_fp16 = real_div(x = x_299_cast_fp16, y = s_269_cast_fp16)[name = string("y_269_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7684_cast_fp16 = mul(x = y_269_cast_fp16, y = y_269_cast_fp16)[name = string("op_7684_cast_fp16")];
            tensor<int32, [1]> var_7689_axes_0 = const()[name = string("op_7689_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7689_keep_dims_0 = const()[name = string("op_7689_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7689_cast_fp16 = reduce_mean(axes = var_7689_axes_0, keep_dims = var_7689_keep_dims_0, x = var_7684_cast_fp16)[name = string("op_7689_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7690_cast_fp16 = mul(x = s_269_cast_fp16, y = s_269_cast_fp16)[name = string("op_7690_cast_fp16")];
            fp32 var_7691_epsilon_0 = const()[name = string("op_7691_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7691_cast_fp16 = inverse(epsilon = var_7691_epsilon_0, x = var_7690_cast_fp16)[name = string("op_7691_cast_fp16")];
            fp16 var_7692_to_fp16 = const()[name = string("op_7692_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7693_cast_fp16 = mul(x = var_7691_cast_fp16, y = var_7692_to_fp16)[name = string("op_7693_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7695_cast_fp16 = add(x = var_7689_cast_fp16, y = var_7693_cast_fp16)[name = string("op_7695_cast_fp16")];
            fp32 var_7696_epsilon_0 = const()[name = string("op_7696_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7696_cast_fp16 = rsqrt(epsilon = var_7696_epsilon_0, x = var_7695_cast_fp16)[name = string("op_7696_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_235_cast_fp16 = mul(x = y_269_cast_fp16, y = var_7696_cast_fp16)[name = string("out_235_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94303040)))];
            tensor<fp16, [1, 512, 512]> input_197_cast_fp16 = mul(x = out_235_cast_fp16, y = tm_layers_16_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94304128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95352768))))[name = string("tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_149_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_149_cast_fp16")];
            string var_7703_mode_0 = const()[name = string("op_7703_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_7703_cast_fp16 = gelu(mode = var_7703_mode_0, x = linear_149_cast_fp16)[name = string("op_7703_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95356928))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96405568))))[name = string("tm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_150_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_199_cast_fp16 = mul(x = var_7703_cast_fp16, y = linear_150_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96409728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97458368))))[name = string("tm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7711_cast_fp16 = abs(x = linear_151_cast_fp16)[name = string("op_7711_cast_fp16")];
            bool var_7714 = const()[name = string("op_7714"), val = bool(true)];
            tensor<int32, [1]> var_7715_axes_0 = const()[name = string("op_7715_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7715_cast_fp16 = reduce_max(axes = var_7715_axes_0, keep_dims = var_7714, x = var_7711_cast_fp16)[name = string("op_7715_cast_fp16")];
            fp16 var_7716_to_fp16 = const()[name = string("op_7716_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_271_cast_fp16 = maximum(x = var_7715_cast_fp16, y = var_7716_to_fp16)[name = string("s_271_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_271_cast_fp16 = real_div(x = linear_151_cast_fp16, y = s_271_cast_fp16)[name = string("y_271_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7719_cast_fp16 = mul(x = y_271_cast_fp16, y = y_271_cast_fp16)[name = string("op_7719_cast_fp16")];
            tensor<int32, [1]> var_7724_axes_0 = const()[name = string("op_7724_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7724_keep_dims_0 = const()[name = string("op_7724_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7724_cast_fp16 = reduce_mean(axes = var_7724_axes_0, keep_dims = var_7724_keep_dims_0, x = var_7719_cast_fp16)[name = string("op_7724_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7725_cast_fp16 = mul(x = s_271_cast_fp16, y = s_271_cast_fp16)[name = string("op_7725_cast_fp16")];
            fp32 var_7726_epsilon_0 = const()[name = string("op_7726_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7726_cast_fp16 = inverse(epsilon = var_7726_epsilon_0, x = var_7725_cast_fp16)[name = string("op_7726_cast_fp16")];
            fp16 var_7727_to_fp16 = const()[name = string("op_7727_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7728_cast_fp16 = mul(x = var_7726_cast_fp16, y = var_7727_to_fp16)[name = string("op_7728_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7730_cast_fp16 = add(x = var_7724_cast_fp16, y = var_7728_cast_fp16)[name = string("op_7730_cast_fp16")];
            fp32 var_7731_epsilon_0 = const()[name = string("op_7731_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7731_cast_fp16 = rsqrt(epsilon = var_7731_epsilon_0, x = var_7730_cast_fp16)[name = string("op_7731_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_237_cast_fp16 = mul(x = y_271_cast_fp16, y = var_7731_cast_fp16)[name = string("out_237_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97459456)))];
            tensor<fp16, [1, 512, 512]> var_7733_cast_fp16 = mul(x = out_237_cast_fp16, y = tm_layers_16_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7733_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_201_cast_fp16 = add(x = x_299_cast_fp16, y = var_7733_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97460544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97722752))))[name = string("tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_152_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_152_cast_fp16")];
            string x_303_mode_0 = const()[name = string("x_303_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_303_cast_fp16 = gelu(mode = x_303_mode_0, x = linear_152_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [4]> var_7753_begin_0 = const()[name = string("op_7753_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7753_end_0 = const()[name = string("op_7753_end_0"), val = tensor<int32, [4]>([1, 512, 17, 512])];
            tensor<bool, [4]> var_7753_end_mask_0 = const()[name = string("op_7753_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7753_squeeze_mask_0 = const()[name = string("op_7753_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_7753_cast_fp16 = slice_by_index(begin = var_7753_begin_0, end = var_7753_end_0, end_mask = var_7753_end_mask_0, squeeze_mask = var_7753_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7753_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_203_cast_fp16 = mul(x = x_303_cast_fp16, y = var_7753_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97723840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97986048))))[name = string("tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7763_cast_fp16 = abs(x = linear_153_cast_fp16)[name = string("op_7763_cast_fp16")];
            bool var_7766 = const()[name = string("op_7766"), val = bool(true)];
            tensor<int32, [1]> var_7767_axes_0 = const()[name = string("op_7767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7767_cast_fp16 = reduce_max(axes = var_7767_axes_0, keep_dims = var_7766, x = var_7763_cast_fp16)[name = string("op_7767_cast_fp16")];
            fp16 var_7768_to_fp16 = const()[name = string("op_7768_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_273_cast_fp16 = maximum(x = var_7767_cast_fp16, y = var_7768_to_fp16)[name = string("s_273_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_273_cast_fp16 = real_div(x = linear_153_cast_fp16, y = s_273_cast_fp16)[name = string("y_273_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7771_cast_fp16 = mul(x = y_273_cast_fp16, y = y_273_cast_fp16)[name = string("op_7771_cast_fp16")];
            tensor<int32, [1]> var_7776_axes_0 = const()[name = string("op_7776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7776_keep_dims_0 = const()[name = string("op_7776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7776_cast_fp16 = reduce_mean(axes = var_7776_axes_0, keep_dims = var_7776_keep_dims_0, x = var_7771_cast_fp16)[name = string("op_7776_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7777_cast_fp16 = mul(x = s_273_cast_fp16, y = s_273_cast_fp16)[name = string("op_7777_cast_fp16")];
            fp32 var_7778_epsilon_0 = const()[name = string("op_7778_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7778_cast_fp16 = inverse(epsilon = var_7778_epsilon_0, x = var_7777_cast_fp16)[name = string("op_7778_cast_fp16")];
            fp16 var_7779_to_fp16 = const()[name = string("op_7779_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7780_cast_fp16 = mul(x = var_7778_cast_fp16, y = var_7779_to_fp16)[name = string("op_7780_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7782_cast_fp16 = add(x = var_7776_cast_fp16, y = var_7780_cast_fp16)[name = string("op_7782_cast_fp16")];
            fp32 var_7783_epsilon_0 = const()[name = string("op_7783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7783_cast_fp16 = rsqrt(epsilon = var_7783_epsilon_0, x = var_7782_cast_fp16)[name = string("op_7783_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_239_cast_fp16 = mul(x = y_273_cast_fp16, y = var_7783_cast_fp16)[name = string("out_239_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97987136)))];
            tensor<fp16, [1, 512, 512]> var_7785_cast_fp16 = mul(x = out_239_cast_fp16, y = tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_33_cast_fp16 = add(x = input_201_cast_fp16, y = var_7785_cast_fp16)[name = string("h_33_cast_fp16")];
            tensor<fp16, [1]> tm_layers_16_layer_scalar_to_fp16 = const()[name = string("tm_layers_16_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7p-2])];
            tensor<fp16, [1, 512, 512]> x_307_cast_fp16 = mul(x = h_33_cast_fp16, y = tm_layers_16_layer_scalar_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7789_cast_fp16 = abs(x = x_307_cast_fp16)[name = string("op_7789_cast_fp16")];
            bool var_7792 = const()[name = string("op_7792"), val = bool(true)];
            tensor<int32, [1]> var_7793_axes_0 = const()[name = string("op_7793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7793_cast_fp16 = reduce_max(axes = var_7793_axes_0, keep_dims = var_7792, x = var_7789_cast_fp16)[name = string("op_7793_cast_fp16")];
            fp16 var_7794_to_fp16 = const()[name = string("op_7794_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_275_cast_fp16 = maximum(x = var_7793_cast_fp16, y = var_7794_to_fp16)[name = string("s_275_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_275_cast_fp16 = real_div(x = x_307_cast_fp16, y = s_275_cast_fp16)[name = string("y_275_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7797_cast_fp16 = mul(x = y_275_cast_fp16, y = y_275_cast_fp16)[name = string("op_7797_cast_fp16")];
            tensor<int32, [1]> var_7802_axes_0 = const()[name = string("op_7802_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7802_keep_dims_0 = const()[name = string("op_7802_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7802_cast_fp16 = reduce_mean(axes = var_7802_axes_0, keep_dims = var_7802_keep_dims_0, x = var_7797_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7803_cast_fp16 = mul(x = s_275_cast_fp16, y = s_275_cast_fp16)[name = string("op_7803_cast_fp16")];
            fp32 var_7804_epsilon_0 = const()[name = string("op_7804_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7804_cast_fp16 = inverse(epsilon = var_7804_epsilon_0, x = var_7803_cast_fp16)[name = string("op_7804_cast_fp16")];
            fp16 var_7805_to_fp16 = const()[name = string("op_7805_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7806_cast_fp16 = mul(x = var_7804_cast_fp16, y = var_7805_to_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7808_cast_fp16 = add(x = var_7802_cast_fp16, y = var_7806_cast_fp16)[name = string("op_7808_cast_fp16")];
            fp32 var_7809_epsilon_0 = const()[name = string("op_7809_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_7809_cast_fp16 = rsqrt(epsilon = var_7809_epsilon_0, x = var_7808_cast_fp16)[name = string("op_7809_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_241_cast_fp16 = mul(x = y_275_cast_fp16, y = var_7809_cast_fp16)[name = string("out_241_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97988224)))];
            tensor<fp16, [1, 512, 512]> input_205_cast_fp16 = mul(x = out_241_cast_fp16, y = tm_layers_17_input_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97989312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99037952))))[name = string("tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_154_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_7819 = const()[name = string("op_7819"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 512, 4, 512]> var_7820_cast_fp16 = reshape(shape = var_7819, x = linear_154_cast_fp16)[name = string("op_7820_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99042112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99304320))))[name = string("tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_7831 = const()[name = string("op_7831"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 512, 1, 512]> var_7832_cast_fp16 = reshape(shape = var_7831, x = linear_155_cast_fp16)[name = string("op_7832_cast_fp16")];
            tensor<int32, [4]> x_311_perm_0 = const()[name = string("x_311_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99305408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99567616))))[name = string("tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_7843 = const()[name = string("op_7843"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 512, 1, 512]> var_7844_cast_fp16 = reshape(shape = var_7843, x = linear_156_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<int32, [4]> x_313_perm_0 = const()[name = string("x_313_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 512]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_7820_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 4, 512, 512]> var_7848_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_7848_cast_fp16")];
            bool var_7851 = const()[name = string("op_7851"), val = bool(true)];
            tensor<int32, [1]> var_7852_axes_0 = const()[name = string("op_7852_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_7852_cast_fp16 = reduce_max(axes = var_7852_axes_0, keep_dims = var_7851, x = var_7848_cast_fp16)[name = string("op_7852_cast_fp16")];
            fp16 var_7853_to_fp16 = const()[name = string("op_7853_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_277_cast_fp16 = maximum(x = var_7852_cast_fp16, y = var_7853_to_fp16)[name = string("s_277_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> y_277_cast_fp16 = real_div(x = x_309_cast_fp16, y = s_277_cast_fp16)[name = string("y_277_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_7856_cast_fp16 = mul(x = y_277_cast_fp16, y = y_277_cast_fp16)[name = string("op_7856_cast_fp16")];
            tensor<int32, [1]> var_7861_axes_0 = const()[name = string("op_7861_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7861_keep_dims_0 = const()[name = string("op_7861_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_7861_cast_fp16 = reduce_mean(axes = var_7861_axes_0, keep_dims = var_7861_keep_dims_0, x = var_7856_cast_fp16)[name = string("op_7861_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_7862_cast_fp16 = mul(x = s_277_cast_fp16, y = s_277_cast_fp16)[name = string("op_7862_cast_fp16")];
            fp32 var_7863_epsilon_0 = const()[name = string("op_7863_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_7863_cast_fp16 = inverse(epsilon = var_7863_epsilon_0, x = var_7862_cast_fp16)[name = string("op_7863_cast_fp16")];
            fp16 var_7864_to_fp16 = const()[name = string("op_7864_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_7865_cast_fp16 = mul(x = var_7863_cast_fp16, y = var_7864_to_fp16)[name = string("op_7865_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_7867_cast_fp16 = add(x = var_7861_cast_fp16, y = var_7865_cast_fp16)[name = string("op_7867_cast_fp16")];
            fp32 var_7868_epsilon_0 = const()[name = string("op_7868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_7868_cast_fp16 = rsqrt(epsilon = var_7868_epsilon_0, x = var_7867_cast_fp16)[name = string("op_7868_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> out_243_cast_fp16 = mul(x = y_277_cast_fp16, y = var_7868_cast_fp16)[name = string("out_243_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99568704)))];
            tensor<fp16, [1, 4, 512, 512]> q_69_cast_fp16 = mul(x = out_243_cast_fp16, y = tm_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("q_69_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> x_311_cast_fp16 = transpose(perm = x_311_perm_0, x = var_7832_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 1, 512, 512]> var_7871_cast_fp16 = abs(x = x_311_cast_fp16)[name = string("op_7871_cast_fp16")];
            bool var_7874 = const()[name = string("op_7874"), val = bool(true)];
            tensor<int32, [1]> var_7875_axes_0 = const()[name = string("op_7875_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_7875_cast_fp16 = reduce_max(axes = var_7875_axes_0, keep_dims = var_7874, x = var_7871_cast_fp16)[name = string("op_7875_cast_fp16")];
            fp16 var_7876_to_fp16 = const()[name = string("op_7876_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 512, 1]> s_279_cast_fp16 = maximum(x = var_7875_cast_fp16, y = var_7876_to_fp16)[name = string("s_279_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> y_279_cast_fp16 = real_div(x = x_311_cast_fp16, y = s_279_cast_fp16)[name = string("y_279_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_7879_cast_fp16 = mul(x = y_279_cast_fp16, y = y_279_cast_fp16)[name = string("op_7879_cast_fp16")];
            tensor<int32, [1]> var_7884_axes_0 = const()[name = string("op_7884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7884_keep_dims_0 = const()[name = string("op_7884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> var_7884_cast_fp16 = reduce_mean(axes = var_7884_axes_0, keep_dims = var_7884_keep_dims_0, x = var_7879_cast_fp16)[name = string("op_7884_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_7885_cast_fp16 = mul(x = s_279_cast_fp16, y = s_279_cast_fp16)[name = string("op_7885_cast_fp16")];
            fp32 var_7886_epsilon_0 = const()[name = string("op_7886_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 512, 1]> var_7886_cast_fp16 = inverse(epsilon = var_7886_epsilon_0, x = var_7885_cast_fp16)[name = string("op_7886_cast_fp16")];
            fp16 var_7887_to_fp16 = const()[name = string("op_7887_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> var_7888_cast_fp16 = mul(x = var_7886_cast_fp16, y = var_7887_to_fp16)[name = string("op_7888_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_7890_cast_fp16 = add(x = var_7884_cast_fp16, y = var_7888_cast_fp16)[name = string("op_7890_cast_fp16")];
            fp32 var_7891_epsilon_0 = const()[name = string("op_7891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> var_7891_cast_fp16 = rsqrt(epsilon = var_7891_epsilon_0, x = var_7890_cast_fp16)[name = string("op_7891_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> out_245_cast_fp16 = mul(x = y_279_cast_fp16, y = var_7891_cast_fp16)[name = string("out_245_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99569792)))];
            tensor<fp16, [1, 1, 512, 512]> k_103_cast_fp16 = mul(x = out_245_cast_fp16, y = tm_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("k_103_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> x_313_cast_fp16 = transpose(perm = x_313_perm_0, x = var_7844_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1, 512, 512]> var_7894_cast_fp16 = abs(x = x_313_cast_fp16)[name = string("op_7894_cast_fp16")];
            bool var_7897 = const()[name = string("op_7897"), val = bool(true)];
            tensor<int32, [1]> var_7898_axes_0 = const()[name = string("op_7898_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_7898_cast_fp16 = reduce_max(axes = var_7898_axes_0, keep_dims = var_7897, x = var_7894_cast_fp16)[name = string("op_7898_cast_fp16")];
            fp16 var_7899_to_fp16 = const()[name = string("op_7899_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 512, 1]> s_281_cast_fp16 = maximum(x = var_7898_cast_fp16, y = var_7899_to_fp16)[name = string("s_281_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> y_281_cast_fp16 = real_div(x = x_313_cast_fp16, y = s_281_cast_fp16)[name = string("y_281_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_7902_cast_fp16 = mul(x = y_281_cast_fp16, y = y_281_cast_fp16)[name = string("op_7902_cast_fp16")];
            tensor<int32, [1]> var_7907_axes_0 = const()[name = string("op_7907_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7907_keep_dims_0 = const()[name = string("op_7907_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> var_7907_cast_fp16 = reduce_mean(axes = var_7907_axes_0, keep_dims = var_7907_keep_dims_0, x = var_7902_cast_fp16)[name = string("op_7907_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_7908_cast_fp16 = mul(x = s_281_cast_fp16, y = s_281_cast_fp16)[name = string("op_7908_cast_fp16")];
            fp32 var_7909_epsilon_0 = const()[name = string("op_7909_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 512, 1]> var_7909_cast_fp16 = inverse(epsilon = var_7909_epsilon_0, x = var_7908_cast_fp16)[name = string("op_7909_cast_fp16")];
            fp16 var_7910_to_fp16 = const()[name = string("op_7910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> var_7911_cast_fp16 = mul(x = var_7909_cast_fp16, y = var_7910_to_fp16)[name = string("op_7911_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_7913_cast_fp16 = add(x = var_7907_cast_fp16, y = var_7911_cast_fp16)[name = string("op_7913_cast_fp16")];
            fp32 var_7914_epsilon_0 = const()[name = string("op_7914_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> var_7914_cast_fp16 = rsqrt(epsilon = var_7914_epsilon_0, x = var_7913_cast_fp16)[name = string("op_7914_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> v_69_cast_fp16 = mul(x = y_281_cast_fp16, y = var_7914_cast_fp16)[name = string("v_69_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_7916_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7916_cast_fp16")];
            tensor<int32, [4]> var_7921_begin_0 = const()[name = string("op_7921_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7921_end_0 = const()[name = string("op_7921_end_0"), val = tensor<int32, [4]>([1, 4, 512, 512])];
            tensor<bool, [4]> var_7921_end_mask_0 = const()[name = string("op_7921_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 256]> var_7921_cast_fp16 = slice_by_index(begin = var_7921_begin_0, end = var_7921_end_0, end_mask = var_7921_end_mask_0, x = q_69_cast_fp16)[name = string("op_7921_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 256]> var_7922_cast_fp16 = mul(x = var_7921_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 256]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = q_69_cast_fp16)[name = string("op_7927_cast_fp16")];
            int32 var_7929 = const()[name = string("op_7929"), val = int32(-1)];
            bool var_7930_interleave_0 = const()[name = string("op_7930_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> var_7930_cast_fp16 = concat(axis = var_7929, interleave = var_7930_interleave_0, values = (var_7922_cast_fp16, var_7927_cast_fp16))[name = string("op_7930_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_7931_cast_fp16 = mul(x = var_7930_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7931_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> q_71_cast_fp16 = add(x = var_7916_cast_fp16, y = var_7931_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_7934_cast_fp16 = mul(x = k_103_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7934_cast_fp16")];
            tensor<int32, [4]> var_7939_begin_0 = const()[name = string("op_7939_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7939_end_0 = const()[name = string("op_7939_end_0"), val = tensor<int32, [4]>([1, 1, 512, 512])];
            tensor<bool, [4]> var_7939_end_mask_0 = const()[name = string("op_7939_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 256]> var_7939_cast_fp16 = slice_by_index(begin = var_7939_begin_0, end = var_7939_end_0, end_mask = var_7939_end_mask_0, x = k_103_cast_fp16)[name = string("op_7939_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 256]> var_7940_cast_fp16 = mul(x = var_7939_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_7940_cast_fp16")];
            tensor<int32, [4]> var_7945_begin_0 = const()[name = string("op_7945_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7945_end_0 = const()[name = string("op_7945_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
            tensor<bool, [4]> var_7945_end_mask_0 = const()[name = string("op_7945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 256]> var_7945_cast_fp16 = slice_by_index(begin = var_7945_begin_0, end = var_7945_end_0, end_mask = var_7945_end_mask_0, x = k_103_cast_fp16)[name = string("op_7945_cast_fp16")];
            int32 var_7947 = const()[name = string("op_7947"), val = int32(-1)];
            bool var_7948_interleave_0 = const()[name = string("op_7948_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 512]> var_7948_cast_fp16 = concat(axis = var_7947, interleave = var_7948_interleave_0, values = (var_7940_cast_fp16, var_7945_cast_fp16))[name = string("op_7948_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_7949_cast_fp16 = mul(x = var_7948_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> k_105_cast_fp16 = add(x = var_7934_cast_fp16, y = var_7949_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_34_reps_0 = const()[name = string("tile_34_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 512, 512]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = k_105_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 1, 512, 512]> tile_34_cast_fp16 = tile(reps = tile_34_reps_0, x = transpose_68_cast_fp16)[name = string("tile_34_cast_fp16")];
            tensor<int32, [5]> concat_68 = const()[name = string("concat_68"), val = tensor<int32, [5]>([4, 1, 1, 512, 512])];
            tensor<fp16, [4, 1, 1, 512, 512]> reshape_68_cast_fp16 = reshape(shape = concat_68, x = tile_34_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [5]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
            tensor<fp16, [1, 4, 1, 512, 512]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 1, 512, 512]> reshape_69_cast_fp16 = reshape(shape = concat_69, x = transpose_69_cast_fp16)[name = string("reshape_69_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_35_reps_0 = const()[name = string("tile_35_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 512, 512]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = v_69_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 1, 512, 512]> tile_35_cast_fp16 = tile(reps = tile_35_reps_0, x = transpose_70_cast_fp16)[name = string("tile_35_cast_fp16")];
            tensor<int32, [5]> concat_70 = const()[name = string("concat_70"), val = tensor<int32, [5]>([4, 1, 1, 512, 512])];
            tensor<fp16, [4, 1, 1, 512, 512]> reshape_70_cast_fp16 = reshape(shape = concat_70, x = tile_35_cast_fp16)[name = string("reshape_70_cast_fp16")];
            tensor<int32, [5]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
            tensor<fp16, [1, 4, 1, 512, 512]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = reshape_70_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 1, 512, 512]> reshape_71_cast_fp16 = reshape(shape = concat_71, x = transpose_71_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7963_transpose_x_0 = const()[name = string("op_7963_transpose_x_0"), val = bool(false)];
            bool var_7963_transpose_y_0 = const()[name = string("op_7963_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = reshape_69_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 4, 512, 512]> var_7963_cast_fp16 = matmul(transpose_x = var_7963_transpose_x_0, transpose_y = var_7963_transpose_y_0, x = q_71_cast_fp16, y = transpose_113_cast_fp16)[name = string("op_7963_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_7965_cast_fp16 = add(x = var_7963_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7965_cast_fp16")];
            int32 var_7966 = const()[name = string("op_7966"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_281_cast_fp16 = softmax(axis = var_7966, x = var_7965_cast_fp16)[name = string("w_281_cast_fp16")];
            bool var_7969_transpose_x_0 = const()[name = string("op_7969_transpose_x_0"), val = bool(false)];
            bool var_7969_transpose_y_0 = const()[name = string("op_7969_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> v_71_cast_fp16 = transpose(perm = v_71_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 4, 512, 512]> var_7969_cast_fp16 = matmul(transpose_x = var_7969_transpose_x_0, transpose_y = var_7969_transpose_y_0, x = w_281_cast_fp16, y = v_71_cast_fp16)[name = string("op_7969_cast_fp16")];
            tensor<int32, [4]> var_7972_perm_0 = const()[name = string("op_7972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7976 = const()[name = string("op_7976"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 512, 4, 512]> var_7972_cast_fp16 = transpose(perm = var_7972_perm_0, x = var_7969_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 512, 2048]> input_207_cast_fp16 = reshape(shape = var_7976, x = var_7972_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99570880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619520))))[name = string("tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7981_cast_fp16 = abs(x = linear_157_cast_fp16)[name = string("op_7981_cast_fp16")];
            bool var_7984 = const()[name = string("op_7984"), val = bool(true)];
            tensor<int32, [1]> var_7985_axes_0 = const()[name = string("op_7985_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_7985_cast_fp16 = reduce_max(axes = var_7985_axes_0, keep_dims = var_7984, x = var_7981_cast_fp16)[name = string("op_7985_cast_fp16")];
            fp16 var_7986_to_fp16 = const()[name = string("op_7986_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_283_cast_fp16 = maximum(x = var_7985_cast_fp16, y = var_7986_to_fp16)[name = string("s_283_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_283_cast_fp16 = real_div(x = linear_157_cast_fp16, y = s_283_cast_fp16)[name = string("y_283_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_7989_cast_fp16 = mul(x = y_283_cast_fp16, y = y_283_cast_fp16)[name = string("op_7989_cast_fp16")];
            tensor<int32, [1]> var_7994_axes_0 = const()[name = string("op_7994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7994_keep_dims_0 = const()[name = string("op_7994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7994_cast_fp16 = reduce_mean(axes = var_7994_axes_0, keep_dims = var_7994_keep_dims_0, x = var_7989_cast_fp16)[name = string("op_7994_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_7995_cast_fp16 = mul(x = s_283_cast_fp16, y = s_283_cast_fp16)[name = string("op_7995_cast_fp16")];
            fp32 var_7996_epsilon_0 = const()[name = string("op_7996_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_7996_cast_fp16 = inverse(epsilon = var_7996_epsilon_0, x = var_7995_cast_fp16)[name = string("op_7996_cast_fp16")];
            fp16 var_7997_to_fp16 = const()[name = string("op_7997_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_7998_cast_fp16 = mul(x = var_7996_cast_fp16, y = var_7997_to_fp16)[name = string("op_7998_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8000_cast_fp16 = add(x = var_7994_cast_fp16, y = var_7998_cast_fp16)[name = string("op_8000_cast_fp16")];
            fp32 var_8001_epsilon_0 = const()[name = string("op_8001_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8001_cast_fp16 = rsqrt(epsilon = var_8001_epsilon_0, x = var_8000_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_247_cast_fp16 = mul(x = y_283_cast_fp16, y = var_8001_cast_fp16)[name = string("out_247_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100620608)))];
            tensor<fp16, [1, 512, 512]> var_8003_cast_fp16 = mul(x = out_247_cast_fp16, y = tm_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("op_8003_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_317_cast_fp16 = add(x = x_307_cast_fp16, y = var_8003_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8006_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_8006_cast_fp16")];
            bool var_8009 = const()[name = string("op_8009"), val = bool(true)];
            tensor<int32, [1]> var_8010_axes_0 = const()[name = string("op_8010_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8010_cast_fp16 = reduce_max(axes = var_8010_axes_0, keep_dims = var_8009, x = var_8006_cast_fp16)[name = string("op_8010_cast_fp16")];
            fp16 var_8011_to_fp16 = const()[name = string("op_8011_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_285_cast_fp16 = maximum(x = var_8010_cast_fp16, y = var_8011_to_fp16)[name = string("s_285_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_285_cast_fp16 = real_div(x = x_317_cast_fp16, y = s_285_cast_fp16)[name = string("y_285_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8014_cast_fp16 = mul(x = y_285_cast_fp16, y = y_285_cast_fp16)[name = string("op_8014_cast_fp16")];
            tensor<int32, [1]> var_8019_axes_0 = const()[name = string("op_8019_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8019_keep_dims_0 = const()[name = string("op_8019_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8019_cast_fp16 = reduce_mean(axes = var_8019_axes_0, keep_dims = var_8019_keep_dims_0, x = var_8014_cast_fp16)[name = string("op_8019_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8020_cast_fp16 = mul(x = s_285_cast_fp16, y = s_285_cast_fp16)[name = string("op_8020_cast_fp16")];
            fp32 var_8021_epsilon_0 = const()[name = string("op_8021_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8021_cast_fp16 = inverse(epsilon = var_8021_epsilon_0, x = var_8020_cast_fp16)[name = string("op_8021_cast_fp16")];
            fp16 var_8022_to_fp16 = const()[name = string("op_8022_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8023_cast_fp16 = mul(x = var_8021_cast_fp16, y = var_8022_to_fp16)[name = string("op_8023_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8025_cast_fp16 = add(x = var_8019_cast_fp16, y = var_8023_cast_fp16)[name = string("op_8025_cast_fp16")];
            fp32 var_8026_epsilon_0 = const()[name = string("op_8026_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8026_cast_fp16 = rsqrt(epsilon = var_8026_epsilon_0, x = var_8025_cast_fp16)[name = string("op_8026_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_249_cast_fp16 = mul(x = y_285_cast_fp16, y = var_8026_cast_fp16)[name = string("out_249_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621696)))];
            tensor<fp16, [1, 512, 512]> input_209_cast_fp16 = mul(x = out_249_cast_fp16, y = tm_layers_17_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100622784))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101671424))))[name = string("tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_158_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_158_cast_fp16")];
            string var_8033_mode_0 = const()[name = string("op_8033_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_8033_cast_fp16 = gelu(mode = var_8033_mode_0, x = linear_158_cast_fp16)[name = string("op_8033_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101675584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102724224))))[name = string("tm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_159_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_211_cast_fp16 = mul(x = var_8033_cast_fp16, y = linear_159_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102728384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103777024))))[name = string("tm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8041_cast_fp16 = abs(x = linear_160_cast_fp16)[name = string("op_8041_cast_fp16")];
            bool var_8044 = const()[name = string("op_8044"), val = bool(true)];
            tensor<int32, [1]> var_8045_axes_0 = const()[name = string("op_8045_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8045_cast_fp16 = reduce_max(axes = var_8045_axes_0, keep_dims = var_8044, x = var_8041_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp16 var_8046_to_fp16 = const()[name = string("op_8046_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_287_cast_fp16 = maximum(x = var_8045_cast_fp16, y = var_8046_to_fp16)[name = string("s_287_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_287_cast_fp16 = real_div(x = linear_160_cast_fp16, y = s_287_cast_fp16)[name = string("y_287_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8049_cast_fp16 = mul(x = y_287_cast_fp16, y = y_287_cast_fp16)[name = string("op_8049_cast_fp16")];
            tensor<int32, [1]> var_8054_axes_0 = const()[name = string("op_8054_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8054_keep_dims_0 = const()[name = string("op_8054_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8054_cast_fp16 = reduce_mean(axes = var_8054_axes_0, keep_dims = var_8054_keep_dims_0, x = var_8049_cast_fp16)[name = string("op_8054_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8055_cast_fp16 = mul(x = s_287_cast_fp16, y = s_287_cast_fp16)[name = string("op_8055_cast_fp16")];
            fp32 var_8056_epsilon_0 = const()[name = string("op_8056_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8056_cast_fp16 = inverse(epsilon = var_8056_epsilon_0, x = var_8055_cast_fp16)[name = string("op_8056_cast_fp16")];
            fp16 var_8057_to_fp16 = const()[name = string("op_8057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8058_cast_fp16 = mul(x = var_8056_cast_fp16, y = var_8057_to_fp16)[name = string("op_8058_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8060_cast_fp16 = add(x = var_8054_cast_fp16, y = var_8058_cast_fp16)[name = string("op_8060_cast_fp16")];
            fp32 var_8061_epsilon_0 = const()[name = string("op_8061_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8061_cast_fp16 = rsqrt(epsilon = var_8061_epsilon_0, x = var_8060_cast_fp16)[name = string("op_8061_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_251_cast_fp16 = mul(x = y_287_cast_fp16, y = var_8061_cast_fp16)[name = string("out_251_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103778112)))];
            tensor<fp16, [1, 512, 512]> var_8063_cast_fp16 = mul(x = out_251_cast_fp16, y = tm_layers_17_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_213_cast_fp16 = add(x = x_317_cast_fp16, y = var_8063_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103779200))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041408))))[name = string("tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_161_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_161_cast_fp16")];
            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_321_cast_fp16 = gelu(mode = x_321_mode_0, x = linear_161_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [4]> var_8083_begin_0 = const()[name = string("op_8083_begin_0"), val = tensor<int32, [4]>([0, 0, 17, 0])];
            tensor<int32, [4]> var_8083_end_0 = const()[name = string("op_8083_end_0"), val = tensor<int32, [4]>([1, 512, 18, 512])];
            tensor<bool, [4]> var_8083_end_mask_0 = const()[name = string("op_8083_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8083_squeeze_mask_0 = const()[name = string("op_8083_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_8083_cast_fp16 = slice_by_index(begin = var_8083_begin_0, end = var_8083_end_0, end_mask = var_8083_end_mask_0, squeeze_mask = var_8083_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8083_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_215_cast_fp16 = mul(x = x_321_cast_fp16, y = var_8083_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104042496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104304704))))[name = string("tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8093_cast_fp16 = abs(x = linear_162_cast_fp16)[name = string("op_8093_cast_fp16")];
            bool var_8096 = const()[name = string("op_8096"), val = bool(true)];
            tensor<int32, [1]> var_8097_axes_0 = const()[name = string("op_8097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8097_cast_fp16 = reduce_max(axes = var_8097_axes_0, keep_dims = var_8096, x = var_8093_cast_fp16)[name = string("op_8097_cast_fp16")];
            fp16 var_8098_to_fp16 = const()[name = string("op_8098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_289_cast_fp16 = maximum(x = var_8097_cast_fp16, y = var_8098_to_fp16)[name = string("s_289_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_289_cast_fp16 = real_div(x = linear_162_cast_fp16, y = s_289_cast_fp16)[name = string("y_289_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8101_cast_fp16 = mul(x = y_289_cast_fp16, y = y_289_cast_fp16)[name = string("op_8101_cast_fp16")];
            tensor<int32, [1]> var_8106_axes_0 = const()[name = string("op_8106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8106_keep_dims_0 = const()[name = string("op_8106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8106_cast_fp16 = reduce_mean(axes = var_8106_axes_0, keep_dims = var_8106_keep_dims_0, x = var_8101_cast_fp16)[name = string("op_8106_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8107_cast_fp16 = mul(x = s_289_cast_fp16, y = s_289_cast_fp16)[name = string("op_8107_cast_fp16")];
            fp32 var_8108_epsilon_0 = const()[name = string("op_8108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8108_cast_fp16 = inverse(epsilon = var_8108_epsilon_0, x = var_8107_cast_fp16)[name = string("op_8108_cast_fp16")];
            fp16 var_8109_to_fp16 = const()[name = string("op_8109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8110_cast_fp16 = mul(x = var_8108_cast_fp16, y = var_8109_to_fp16)[name = string("op_8110_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8112_cast_fp16 = add(x = var_8106_cast_fp16, y = var_8110_cast_fp16)[name = string("op_8112_cast_fp16")];
            fp32 var_8113_epsilon_0 = const()[name = string("op_8113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8113_cast_fp16 = rsqrt(epsilon = var_8113_epsilon_0, x = var_8112_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_253_cast_fp16 = mul(x = y_289_cast_fp16, y = var_8113_cast_fp16)[name = string("out_253_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104305792)))];
            tensor<fp16, [1, 512, 512]> var_8115_cast_fp16 = mul(x = out_253_cast_fp16, y = tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8115_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_35_cast_fp16 = add(x = input_213_cast_fp16, y = var_8115_cast_fp16)[name = string("h_35_cast_fp16")];
            tensor<fp16, [1]> tm_layers_17_layer_scalar_to_fp16 = const()[name = string("tm_layers_17_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e6p-1])];
            tensor<fp16, [1, 512, 512]> x_325_cast_fp16 = mul(x = h_35_cast_fp16, y = tm_layers_17_layer_scalar_to_fp16)[name = string("x_325_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8119_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_8119_cast_fp16")];
            bool var_8122 = const()[name = string("op_8122"), val = bool(true)];
            tensor<int32, [1]> var_8123_axes_0 = const()[name = string("op_8123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8123_cast_fp16 = reduce_max(axes = var_8123_axes_0, keep_dims = var_8122, x = var_8119_cast_fp16)[name = string("op_8123_cast_fp16")];
            fp16 var_8124_to_fp16 = const()[name = string("op_8124_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_291_cast_fp16 = maximum(x = var_8123_cast_fp16, y = var_8124_to_fp16)[name = string("s_291_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_291_cast_fp16 = real_div(x = x_325_cast_fp16, y = s_291_cast_fp16)[name = string("y_291_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8127_cast_fp16 = mul(x = y_291_cast_fp16, y = y_291_cast_fp16)[name = string("op_8127_cast_fp16")];
            tensor<int32, [1]> var_8132_axes_0 = const()[name = string("op_8132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8132_keep_dims_0 = const()[name = string("op_8132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8132_cast_fp16 = reduce_mean(axes = var_8132_axes_0, keep_dims = var_8132_keep_dims_0, x = var_8127_cast_fp16)[name = string("op_8132_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8133_cast_fp16 = mul(x = s_291_cast_fp16, y = s_291_cast_fp16)[name = string("op_8133_cast_fp16")];
            fp32 var_8134_epsilon_0 = const()[name = string("op_8134_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8134_cast_fp16 = inverse(epsilon = var_8134_epsilon_0, x = var_8133_cast_fp16)[name = string("op_8134_cast_fp16")];
            fp16 var_8135_to_fp16 = const()[name = string("op_8135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = string("op_8136_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8138_cast_fp16 = add(x = var_8132_cast_fp16, y = var_8136_cast_fp16)[name = string("op_8138_cast_fp16")];
            fp32 var_8139_epsilon_0 = const()[name = string("op_8139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8139_cast_fp16 = rsqrt(epsilon = var_8139_epsilon_0, x = var_8138_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_255_cast_fp16 = mul(x = y_291_cast_fp16, y = var_8139_cast_fp16)[name = string("out_255_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104306880)))];
            tensor<fp16, [1, 512, 512]> input_217_cast_fp16 = mul(x = out_255_cast_fp16, y = tm_layers_18_input_layernorm_weight_to_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104307968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104832320))))[name = string("tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8149 = const()[name = string("op_8149"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_8150_cast_fp16 = reshape(shape = var_8149, x = linear_163_cast_fp16)[name = string("op_8150_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104834432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105096640))))[name = string("tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [4]> var_8161 = const()[name = string("op_8161"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_8162_cast_fp16 = reshape(shape = var_8161, x = linear_164_cast_fp16)[name = string("op_8162_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105097728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105359936))))[name = string("tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<int32, [4]> var_8173 = const()[name = string("op_8173"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_8174_cast_fp16 = reshape(shape = var_8173, x = linear_165_cast_fp16)[name = string("op_8174_cast_fp16")];
            tensor<int32, [4]> x_331_perm_0 = const()[name = string("x_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_8150_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 4, 512, 256]> var_8178_cast_fp16 = abs(x = x_327_cast_fp16)[name = string("op_8178_cast_fp16")];
            bool var_8181 = const()[name = string("op_8181"), val = bool(true)];
            tensor<int32, [1]> var_8182_axes_0 = const()[name = string("op_8182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_8182_cast_fp16 = reduce_max(axes = var_8182_axes_0, keep_dims = var_8181, x = var_8178_cast_fp16)[name = string("op_8182_cast_fp16")];
            fp16 var_8183_to_fp16 = const()[name = string("op_8183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_293_cast_fp16 = maximum(x = var_8182_cast_fp16, y = var_8183_to_fp16)[name = string("s_293_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_293_cast_fp16 = real_div(x = x_327_cast_fp16, y = s_293_cast_fp16)[name = string("y_293_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8186_cast_fp16 = mul(x = y_293_cast_fp16, y = y_293_cast_fp16)[name = string("op_8186_cast_fp16")];
            tensor<int32, [1]> var_8191_axes_0 = const()[name = string("op_8191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8191_keep_dims_0 = const()[name = string("op_8191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_8191_cast_fp16 = reduce_mean(axes = var_8191_axes_0, keep_dims = var_8191_keep_dims_0, x = var_8186_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_8192_cast_fp16 = mul(x = s_293_cast_fp16, y = s_293_cast_fp16)[name = string("op_8192_cast_fp16")];
            fp32 var_8193_epsilon_0 = const()[name = string("op_8193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_8193_cast_fp16 = inverse(epsilon = var_8193_epsilon_0, x = var_8192_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_8197_cast_fp16 = add(x = var_8191_cast_fp16, y = var_8195_cast_fp16)[name = string("op_8197_cast_fp16")];
            fp32 var_8198_epsilon_0 = const()[name = string("op_8198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_8198_cast_fp16 = rsqrt(epsilon = var_8198_epsilon_0, x = var_8197_cast_fp16)[name = string("op_8198_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_257_cast_fp16 = mul(x = y_293_cast_fp16, y = var_8198_cast_fp16)[name = string("out_257_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_73_cast_fp16 = mul(x = out_257_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_73_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_8162_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 512, 256]> var_8201_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_8201_cast_fp16")];
            bool var_8204 = const()[name = string("op_8204"), val = bool(true)];
            tensor<int32, [1]> var_8205_axes_0 = const()[name = string("op_8205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_8205_cast_fp16 = reduce_max(axes = var_8205_axes_0, keep_dims = var_8204, x = var_8201_cast_fp16)[name = string("op_8205_cast_fp16")];
            fp16 var_8206_to_fp16 = const()[name = string("op_8206_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_295_cast_fp16 = maximum(x = var_8205_cast_fp16, y = var_8206_to_fp16)[name = string("s_295_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_295_cast_fp16 = real_div(x = x_329_cast_fp16, y = s_295_cast_fp16)[name = string("y_295_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8209_cast_fp16 = mul(x = y_295_cast_fp16, y = y_295_cast_fp16)[name = string("op_8209_cast_fp16")];
            tensor<int32, [1]> var_8214_axes_0 = const()[name = string("op_8214_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8214_keep_dims_0 = const()[name = string("op_8214_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_8214_cast_fp16 = reduce_mean(axes = var_8214_axes_0, keep_dims = var_8214_keep_dims_0, x = var_8209_cast_fp16)[name = string("op_8214_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8215_cast_fp16 = mul(x = s_295_cast_fp16, y = s_295_cast_fp16)[name = string("op_8215_cast_fp16")];
            fp32 var_8216_epsilon_0 = const()[name = string("op_8216_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_8216_cast_fp16 = inverse(epsilon = var_8216_epsilon_0, x = var_8215_cast_fp16)[name = string("op_8216_cast_fp16")];
            fp16 var_8217_to_fp16 = const()[name = string("op_8217_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_8218_cast_fp16 = mul(x = var_8216_cast_fp16, y = var_8217_to_fp16)[name = string("op_8218_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8220_cast_fp16 = add(x = var_8214_cast_fp16, y = var_8218_cast_fp16)[name = string("op_8220_cast_fp16")];
            fp32 var_8221_epsilon_0 = const()[name = string("op_8221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_8221_cast_fp16 = rsqrt(epsilon = var_8221_epsilon_0, x = var_8220_cast_fp16)[name = string("op_8221_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_259_cast_fp16 = mul(x = y_295_cast_fp16, y = var_8221_cast_fp16)[name = string("out_259_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_109_cast_fp16 = mul(x = out_259_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_109_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_331_cast_fp16 = transpose(perm = x_331_perm_0, x = var_8174_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 512, 256]> var_8224_cast_fp16 = abs(x = x_331_cast_fp16)[name = string("op_8224_cast_fp16")];
            bool var_8227 = const()[name = string("op_8227"), val = bool(true)];
            tensor<int32, [1]> var_8228_axes_0 = const()[name = string("op_8228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_8228_cast_fp16 = reduce_max(axes = var_8228_axes_0, keep_dims = var_8227, x = var_8224_cast_fp16)[name = string("op_8228_cast_fp16")];
            fp16 var_8229_to_fp16 = const()[name = string("op_8229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_297_cast_fp16 = maximum(x = var_8228_cast_fp16, y = var_8229_to_fp16)[name = string("s_297_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_297_cast_fp16 = real_div(x = x_331_cast_fp16, y = s_297_cast_fp16)[name = string("y_297_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8232_cast_fp16 = mul(x = y_297_cast_fp16, y = y_297_cast_fp16)[name = string("op_8232_cast_fp16")];
            tensor<int32, [1]> var_8237_axes_0 = const()[name = string("op_8237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8237_keep_dims_0 = const()[name = string("op_8237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_8237_cast_fp16 = reduce_mean(axes = var_8237_axes_0, keep_dims = var_8237_keep_dims_0, x = var_8232_cast_fp16)[name = string("op_8237_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8238_cast_fp16 = mul(x = s_297_cast_fp16, y = s_297_cast_fp16)[name = string("op_8238_cast_fp16")];
            fp32 var_8239_epsilon_0 = const()[name = string("op_8239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_8239_cast_fp16 = inverse(epsilon = var_8239_epsilon_0, x = var_8238_cast_fp16)[name = string("op_8239_cast_fp16")];
            fp16 var_8240_to_fp16 = const()[name = string("op_8240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_8241_cast_fp16 = mul(x = var_8239_cast_fp16, y = var_8240_to_fp16)[name = string("op_8241_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8243_cast_fp16 = add(x = var_8237_cast_fp16, y = var_8241_cast_fp16)[name = string("op_8243_cast_fp16")];
            fp32 var_8244_epsilon_0 = const()[name = string("op_8244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_8244_cast_fp16 = rsqrt(epsilon = var_8244_epsilon_0, x = var_8243_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_73_cast_fp16 = mul(x = y_297_cast_fp16, y = var_8244_cast_fp16)[name = string("v_73_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8246_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8246_cast_fp16")];
            tensor<int32, [4]> var_8251_begin_0 = const()[name = string("op_8251_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8251_end_0 = const()[name = string("op_8251_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_8251_end_mask_0 = const()[name = string("op_8251_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_8251_cast_fp16 = slice_by_index(begin = var_8251_begin_0, end = var_8251_end_0, end_mask = var_8251_end_mask_0, x = q_73_cast_fp16)[name = string("op_8251_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_8252_cast_fp16 = mul(x = var_8251_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8257_begin_0 = const()[name = string("op_8257_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8257_end_0 = const()[name = string("op_8257_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_8257_end_mask_0 = const()[name = string("op_8257_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_8257_cast_fp16 = slice_by_index(begin = var_8257_begin_0, end = var_8257_end_0, end_mask = var_8257_end_mask_0, x = q_73_cast_fp16)[name = string("op_8257_cast_fp16")];
            int32 var_8259 = const()[name = string("op_8259"), val = int32(-1)];
            bool var_8260_interleave_0 = const()[name = string("op_8260_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_8260_cast_fp16 = concat(axis = var_8259, interleave = var_8260_interleave_0, values = (var_8252_cast_fp16, var_8257_cast_fp16))[name = string("op_8260_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8261_cast_fp16 = mul(x = var_8260_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8261_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_75_cast_fp16 = add(x = var_8246_cast_fp16, y = var_8261_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8264_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8264_cast_fp16")];
            tensor<int32, [4]> var_8269_begin_0 = const()[name = string("op_8269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8269_end_0 = const()[name = string("op_8269_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_8269_end_mask_0 = const()[name = string("op_8269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_8269_cast_fp16 = slice_by_index(begin = var_8269_begin_0, end = var_8269_end_0, end_mask = var_8269_end_mask_0, x = k_109_cast_fp16)[name = string("op_8269_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_8270_cast_fp16 = mul(x = var_8269_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_8270_cast_fp16")];
            tensor<int32, [4]> var_8275_begin_0 = const()[name = string("op_8275_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8275_end_0 = const()[name = string("op_8275_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_8275_end_mask_0 = const()[name = string("op_8275_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_8275_cast_fp16 = slice_by_index(begin = var_8275_begin_0, end = var_8275_end_0, end_mask = var_8275_end_mask_0, x = k_109_cast_fp16)[name = string("op_8275_cast_fp16")];
            int32 var_8277 = const()[name = string("op_8277"), val = int32(-1)];
            bool var_8278_interleave_0 = const()[name = string("op_8278_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_8278_cast_fp16 = concat(axis = var_8277, interleave = var_8278_interleave_0, values = (var_8270_cast_fp16, var_8275_cast_fp16))[name = string("op_8278_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8279_cast_fp16 = mul(x = var_8278_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8279_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_111_cast_fp16 = add(x = var_8264_cast_fp16, y = var_8279_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_36_reps_0 = const()[name = string("tile_36_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_72_cast_fp16 = transpose(perm = transpose_72_perm_0, x = k_111_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 1, 512, 256]> tile_36_cast_fp16 = tile(reps = tile_36_reps_0, x = transpose_72_cast_fp16)[name = string("tile_36_cast_fp16")];
            tensor<int32, [5]> concat_72 = const()[name = string("concat_72"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_72_cast_fp16 = reshape(shape = concat_72, x = tile_36_cast_fp16)[name = string("reshape_72_cast_fp16")];
            tensor<int32, [5]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_73 = const()[name = string("concat_73"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = reshape_72_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 1, 512, 256]> reshape_73_cast_fp16 = reshape(shape = concat_73, x = transpose_73_cast_fp16)[name = string("reshape_73_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_37_reps_0 = const()[name = string("tile_37_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = v_73_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 1, 512, 256]> tile_37_cast_fp16 = tile(reps = tile_37_reps_0, x = transpose_74_cast_fp16)[name = string("tile_37_cast_fp16")];
            tensor<int32, [5]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_74_cast_fp16 = reshape(shape = concat_74, x = tile_37_cast_fp16)[name = string("reshape_74_cast_fp16")];
            tensor<int32, [5]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = reshape_74_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 1, 512, 256]> reshape_75_cast_fp16 = reshape(shape = concat_75, x = transpose_75_cast_fp16)[name = string("reshape_75_cast_fp16")];
            tensor<int32, [4]> v_75_perm_0 = const()[name = string("v_75_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8293_transpose_x_0 = const()[name = string("op_8293_transpose_x_0"), val = bool(false)];
            bool var_8293_transpose_y_0 = const()[name = string("op_8293_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = reshape_73_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 4, 512, 512]> var_8293_cast_fp16 = matmul(transpose_x = var_8293_transpose_x_0, transpose_y = var_8293_transpose_y_0, x = q_75_cast_fp16, y = transpose_114_cast_fp16)[name = string("op_8293_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_8295_cast_fp16 = add(x = var_8293_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8295_cast_fp16")];
            int32 var_8296 = const()[name = string("op_8296"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_297_cast_fp16 = softmax(axis = var_8296, x = var_8295_cast_fp16)[name = string("w_297_cast_fp16")];
            bool var_8299_transpose_x_0 = const()[name = string("op_8299_transpose_x_0"), val = bool(false)];
            bool var_8299_transpose_y_0 = const()[name = string("op_8299_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_75_cast_fp16 = transpose(perm = v_75_perm_0, x = reshape_75_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 4, 512, 256]> var_8299_cast_fp16 = matmul(transpose_x = var_8299_transpose_x_0, transpose_y = var_8299_transpose_y_0, x = w_297_cast_fp16, y = v_75_cast_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [4]> var_8302_perm_0 = const()[name = string("op_8302_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8306 = const()[name = string("op_8306"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_8302_cast_fp16 = transpose(perm = var_8302_perm_0, x = var_8299_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 512, 1024]> input_219_cast_fp16 = reshape(shape = var_8306, x = var_8302_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105361024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105885376))))[name = string("tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8311_cast_fp16 = abs(x = linear_166_cast_fp16)[name = string("op_8311_cast_fp16")];
            bool var_8314 = const()[name = string("op_8314"), val = bool(true)];
            tensor<int32, [1]> var_8315_axes_0 = const()[name = string("op_8315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8315_cast_fp16 = reduce_max(axes = var_8315_axes_0, keep_dims = var_8314, x = var_8311_cast_fp16)[name = string("op_8315_cast_fp16")];
            fp16 var_8316_to_fp16 = const()[name = string("op_8316_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_299_cast_fp16 = maximum(x = var_8315_cast_fp16, y = var_8316_to_fp16)[name = string("s_299_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_299_cast_fp16 = real_div(x = linear_166_cast_fp16, y = s_299_cast_fp16)[name = string("y_299_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8319_cast_fp16 = mul(x = y_299_cast_fp16, y = y_299_cast_fp16)[name = string("op_8319_cast_fp16")];
            tensor<int32, [1]> var_8324_axes_0 = const()[name = string("op_8324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8324_keep_dims_0 = const()[name = string("op_8324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8324_cast_fp16 = reduce_mean(axes = var_8324_axes_0, keep_dims = var_8324_keep_dims_0, x = var_8319_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8325_cast_fp16 = mul(x = s_299_cast_fp16, y = s_299_cast_fp16)[name = string("op_8325_cast_fp16")];
            fp32 var_8326_epsilon_0 = const()[name = string("op_8326_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8326_cast_fp16 = inverse(epsilon = var_8326_epsilon_0, x = var_8325_cast_fp16)[name = string("op_8326_cast_fp16")];
            fp16 var_8327_to_fp16 = const()[name = string("op_8327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8328_cast_fp16 = mul(x = var_8326_cast_fp16, y = var_8327_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8330_cast_fp16 = add(x = var_8324_cast_fp16, y = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_261_cast_fp16 = mul(x = y_299_cast_fp16, y = var_8331_cast_fp16)[name = string("out_261_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105886464)))];
            tensor<fp16, [1, 512, 512]> var_8333_cast_fp16 = mul(x = out_261_cast_fp16, y = tm_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("op_8333_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = var_8333_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8336_cast_fp16 = abs(x = x_335_cast_fp16)[name = string("op_8336_cast_fp16")];
            bool var_8339 = const()[name = string("op_8339"), val = bool(true)];
            tensor<int32, [1]> var_8340_axes_0 = const()[name = string("op_8340_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8340_cast_fp16 = reduce_max(axes = var_8340_axes_0, keep_dims = var_8339, x = var_8336_cast_fp16)[name = string("op_8340_cast_fp16")];
            fp16 var_8341_to_fp16 = const()[name = string("op_8341_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_301_cast_fp16 = maximum(x = var_8340_cast_fp16, y = var_8341_to_fp16)[name = string("s_301_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_301_cast_fp16 = real_div(x = x_335_cast_fp16, y = s_301_cast_fp16)[name = string("y_301_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8344_cast_fp16 = mul(x = y_301_cast_fp16, y = y_301_cast_fp16)[name = string("op_8344_cast_fp16")];
            tensor<int32, [1]> var_8349_axes_0 = const()[name = string("op_8349_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8349_keep_dims_0 = const()[name = string("op_8349_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8349_cast_fp16 = reduce_mean(axes = var_8349_axes_0, keep_dims = var_8349_keep_dims_0, x = var_8344_cast_fp16)[name = string("op_8349_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8350_cast_fp16 = mul(x = s_301_cast_fp16, y = s_301_cast_fp16)[name = string("op_8350_cast_fp16")];
            fp32 var_8351_epsilon_0 = const()[name = string("op_8351_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8351_cast_fp16 = inverse(epsilon = var_8351_epsilon_0, x = var_8350_cast_fp16)[name = string("op_8351_cast_fp16")];
            fp16 var_8352_to_fp16 = const()[name = string("op_8352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8353_cast_fp16 = mul(x = var_8351_cast_fp16, y = var_8352_to_fp16)[name = string("op_8353_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8355_cast_fp16 = add(x = var_8349_cast_fp16, y = var_8353_cast_fp16)[name = string("op_8355_cast_fp16")];
            fp32 var_8356_epsilon_0 = const()[name = string("op_8356_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8356_cast_fp16 = rsqrt(epsilon = var_8356_epsilon_0, x = var_8355_cast_fp16)[name = string("op_8356_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_263_cast_fp16 = mul(x = y_301_cast_fp16, y = var_8356_cast_fp16)[name = string("out_263_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105887552)))];
            tensor<fp16, [1, 512, 512]> input_221_cast_fp16 = mul(x = out_263_cast_fp16, y = tm_layers_18_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105888640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106937280))))[name = string("tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_167_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_167_cast_fp16")];
            string var_8363_mode_0 = const()[name = string("op_8363_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_8363_cast_fp16 = gelu(mode = var_8363_mode_0, x = linear_167_cast_fp16)[name = string("op_8363_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106941440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107990080))))[name = string("tm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_168_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_223_cast_fp16 = mul(x = var_8363_cast_fp16, y = linear_168_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107994240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109042880))))[name = string("tm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8371_cast_fp16 = abs(x = linear_169_cast_fp16)[name = string("op_8371_cast_fp16")];
            bool var_8374 = const()[name = string("op_8374"), val = bool(true)];
            tensor<int32, [1]> var_8375_axes_0 = const()[name = string("op_8375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8375_cast_fp16 = reduce_max(axes = var_8375_axes_0, keep_dims = var_8374, x = var_8371_cast_fp16)[name = string("op_8375_cast_fp16")];
            fp16 var_8376_to_fp16 = const()[name = string("op_8376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_303_cast_fp16 = maximum(x = var_8375_cast_fp16, y = var_8376_to_fp16)[name = string("s_303_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_303_cast_fp16 = real_div(x = linear_169_cast_fp16, y = s_303_cast_fp16)[name = string("y_303_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8379_cast_fp16 = mul(x = y_303_cast_fp16, y = y_303_cast_fp16)[name = string("op_8379_cast_fp16")];
            tensor<int32, [1]> var_8384_axes_0 = const()[name = string("op_8384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8384_keep_dims_0 = const()[name = string("op_8384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8384_cast_fp16 = reduce_mean(axes = var_8384_axes_0, keep_dims = var_8384_keep_dims_0, x = var_8379_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8385_cast_fp16 = mul(x = s_303_cast_fp16, y = s_303_cast_fp16)[name = string("op_8385_cast_fp16")];
            fp32 var_8386_epsilon_0 = const()[name = string("op_8386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8386_cast_fp16 = inverse(epsilon = var_8386_epsilon_0, x = var_8385_cast_fp16)[name = string("op_8386_cast_fp16")];
            fp16 var_8387_to_fp16 = const()[name = string("op_8387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8388_cast_fp16 = mul(x = var_8386_cast_fp16, y = var_8387_to_fp16)[name = string("op_8388_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8390_cast_fp16 = add(x = var_8384_cast_fp16, y = var_8388_cast_fp16)[name = string("op_8390_cast_fp16")];
            fp32 var_8391_epsilon_0 = const()[name = string("op_8391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8391_cast_fp16 = rsqrt(epsilon = var_8391_epsilon_0, x = var_8390_cast_fp16)[name = string("op_8391_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_265_cast_fp16 = mul(x = y_303_cast_fp16, y = var_8391_cast_fp16)[name = string("out_265_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109043968)))];
            tensor<fp16, [1, 512, 512]> var_8393_cast_fp16 = mul(x = out_265_cast_fp16, y = tm_layers_18_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8393_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_225_cast_fp16 = add(x = x_335_cast_fp16, y = var_8393_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109045056))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109307264))))[name = string("tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_170_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_170_cast_fp16")];
            string x_339_mode_0 = const()[name = string("x_339_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_339_cast_fp16 = gelu(mode = x_339_mode_0, x = linear_170_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [4]> var_8413_begin_0 = const()[name = string("op_8413_begin_0"), val = tensor<int32, [4]>([0, 0, 18, 0])];
            tensor<int32, [4]> var_8413_end_0 = const()[name = string("op_8413_end_0"), val = tensor<int32, [4]>([1, 512, 19, 512])];
            tensor<bool, [4]> var_8413_end_mask_0 = const()[name = string("op_8413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8413_squeeze_mask_0 = const()[name = string("op_8413_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_8413_cast_fp16 = slice_by_index(begin = var_8413_begin_0, end = var_8413_end_0, end_mask = var_8413_end_mask_0, squeeze_mask = var_8413_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_227_cast_fp16 = mul(x = x_339_cast_fp16, y = var_8413_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109308352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109570560))))[name = string("tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8423_cast_fp16 = abs(x = linear_171_cast_fp16)[name = string("op_8423_cast_fp16")];
            bool var_8426 = const()[name = string("op_8426"), val = bool(true)];
            tensor<int32, [1]> var_8427_axes_0 = const()[name = string("op_8427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8427_cast_fp16 = reduce_max(axes = var_8427_axes_0, keep_dims = var_8426, x = var_8423_cast_fp16)[name = string("op_8427_cast_fp16")];
            fp16 var_8428_to_fp16 = const()[name = string("op_8428_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_305_cast_fp16 = maximum(x = var_8427_cast_fp16, y = var_8428_to_fp16)[name = string("s_305_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_305_cast_fp16 = real_div(x = linear_171_cast_fp16, y = s_305_cast_fp16)[name = string("y_305_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8431_cast_fp16 = mul(x = y_305_cast_fp16, y = y_305_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<int32, [1]> var_8436_axes_0 = const()[name = string("op_8436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8436_keep_dims_0 = const()[name = string("op_8436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8436_cast_fp16 = reduce_mean(axes = var_8436_axes_0, keep_dims = var_8436_keep_dims_0, x = var_8431_cast_fp16)[name = string("op_8436_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8437_cast_fp16 = mul(x = s_305_cast_fp16, y = s_305_cast_fp16)[name = string("op_8437_cast_fp16")];
            fp32 var_8438_epsilon_0 = const()[name = string("op_8438_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8438_cast_fp16 = inverse(epsilon = var_8438_epsilon_0, x = var_8437_cast_fp16)[name = string("op_8438_cast_fp16")];
            fp16 var_8439_to_fp16 = const()[name = string("op_8439_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8440_cast_fp16 = mul(x = var_8438_cast_fp16, y = var_8439_to_fp16)[name = string("op_8440_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8442_cast_fp16 = add(x = var_8436_cast_fp16, y = var_8440_cast_fp16)[name = string("op_8442_cast_fp16")];
            fp32 var_8443_epsilon_0 = const()[name = string("op_8443_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8443_cast_fp16 = rsqrt(epsilon = var_8443_epsilon_0, x = var_8442_cast_fp16)[name = string("op_8443_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_267_cast_fp16 = mul(x = y_305_cast_fp16, y = var_8443_cast_fp16)[name = string("out_267_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109571648)))];
            tensor<fp16, [1, 512, 512]> var_8445_cast_fp16 = mul(x = out_267_cast_fp16, y = tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8445_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_37_cast_fp16 = add(x = input_225_cast_fp16, y = var_8445_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<fp16, [1]> tm_layers_18_layer_scalar_to_fp16 = const()[name = string("tm_layers_18_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.68p-1])];
            tensor<fp16, [1, 512, 512]> x_343_cast_fp16 = mul(x = h_37_cast_fp16, y = tm_layers_18_layer_scalar_to_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8449_cast_fp16 = abs(x = x_343_cast_fp16)[name = string("op_8449_cast_fp16")];
            bool var_8452 = const()[name = string("op_8452"), val = bool(true)];
            tensor<int32, [1]> var_8453_axes_0 = const()[name = string("op_8453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8453_cast_fp16 = reduce_max(axes = var_8453_axes_0, keep_dims = var_8452, x = var_8449_cast_fp16)[name = string("op_8453_cast_fp16")];
            fp16 var_8454_to_fp16 = const()[name = string("op_8454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_307_cast_fp16 = maximum(x = var_8453_cast_fp16, y = var_8454_to_fp16)[name = string("s_307_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_307_cast_fp16 = real_div(x = x_343_cast_fp16, y = s_307_cast_fp16)[name = string("y_307_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8457_cast_fp16 = mul(x = y_307_cast_fp16, y = y_307_cast_fp16)[name = string("op_8457_cast_fp16")];
            tensor<int32, [1]> var_8462_axes_0 = const()[name = string("op_8462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8462_keep_dims_0 = const()[name = string("op_8462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8462_cast_fp16 = reduce_mean(axes = var_8462_axes_0, keep_dims = var_8462_keep_dims_0, x = var_8457_cast_fp16)[name = string("op_8462_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8463_cast_fp16 = mul(x = s_307_cast_fp16, y = s_307_cast_fp16)[name = string("op_8463_cast_fp16")];
            fp32 var_8464_epsilon_0 = const()[name = string("op_8464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8464_cast_fp16 = inverse(epsilon = var_8464_epsilon_0, x = var_8463_cast_fp16)[name = string("op_8464_cast_fp16")];
            fp16 var_8465_to_fp16 = const()[name = string("op_8465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8466_cast_fp16 = mul(x = var_8464_cast_fp16, y = var_8465_to_fp16)[name = string("op_8466_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8468_cast_fp16 = add(x = var_8462_cast_fp16, y = var_8466_cast_fp16)[name = string("op_8468_cast_fp16")];
            fp32 var_8469_epsilon_0 = const()[name = string("op_8469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8469_cast_fp16 = rsqrt(epsilon = var_8469_epsilon_0, x = var_8468_cast_fp16)[name = string("op_8469_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_269_cast_fp16 = mul(x = y_307_cast_fp16, y = var_8469_cast_fp16)[name = string("out_269_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109572736)))];
            tensor<fp16, [1, 512, 512]> input_229_cast_fp16 = mul(x = out_269_cast_fp16, y = tm_layers_19_input_layernorm_weight_to_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109573824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110098176))))[name = string("tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [4]> var_8479 = const()[name = string("op_8479"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_8480_cast_fp16 = reshape(shape = var_8479, x = linear_172_cast_fp16)[name = string("op_8480_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110100288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110362496))))[name = string("tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<int32, [4]> var_8491 = const()[name = string("op_8491"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_8492_cast_fp16 = reshape(shape = var_8491, x = linear_173_cast_fp16)[name = string("op_8492_cast_fp16")];
            tensor<int32, [4]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110363584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110625792))))[name = string("tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<int32, [4]> var_8503 = const()[name = string("op_8503"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_8504_cast_fp16 = reshape(shape = var_8503, x = linear_174_cast_fp16)[name = string("op_8504_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_8480_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 4, 512, 256]> var_8508_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_8508_cast_fp16")];
            bool var_8511 = const()[name = string("op_8511"), val = bool(true)];
            tensor<int32, [1]> var_8512_axes_0 = const()[name = string("op_8512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_8512_cast_fp16 = reduce_max(axes = var_8512_axes_0, keep_dims = var_8511, x = var_8508_cast_fp16)[name = string("op_8512_cast_fp16")];
            fp16 var_8513_to_fp16 = const()[name = string("op_8513_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_309_cast_fp16 = maximum(x = var_8512_cast_fp16, y = var_8513_to_fp16)[name = string("s_309_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_309_cast_fp16 = real_div(x = x_345_cast_fp16, y = s_309_cast_fp16)[name = string("y_309_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8516_cast_fp16 = mul(x = y_309_cast_fp16, y = y_309_cast_fp16)[name = string("op_8516_cast_fp16")];
            tensor<int32, [1]> var_8521_axes_0 = const()[name = string("op_8521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8521_keep_dims_0 = const()[name = string("op_8521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_8521_cast_fp16 = reduce_mean(axes = var_8521_axes_0, keep_dims = var_8521_keep_dims_0, x = var_8516_cast_fp16)[name = string("op_8521_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_8522_cast_fp16 = mul(x = s_309_cast_fp16, y = s_309_cast_fp16)[name = string("op_8522_cast_fp16")];
            fp32 var_8523_epsilon_0 = const()[name = string("op_8523_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_8523_cast_fp16 = inverse(epsilon = var_8523_epsilon_0, x = var_8522_cast_fp16)[name = string("op_8523_cast_fp16")];
            fp16 var_8524_to_fp16 = const()[name = string("op_8524_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_8525_cast_fp16 = mul(x = var_8523_cast_fp16, y = var_8524_to_fp16)[name = string("op_8525_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_8527_cast_fp16 = add(x = var_8521_cast_fp16, y = var_8525_cast_fp16)[name = string("op_8527_cast_fp16")];
            fp32 var_8528_epsilon_0 = const()[name = string("op_8528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_8528_cast_fp16 = rsqrt(epsilon = var_8528_epsilon_0, x = var_8527_cast_fp16)[name = string("op_8528_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_271_cast_fp16 = mul(x = y_309_cast_fp16, y = var_8528_cast_fp16)[name = string("out_271_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_77_cast_fp16 = mul(x = out_271_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_77_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = var_8492_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 512, 256]> var_8531_cast_fp16 = abs(x = x_347_cast_fp16)[name = string("op_8531_cast_fp16")];
            bool var_8534 = const()[name = string("op_8534"), val = bool(true)];
            tensor<int32, [1]> var_8535_axes_0 = const()[name = string("op_8535_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_8535_cast_fp16 = reduce_max(axes = var_8535_axes_0, keep_dims = var_8534, x = var_8531_cast_fp16)[name = string("op_8535_cast_fp16")];
            fp16 var_8536_to_fp16 = const()[name = string("op_8536_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_311_cast_fp16 = maximum(x = var_8535_cast_fp16, y = var_8536_to_fp16)[name = string("s_311_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_311_cast_fp16 = real_div(x = x_347_cast_fp16, y = s_311_cast_fp16)[name = string("y_311_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8539_cast_fp16 = mul(x = y_311_cast_fp16, y = y_311_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<int32, [1]> var_8544_axes_0 = const()[name = string("op_8544_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8544_keep_dims_0 = const()[name = string("op_8544_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_8544_cast_fp16 = reduce_mean(axes = var_8544_axes_0, keep_dims = var_8544_keep_dims_0, x = var_8539_cast_fp16)[name = string("op_8544_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8545_cast_fp16 = mul(x = s_311_cast_fp16, y = s_311_cast_fp16)[name = string("op_8545_cast_fp16")];
            fp32 var_8546_epsilon_0 = const()[name = string("op_8546_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_8546_cast_fp16 = inverse(epsilon = var_8546_epsilon_0, x = var_8545_cast_fp16)[name = string("op_8546_cast_fp16")];
            fp16 var_8547_to_fp16 = const()[name = string("op_8547_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_8548_cast_fp16 = mul(x = var_8546_cast_fp16, y = var_8547_to_fp16)[name = string("op_8548_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8550_cast_fp16 = add(x = var_8544_cast_fp16, y = var_8548_cast_fp16)[name = string("op_8550_cast_fp16")];
            fp32 var_8551_epsilon_0 = const()[name = string("op_8551_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_8551_cast_fp16 = rsqrt(epsilon = var_8551_epsilon_0, x = var_8550_cast_fp16)[name = string("op_8551_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_273_cast_fp16 = mul(x = y_311_cast_fp16, y = var_8551_cast_fp16)[name = string("out_273_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_115_cast_fp16 = mul(x = out_273_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_115_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_8504_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 2, 512, 256]> var_8554_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_8554_cast_fp16")];
            bool var_8557 = const()[name = string("op_8557"), val = bool(true)];
            tensor<int32, [1]> var_8558_axes_0 = const()[name = string("op_8558_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_8558_cast_fp16 = reduce_max(axes = var_8558_axes_0, keep_dims = var_8557, x = var_8554_cast_fp16)[name = string("op_8558_cast_fp16")];
            fp16 var_8559_to_fp16 = const()[name = string("op_8559_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_313_cast_fp16 = maximum(x = var_8558_cast_fp16, y = var_8559_to_fp16)[name = string("s_313_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_313_cast_fp16 = real_div(x = x_349_cast_fp16, y = s_313_cast_fp16)[name = string("y_313_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8562_cast_fp16 = mul(x = y_313_cast_fp16, y = y_313_cast_fp16)[name = string("op_8562_cast_fp16")];
            tensor<int32, [1]> var_8567_axes_0 = const()[name = string("op_8567_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8567_keep_dims_0 = const()[name = string("op_8567_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_8567_cast_fp16 = reduce_mean(axes = var_8567_axes_0, keep_dims = var_8567_keep_dims_0, x = var_8562_cast_fp16)[name = string("op_8567_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8568_cast_fp16 = mul(x = s_313_cast_fp16, y = s_313_cast_fp16)[name = string("op_8568_cast_fp16")];
            fp32 var_8569_epsilon_0 = const()[name = string("op_8569_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_8569_cast_fp16 = inverse(epsilon = var_8569_epsilon_0, x = var_8568_cast_fp16)[name = string("op_8569_cast_fp16")];
            fp16 var_8570_to_fp16 = const()[name = string("op_8570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_8571_cast_fp16 = mul(x = var_8569_cast_fp16, y = var_8570_to_fp16)[name = string("op_8571_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8573_cast_fp16 = add(x = var_8567_cast_fp16, y = var_8571_cast_fp16)[name = string("op_8573_cast_fp16")];
            fp32 var_8574_epsilon_0 = const()[name = string("op_8574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_8574_cast_fp16 = rsqrt(epsilon = var_8574_epsilon_0, x = var_8573_cast_fp16)[name = string("op_8574_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_77_cast_fp16 = mul(x = y_313_cast_fp16, y = var_8574_cast_fp16)[name = string("v_77_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8576_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8576_cast_fp16")];
            tensor<int32, [4]> var_8581_begin_0 = const()[name = string("op_8581_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8581_end_0 = const()[name = string("op_8581_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_8581_end_mask_0 = const()[name = string("op_8581_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_8581_cast_fp16 = slice_by_index(begin = var_8581_begin_0, end = var_8581_end_0, end_mask = var_8581_end_mask_0, x = q_77_cast_fp16)[name = string("op_8581_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_8582_cast_fp16 = mul(x = var_8581_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_8582_cast_fp16")];
            tensor<int32, [4]> var_8587_begin_0 = const()[name = string("op_8587_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8587_end_0 = const()[name = string("op_8587_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_8587_end_mask_0 = const()[name = string("op_8587_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_8587_cast_fp16 = slice_by_index(begin = var_8587_begin_0, end = var_8587_end_0, end_mask = var_8587_end_mask_0, x = q_77_cast_fp16)[name = string("op_8587_cast_fp16")];
            int32 var_8589 = const()[name = string("op_8589"), val = int32(-1)];
            bool var_8590_interleave_0 = const()[name = string("op_8590_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_8590_cast_fp16 = concat(axis = var_8589, interleave = var_8590_interleave_0, values = (var_8582_cast_fp16, var_8587_cast_fp16))[name = string("op_8590_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8591_cast_fp16 = mul(x = var_8590_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8591_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_79_cast_fp16 = add(x = var_8576_cast_fp16, y = var_8591_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8594_cast_fp16 = mul(x = k_115_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8594_cast_fp16")];
            tensor<int32, [4]> var_8599_begin_0 = const()[name = string("op_8599_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8599_end_0 = const()[name = string("op_8599_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_8599_end_mask_0 = const()[name = string("op_8599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_8599_cast_fp16 = slice_by_index(begin = var_8599_begin_0, end = var_8599_end_0, end_mask = var_8599_end_mask_0, x = k_115_cast_fp16)[name = string("op_8599_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_8600_cast_fp16 = mul(x = var_8599_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_8600_cast_fp16")];
            tensor<int32, [4]> var_8605_begin_0 = const()[name = string("op_8605_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8605_end_0 = const()[name = string("op_8605_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_8605_end_mask_0 = const()[name = string("op_8605_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_8605_cast_fp16 = slice_by_index(begin = var_8605_begin_0, end = var_8605_end_0, end_mask = var_8605_end_mask_0, x = k_115_cast_fp16)[name = string("op_8605_cast_fp16")];
            int32 var_8607 = const()[name = string("op_8607"), val = int32(-1)];
            bool var_8608_interleave_0 = const()[name = string("op_8608_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_8608_cast_fp16 = concat(axis = var_8607, interleave = var_8608_interleave_0, values = (var_8600_cast_fp16, var_8605_cast_fp16))[name = string("op_8608_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8609_cast_fp16 = mul(x = var_8608_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8609_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_117_cast_fp16 = add(x = var_8594_cast_fp16, y = var_8609_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_38_reps_0 = const()[name = string("tile_38_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = k_117_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 1, 512, 256]> tile_38_cast_fp16 = tile(reps = tile_38_reps_0, x = transpose_76_cast_fp16)[name = string("tile_38_cast_fp16")];
            tensor<int32, [5]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_76_cast_fp16 = reshape(shape = concat_76, x = tile_38_cast_fp16)[name = string("reshape_76_cast_fp16")];
            tensor<int32, [5]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_77 = const()[name = string("concat_77"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_77_cast_fp16 = transpose(perm = transpose_77_perm_0, x = reshape_76_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 1, 512, 256]> reshape_77_cast_fp16 = reshape(shape = concat_77, x = transpose_77_cast_fp16)[name = string("reshape_77_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_39_reps_0 = const()[name = string("tile_39_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_78_cast_fp16 = transpose(perm = transpose_78_perm_0, x = v_77_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 1, 512, 256]> tile_39_cast_fp16 = tile(reps = tile_39_reps_0, x = transpose_78_cast_fp16)[name = string("tile_39_cast_fp16")];
            tensor<int32, [5]> concat_78 = const()[name = string("concat_78"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_78_cast_fp16 = reshape(shape = concat_78, x = tile_39_cast_fp16)[name = string("reshape_78_cast_fp16")];
            tensor<int32, [5]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = reshape_78_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 1, 512, 256]> reshape_79_cast_fp16 = reshape(shape = concat_79, x = transpose_79_cast_fp16)[name = string("reshape_79_cast_fp16")];
            tensor<int32, [4]> v_79_perm_0 = const()[name = string("v_79_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8623_transpose_x_0 = const()[name = string("op_8623_transpose_x_0"), val = bool(false)];
            bool var_8623_transpose_y_0 = const()[name = string("op_8623_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = reshape_77_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4, 512, 512]> var_8623_cast_fp16 = matmul(transpose_x = var_8623_transpose_x_0, transpose_y = var_8623_transpose_y_0, x = q_79_cast_fp16, y = transpose_115_cast_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8625_cast_fp16")];
            int32 var_8626 = const()[name = string("op_8626"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_313_cast_fp16 = softmax(axis = var_8626, x = var_8625_cast_fp16)[name = string("w_313_cast_fp16")];
            bool var_8629_transpose_x_0 = const()[name = string("op_8629_transpose_x_0"), val = bool(false)];
            bool var_8629_transpose_y_0 = const()[name = string("op_8629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = reshape_79_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 4, 512, 256]> var_8629_cast_fp16 = matmul(transpose_x = var_8629_transpose_x_0, transpose_y = var_8629_transpose_y_0, x = w_313_cast_fp16, y = v_79_cast_fp16)[name = string("op_8629_cast_fp16")];
            tensor<int32, [4]> var_8632_perm_0 = const()[name = string("op_8632_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_8632_cast_fp16 = transpose(perm = var_8632_perm_0, x = var_8629_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 512, 1024]> input_231_cast_fp16 = reshape(shape = var_8636, x = var_8632_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110626880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111151232))))[name = string("tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8641_cast_fp16 = abs(x = linear_175_cast_fp16)[name = string("op_8641_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<int32, [1]> var_8645_axes_0 = const()[name = string("op_8645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8645_cast_fp16 = reduce_max(axes = var_8645_axes_0, keep_dims = var_8644, x = var_8641_cast_fp16)[name = string("op_8645_cast_fp16")];
            fp16 var_8646_to_fp16 = const()[name = string("op_8646_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_315_cast_fp16 = maximum(x = var_8645_cast_fp16, y = var_8646_to_fp16)[name = string("s_315_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_315_cast_fp16 = real_div(x = linear_175_cast_fp16, y = s_315_cast_fp16)[name = string("y_315_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8649_cast_fp16 = mul(x = y_315_cast_fp16, y = y_315_cast_fp16)[name = string("op_8649_cast_fp16")];
            tensor<int32, [1]> var_8654_axes_0 = const()[name = string("op_8654_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8654_keep_dims_0 = const()[name = string("op_8654_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8654_cast_fp16 = reduce_mean(axes = var_8654_axes_0, keep_dims = var_8654_keep_dims_0, x = var_8649_cast_fp16)[name = string("op_8654_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8655_cast_fp16 = mul(x = s_315_cast_fp16, y = s_315_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp32 var_8656_epsilon_0 = const()[name = string("op_8656_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8656_cast_fp16 = inverse(epsilon = var_8656_epsilon_0, x = var_8655_cast_fp16)[name = string("op_8656_cast_fp16")];
            fp16 var_8657_to_fp16 = const()[name = string("op_8657_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8658_cast_fp16 = mul(x = var_8656_cast_fp16, y = var_8657_to_fp16)[name = string("op_8658_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8660_cast_fp16 = add(x = var_8654_cast_fp16, y = var_8658_cast_fp16)[name = string("op_8660_cast_fp16")];
            fp32 var_8661_epsilon_0 = const()[name = string("op_8661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8661_cast_fp16 = rsqrt(epsilon = var_8661_epsilon_0, x = var_8660_cast_fp16)[name = string("op_8661_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_275_cast_fp16 = mul(x = y_315_cast_fp16, y = var_8661_cast_fp16)[name = string("out_275_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111152320)))];
            tensor<fp16, [1, 512, 512]> var_8663_cast_fp16 = mul(x = out_275_cast_fp16, y = tm_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("op_8663_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_353_cast_fp16 = add(x = x_343_cast_fp16, y = var_8663_cast_fp16)[name = string("x_353_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8666_cast_fp16 = abs(x = x_353_cast_fp16)[name = string("op_8666_cast_fp16")];
            bool var_8669 = const()[name = string("op_8669"), val = bool(true)];
            tensor<int32, [1]> var_8670_axes_0 = const()[name = string("op_8670_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8670_cast_fp16 = reduce_max(axes = var_8670_axes_0, keep_dims = var_8669, x = var_8666_cast_fp16)[name = string("op_8670_cast_fp16")];
            fp16 var_8671_to_fp16 = const()[name = string("op_8671_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_317_cast_fp16 = maximum(x = var_8670_cast_fp16, y = var_8671_to_fp16)[name = string("s_317_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_317_cast_fp16 = real_div(x = x_353_cast_fp16, y = s_317_cast_fp16)[name = string("y_317_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8674_cast_fp16 = mul(x = y_317_cast_fp16, y = y_317_cast_fp16)[name = string("op_8674_cast_fp16")];
            tensor<int32, [1]> var_8679_axes_0 = const()[name = string("op_8679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8679_keep_dims_0 = const()[name = string("op_8679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8679_cast_fp16 = reduce_mean(axes = var_8679_axes_0, keep_dims = var_8679_keep_dims_0, x = var_8674_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8680_cast_fp16 = mul(x = s_317_cast_fp16, y = s_317_cast_fp16)[name = string("op_8680_cast_fp16")];
            fp32 var_8681_epsilon_0 = const()[name = string("op_8681_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8681_cast_fp16 = inverse(epsilon = var_8681_epsilon_0, x = var_8680_cast_fp16)[name = string("op_8681_cast_fp16")];
            fp16 var_8682_to_fp16 = const()[name = string("op_8682_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8683_cast_fp16 = mul(x = var_8681_cast_fp16, y = var_8682_to_fp16)[name = string("op_8683_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8685_cast_fp16 = add(x = var_8679_cast_fp16, y = var_8683_cast_fp16)[name = string("op_8685_cast_fp16")];
            fp32 var_8686_epsilon_0 = const()[name = string("op_8686_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8686_cast_fp16 = rsqrt(epsilon = var_8686_epsilon_0, x = var_8685_cast_fp16)[name = string("op_8686_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_277_cast_fp16 = mul(x = y_317_cast_fp16, y = var_8686_cast_fp16)[name = string("out_277_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111153408)))];
            tensor<fp16, [1, 512, 512]> input_233_cast_fp16 = mul(x = out_277_cast_fp16, y = tm_layers_19_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111154496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112203136))))[name = string("tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_176_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_176_cast_fp16")];
            string var_8693_mode_0 = const()[name = string("op_8693_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_8693_cast_fp16 = gelu(mode = var_8693_mode_0, x = linear_176_cast_fp16)[name = string("op_8693_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112207296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113255936))))[name = string("tm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_177_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_235_cast_fp16 = mul(x = var_8693_cast_fp16, y = linear_177_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113260096))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114308736))))[name = string("tm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8701_cast_fp16 = abs(x = linear_178_cast_fp16)[name = string("op_8701_cast_fp16")];
            bool var_8704 = const()[name = string("op_8704"), val = bool(true)];
            tensor<int32, [1]> var_8705_axes_0 = const()[name = string("op_8705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8705_cast_fp16 = reduce_max(axes = var_8705_axes_0, keep_dims = var_8704, x = var_8701_cast_fp16)[name = string("op_8705_cast_fp16")];
            fp16 var_8706_to_fp16 = const()[name = string("op_8706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_319_cast_fp16 = maximum(x = var_8705_cast_fp16, y = var_8706_to_fp16)[name = string("s_319_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_319_cast_fp16 = real_div(x = linear_178_cast_fp16, y = s_319_cast_fp16)[name = string("y_319_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8709_cast_fp16 = mul(x = y_319_cast_fp16, y = y_319_cast_fp16)[name = string("op_8709_cast_fp16")];
            tensor<int32, [1]> var_8714_axes_0 = const()[name = string("op_8714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8714_keep_dims_0 = const()[name = string("op_8714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8714_cast_fp16 = reduce_mean(axes = var_8714_axes_0, keep_dims = var_8714_keep_dims_0, x = var_8709_cast_fp16)[name = string("op_8714_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8715_cast_fp16 = mul(x = s_319_cast_fp16, y = s_319_cast_fp16)[name = string("op_8715_cast_fp16")];
            fp32 var_8716_epsilon_0 = const()[name = string("op_8716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8716_cast_fp16 = inverse(epsilon = var_8716_epsilon_0, x = var_8715_cast_fp16)[name = string("op_8716_cast_fp16")];
            fp16 var_8717_to_fp16 = const()[name = string("op_8717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8718_cast_fp16 = mul(x = var_8716_cast_fp16, y = var_8717_to_fp16)[name = string("op_8718_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8720_cast_fp16 = add(x = var_8714_cast_fp16, y = var_8718_cast_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8721_cast_fp16 = rsqrt(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_279_cast_fp16 = mul(x = y_319_cast_fp16, y = var_8721_cast_fp16)[name = string("out_279_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114309824)))];
            tensor<fp16, [1, 512, 512]> var_8723_cast_fp16 = mul(x = out_279_cast_fp16, y = tm_layers_19_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_237_cast_fp16 = add(x = x_353_cast_fp16, y = var_8723_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114310912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114573120))))[name = string("tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_179_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_237_cast_fp16)[name = string("linear_179_cast_fp16")];
            string x_357_mode_0 = const()[name = string("x_357_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_357_cast_fp16 = gelu(mode = x_357_mode_0, x = linear_179_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [4]> var_8743_begin_0 = const()[name = string("op_8743_begin_0"), val = tensor<int32, [4]>([0, 0, 19, 0])];
            tensor<int32, [4]> var_8743_end_0 = const()[name = string("op_8743_end_0"), val = tensor<int32, [4]>([1, 512, 20, 512])];
            tensor<bool, [4]> var_8743_end_mask_0 = const()[name = string("op_8743_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8743_squeeze_mask_0 = const()[name = string("op_8743_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_8743_cast_fp16 = slice_by_index(begin = var_8743_begin_0, end = var_8743_end_0, end_mask = var_8743_end_mask_0, squeeze_mask = var_8743_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8743_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_239_cast_fp16 = mul(x = x_357_cast_fp16, y = var_8743_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114574208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114836416))))[name = string("tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8753_cast_fp16 = abs(x = linear_180_cast_fp16)[name = string("op_8753_cast_fp16")];
            bool var_8756 = const()[name = string("op_8756"), val = bool(true)];
            tensor<int32, [1]> var_8757_axes_0 = const()[name = string("op_8757_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8757_cast_fp16 = reduce_max(axes = var_8757_axes_0, keep_dims = var_8756, x = var_8753_cast_fp16)[name = string("op_8757_cast_fp16")];
            fp16 var_8758_to_fp16 = const()[name = string("op_8758_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_321_cast_fp16 = maximum(x = var_8757_cast_fp16, y = var_8758_to_fp16)[name = string("s_321_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_321_cast_fp16 = real_div(x = linear_180_cast_fp16, y = s_321_cast_fp16)[name = string("y_321_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8761_cast_fp16 = mul(x = y_321_cast_fp16, y = y_321_cast_fp16)[name = string("op_8761_cast_fp16")];
            tensor<int32, [1]> var_8766_axes_0 = const()[name = string("op_8766_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8766_keep_dims_0 = const()[name = string("op_8766_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8766_cast_fp16 = reduce_mean(axes = var_8766_axes_0, keep_dims = var_8766_keep_dims_0, x = var_8761_cast_fp16)[name = string("op_8766_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8767_cast_fp16 = mul(x = s_321_cast_fp16, y = s_321_cast_fp16)[name = string("op_8767_cast_fp16")];
            fp32 var_8768_epsilon_0 = const()[name = string("op_8768_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8768_cast_fp16 = inverse(epsilon = var_8768_epsilon_0, x = var_8767_cast_fp16)[name = string("op_8768_cast_fp16")];
            fp16 var_8769_to_fp16 = const()[name = string("op_8769_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8770_cast_fp16 = mul(x = var_8768_cast_fp16, y = var_8769_to_fp16)[name = string("op_8770_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8772_cast_fp16 = add(x = var_8766_cast_fp16, y = var_8770_cast_fp16)[name = string("op_8772_cast_fp16")];
            fp32 var_8773_epsilon_0 = const()[name = string("op_8773_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8773_cast_fp16 = rsqrt(epsilon = var_8773_epsilon_0, x = var_8772_cast_fp16)[name = string("op_8773_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_281_cast_fp16 = mul(x = y_321_cast_fp16, y = var_8773_cast_fp16)[name = string("out_281_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114837504)))];
            tensor<fp16, [1, 512, 512]> var_8775_cast_fp16 = mul(x = out_281_cast_fp16, y = tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_39_cast_fp16 = add(x = input_237_cast_fp16, y = var_8775_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<fp16, [1]> tm_layers_19_layer_scalar_to_fp16 = const()[name = string("tm_layers_19_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fp-1])];
            tensor<fp16, [1, 512, 512]> x_361_cast_fp16 = mul(x = h_39_cast_fp16, y = tm_layers_19_layer_scalar_to_fp16)[name = string("x_361_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8779_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_8779_cast_fp16")];
            bool var_8782 = const()[name = string("op_8782"), val = bool(true)];
            tensor<int32, [1]> var_8783_axes_0 = const()[name = string("op_8783_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8783_cast_fp16 = reduce_max(axes = var_8783_axes_0, keep_dims = var_8782, x = var_8779_cast_fp16)[name = string("op_8783_cast_fp16")];
            fp16 var_8784_to_fp16 = const()[name = string("op_8784_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_323_cast_fp16 = maximum(x = var_8783_cast_fp16, y = var_8784_to_fp16)[name = string("s_323_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_323_cast_fp16 = real_div(x = x_361_cast_fp16, y = s_323_cast_fp16)[name = string("y_323_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8787_cast_fp16 = mul(x = y_323_cast_fp16, y = y_323_cast_fp16)[name = string("op_8787_cast_fp16")];
            tensor<int32, [1]> var_8792_axes_0 = const()[name = string("op_8792_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8792_keep_dims_0 = const()[name = string("op_8792_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8792_cast_fp16 = reduce_mean(axes = var_8792_axes_0, keep_dims = var_8792_keep_dims_0, x = var_8787_cast_fp16)[name = string("op_8792_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8793_cast_fp16 = mul(x = s_323_cast_fp16, y = s_323_cast_fp16)[name = string("op_8793_cast_fp16")];
            fp32 var_8794_epsilon_0 = const()[name = string("op_8794_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8794_cast_fp16 = inverse(epsilon = var_8794_epsilon_0, x = var_8793_cast_fp16)[name = string("op_8794_cast_fp16")];
            fp16 var_8795_to_fp16 = const()[name = string("op_8795_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8796_cast_fp16 = mul(x = var_8794_cast_fp16, y = var_8795_to_fp16)[name = string("op_8796_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8798_cast_fp16 = add(x = var_8792_cast_fp16, y = var_8796_cast_fp16)[name = string("op_8798_cast_fp16")];
            fp32 var_8799_epsilon_0 = const()[name = string("op_8799_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8799_cast_fp16 = rsqrt(epsilon = var_8799_epsilon_0, x = var_8798_cast_fp16)[name = string("op_8799_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_283_cast_fp16 = mul(x = y_323_cast_fp16, y = var_8799_cast_fp16)[name = string("out_283_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114838592)))];
            tensor<fp16, [1, 512, 512]> input_241_cast_fp16 = mul(x = out_283_cast_fp16, y = tm_layers_20_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114839680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115364032))))[name = string("tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<int32, [4]> var_8809 = const()[name = string("op_8809"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_8810_cast_fp16 = reshape(shape = var_8809, x = linear_181_cast_fp16)[name = string("op_8810_cast_fp16")];
            tensor<int32, [4]> x_363_perm_0 = const()[name = string("x_363_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115366144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115628352))))[name = string("tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_8821 = const()[name = string("op_8821"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_8822_cast_fp16 = reshape(shape = var_8821, x = linear_182_cast_fp16)[name = string("op_8822_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115629440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115891648))))[name = string("tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_8833 = const()[name = string("op_8833"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_8834_cast_fp16 = reshape(shape = var_8833, x = linear_183_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<int32, [4]> x_367_perm_0 = const()[name = string("x_367_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_363_cast_fp16 = transpose(perm = x_363_perm_0, x = var_8810_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 4, 512, 256]> var_8838_cast_fp16 = abs(x = x_363_cast_fp16)[name = string("op_8838_cast_fp16")];
            bool var_8841 = const()[name = string("op_8841"), val = bool(true)];
            tensor<int32, [1]> var_8842_axes_0 = const()[name = string("op_8842_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_8842_cast_fp16 = reduce_max(axes = var_8842_axes_0, keep_dims = var_8841, x = var_8838_cast_fp16)[name = string("op_8842_cast_fp16")];
            fp16 var_8843_to_fp16 = const()[name = string("op_8843_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_325_cast_fp16 = maximum(x = var_8842_cast_fp16, y = var_8843_to_fp16)[name = string("s_325_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_325_cast_fp16 = real_div(x = x_363_cast_fp16, y = s_325_cast_fp16)[name = string("y_325_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8846_cast_fp16 = mul(x = y_325_cast_fp16, y = y_325_cast_fp16)[name = string("op_8846_cast_fp16")];
            tensor<int32, [1]> var_8851_axes_0 = const()[name = string("op_8851_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8851_keep_dims_0 = const()[name = string("op_8851_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_8851_cast_fp16 = reduce_mean(axes = var_8851_axes_0, keep_dims = var_8851_keep_dims_0, x = var_8846_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_8852_cast_fp16 = mul(x = s_325_cast_fp16, y = s_325_cast_fp16)[name = string("op_8852_cast_fp16")];
            fp32 var_8853_epsilon_0 = const()[name = string("op_8853_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_8853_cast_fp16 = inverse(epsilon = var_8853_epsilon_0, x = var_8852_cast_fp16)[name = string("op_8853_cast_fp16")];
            fp16 var_8854_to_fp16 = const()[name = string("op_8854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_8855_cast_fp16 = mul(x = var_8853_cast_fp16, y = var_8854_to_fp16)[name = string("op_8855_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_8857_cast_fp16 = add(x = var_8851_cast_fp16, y = var_8855_cast_fp16)[name = string("op_8857_cast_fp16")];
            fp32 var_8858_epsilon_0 = const()[name = string("op_8858_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_8858_cast_fp16 = rsqrt(epsilon = var_8858_epsilon_0, x = var_8857_cast_fp16)[name = string("op_8858_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_285_cast_fp16 = mul(x = y_325_cast_fp16, y = var_8858_cast_fp16)[name = string("out_285_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_81_cast_fp16 = mul(x = out_285_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_81_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_8822_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 512, 256]> var_8861_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_8861_cast_fp16")];
            bool var_8864 = const()[name = string("op_8864"), val = bool(true)];
            tensor<int32, [1]> var_8865_axes_0 = const()[name = string("op_8865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_8865_cast_fp16 = reduce_max(axes = var_8865_axes_0, keep_dims = var_8864, x = var_8861_cast_fp16)[name = string("op_8865_cast_fp16")];
            fp16 var_8866_to_fp16 = const()[name = string("op_8866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_327_cast_fp16 = maximum(x = var_8865_cast_fp16, y = var_8866_to_fp16)[name = string("s_327_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_327_cast_fp16 = real_div(x = x_365_cast_fp16, y = s_327_cast_fp16)[name = string("y_327_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8869_cast_fp16 = mul(x = y_327_cast_fp16, y = y_327_cast_fp16)[name = string("op_8869_cast_fp16")];
            tensor<int32, [1]> var_8874_axes_0 = const()[name = string("op_8874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8874_keep_dims_0 = const()[name = string("op_8874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_8874_cast_fp16 = reduce_mean(axes = var_8874_axes_0, keep_dims = var_8874_keep_dims_0, x = var_8869_cast_fp16)[name = string("op_8874_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8875_cast_fp16 = mul(x = s_327_cast_fp16, y = s_327_cast_fp16)[name = string("op_8875_cast_fp16")];
            fp32 var_8876_epsilon_0 = const()[name = string("op_8876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_8876_cast_fp16 = inverse(epsilon = var_8876_epsilon_0, x = var_8875_cast_fp16)[name = string("op_8876_cast_fp16")];
            fp16 var_8877_to_fp16 = const()[name = string("op_8877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_8878_cast_fp16 = mul(x = var_8876_cast_fp16, y = var_8877_to_fp16)[name = string("op_8878_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8880_cast_fp16 = add(x = var_8874_cast_fp16, y = var_8878_cast_fp16)[name = string("op_8880_cast_fp16")];
            fp32 var_8881_epsilon_0 = const()[name = string("op_8881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_8881_cast_fp16 = rsqrt(epsilon = var_8881_epsilon_0, x = var_8880_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_287_cast_fp16 = mul(x = y_327_cast_fp16, y = var_8881_cast_fp16)[name = string("out_287_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_121_cast_fp16 = mul(x = out_287_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_121_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_367_cast_fp16 = transpose(perm = x_367_perm_0, x = var_8834_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 512, 256]> var_8884_cast_fp16 = abs(x = x_367_cast_fp16)[name = string("op_8884_cast_fp16")];
            bool var_8887 = const()[name = string("op_8887"), val = bool(true)];
            tensor<int32, [1]> var_8888_axes_0 = const()[name = string("op_8888_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_8888_cast_fp16 = reduce_max(axes = var_8888_axes_0, keep_dims = var_8887, x = var_8884_cast_fp16)[name = string("op_8888_cast_fp16")];
            fp16 var_8889_to_fp16 = const()[name = string("op_8889_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_329_cast_fp16 = maximum(x = var_8888_cast_fp16, y = var_8889_to_fp16)[name = string("s_329_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_329_cast_fp16 = real_div(x = x_367_cast_fp16, y = s_329_cast_fp16)[name = string("y_329_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8892_cast_fp16 = mul(x = y_329_cast_fp16, y = y_329_cast_fp16)[name = string("op_8892_cast_fp16")];
            tensor<int32, [1]> var_8897_axes_0 = const()[name = string("op_8897_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8897_keep_dims_0 = const()[name = string("op_8897_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_8897_cast_fp16 = reduce_mean(axes = var_8897_axes_0, keep_dims = var_8897_keep_dims_0, x = var_8892_cast_fp16)[name = string("op_8897_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8898_cast_fp16 = mul(x = s_329_cast_fp16, y = s_329_cast_fp16)[name = string("op_8898_cast_fp16")];
            fp32 var_8899_epsilon_0 = const()[name = string("op_8899_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_8899_cast_fp16 = inverse(epsilon = var_8899_epsilon_0, x = var_8898_cast_fp16)[name = string("op_8899_cast_fp16")];
            fp16 var_8900_to_fp16 = const()[name = string("op_8900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_8901_cast_fp16 = mul(x = var_8899_cast_fp16, y = var_8900_to_fp16)[name = string("op_8901_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_8903_cast_fp16 = add(x = var_8897_cast_fp16, y = var_8901_cast_fp16)[name = string("op_8903_cast_fp16")];
            fp32 var_8904_epsilon_0 = const()[name = string("op_8904_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_8904_cast_fp16 = rsqrt(epsilon = var_8904_epsilon_0, x = var_8903_cast_fp16)[name = string("op_8904_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_81_cast_fp16 = mul(x = y_329_cast_fp16, y = var_8904_cast_fp16)[name = string("v_81_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8906_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8906_cast_fp16")];
            tensor<int32, [4]> var_8911_begin_0 = const()[name = string("op_8911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8911_end_0 = const()[name = string("op_8911_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_8911_end_mask_0 = const()[name = string("op_8911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_8911_cast_fp16 = slice_by_index(begin = var_8911_begin_0, end = var_8911_end_0, end_mask = var_8911_end_mask_0, x = q_81_cast_fp16)[name = string("op_8911_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_8912_cast_fp16 = mul(x = var_8911_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_8912_cast_fp16")];
            tensor<int32, [4]> var_8917_begin_0 = const()[name = string("op_8917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8917_end_0 = const()[name = string("op_8917_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_8917_end_mask_0 = const()[name = string("op_8917_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_8917_cast_fp16 = slice_by_index(begin = var_8917_begin_0, end = var_8917_end_0, end_mask = var_8917_end_mask_0, x = q_81_cast_fp16)[name = string("op_8917_cast_fp16")];
            int32 var_8919 = const()[name = string("op_8919"), val = int32(-1)];
            bool var_8920_interleave_0 = const()[name = string("op_8920_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_8920_cast_fp16 = concat(axis = var_8919, interleave = var_8920_interleave_0, values = (var_8912_cast_fp16, var_8917_cast_fp16))[name = string("op_8920_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_8921_cast_fp16 = mul(x = var_8920_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_83_cast_fp16 = add(x = var_8906_cast_fp16, y = var_8921_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8924_cast_fp16 = mul(x = k_121_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8924_cast_fp16")];
            tensor<int32, [4]> var_8929_begin_0 = const()[name = string("op_8929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8929_end_0 = const()[name = string("op_8929_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_8929_end_mask_0 = const()[name = string("op_8929_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, x = k_121_cast_fp16)[name = string("op_8929_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_8930_cast_fp16 = mul(x = var_8929_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_8930_cast_fp16")];
            tensor<int32, [4]> var_8935_begin_0 = const()[name = string("op_8935_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8935_end_0 = const()[name = string("op_8935_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_8935_end_mask_0 = const()[name = string("op_8935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, x = k_121_cast_fp16)[name = string("op_8935_cast_fp16")];
            int32 var_8937 = const()[name = string("op_8937"), val = int32(-1)];
            bool var_8938_interleave_0 = const()[name = string("op_8938_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_8938_cast_fp16 = concat(axis = var_8937, interleave = var_8938_interleave_0, values = (var_8930_cast_fp16, var_8935_cast_fp16))[name = string("op_8938_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_8939_cast_fp16 = mul(x = var_8938_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8939_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_123_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8939_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_40_reps_0 = const()[name = string("tile_40_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = k_123_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 1, 512, 256]> tile_40_cast_fp16 = tile(reps = tile_40_reps_0, x = transpose_80_cast_fp16)[name = string("tile_40_cast_fp16")];
            tensor<int32, [5]> concat_80 = const()[name = string("concat_80"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_80_cast_fp16 = reshape(shape = concat_80, x = tile_40_cast_fp16)[name = string("reshape_80_cast_fp16")];
            tensor<int32, [5]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_81 = const()[name = string("concat_81"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = reshape_80_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 1, 512, 256]> reshape_81_cast_fp16 = reshape(shape = concat_81, x = transpose_81_cast_fp16)[name = string("reshape_81_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_41_reps_0 = const()[name = string("tile_41_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_82_cast_fp16 = transpose(perm = transpose_82_perm_0, x = v_81_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 1, 512, 256]> tile_41_cast_fp16 = tile(reps = tile_41_reps_0, x = transpose_82_cast_fp16)[name = string("tile_41_cast_fp16")];
            tensor<int32, [5]> concat_82 = const()[name = string("concat_82"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_82_cast_fp16 = reshape(shape = concat_82, x = tile_41_cast_fp16)[name = string("reshape_82_cast_fp16")];
            tensor<int32, [5]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_83 = const()[name = string("concat_83"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = reshape_82_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 1, 512, 256]> reshape_83_cast_fp16 = reshape(shape = concat_83, x = transpose_83_cast_fp16)[name = string("reshape_83_cast_fp16")];
            tensor<int32, [4]> v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8953_transpose_x_0 = const()[name = string("op_8953_transpose_x_0"), val = bool(false)];
            bool var_8953_transpose_y_0 = const()[name = string("op_8953_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = reshape_81_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 4, 512, 512]> var_8953_cast_fp16 = matmul(transpose_x = var_8953_transpose_x_0, transpose_y = var_8953_transpose_y_0, x = q_83_cast_fp16, y = transpose_116_cast_fp16)[name = string("op_8953_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_8955_cast_fp16 = add(x = var_8953_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8955_cast_fp16")];
            int32 var_8956 = const()[name = string("op_8956"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_329_cast_fp16 = softmax(axis = var_8956, x = var_8955_cast_fp16)[name = string("w_329_cast_fp16")];
            bool var_8959_transpose_x_0 = const()[name = string("op_8959_transpose_x_0"), val = bool(false)];
            bool var_8959_transpose_y_0 = const()[name = string("op_8959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_83_cast_fp16 = transpose(perm = v_83_perm_0, x = reshape_83_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 4, 512, 256]> var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = w_329_cast_fp16, y = v_83_cast_fp16)[name = string("op_8959_cast_fp16")];
            tensor<int32, [4]> var_8962_perm_0 = const()[name = string("op_8962_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8966 = const()[name = string("op_8966"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_8962_cast_fp16 = transpose(perm = var_8962_perm_0, x = var_8959_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 512, 1024]> input_243_cast_fp16 = reshape(shape = var_8966, x = var_8962_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115892736))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116417088))))[name = string("tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8971_cast_fp16 = abs(x = linear_184_cast_fp16)[name = string("op_8971_cast_fp16")];
            bool var_8974 = const()[name = string("op_8974"), val = bool(true)];
            tensor<int32, [1]> var_8975_axes_0 = const()[name = string("op_8975_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_8975_cast_fp16 = reduce_max(axes = var_8975_axes_0, keep_dims = var_8974, x = var_8971_cast_fp16)[name = string("op_8975_cast_fp16")];
            fp16 var_8976_to_fp16 = const()[name = string("op_8976_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_331_cast_fp16 = maximum(x = var_8975_cast_fp16, y = var_8976_to_fp16)[name = string("s_331_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_331_cast_fp16 = real_div(x = linear_184_cast_fp16, y = s_331_cast_fp16)[name = string("y_331_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8979_cast_fp16 = mul(x = y_331_cast_fp16, y = y_331_cast_fp16)[name = string("op_8979_cast_fp16")];
            tensor<int32, [1]> var_8984_axes_0 = const()[name = string("op_8984_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8984_keep_dims_0 = const()[name = string("op_8984_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_8984_cast_fp16 = reduce_mean(axes = var_8984_axes_0, keep_dims = var_8984_keep_dims_0, x = var_8979_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8985_cast_fp16 = mul(x = s_331_cast_fp16, y = s_331_cast_fp16)[name = string("op_8985_cast_fp16")];
            fp32 var_8986_epsilon_0 = const()[name = string("op_8986_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_8986_cast_fp16 = inverse(epsilon = var_8986_epsilon_0, x = var_8985_cast_fp16)[name = string("op_8986_cast_fp16")];
            fp16 var_8987_to_fp16 = const()[name = string("op_8987_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_8988_cast_fp16 = mul(x = var_8986_cast_fp16, y = var_8987_to_fp16)[name = string("op_8988_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_8990_cast_fp16 = add(x = var_8984_cast_fp16, y = var_8988_cast_fp16)[name = string("op_8990_cast_fp16")];
            fp32 var_8991_epsilon_0 = const()[name = string("op_8991_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_8991_cast_fp16 = rsqrt(epsilon = var_8991_epsilon_0, x = var_8990_cast_fp16)[name = string("op_8991_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_289_cast_fp16 = mul(x = y_331_cast_fp16, y = var_8991_cast_fp16)[name = string("out_289_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116418176)))];
            tensor<fp16, [1, 512, 512]> var_8993_cast_fp16 = mul(x = out_289_cast_fp16, y = tm_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("op_8993_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_371_cast_fp16 = add(x = x_361_cast_fp16, y = var_8993_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_8996_cast_fp16 = abs(x = x_371_cast_fp16)[name = string("op_8996_cast_fp16")];
            bool var_8999 = const()[name = string("op_8999"), val = bool(true)];
            tensor<int32, [1]> var_9000_axes_0 = const()[name = string("op_9000_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9000_cast_fp16 = reduce_max(axes = var_9000_axes_0, keep_dims = var_8999, x = var_8996_cast_fp16)[name = string("op_9000_cast_fp16")];
            fp16 var_9001_to_fp16 = const()[name = string("op_9001_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_333_cast_fp16 = maximum(x = var_9000_cast_fp16, y = var_9001_to_fp16)[name = string("s_333_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_333_cast_fp16 = real_div(x = x_371_cast_fp16, y = s_333_cast_fp16)[name = string("y_333_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9004_cast_fp16 = mul(x = y_333_cast_fp16, y = y_333_cast_fp16)[name = string("op_9004_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9004_cast_fp16)[name = string("op_9009_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9010_cast_fp16 = mul(x = s_333_cast_fp16, y = s_333_cast_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9015_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9015_cast_fp16")];
            fp32 var_9016_epsilon_0 = const()[name = string("op_9016_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9016_cast_fp16 = rsqrt(epsilon = var_9016_epsilon_0, x = var_9015_cast_fp16)[name = string("op_9016_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_291_cast_fp16 = mul(x = y_333_cast_fp16, y = var_9016_cast_fp16)[name = string("out_291_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116419264)))];
            tensor<fp16, [1, 512, 512]> input_245_cast_fp16 = mul(x = out_291_cast_fp16, y = tm_layers_20_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116420352))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117468992))))[name = string("tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_185_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_185_cast_fp16")];
            string var_9023_mode_0 = const()[name = string("op_9023_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_9023_cast_fp16 = gelu(mode = var_9023_mode_0, x = linear_185_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117473152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118521792))))[name = string("tm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_186_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_247_cast_fp16 = mul(x = var_9023_cast_fp16, y = linear_186_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118525952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119574592))))[name = string("tm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9031_cast_fp16 = abs(x = linear_187_cast_fp16)[name = string("op_9031_cast_fp16")];
            bool var_9034 = const()[name = string("op_9034"), val = bool(true)];
            tensor<int32, [1]> var_9035_axes_0 = const()[name = string("op_9035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9035_cast_fp16 = reduce_max(axes = var_9035_axes_0, keep_dims = var_9034, x = var_9031_cast_fp16)[name = string("op_9035_cast_fp16")];
            fp16 var_9036_to_fp16 = const()[name = string("op_9036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_335_cast_fp16 = maximum(x = var_9035_cast_fp16, y = var_9036_to_fp16)[name = string("s_335_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_335_cast_fp16 = real_div(x = linear_187_cast_fp16, y = s_335_cast_fp16)[name = string("y_335_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9039_cast_fp16 = mul(x = y_335_cast_fp16, y = y_335_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [1]> var_9044_axes_0 = const()[name = string("op_9044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9044_keep_dims_0 = const()[name = string("op_9044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9044_cast_fp16 = reduce_mean(axes = var_9044_axes_0, keep_dims = var_9044_keep_dims_0, x = var_9039_cast_fp16)[name = string("op_9044_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9045_cast_fp16 = mul(x = s_335_cast_fp16, y = s_335_cast_fp16)[name = string("op_9045_cast_fp16")];
            fp32 var_9046_epsilon_0 = const()[name = string("op_9046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9046_cast_fp16 = inverse(epsilon = var_9046_epsilon_0, x = var_9045_cast_fp16)[name = string("op_9046_cast_fp16")];
            fp16 var_9047_to_fp16 = const()[name = string("op_9047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9048_cast_fp16 = mul(x = var_9046_cast_fp16, y = var_9047_to_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9050_cast_fp16 = add(x = var_9044_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9050_cast_fp16")];
            fp32 var_9051_epsilon_0 = const()[name = string("op_9051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9051_cast_fp16 = rsqrt(epsilon = var_9051_epsilon_0, x = var_9050_cast_fp16)[name = string("op_9051_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_293_cast_fp16 = mul(x = y_335_cast_fp16, y = var_9051_cast_fp16)[name = string("out_293_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119575680)))];
            tensor<fp16, [1, 512, 512]> var_9053_cast_fp16 = mul(x = out_293_cast_fp16, y = tm_layers_20_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9053_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_249_cast_fp16 = add(x = x_371_cast_fp16, y = var_9053_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119576768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119838976))))[name = string("tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_188_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_188_cast_fp16")];
            string x_375_mode_0 = const()[name = string("x_375_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_375_cast_fp16 = gelu(mode = x_375_mode_0, x = linear_188_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [4]> var_9073_begin_0 = const()[name = string("op_9073_begin_0"), val = tensor<int32, [4]>([0, 0, 20, 0])];
            tensor<int32, [4]> var_9073_end_0 = const()[name = string("op_9073_end_0"), val = tensor<int32, [4]>([1, 512, 21, 512])];
            tensor<bool, [4]> var_9073_end_mask_0 = const()[name = string("op_9073_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9073_squeeze_mask_0 = const()[name = string("op_9073_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_9073_cast_fp16 = slice_by_index(begin = var_9073_begin_0, end = var_9073_end_0, end_mask = var_9073_end_mask_0, squeeze_mask = var_9073_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_251_cast_fp16 = mul(x = x_375_cast_fp16, y = var_9073_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119840064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120102272))))[name = string("tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9083_cast_fp16 = abs(x = linear_189_cast_fp16)[name = string("op_9083_cast_fp16")];
            bool var_9086 = const()[name = string("op_9086"), val = bool(true)];
            tensor<int32, [1]> var_9087_axes_0 = const()[name = string("op_9087_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9087_cast_fp16 = reduce_max(axes = var_9087_axes_0, keep_dims = var_9086, x = var_9083_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_337_cast_fp16 = maximum(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("s_337_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_337_cast_fp16 = real_div(x = linear_189_cast_fp16, y = s_337_cast_fp16)[name = string("y_337_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9091_cast_fp16 = mul(x = y_337_cast_fp16, y = y_337_cast_fp16)[name = string("op_9091_cast_fp16")];
            tensor<int32, [1]> var_9096_axes_0 = const()[name = string("op_9096_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9096_keep_dims_0 = const()[name = string("op_9096_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9096_cast_fp16 = reduce_mean(axes = var_9096_axes_0, keep_dims = var_9096_keep_dims_0, x = var_9091_cast_fp16)[name = string("op_9096_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9097_cast_fp16 = mul(x = s_337_cast_fp16, y = s_337_cast_fp16)[name = string("op_9097_cast_fp16")];
            fp32 var_9098_epsilon_0 = const()[name = string("op_9098_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9098_cast_fp16 = inverse(epsilon = var_9098_epsilon_0, x = var_9097_cast_fp16)[name = string("op_9098_cast_fp16")];
            fp16 var_9099_to_fp16 = const()[name = string("op_9099_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9100_cast_fp16 = mul(x = var_9098_cast_fp16, y = var_9099_to_fp16)[name = string("op_9100_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9102_cast_fp16 = add(x = var_9096_cast_fp16, y = var_9100_cast_fp16)[name = string("op_9102_cast_fp16")];
            fp32 var_9103_epsilon_0 = const()[name = string("op_9103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9103_cast_fp16 = rsqrt(epsilon = var_9103_epsilon_0, x = var_9102_cast_fp16)[name = string("op_9103_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_295_cast_fp16 = mul(x = y_337_cast_fp16, y = var_9103_cast_fp16)[name = string("out_295_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120103360)))];
            tensor<fp16, [1, 512, 512]> var_9105_cast_fp16 = mul(x = out_295_cast_fp16, y = tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9105_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_41_cast_fp16 = add(x = input_249_cast_fp16, y = var_9105_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<fp16, [1]> tm_layers_20_layer_scalar_to_fp16 = const()[name = string("tm_layers_20_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ecp-1])];
            tensor<fp16, [1, 512, 512]> x_379_cast_fp16 = mul(x = h_41_cast_fp16, y = tm_layers_20_layer_scalar_to_fp16)[name = string("x_379_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9109_cast_fp16 = abs(x = x_379_cast_fp16)[name = string("op_9109_cast_fp16")];
            bool var_9112 = const()[name = string("op_9112"), val = bool(true)];
            tensor<int32, [1]> var_9113_axes_0 = const()[name = string("op_9113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9113_cast_fp16 = reduce_max(axes = var_9113_axes_0, keep_dims = var_9112, x = var_9109_cast_fp16)[name = string("op_9113_cast_fp16")];
            fp16 var_9114_to_fp16 = const()[name = string("op_9114_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_339_cast_fp16 = maximum(x = var_9113_cast_fp16, y = var_9114_to_fp16)[name = string("s_339_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_339_cast_fp16 = real_div(x = x_379_cast_fp16, y = s_339_cast_fp16)[name = string("y_339_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9117_cast_fp16 = mul(x = y_339_cast_fp16, y = y_339_cast_fp16)[name = string("op_9117_cast_fp16")];
            tensor<int32, [1]> var_9122_axes_0 = const()[name = string("op_9122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9122_keep_dims_0 = const()[name = string("op_9122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9122_cast_fp16 = reduce_mean(axes = var_9122_axes_0, keep_dims = var_9122_keep_dims_0, x = var_9117_cast_fp16)[name = string("op_9122_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9123_cast_fp16 = mul(x = s_339_cast_fp16, y = s_339_cast_fp16)[name = string("op_9123_cast_fp16")];
            fp32 var_9124_epsilon_0 = const()[name = string("op_9124_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9124_cast_fp16 = inverse(epsilon = var_9124_epsilon_0, x = var_9123_cast_fp16)[name = string("op_9124_cast_fp16")];
            fp16 var_9125_to_fp16 = const()[name = string("op_9125_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9126_cast_fp16 = mul(x = var_9124_cast_fp16, y = var_9125_to_fp16)[name = string("op_9126_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9128_cast_fp16 = add(x = var_9122_cast_fp16, y = var_9126_cast_fp16)[name = string("op_9128_cast_fp16")];
            fp32 var_9129_epsilon_0 = const()[name = string("op_9129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9129_cast_fp16 = rsqrt(epsilon = var_9129_epsilon_0, x = var_9128_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_297_cast_fp16 = mul(x = y_339_cast_fp16, y = var_9129_cast_fp16)[name = string("out_297_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120104448)))];
            tensor<fp16, [1, 512, 512]> input_253_cast_fp16 = mul(x = out_297_cast_fp16, y = tm_layers_21_input_layernorm_weight_to_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120105536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120629888))))[name = string("tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9139 = const()[name = string("op_9139"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_9140_cast_fp16 = reshape(shape = var_9139, x = linear_190_cast_fp16)[name = string("op_9140_cast_fp16")];
            tensor<int32, [4]> x_381_perm_0 = const()[name = string("x_381_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120632000))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120894208))))[name = string("tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9151 = const()[name = string("op_9151"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_9152_cast_fp16 = reshape(shape = var_9151, x = linear_191_cast_fp16)[name = string("op_9152_cast_fp16")];
            tensor<int32, [4]> x_383_perm_0 = const()[name = string("x_383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120895296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121157504))))[name = string("tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_9163 = const()[name = string("op_9163"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_9164_cast_fp16 = reshape(shape = var_9163, x = linear_192_cast_fp16)[name = string("op_9164_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_381_cast_fp16 = transpose(perm = x_381_perm_0, x = var_9140_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 4, 512, 256]> var_9168_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_9168_cast_fp16")];
            bool var_9171 = const()[name = string("op_9171"), val = bool(true)];
            tensor<int32, [1]> var_9172_axes_0 = const()[name = string("op_9172_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_9172_cast_fp16 = reduce_max(axes = var_9172_axes_0, keep_dims = var_9171, x = var_9168_cast_fp16)[name = string("op_9172_cast_fp16")];
            fp16 var_9173_to_fp16 = const()[name = string("op_9173_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_341_cast_fp16 = maximum(x = var_9172_cast_fp16, y = var_9173_to_fp16)[name = string("s_341_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_341_cast_fp16 = real_div(x = x_381_cast_fp16, y = s_341_cast_fp16)[name = string("y_341_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_9176_cast_fp16 = mul(x = y_341_cast_fp16, y = y_341_cast_fp16)[name = string("op_9176_cast_fp16")];
            tensor<int32, [1]> var_9181_axes_0 = const()[name = string("op_9181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9181_keep_dims_0 = const()[name = string("op_9181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_9181_cast_fp16 = reduce_mean(axes = var_9181_axes_0, keep_dims = var_9181_keep_dims_0, x = var_9176_cast_fp16)[name = string("op_9181_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_9182_cast_fp16 = mul(x = s_341_cast_fp16, y = s_341_cast_fp16)[name = string("op_9182_cast_fp16")];
            fp32 var_9183_epsilon_0 = const()[name = string("op_9183_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_9183_cast_fp16 = inverse(epsilon = var_9183_epsilon_0, x = var_9182_cast_fp16)[name = string("op_9183_cast_fp16")];
            fp16 var_9184_to_fp16 = const()[name = string("op_9184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_9185_cast_fp16 = mul(x = var_9183_cast_fp16, y = var_9184_to_fp16)[name = string("op_9185_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_9187_cast_fp16 = add(x = var_9181_cast_fp16, y = var_9185_cast_fp16)[name = string("op_9187_cast_fp16")];
            fp32 var_9188_epsilon_0 = const()[name = string("op_9188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_9188_cast_fp16 = rsqrt(epsilon = var_9188_epsilon_0, x = var_9187_cast_fp16)[name = string("op_9188_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_299_cast_fp16 = mul(x = y_341_cast_fp16, y = var_9188_cast_fp16)[name = string("out_299_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_85_cast_fp16 = mul(x = out_299_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_85_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_383_cast_fp16 = transpose(perm = x_383_perm_0, x = var_9152_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2, 512, 256]> var_9191_cast_fp16 = abs(x = x_383_cast_fp16)[name = string("op_9191_cast_fp16")];
            bool var_9194 = const()[name = string("op_9194"), val = bool(true)];
            tensor<int32, [1]> var_9195_axes_0 = const()[name = string("op_9195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_9195_cast_fp16 = reduce_max(axes = var_9195_axes_0, keep_dims = var_9194, x = var_9191_cast_fp16)[name = string("op_9195_cast_fp16")];
            fp16 var_9196_to_fp16 = const()[name = string("op_9196_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_343_cast_fp16 = maximum(x = var_9195_cast_fp16, y = var_9196_to_fp16)[name = string("s_343_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_343_cast_fp16 = real_div(x = x_383_cast_fp16, y = s_343_cast_fp16)[name = string("y_343_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_9199_cast_fp16 = mul(x = y_343_cast_fp16, y = y_343_cast_fp16)[name = string("op_9199_cast_fp16")];
            tensor<int32, [1]> var_9204_axes_0 = const()[name = string("op_9204_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9204_keep_dims_0 = const()[name = string("op_9204_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_9204_cast_fp16 = reduce_mean(axes = var_9204_axes_0, keep_dims = var_9204_keep_dims_0, x = var_9199_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_9205_cast_fp16 = mul(x = s_343_cast_fp16, y = s_343_cast_fp16)[name = string("op_9205_cast_fp16")];
            fp32 var_9206_epsilon_0 = const()[name = string("op_9206_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_9206_cast_fp16 = inverse(epsilon = var_9206_epsilon_0, x = var_9205_cast_fp16)[name = string("op_9206_cast_fp16")];
            fp16 var_9207_to_fp16 = const()[name = string("op_9207_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_9208_cast_fp16 = mul(x = var_9206_cast_fp16, y = var_9207_to_fp16)[name = string("op_9208_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_9210_cast_fp16 = add(x = var_9204_cast_fp16, y = var_9208_cast_fp16)[name = string("op_9210_cast_fp16")];
            fp32 var_9211_epsilon_0 = const()[name = string("op_9211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_9211_cast_fp16 = rsqrt(epsilon = var_9211_epsilon_0, x = var_9210_cast_fp16)[name = string("op_9211_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_301_cast_fp16 = mul(x = y_343_cast_fp16, y = var_9211_cast_fp16)[name = string("out_301_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_127_cast_fp16 = mul(x = out_301_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_127_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_9164_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2, 512, 256]> var_9214_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_9214_cast_fp16")];
            bool var_9217 = const()[name = string("op_9217"), val = bool(true)];
            tensor<int32, [1]> var_9218_axes_0 = const()[name = string("op_9218_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_9218_cast_fp16 = reduce_max(axes = var_9218_axes_0, keep_dims = var_9217, x = var_9214_cast_fp16)[name = string("op_9218_cast_fp16")];
            fp16 var_9219_to_fp16 = const()[name = string("op_9219_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_345_cast_fp16 = maximum(x = var_9218_cast_fp16, y = var_9219_to_fp16)[name = string("s_345_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_345_cast_fp16 = real_div(x = x_385_cast_fp16, y = s_345_cast_fp16)[name = string("y_345_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_9222_cast_fp16 = mul(x = y_345_cast_fp16, y = y_345_cast_fp16)[name = string("op_9222_cast_fp16")];
            tensor<int32, [1]> var_9227_axes_0 = const()[name = string("op_9227_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9227_keep_dims_0 = const()[name = string("op_9227_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_9227_cast_fp16 = reduce_mean(axes = var_9227_axes_0, keep_dims = var_9227_keep_dims_0, x = var_9222_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_9228_cast_fp16 = mul(x = s_345_cast_fp16, y = s_345_cast_fp16)[name = string("op_9228_cast_fp16")];
            fp32 var_9229_epsilon_0 = const()[name = string("op_9229_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_9229_cast_fp16 = inverse(epsilon = var_9229_epsilon_0, x = var_9228_cast_fp16)[name = string("op_9229_cast_fp16")];
            fp16 var_9230_to_fp16 = const()[name = string("op_9230_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_9231_cast_fp16 = mul(x = var_9229_cast_fp16, y = var_9230_to_fp16)[name = string("op_9231_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_9233_cast_fp16 = add(x = var_9227_cast_fp16, y = var_9231_cast_fp16)[name = string("op_9233_cast_fp16")];
            fp32 var_9234_epsilon_0 = const()[name = string("op_9234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_9234_cast_fp16 = rsqrt(epsilon = var_9234_epsilon_0, x = var_9233_cast_fp16)[name = string("op_9234_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_85_cast_fp16 = mul(x = y_345_cast_fp16, y = var_9234_cast_fp16)[name = string("v_85_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_9236_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9236_cast_fp16")];
            tensor<int32, [4]> var_9241_begin_0 = const()[name = string("op_9241_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9241_end_0 = const()[name = string("op_9241_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_9241_end_mask_0 = const()[name = string("op_9241_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_9241_cast_fp16 = slice_by_index(begin = var_9241_begin_0, end = var_9241_end_0, end_mask = var_9241_end_mask_0, x = q_85_cast_fp16)[name = string("op_9241_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_9242_cast_fp16 = mul(x = var_9241_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_9242_cast_fp16")];
            tensor<int32, [4]> var_9247_begin_0 = const()[name = string("op_9247_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9247_end_0 = const()[name = string("op_9247_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_9247_end_mask_0 = const()[name = string("op_9247_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_9247_cast_fp16 = slice_by_index(begin = var_9247_begin_0, end = var_9247_end_0, end_mask = var_9247_end_mask_0, x = q_85_cast_fp16)[name = string("op_9247_cast_fp16")];
            int32 var_9249 = const()[name = string("op_9249"), val = int32(-1)];
            bool var_9250_interleave_0 = const()[name = string("op_9250_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_9250_cast_fp16 = concat(axis = var_9249, interleave = var_9250_interleave_0, values = (var_9242_cast_fp16, var_9247_cast_fp16))[name = string("op_9250_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_9251_cast_fp16 = mul(x = var_9250_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9251_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_87_cast_fp16 = add(x = var_9236_cast_fp16, y = var_9251_cast_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_9254_cast_fp16 = mul(x = k_127_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9254_cast_fp16")];
            tensor<int32, [4]> var_9259_begin_0 = const()[name = string("op_9259_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9259_end_0 = const()[name = string("op_9259_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_9259_end_mask_0 = const()[name = string("op_9259_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_9259_cast_fp16 = slice_by_index(begin = var_9259_begin_0, end = var_9259_end_0, end_mask = var_9259_end_mask_0, x = k_127_cast_fp16)[name = string("op_9259_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_9260_cast_fp16 = mul(x = var_9259_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_9260_cast_fp16")];
            tensor<int32, [4]> var_9265_begin_0 = const()[name = string("op_9265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9265_end_0 = const()[name = string("op_9265_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_9265_end_mask_0 = const()[name = string("op_9265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_9265_cast_fp16 = slice_by_index(begin = var_9265_begin_0, end = var_9265_end_0, end_mask = var_9265_end_mask_0, x = k_127_cast_fp16)[name = string("op_9265_cast_fp16")];
            int32 var_9267 = const()[name = string("op_9267"), val = int32(-1)];
            bool var_9268_interleave_0 = const()[name = string("op_9268_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_9268_cast_fp16 = concat(axis = var_9267, interleave = var_9268_interleave_0, values = (var_9260_cast_fp16, var_9265_cast_fp16))[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_9269_cast_fp16 = mul(x = var_9268_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_129_cast_fp16 = add(x = var_9254_cast_fp16, y = var_9269_cast_fp16)[name = string("k_129_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_42_reps_0 = const()[name = string("tile_42_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = k_129_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 1, 512, 256]> tile_42_cast_fp16 = tile(reps = tile_42_reps_0, x = transpose_84_cast_fp16)[name = string("tile_42_cast_fp16")];
            tensor<int32, [5]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_84_cast_fp16 = reshape(shape = concat_84, x = tile_42_cast_fp16)[name = string("reshape_84_cast_fp16")];
            tensor<int32, [5]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = reshape_84_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 1, 512, 256]> reshape_85_cast_fp16 = reshape(shape = concat_85, x = transpose_85_cast_fp16)[name = string("reshape_85_cast_fp16")];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_43_reps_0 = const()[name = string("tile_43_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_86_cast_fp16 = transpose(perm = transpose_86_perm_0, x = v_85_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 1, 512, 256]> tile_43_cast_fp16 = tile(reps = tile_43_reps_0, x = transpose_86_cast_fp16)[name = string("tile_43_cast_fp16")];
            tensor<int32, [5]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_86_cast_fp16 = reshape(shape = concat_86, x = tile_43_cast_fp16)[name = string("reshape_86_cast_fp16")];
            tensor<int32, [5]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_87_cast_fp16 = transpose(perm = transpose_87_perm_0, x = reshape_86_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 1, 512, 256]> reshape_87_cast_fp16 = reshape(shape = concat_87, x = transpose_87_cast_fp16)[name = string("reshape_87_cast_fp16")];
            tensor<int32, [4]> v_87_perm_0 = const()[name = string("v_87_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9283_transpose_x_0 = const()[name = string("op_9283_transpose_x_0"), val = bool(false)];
            bool var_9283_transpose_y_0 = const()[name = string("op_9283_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_117_cast_fp16 = transpose(perm = transpose_117_perm_0, x = reshape_85_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 4, 512, 512]> var_9283_cast_fp16 = matmul(transpose_x = var_9283_transpose_x_0, transpose_y = var_9283_transpose_y_0, x = q_87_cast_fp16, y = transpose_117_cast_fp16)[name = string("op_9283_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_9285_cast_fp16 = add(x = var_9283_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9285_cast_fp16")];
            int32 var_9286 = const()[name = string("op_9286"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_345_cast_fp16 = softmax(axis = var_9286, x = var_9285_cast_fp16)[name = string("w_345_cast_fp16")];
            bool var_9289_transpose_x_0 = const()[name = string("op_9289_transpose_x_0"), val = bool(false)];
            bool var_9289_transpose_y_0 = const()[name = string("op_9289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_87_cast_fp16 = transpose(perm = v_87_perm_0, x = reshape_87_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 4, 512, 256]> var_9289_cast_fp16 = matmul(transpose_x = var_9289_transpose_x_0, transpose_y = var_9289_transpose_y_0, x = w_345_cast_fp16, y = v_87_cast_fp16)[name = string("op_9289_cast_fp16")];
            tensor<int32, [4]> var_9292_perm_0 = const()[name = string("op_9292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9296 = const()[name = string("op_9296"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_9292_cast_fp16 = transpose(perm = var_9292_perm_0, x = var_9289_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 512, 1024]> input_255_cast_fp16 = reshape(shape = var_9296, x = var_9292_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121158592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121682944))))[name = string("tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9301_cast_fp16 = abs(x = linear_193_cast_fp16)[name = string("op_9301_cast_fp16")];
            bool var_9304 = const()[name = string("op_9304"), val = bool(true)];
            tensor<int32, [1]> var_9305_axes_0 = const()[name = string("op_9305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9305_cast_fp16 = reduce_max(axes = var_9305_axes_0, keep_dims = var_9304, x = var_9301_cast_fp16)[name = string("op_9305_cast_fp16")];
            fp16 var_9306_to_fp16 = const()[name = string("op_9306_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_347_cast_fp16 = maximum(x = var_9305_cast_fp16, y = var_9306_to_fp16)[name = string("s_347_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_347_cast_fp16 = real_div(x = linear_193_cast_fp16, y = s_347_cast_fp16)[name = string("y_347_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9309_cast_fp16 = mul(x = y_347_cast_fp16, y = y_347_cast_fp16)[name = string("op_9309_cast_fp16")];
            tensor<int32, [1]> var_9314_axes_0 = const()[name = string("op_9314_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9314_keep_dims_0 = const()[name = string("op_9314_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9314_cast_fp16 = reduce_mean(axes = var_9314_axes_0, keep_dims = var_9314_keep_dims_0, x = var_9309_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9315_cast_fp16 = mul(x = s_347_cast_fp16, y = s_347_cast_fp16)[name = string("op_9315_cast_fp16")];
            fp32 var_9316_epsilon_0 = const()[name = string("op_9316_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9316_cast_fp16 = inverse(epsilon = var_9316_epsilon_0, x = var_9315_cast_fp16)[name = string("op_9316_cast_fp16")];
            fp16 var_9317_to_fp16 = const()[name = string("op_9317_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9318_cast_fp16 = mul(x = var_9316_cast_fp16, y = var_9317_to_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9320_cast_fp16 = add(x = var_9314_cast_fp16, y = var_9318_cast_fp16)[name = string("op_9320_cast_fp16")];
            fp32 var_9321_epsilon_0 = const()[name = string("op_9321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9321_cast_fp16 = rsqrt(epsilon = var_9321_epsilon_0, x = var_9320_cast_fp16)[name = string("op_9321_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_303_cast_fp16 = mul(x = y_347_cast_fp16, y = var_9321_cast_fp16)[name = string("out_303_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121684032)))];
            tensor<fp16, [1, 512, 512]> var_9323_cast_fp16 = mul(x = out_303_cast_fp16, y = tm_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("op_9323_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_389_cast_fp16 = add(x = x_379_cast_fp16, y = var_9323_cast_fp16)[name = string("x_389_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9326_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_9326_cast_fp16")];
            bool var_9329 = const()[name = string("op_9329"), val = bool(true)];
            tensor<int32, [1]> var_9330_axes_0 = const()[name = string("op_9330_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9330_cast_fp16 = reduce_max(axes = var_9330_axes_0, keep_dims = var_9329, x = var_9326_cast_fp16)[name = string("op_9330_cast_fp16")];
            fp16 var_9331_to_fp16 = const()[name = string("op_9331_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_349_cast_fp16 = maximum(x = var_9330_cast_fp16, y = var_9331_to_fp16)[name = string("s_349_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_349_cast_fp16 = real_div(x = x_389_cast_fp16, y = s_349_cast_fp16)[name = string("y_349_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9334_cast_fp16 = mul(x = y_349_cast_fp16, y = y_349_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [1]> var_9339_axes_0 = const()[name = string("op_9339_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9339_keep_dims_0 = const()[name = string("op_9339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9339_cast_fp16 = reduce_mean(axes = var_9339_axes_0, keep_dims = var_9339_keep_dims_0, x = var_9334_cast_fp16)[name = string("op_9339_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9340_cast_fp16 = mul(x = s_349_cast_fp16, y = s_349_cast_fp16)[name = string("op_9340_cast_fp16")];
            fp32 var_9341_epsilon_0 = const()[name = string("op_9341_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9341_cast_fp16 = inverse(epsilon = var_9341_epsilon_0, x = var_9340_cast_fp16)[name = string("op_9341_cast_fp16")];
            fp16 var_9342_to_fp16 = const()[name = string("op_9342_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9343_cast_fp16 = mul(x = var_9341_cast_fp16, y = var_9342_to_fp16)[name = string("op_9343_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9345_cast_fp16 = add(x = var_9339_cast_fp16, y = var_9343_cast_fp16)[name = string("op_9345_cast_fp16")];
            fp32 var_9346_epsilon_0 = const()[name = string("op_9346_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9346_cast_fp16 = rsqrt(epsilon = var_9346_epsilon_0, x = var_9345_cast_fp16)[name = string("op_9346_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_305_cast_fp16 = mul(x = y_349_cast_fp16, y = var_9346_cast_fp16)[name = string("out_305_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121685120)))];
            tensor<fp16, [1, 512, 512]> input_257_cast_fp16 = mul(x = out_305_cast_fp16, y = tm_layers_21_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121686208))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122734848))))[name = string("tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_194_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_194_cast_fp16")];
            string var_9353_mode_0 = const()[name = string("op_9353_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_9353_cast_fp16 = gelu(mode = var_9353_mode_0, x = linear_194_cast_fp16)[name = string("op_9353_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122739008))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123787648))))[name = string("tm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_195_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_259_cast_fp16 = mul(x = var_9353_cast_fp16, y = linear_195_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123791808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124840448))))[name = string("tm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9361_cast_fp16 = abs(x = linear_196_cast_fp16)[name = string("op_9361_cast_fp16")];
            bool var_9364 = const()[name = string("op_9364"), val = bool(true)];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9365_cast_fp16 = reduce_max(axes = var_9365_axes_0, keep_dims = var_9364, x = var_9361_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9366_to_fp16 = const()[name = string("op_9366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_351_cast_fp16 = maximum(x = var_9365_cast_fp16, y = var_9366_to_fp16)[name = string("s_351_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_351_cast_fp16 = real_div(x = linear_196_cast_fp16, y = s_351_cast_fp16)[name = string("y_351_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9369_cast_fp16 = mul(x = y_351_cast_fp16, y = y_351_cast_fp16)[name = string("op_9369_cast_fp16")];
            tensor<int32, [1]> var_9374_axes_0 = const()[name = string("op_9374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9374_keep_dims_0 = const()[name = string("op_9374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9374_cast_fp16 = reduce_mean(axes = var_9374_axes_0, keep_dims = var_9374_keep_dims_0, x = var_9369_cast_fp16)[name = string("op_9374_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9375_cast_fp16 = mul(x = s_351_cast_fp16, y = s_351_cast_fp16)[name = string("op_9375_cast_fp16")];
            fp32 var_9376_epsilon_0 = const()[name = string("op_9376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9376_cast_fp16 = inverse(epsilon = var_9376_epsilon_0, x = var_9375_cast_fp16)[name = string("op_9376_cast_fp16")];
            fp16 var_9377_to_fp16 = const()[name = string("op_9377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9378_cast_fp16 = mul(x = var_9376_cast_fp16, y = var_9377_to_fp16)[name = string("op_9378_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9380_cast_fp16 = add(x = var_9374_cast_fp16, y = var_9378_cast_fp16)[name = string("op_9380_cast_fp16")];
            fp32 var_9381_epsilon_0 = const()[name = string("op_9381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9381_cast_fp16 = rsqrt(epsilon = var_9381_epsilon_0, x = var_9380_cast_fp16)[name = string("op_9381_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_307_cast_fp16 = mul(x = y_351_cast_fp16, y = var_9381_cast_fp16)[name = string("out_307_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124841536)))];
            tensor<fp16, [1, 512, 512]> var_9383_cast_fp16 = mul(x = out_307_cast_fp16, y = tm_layers_21_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9383_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_261_cast_fp16 = add(x = x_389_cast_fp16, y = var_9383_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124842624))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125104832))))[name = string("tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_197_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_197_cast_fp16")];
            string x_393_mode_0 = const()[name = string("x_393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_393_cast_fp16 = gelu(mode = x_393_mode_0, x = linear_197_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<int32, [4]> var_9403_begin_0 = const()[name = string("op_9403_begin_0"), val = tensor<int32, [4]>([0, 0, 21, 0])];
            tensor<int32, [4]> var_9403_end_0 = const()[name = string("op_9403_end_0"), val = tensor<int32, [4]>([1, 512, 22, 512])];
            tensor<bool, [4]> var_9403_end_mask_0 = const()[name = string("op_9403_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9403_squeeze_mask_0 = const()[name = string("op_9403_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_9403_cast_fp16 = slice_by_index(begin = var_9403_begin_0, end = var_9403_end_0, end_mask = var_9403_end_mask_0, squeeze_mask = var_9403_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9403_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_263_cast_fp16 = mul(x = x_393_cast_fp16, y = var_9403_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125105920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125368128))))[name = string("tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9413_cast_fp16 = abs(x = linear_198_cast_fp16)[name = string("op_9413_cast_fp16")];
            bool var_9416 = const()[name = string("op_9416"), val = bool(true)];
            tensor<int32, [1]> var_9417_axes_0 = const()[name = string("op_9417_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9417_cast_fp16 = reduce_max(axes = var_9417_axes_0, keep_dims = var_9416, x = var_9413_cast_fp16)[name = string("op_9417_cast_fp16")];
            fp16 var_9418_to_fp16 = const()[name = string("op_9418_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_353_cast_fp16 = maximum(x = var_9417_cast_fp16, y = var_9418_to_fp16)[name = string("s_353_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_353_cast_fp16 = real_div(x = linear_198_cast_fp16, y = s_353_cast_fp16)[name = string("y_353_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9421_cast_fp16 = mul(x = y_353_cast_fp16, y = y_353_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<int32, [1]> var_9426_axes_0 = const()[name = string("op_9426_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9426_keep_dims_0 = const()[name = string("op_9426_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9426_cast_fp16 = reduce_mean(axes = var_9426_axes_0, keep_dims = var_9426_keep_dims_0, x = var_9421_cast_fp16)[name = string("op_9426_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9427_cast_fp16 = mul(x = s_353_cast_fp16, y = s_353_cast_fp16)[name = string("op_9427_cast_fp16")];
            fp32 var_9428_epsilon_0 = const()[name = string("op_9428_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9428_cast_fp16 = inverse(epsilon = var_9428_epsilon_0, x = var_9427_cast_fp16)[name = string("op_9428_cast_fp16")];
            fp16 var_9429_to_fp16 = const()[name = string("op_9429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9430_cast_fp16 = mul(x = var_9428_cast_fp16, y = var_9429_to_fp16)[name = string("op_9430_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9432_cast_fp16 = add(x = var_9426_cast_fp16, y = var_9430_cast_fp16)[name = string("op_9432_cast_fp16")];
            fp32 var_9433_epsilon_0 = const()[name = string("op_9433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9433_cast_fp16 = rsqrt(epsilon = var_9433_epsilon_0, x = var_9432_cast_fp16)[name = string("op_9433_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_309_cast_fp16 = mul(x = y_353_cast_fp16, y = var_9433_cast_fp16)[name = string("out_309_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125369216)))];
            tensor<fp16, [1, 512, 512]> var_9435_cast_fp16 = mul(x = out_309_cast_fp16, y = tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9435_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_43_cast_fp16 = add(x = input_261_cast_fp16, y = var_9435_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<fp16, [1]> tm_layers_21_layer_scalar_to_fp16 = const()[name = string("tm_layers_21_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e8p-1])];
            tensor<fp16, [1, 512, 512]> x_397_cast_fp16 = mul(x = h_43_cast_fp16, y = tm_layers_21_layer_scalar_to_fp16)[name = string("x_397_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9439_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_9439_cast_fp16")];
            bool var_9442 = const()[name = string("op_9442"), val = bool(true)];
            tensor<int32, [1]> var_9443_axes_0 = const()[name = string("op_9443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9443_cast_fp16 = reduce_max(axes = var_9443_axes_0, keep_dims = var_9442, x = var_9439_cast_fp16)[name = string("op_9443_cast_fp16")];
            fp16 var_9444_to_fp16 = const()[name = string("op_9444_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_355_cast_fp16 = maximum(x = var_9443_cast_fp16, y = var_9444_to_fp16)[name = string("s_355_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_355_cast_fp16 = real_div(x = x_397_cast_fp16, y = s_355_cast_fp16)[name = string("y_355_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9447_cast_fp16 = mul(x = y_355_cast_fp16, y = y_355_cast_fp16)[name = string("op_9447_cast_fp16")];
            tensor<int32, [1]> var_9452_axes_0 = const()[name = string("op_9452_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9452_keep_dims_0 = const()[name = string("op_9452_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9452_cast_fp16 = reduce_mean(axes = var_9452_axes_0, keep_dims = var_9452_keep_dims_0, x = var_9447_cast_fp16)[name = string("op_9452_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9453_cast_fp16 = mul(x = s_355_cast_fp16, y = s_355_cast_fp16)[name = string("op_9453_cast_fp16")];
            fp32 var_9454_epsilon_0 = const()[name = string("op_9454_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9454_cast_fp16 = inverse(epsilon = var_9454_epsilon_0, x = var_9453_cast_fp16)[name = string("op_9454_cast_fp16")];
            fp16 var_9455_to_fp16 = const()[name = string("op_9455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9456_cast_fp16 = mul(x = var_9454_cast_fp16, y = var_9455_to_fp16)[name = string("op_9456_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9458_cast_fp16 = add(x = var_9452_cast_fp16, y = var_9456_cast_fp16)[name = string("op_9458_cast_fp16")];
            fp32 var_9459_epsilon_0 = const()[name = string("op_9459_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9459_cast_fp16 = rsqrt(epsilon = var_9459_epsilon_0, x = var_9458_cast_fp16)[name = string("op_9459_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_311_cast_fp16 = mul(x = y_355_cast_fp16, y = var_9459_cast_fp16)[name = string("out_311_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125370304)))];
            tensor<fp16, [1, 512, 512]> input_265_cast_fp16 = mul(x = out_311_cast_fp16, y = tm_layers_22_input_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125371392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125895744))))[name = string("tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 1024]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<int32, [4]> var_9469 = const()[name = string("op_9469"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 512, 4, 256]> var_9470_cast_fp16 = reshape(shape = var_9469, x = linear_199_cast_fp16)[name = string("op_9470_cast_fp16")];
            tensor<int32, [4]> x_399_perm_0 = const()[name = string("x_399_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125897856))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126160064))))[name = string("tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<int32, [4]> var_9481 = const()[name = string("op_9481"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_9482_cast_fp16 = reshape(shape = var_9481, x = linear_200_cast_fp16)[name = string("op_9482_cast_fp16")];
            tensor<int32, [4]> x_401_perm_0 = const()[name = string("x_401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126161152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126423360))))[name = string("tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<int32, [4]> var_9493 = const()[name = string("op_9493"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 512, 2, 256]> var_9494_cast_fp16 = reshape(shape = var_9493, x = linear_201_cast_fp16)[name = string("op_9494_cast_fp16")];
            tensor<int32, [4]> x_403_perm_0 = const()[name = string("x_403_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 256]> x_399_cast_fp16 = transpose(perm = x_399_perm_0, x = var_9470_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 4, 512, 256]> var_9498_cast_fp16 = abs(x = x_399_cast_fp16)[name = string("op_9498_cast_fp16")];
            bool var_9501 = const()[name = string("op_9501"), val = bool(true)];
            tensor<int32, [1]> var_9502_axes_0 = const()[name = string("op_9502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_9502_cast_fp16 = reduce_max(axes = var_9502_axes_0, keep_dims = var_9501, x = var_9498_cast_fp16)[name = string("op_9502_cast_fp16")];
            fp16 var_9503_to_fp16 = const()[name = string("op_9503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_357_cast_fp16 = maximum(x = var_9502_cast_fp16, y = var_9503_to_fp16)[name = string("s_357_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> y_357_cast_fp16 = real_div(x = x_399_cast_fp16, y = s_357_cast_fp16)[name = string("y_357_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_9506_cast_fp16 = mul(x = y_357_cast_fp16, y = y_357_cast_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9511_axes_0 = const()[name = string("op_9511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9511_keep_dims_0 = const()[name = string("op_9511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_9511_cast_fp16 = reduce_mean(axes = var_9511_axes_0, keep_dims = var_9511_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_9512_cast_fp16 = mul(x = s_357_cast_fp16, y = s_357_cast_fp16)[name = string("op_9512_cast_fp16")];
            fp32 var_9513_epsilon_0 = const()[name = string("op_9513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_9513_cast_fp16 = inverse(epsilon = var_9513_epsilon_0, x = var_9512_cast_fp16)[name = string("op_9513_cast_fp16")];
            fp16 var_9514_to_fp16 = const()[name = string("op_9514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_9515_cast_fp16 = mul(x = var_9513_cast_fp16, y = var_9514_to_fp16)[name = string("op_9515_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_9517_cast_fp16 = add(x = var_9511_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9517_cast_fp16")];
            fp32 var_9518_epsilon_0 = const()[name = string("op_9518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_9518_cast_fp16 = rsqrt(epsilon = var_9518_epsilon_0, x = var_9517_cast_fp16)[name = string("op_9518_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> out_313_cast_fp16 = mul(x = y_357_cast_fp16, y = var_9518_cast_fp16)[name = string("out_313_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_89_cast_fp16 = mul(x = out_313_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_89_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_401_cast_fp16 = transpose(perm = x_401_perm_0, x = var_9482_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 512, 256]> var_9521_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_9521_cast_fp16")];
            bool var_9524 = const()[name = string("op_9524"), val = bool(true)];
            tensor<int32, [1]> var_9525_axes_0 = const()[name = string("op_9525_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_9525_cast_fp16 = reduce_max(axes = var_9525_axes_0, keep_dims = var_9524, x = var_9521_cast_fp16)[name = string("op_9525_cast_fp16")];
            fp16 var_9526_to_fp16 = const()[name = string("op_9526_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_359_cast_fp16 = maximum(x = var_9525_cast_fp16, y = var_9526_to_fp16)[name = string("s_359_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_359_cast_fp16 = real_div(x = x_401_cast_fp16, y = s_359_cast_fp16)[name = string("y_359_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_9529_cast_fp16 = mul(x = y_359_cast_fp16, y = y_359_cast_fp16)[name = string("op_9529_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9529_cast_fp16)[name = string("op_9534_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_9535_cast_fp16 = mul(x = s_359_cast_fp16, y = s_359_cast_fp16)[name = string("op_9535_cast_fp16")];
            fp32 var_9536_epsilon_0 = const()[name = string("op_9536_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_9536_cast_fp16 = inverse(epsilon = var_9536_epsilon_0, x = var_9535_cast_fp16)[name = string("op_9536_cast_fp16")];
            fp16 var_9537_to_fp16 = const()[name = string("op_9537_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_9538_cast_fp16 = mul(x = var_9536_cast_fp16, y = var_9537_to_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_9540_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9538_cast_fp16)[name = string("op_9540_cast_fp16")];
            fp32 var_9541_epsilon_0 = const()[name = string("op_9541_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_9541_cast_fp16 = rsqrt(epsilon = var_9541_epsilon_0, x = var_9540_cast_fp16)[name = string("op_9541_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> out_315_cast_fp16 = mul(x = y_359_cast_fp16, y = var_9541_cast_fp16)[name = string("out_315_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_133_cast_fp16 = mul(x = out_315_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_133_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> x_403_cast_fp16 = transpose(perm = x_403_perm_0, x = var_9494_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 512, 256]> var_9544_cast_fp16 = abs(x = x_403_cast_fp16)[name = string("op_9544_cast_fp16")];
            bool var_9547 = const()[name = string("op_9547"), val = bool(true)];
            tensor<int32, [1]> var_9548_axes_0 = const()[name = string("op_9548_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 512, 1]> var_9548_cast_fp16 = reduce_max(axes = var_9548_axes_0, keep_dims = var_9547, x = var_9544_cast_fp16)[name = string("op_9548_cast_fp16")];
            fp16 var_9549_to_fp16 = const()[name = string("op_9549_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 512, 1]> s_361_cast_fp16 = maximum(x = var_9548_cast_fp16, y = var_9549_to_fp16)[name = string("s_361_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> y_361_cast_fp16 = real_div(x = x_403_cast_fp16, y = s_361_cast_fp16)[name = string("y_361_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_9552_cast_fp16 = mul(x = y_361_cast_fp16, y = y_361_cast_fp16)[name = string("op_9552_cast_fp16")];
            tensor<int32, [1]> var_9557_axes_0 = const()[name = string("op_9557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9557_keep_dims_0 = const()[name = string("op_9557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 512, 1]> var_9557_cast_fp16 = reduce_mean(axes = var_9557_axes_0, keep_dims = var_9557_keep_dims_0, x = var_9552_cast_fp16)[name = string("op_9557_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_9558_cast_fp16 = mul(x = s_361_cast_fp16, y = s_361_cast_fp16)[name = string("op_9558_cast_fp16")];
            fp32 var_9559_epsilon_0 = const()[name = string("op_9559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 512, 1]> var_9559_cast_fp16 = inverse(epsilon = var_9559_epsilon_0, x = var_9558_cast_fp16)[name = string("op_9559_cast_fp16")];
            fp16 var_9560_to_fp16 = const()[name = string("op_9560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 512, 1]> var_9561_cast_fp16 = mul(x = var_9559_cast_fp16, y = var_9560_to_fp16)[name = string("op_9561_cast_fp16")];
            tensor<fp16, [1, 2, 512, 1]> var_9563_cast_fp16 = add(x = var_9557_cast_fp16, y = var_9561_cast_fp16)[name = string("op_9563_cast_fp16")];
            fp32 var_9564_epsilon_0 = const()[name = string("op_9564_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 512, 1]> var_9564_cast_fp16 = rsqrt(epsilon = var_9564_epsilon_0, x = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> v_89_cast_fp16 = mul(x = y_361_cast_fp16, y = var_9564_cast_fp16)[name = string("v_89_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_9566_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9566_cast_fp16")];
            tensor<int32, [4]> var_9571_begin_0 = const()[name = string("op_9571_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9571_end_0 = const()[name = string("op_9571_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_9571_end_mask_0 = const()[name = string("op_9571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 128]> var_9571_cast_fp16 = slice_by_index(begin = var_9571_begin_0, end = var_9571_end_0, end_mask = var_9571_end_mask_0, x = q_89_cast_fp16)[name = string("op_9571_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 128]> var_9572_cast_fp16 = mul(x = var_9571_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_9572_cast_fp16")];
            tensor<int32, [4]> var_9577_begin_0 = const()[name = string("op_9577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9577_end_0 = const()[name = string("op_9577_end_0"), val = tensor<int32, [4]>([1, 4, 512, 128])];
            tensor<bool, [4]> var_9577_end_mask_0 = const()[name = string("op_9577_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 128]> var_9577_cast_fp16 = slice_by_index(begin = var_9577_begin_0, end = var_9577_end_0, end_mask = var_9577_end_mask_0, x = q_89_cast_fp16)[name = string("op_9577_cast_fp16")];
            int32 var_9579 = const()[name = string("op_9579"), val = int32(-1)];
            bool var_9580_interleave_0 = const()[name = string("op_9580_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> var_9580_cast_fp16 = concat(axis = var_9579, interleave = var_9580_interleave_0, values = (var_9572_cast_fp16, var_9577_cast_fp16))[name = string("op_9580_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> var_9581_cast_fp16 = mul(x = var_9580_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9581_cast_fp16")];
            tensor<fp16, [1, 4, 512, 256]> q_91_cast_fp16 = add(x = var_9566_cast_fp16, y = var_9581_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_9584_cast_fp16 = mul(x = k_133_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9584_cast_fp16")];
            tensor<int32, [4]> var_9589_begin_0 = const()[name = string("op_9589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9589_end_0 = const()[name = string("op_9589_end_0"), val = tensor<int32, [4]>([1, 2, 512, 256])];
            tensor<bool, [4]> var_9589_end_mask_0 = const()[name = string("op_9589_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 128]> var_9589_cast_fp16 = slice_by_index(begin = var_9589_begin_0, end = var_9589_end_0, end_mask = var_9589_end_mask_0, x = k_133_cast_fp16)[name = string("op_9589_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 128]> var_9590_cast_fp16 = mul(x = var_9589_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_9590_cast_fp16")];
            tensor<int32, [4]> var_9595_begin_0 = const()[name = string("op_9595_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9595_end_0 = const()[name = string("op_9595_end_0"), val = tensor<int32, [4]>([1, 2, 512, 128])];
            tensor<bool, [4]> var_9595_end_mask_0 = const()[name = string("op_9595_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 128]> var_9595_cast_fp16 = slice_by_index(begin = var_9595_begin_0, end = var_9595_end_0, end_mask = var_9595_end_mask_0, x = k_133_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool var_9598_interleave_0 = const()[name = string("op_9598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 256]> var_9598_cast_fp16 = concat(axis = var_9597, interleave = var_9598_interleave_0, values = (var_9590_cast_fp16, var_9595_cast_fp16))[name = string("op_9598_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> var_9599_cast_fp16 = mul(x = var_9598_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9599_cast_fp16")];
            tensor<fp16, [1, 2, 512, 256]> k_135_cast_fp16 = add(x = var_9584_cast_fp16, y = var_9599_cast_fp16)[name = string("k_135_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_44_reps_0 = const()[name = string("tile_44_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = k_135_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 1, 512, 256]> tile_44_cast_fp16 = tile(reps = tile_44_reps_0, x = transpose_88_cast_fp16)[name = string("tile_44_cast_fp16")];
            tensor<int32, [5]> concat_88 = const()[name = string("concat_88"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_88_cast_fp16 = reshape(shape = concat_88, x = tile_44_cast_fp16)[name = string("reshape_88_cast_fp16")];
            tensor<int32, [5]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = reshape_88_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 1, 512, 256]> reshape_89_cast_fp16 = reshape(shape = concat_89, x = transpose_89_cast_fp16)[name = string("reshape_89_cast_fp16")];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_45_reps_0 = const()[name = string("tile_45_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 512, 256]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = v_89_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 1, 512, 256]> tile_45_cast_fp16 = tile(reps = tile_45_reps_0, x = transpose_90_cast_fp16)[name = string("tile_45_cast_fp16")];
            tensor<int32, [5]> concat_90 = const()[name = string("concat_90"), val = tensor<int32, [5]>([2, 2, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> reshape_90_cast_fp16 = reshape(shape = concat_90, x = tile_45_cast_fp16)[name = string("reshape_90_cast_fp16")];
            tensor<int32, [5]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_91 = const()[name = string("concat_91"), val = tensor<int32, [4]>([-1, 1, 512, 256])];
            tensor<fp16, [2, 2, 1, 512, 256]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = reshape_90_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 1, 512, 256]> reshape_91_cast_fp16 = reshape(shape = concat_91, x = transpose_91_cast_fp16)[name = string("reshape_91_cast_fp16")];
            tensor<int32, [4]> v_91_perm_0 = const()[name = string("v_91_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9613_transpose_x_0 = const()[name = string("op_9613_transpose_x_0"), val = bool(false)];
            bool var_9613_transpose_y_0 = const()[name = string("op_9613_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> transpose_118_cast_fp16 = transpose(perm = transpose_118_perm_0, x = reshape_89_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 4, 512, 512]> var_9613_cast_fp16 = matmul(transpose_x = var_9613_transpose_x_0, transpose_y = var_9613_transpose_y_0, x = q_91_cast_fp16, y = transpose_118_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_9615_cast_fp16 = add(x = var_9613_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9615_cast_fp16")];
            int32 var_9616 = const()[name = string("op_9616"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_361_cast_fp16 = softmax(axis = var_9616, x = var_9615_cast_fp16)[name = string("w_361_cast_fp16")];
            bool var_9619_transpose_x_0 = const()[name = string("op_9619_transpose_x_0"), val = bool(false)];
            bool var_9619_transpose_y_0 = const()[name = string("op_9619_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = reshape_91_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 4, 512, 256]> var_9619_cast_fp16 = matmul(transpose_x = var_9619_transpose_x_0, transpose_y = var_9619_transpose_y_0, x = w_361_cast_fp16, y = v_91_cast_fp16)[name = string("op_9619_cast_fp16")];
            tensor<int32, [4]> var_9622_perm_0 = const()[name = string("op_9622_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9626 = const()[name = string("op_9626"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 512, 4, 256]> var_9622_cast_fp16 = transpose(perm = var_9622_perm_0, x = var_9619_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 512, 1024]> input_267_cast_fp16 = reshape(shape = var_9626, x = var_9622_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126424448))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126948800))))[name = string("tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9631_cast_fp16 = abs(x = linear_202_cast_fp16)[name = string("op_9631_cast_fp16")];
            bool var_9634 = const()[name = string("op_9634"), val = bool(true)];
            tensor<int32, [1]> var_9635_axes_0 = const()[name = string("op_9635_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9635_cast_fp16 = reduce_max(axes = var_9635_axes_0, keep_dims = var_9634, x = var_9631_cast_fp16)[name = string("op_9635_cast_fp16")];
            fp16 var_9636_to_fp16 = const()[name = string("op_9636_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_363_cast_fp16 = maximum(x = var_9635_cast_fp16, y = var_9636_to_fp16)[name = string("s_363_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_363_cast_fp16 = real_div(x = linear_202_cast_fp16, y = s_363_cast_fp16)[name = string("y_363_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9639_cast_fp16 = mul(x = y_363_cast_fp16, y = y_363_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [1]> var_9644_axes_0 = const()[name = string("op_9644_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9644_keep_dims_0 = const()[name = string("op_9644_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9644_cast_fp16 = reduce_mean(axes = var_9644_axes_0, keep_dims = var_9644_keep_dims_0, x = var_9639_cast_fp16)[name = string("op_9644_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9645_cast_fp16 = mul(x = s_363_cast_fp16, y = s_363_cast_fp16)[name = string("op_9645_cast_fp16")];
            fp32 var_9646_epsilon_0 = const()[name = string("op_9646_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9646_cast_fp16 = inverse(epsilon = var_9646_epsilon_0, x = var_9645_cast_fp16)[name = string("op_9646_cast_fp16")];
            fp16 var_9647_to_fp16 = const()[name = string("op_9647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9648_cast_fp16 = mul(x = var_9646_cast_fp16, y = var_9647_to_fp16)[name = string("op_9648_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9650_cast_fp16 = add(x = var_9644_cast_fp16, y = var_9648_cast_fp16)[name = string("op_9650_cast_fp16")];
            fp32 var_9651_epsilon_0 = const()[name = string("op_9651_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9651_cast_fp16 = rsqrt(epsilon = var_9651_epsilon_0, x = var_9650_cast_fp16)[name = string("op_9651_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_317_cast_fp16 = mul(x = y_363_cast_fp16, y = var_9651_cast_fp16)[name = string("out_317_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126949888)))];
            tensor<fp16, [1, 512, 512]> var_9653_cast_fp16 = mul(x = out_317_cast_fp16, y = tm_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_407_cast_fp16 = add(x = x_397_cast_fp16, y = var_9653_cast_fp16)[name = string("x_407_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9656_cast_fp16 = abs(x = x_407_cast_fp16)[name = string("op_9656_cast_fp16")];
            bool var_9659 = const()[name = string("op_9659"), val = bool(true)];
            tensor<int32, [1]> var_9660_axes_0 = const()[name = string("op_9660_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9660_cast_fp16 = reduce_max(axes = var_9660_axes_0, keep_dims = var_9659, x = var_9656_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp16 var_9661_to_fp16 = const()[name = string("op_9661_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_365_cast_fp16 = maximum(x = var_9660_cast_fp16, y = var_9661_to_fp16)[name = string("s_365_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_365_cast_fp16 = real_div(x = x_407_cast_fp16, y = s_365_cast_fp16)[name = string("y_365_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9664_cast_fp16 = mul(x = y_365_cast_fp16, y = y_365_cast_fp16)[name = string("op_9664_cast_fp16")];
            tensor<int32, [1]> var_9669_axes_0 = const()[name = string("op_9669_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9669_keep_dims_0 = const()[name = string("op_9669_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9669_cast_fp16 = reduce_mean(axes = var_9669_axes_0, keep_dims = var_9669_keep_dims_0, x = var_9664_cast_fp16)[name = string("op_9669_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9670_cast_fp16 = mul(x = s_365_cast_fp16, y = s_365_cast_fp16)[name = string("op_9670_cast_fp16")];
            fp32 var_9671_epsilon_0 = const()[name = string("op_9671_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9671_cast_fp16 = inverse(epsilon = var_9671_epsilon_0, x = var_9670_cast_fp16)[name = string("op_9671_cast_fp16")];
            fp16 var_9672_to_fp16 = const()[name = string("op_9672_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9673_cast_fp16 = mul(x = var_9671_cast_fp16, y = var_9672_to_fp16)[name = string("op_9673_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9675_cast_fp16 = add(x = var_9669_cast_fp16, y = var_9673_cast_fp16)[name = string("op_9675_cast_fp16")];
            fp32 var_9676_epsilon_0 = const()[name = string("op_9676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9676_cast_fp16 = rsqrt(epsilon = var_9676_epsilon_0, x = var_9675_cast_fp16)[name = string("op_9676_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_319_cast_fp16 = mul(x = y_365_cast_fp16, y = var_9676_cast_fp16)[name = string("out_319_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126950976)))];
            tensor<fp16, [1, 512, 512]> input_269_cast_fp16 = mul(x = out_319_cast_fp16, y = tm_layers_22_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126952064))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128000704))))[name = string("tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_203_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_203_cast_fp16")];
            string var_9683_mode_0 = const()[name = string("op_9683_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_9683_cast_fp16 = gelu(mode = var_9683_mode_0, x = linear_203_cast_fp16)[name = string("op_9683_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128004864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129053504))))[name = string("tm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_204_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_271_cast_fp16 = mul(x = var_9683_cast_fp16, y = linear_204_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129057664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130106304))))[name = string("tm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9691_cast_fp16 = abs(x = linear_205_cast_fp16)[name = string("op_9691_cast_fp16")];
            bool var_9694 = const()[name = string("op_9694"), val = bool(true)];
            tensor<int32, [1]> var_9695_axes_0 = const()[name = string("op_9695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9695_cast_fp16 = reduce_max(axes = var_9695_axes_0, keep_dims = var_9694, x = var_9691_cast_fp16)[name = string("op_9695_cast_fp16")];
            fp16 var_9696_to_fp16 = const()[name = string("op_9696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_367_cast_fp16 = maximum(x = var_9695_cast_fp16, y = var_9696_to_fp16)[name = string("s_367_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_367_cast_fp16 = real_div(x = linear_205_cast_fp16, y = s_367_cast_fp16)[name = string("y_367_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9699_cast_fp16 = mul(x = y_367_cast_fp16, y = y_367_cast_fp16)[name = string("op_9699_cast_fp16")];
            tensor<int32, [1]> var_9704_axes_0 = const()[name = string("op_9704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9704_keep_dims_0 = const()[name = string("op_9704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9704_cast_fp16 = reduce_mean(axes = var_9704_axes_0, keep_dims = var_9704_keep_dims_0, x = var_9699_cast_fp16)[name = string("op_9704_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9705_cast_fp16 = mul(x = s_367_cast_fp16, y = s_367_cast_fp16)[name = string("op_9705_cast_fp16")];
            fp32 var_9706_epsilon_0 = const()[name = string("op_9706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9706_cast_fp16 = inverse(epsilon = var_9706_epsilon_0, x = var_9705_cast_fp16)[name = string("op_9706_cast_fp16")];
            fp16 var_9707_to_fp16 = const()[name = string("op_9707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9708_cast_fp16 = mul(x = var_9706_cast_fp16, y = var_9707_to_fp16)[name = string("op_9708_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9710_cast_fp16 = add(x = var_9704_cast_fp16, y = var_9708_cast_fp16)[name = string("op_9710_cast_fp16")];
            fp32 var_9711_epsilon_0 = const()[name = string("op_9711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9711_cast_fp16 = rsqrt(epsilon = var_9711_epsilon_0, x = var_9710_cast_fp16)[name = string("op_9711_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_321_cast_fp16 = mul(x = y_367_cast_fp16, y = var_9711_cast_fp16)[name = string("out_321_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130107392)))];
            tensor<fp16, [1, 512, 512]> var_9713_cast_fp16 = mul(x = out_321_cast_fp16, y = tm_layers_22_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9713_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_273_cast_fp16 = add(x = x_407_cast_fp16, y = var_9713_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130108480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130370688))))[name = string("tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_206_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_206_cast_fp16")];
            string x_411_mode_0 = const()[name = string("x_411_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_411_cast_fp16 = gelu(mode = x_411_mode_0, x = linear_206_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [4]> var_9733_begin_0 = const()[name = string("op_9733_begin_0"), val = tensor<int32, [4]>([0, 0, 22, 0])];
            tensor<int32, [4]> var_9733_end_0 = const()[name = string("op_9733_end_0"), val = tensor<int32, [4]>([1, 512, 23, 512])];
            tensor<bool, [4]> var_9733_end_mask_0 = const()[name = string("op_9733_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9733_squeeze_mask_0 = const()[name = string("op_9733_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_9733_cast_fp16 = slice_by_index(begin = var_9733_begin_0, end = var_9733_end_0, end_mask = var_9733_end_mask_0, squeeze_mask = var_9733_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9733_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_275_cast_fp16 = mul(x = x_411_cast_fp16, y = var_9733_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130371776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130633984))))[name = string("tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9743_cast_fp16 = abs(x = linear_207_cast_fp16)[name = string("op_9743_cast_fp16")];
            bool var_9746 = const()[name = string("op_9746"), val = bool(true)];
            tensor<int32, [1]> var_9747_axes_0 = const()[name = string("op_9747_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9747_cast_fp16 = reduce_max(axes = var_9747_axes_0, keep_dims = var_9746, x = var_9743_cast_fp16)[name = string("op_9747_cast_fp16")];
            fp16 var_9748_to_fp16 = const()[name = string("op_9748_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_369_cast_fp16 = maximum(x = var_9747_cast_fp16, y = var_9748_to_fp16)[name = string("s_369_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_369_cast_fp16 = real_div(x = linear_207_cast_fp16, y = s_369_cast_fp16)[name = string("y_369_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9751_cast_fp16 = mul(x = y_369_cast_fp16, y = y_369_cast_fp16)[name = string("op_9751_cast_fp16")];
            tensor<int32, [1]> var_9756_axes_0 = const()[name = string("op_9756_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9756_keep_dims_0 = const()[name = string("op_9756_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9756_cast_fp16 = reduce_mean(axes = var_9756_axes_0, keep_dims = var_9756_keep_dims_0, x = var_9751_cast_fp16)[name = string("op_9756_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9757_cast_fp16 = mul(x = s_369_cast_fp16, y = s_369_cast_fp16)[name = string("op_9757_cast_fp16")];
            fp32 var_9758_epsilon_0 = const()[name = string("op_9758_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9758_cast_fp16 = inverse(epsilon = var_9758_epsilon_0, x = var_9757_cast_fp16)[name = string("op_9758_cast_fp16")];
            fp16 var_9759_to_fp16 = const()[name = string("op_9759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9760_cast_fp16 = mul(x = var_9758_cast_fp16, y = var_9759_to_fp16)[name = string("op_9760_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9762_cast_fp16 = add(x = var_9756_cast_fp16, y = var_9760_cast_fp16)[name = string("op_9762_cast_fp16")];
            fp32 var_9763_epsilon_0 = const()[name = string("op_9763_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9763_cast_fp16 = rsqrt(epsilon = var_9763_epsilon_0, x = var_9762_cast_fp16)[name = string("op_9763_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_323_cast_fp16 = mul(x = y_369_cast_fp16, y = var_9763_cast_fp16)[name = string("out_323_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130635072)))];
            tensor<fp16, [1, 512, 512]> var_9765_cast_fp16 = mul(x = out_323_cast_fp16, y = tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9765_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_45_cast_fp16 = add(x = input_273_cast_fp16, y = var_9765_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<fp16, [1]> tm_layers_22_layer_scalar_to_fp16 = const()[name = string("tm_layers_22_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 512, 512]> x_415_cast_fp16 = mul(x = h_45_cast_fp16, y = tm_layers_22_layer_scalar_to_fp16)[name = string("x_415_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9769_cast_fp16 = abs(x = x_415_cast_fp16)[name = string("op_9769_cast_fp16")];
            bool var_9772 = const()[name = string("op_9772"), val = bool(true)];
            tensor<int32, [1]> var_9773_axes_0 = const()[name = string("op_9773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9773_cast_fp16 = reduce_max(axes = var_9773_axes_0, keep_dims = var_9772, x = var_9769_cast_fp16)[name = string("op_9773_cast_fp16")];
            fp16 var_9774_to_fp16 = const()[name = string("op_9774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_371_cast_fp16 = maximum(x = var_9773_cast_fp16, y = var_9774_to_fp16)[name = string("s_371_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_371_cast_fp16 = real_div(x = x_415_cast_fp16, y = s_371_cast_fp16)[name = string("y_371_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9777_cast_fp16 = mul(x = y_371_cast_fp16, y = y_371_cast_fp16)[name = string("op_9777_cast_fp16")];
            tensor<int32, [1]> var_9782_axes_0 = const()[name = string("op_9782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9782_keep_dims_0 = const()[name = string("op_9782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9782_cast_fp16 = reduce_mean(axes = var_9782_axes_0, keep_dims = var_9782_keep_dims_0, x = var_9777_cast_fp16)[name = string("op_9782_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9783_cast_fp16 = mul(x = s_371_cast_fp16, y = s_371_cast_fp16)[name = string("op_9783_cast_fp16")];
            fp32 var_9784_epsilon_0 = const()[name = string("op_9784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9784_cast_fp16 = inverse(epsilon = var_9784_epsilon_0, x = var_9783_cast_fp16)[name = string("op_9784_cast_fp16")];
            fp16 var_9785_to_fp16 = const()[name = string("op_9785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9786_cast_fp16 = mul(x = var_9784_cast_fp16, y = var_9785_to_fp16)[name = string("op_9786_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9788_cast_fp16 = add(x = var_9782_cast_fp16, y = var_9786_cast_fp16)[name = string("op_9788_cast_fp16")];
            fp32 var_9789_epsilon_0 = const()[name = string("op_9789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9789_cast_fp16 = rsqrt(epsilon = var_9789_epsilon_0, x = var_9788_cast_fp16)[name = string("op_9789_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_325_cast_fp16 = mul(x = y_371_cast_fp16, y = var_9789_cast_fp16)[name = string("out_325_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130636160)))];
            tensor<fp16, [1, 512, 512]> input_277_cast_fp16 = mul(x = out_325_cast_fp16, y = tm_layers_23_input_layernorm_weight_to_fp16)[name = string("input_277_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130637248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131685888))))[name = string("tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_208_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<int32, [4]> var_9799 = const()[name = string("op_9799"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 512, 4, 512]> var_9800_cast_fp16 = reshape(shape = var_9799, x = linear_208_cast_fp16)[name = string("op_9800_cast_fp16")];
            tensor<int32, [4]> x_417_perm_0 = const()[name = string("x_417_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131690048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131952256))))[name = string("tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<int32, [4]> var_9811 = const()[name = string("op_9811"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 512, 1, 512]> var_9812_cast_fp16 = reshape(shape = var_9811, x = linear_209_cast_fp16)[name = string("op_9812_cast_fp16")];
            tensor<int32, [4]> x_419_perm_0 = const()[name = string("x_419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131953344))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132215552))))[name = string("tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<int32, [4]> var_9823 = const()[name = string("op_9823"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 512, 1, 512]> var_9824_cast_fp16 = reshape(shape = var_9823, x = linear_210_cast_fp16)[name = string("op_9824_cast_fp16")];
            tensor<int32, [4]> x_421_perm_0 = const()[name = string("x_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 512, 512]> x_417_cast_fp16 = transpose(perm = x_417_perm_0, x = var_9800_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 4, 512, 512]> var_9828_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_9828_cast_fp16")];
            bool var_9831 = const()[name = string("op_9831"), val = bool(true)];
            tensor<int32, [1]> var_9832_axes_0 = const()[name = string("op_9832_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 512, 1]> var_9832_cast_fp16 = reduce_max(axes = var_9832_axes_0, keep_dims = var_9831, x = var_9828_cast_fp16)[name = string("op_9832_cast_fp16")];
            fp16 var_9833_to_fp16 = const()[name = string("op_9833_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 512, 1]> s_373_cast_fp16 = maximum(x = var_9832_cast_fp16, y = var_9833_to_fp16)[name = string("s_373_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> y_373_cast_fp16 = real_div(x = x_417_cast_fp16, y = s_373_cast_fp16)[name = string("y_373_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_9836_cast_fp16 = mul(x = y_373_cast_fp16, y = y_373_cast_fp16)[name = string("op_9836_cast_fp16")];
            tensor<int32, [1]> var_9841_axes_0 = const()[name = string("op_9841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9841_keep_dims_0 = const()[name = string("op_9841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 512, 1]> var_9841_cast_fp16 = reduce_mean(axes = var_9841_axes_0, keep_dims = var_9841_keep_dims_0, x = var_9836_cast_fp16)[name = string("op_9841_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_9842_cast_fp16 = mul(x = s_373_cast_fp16, y = s_373_cast_fp16)[name = string("op_9842_cast_fp16")];
            fp32 var_9843_epsilon_0 = const()[name = string("op_9843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 512, 1]> var_9843_cast_fp16 = inverse(epsilon = var_9843_epsilon_0, x = var_9842_cast_fp16)[name = string("op_9843_cast_fp16")];
            fp16 var_9844_to_fp16 = const()[name = string("op_9844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 512, 1]> var_9845_cast_fp16 = mul(x = var_9843_cast_fp16, y = var_9844_to_fp16)[name = string("op_9845_cast_fp16")];
            tensor<fp16, [1, 4, 512, 1]> var_9847_cast_fp16 = add(x = var_9841_cast_fp16, y = var_9845_cast_fp16)[name = string("op_9847_cast_fp16")];
            fp32 var_9848_epsilon_0 = const()[name = string("op_9848_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 512, 1]> var_9848_cast_fp16 = rsqrt(epsilon = var_9848_epsilon_0, x = var_9847_cast_fp16)[name = string("op_9848_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> out_327_cast_fp16 = mul(x = y_373_cast_fp16, y = var_9848_cast_fp16)[name = string("out_327_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132216640)))];
            tensor<fp16, [1, 4, 512, 512]> q_93_cast_fp16 = mul(x = out_327_cast_fp16, y = tm_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("q_93_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> x_419_cast_fp16 = transpose(perm = x_419_perm_0, x = var_9812_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1, 512, 512]> var_9851_cast_fp16 = abs(x = x_419_cast_fp16)[name = string("op_9851_cast_fp16")];
            bool var_9854 = const()[name = string("op_9854"), val = bool(true)];
            tensor<int32, [1]> var_9855_axes_0 = const()[name = string("op_9855_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_9855_cast_fp16 = reduce_max(axes = var_9855_axes_0, keep_dims = var_9854, x = var_9851_cast_fp16)[name = string("op_9855_cast_fp16")];
            fp16 var_9856_to_fp16 = const()[name = string("op_9856_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 512, 1]> s_375_cast_fp16 = maximum(x = var_9855_cast_fp16, y = var_9856_to_fp16)[name = string("s_375_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> y_375_cast_fp16 = real_div(x = x_419_cast_fp16, y = s_375_cast_fp16)[name = string("y_375_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_9859_cast_fp16 = mul(x = y_375_cast_fp16, y = y_375_cast_fp16)[name = string("op_9859_cast_fp16")];
            tensor<int32, [1]> var_9864_axes_0 = const()[name = string("op_9864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9864_keep_dims_0 = const()[name = string("op_9864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> var_9864_cast_fp16 = reduce_mean(axes = var_9864_axes_0, keep_dims = var_9864_keep_dims_0, x = var_9859_cast_fp16)[name = string("op_9864_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_9865_cast_fp16 = mul(x = s_375_cast_fp16, y = s_375_cast_fp16)[name = string("op_9865_cast_fp16")];
            fp32 var_9866_epsilon_0 = const()[name = string("op_9866_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 512, 1]> var_9866_cast_fp16 = inverse(epsilon = var_9866_epsilon_0, x = var_9865_cast_fp16)[name = string("op_9866_cast_fp16")];
            fp16 var_9867_to_fp16 = const()[name = string("op_9867_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> var_9868_cast_fp16 = mul(x = var_9866_cast_fp16, y = var_9867_to_fp16)[name = string("op_9868_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_9870_cast_fp16 = add(x = var_9864_cast_fp16, y = var_9868_cast_fp16)[name = string("op_9870_cast_fp16")];
            fp32 var_9871_epsilon_0 = const()[name = string("op_9871_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> var_9871_cast_fp16 = rsqrt(epsilon = var_9871_epsilon_0, x = var_9870_cast_fp16)[name = string("op_9871_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> out_329_cast_fp16 = mul(x = y_375_cast_fp16, y = var_9871_cast_fp16)[name = string("out_329_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132217728)))];
            tensor<fp16, [1, 1, 512, 512]> k_139_cast_fp16 = mul(x = out_329_cast_fp16, y = tm_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("k_139_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> x_421_cast_fp16 = transpose(perm = x_421_perm_0, x = var_9824_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 1, 512, 512]> var_9874_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_9874_cast_fp16")];
            bool var_9877 = const()[name = string("op_9877"), val = bool(true)];
            tensor<int32, [1]> var_9878_axes_0 = const()[name = string("op_9878_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 512, 1]> var_9878_cast_fp16 = reduce_max(axes = var_9878_axes_0, keep_dims = var_9877, x = var_9874_cast_fp16)[name = string("op_9878_cast_fp16")];
            fp16 var_9879_to_fp16 = const()[name = string("op_9879_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 512, 1]> s_377_cast_fp16 = maximum(x = var_9878_cast_fp16, y = var_9879_to_fp16)[name = string("s_377_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> y_377_cast_fp16 = real_div(x = x_421_cast_fp16, y = s_377_cast_fp16)[name = string("y_377_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_9882_cast_fp16 = mul(x = y_377_cast_fp16, y = y_377_cast_fp16)[name = string("op_9882_cast_fp16")];
            tensor<int32, [1]> var_9887_axes_0 = const()[name = string("op_9887_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9887_keep_dims_0 = const()[name = string("op_9887_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 512, 1]> var_9887_cast_fp16 = reduce_mean(axes = var_9887_axes_0, keep_dims = var_9887_keep_dims_0, x = var_9882_cast_fp16)[name = string("op_9887_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_9888_cast_fp16 = mul(x = s_377_cast_fp16, y = s_377_cast_fp16)[name = string("op_9888_cast_fp16")];
            fp32 var_9889_epsilon_0 = const()[name = string("op_9889_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 512, 1]> var_9889_cast_fp16 = inverse(epsilon = var_9889_epsilon_0, x = var_9888_cast_fp16)[name = string("op_9889_cast_fp16")];
            fp16 var_9890_to_fp16 = const()[name = string("op_9890_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 512, 1]> var_9891_cast_fp16 = mul(x = var_9889_cast_fp16, y = var_9890_to_fp16)[name = string("op_9891_cast_fp16")];
            tensor<fp16, [1, 1, 512, 1]> var_9893_cast_fp16 = add(x = var_9887_cast_fp16, y = var_9891_cast_fp16)[name = string("op_9893_cast_fp16")];
            fp32 var_9894_epsilon_0 = const()[name = string("op_9894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 512, 1]> var_9894_cast_fp16 = rsqrt(epsilon = var_9894_epsilon_0, x = var_9893_cast_fp16)[name = string("op_9894_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> v_93_cast_fp16 = mul(x = y_377_cast_fp16, y = var_9894_cast_fp16)[name = string("v_93_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_9896_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9896_cast_fp16")];
            tensor<int32, [4]> var_9901_begin_0 = const()[name = string("op_9901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9901_end_0 = const()[name = string("op_9901_end_0"), val = tensor<int32, [4]>([1, 4, 512, 512])];
            tensor<bool, [4]> var_9901_end_mask_0 = const()[name = string("op_9901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 512, 256]> var_9901_cast_fp16 = slice_by_index(begin = var_9901_begin_0, end = var_9901_end_0, end_mask = var_9901_end_mask_0, x = q_93_cast_fp16)[name = string("op_9901_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 512, 256]> var_9902_cast_fp16 = mul(x = var_9901_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_9902_cast_fp16")];
            tensor<int32, [4]> var_9907_begin_0 = const()[name = string("op_9907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9907_end_0 = const()[name = string("op_9907_end_0"), val = tensor<int32, [4]>([1, 4, 512, 256])];
            tensor<bool, [4]> var_9907_end_mask_0 = const()[name = string("op_9907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 512, 256]> var_9907_cast_fp16 = slice_by_index(begin = var_9907_begin_0, end = var_9907_end_0, end_mask = var_9907_end_mask_0, x = q_93_cast_fp16)[name = string("op_9907_cast_fp16")];
            int32 var_9909 = const()[name = string("op_9909"), val = int32(-1)];
            bool var_9910_interleave_0 = const()[name = string("op_9910_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> var_9910_cast_fp16 = concat(axis = var_9909, interleave = var_9910_interleave_0, values = (var_9902_cast_fp16, var_9907_cast_fp16))[name = string("op_9910_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_9911_cast_fp16 = mul(x = var_9910_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9911_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> q_cast_fp16 = add(x = var_9896_cast_fp16, y = var_9911_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_9914_cast_fp16 = mul(x = k_139_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9914_cast_fp16")];
            tensor<int32, [4]> var_9919_begin_0 = const()[name = string("op_9919_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9919_end_0 = const()[name = string("op_9919_end_0"), val = tensor<int32, [4]>([1, 1, 512, 512])];
            tensor<bool, [4]> var_9919_end_mask_0 = const()[name = string("op_9919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 512, 256]> var_9919_cast_fp16 = slice_by_index(begin = var_9919_begin_0, end = var_9919_end_0, end_mask = var_9919_end_mask_0, x = k_139_cast_fp16)[name = string("op_9919_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 512, 256]> var_9920_cast_fp16 = mul(x = var_9919_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_9920_cast_fp16")];
            tensor<int32, [4]> var_9925_begin_0 = const()[name = string("op_9925_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9925_end_0 = const()[name = string("op_9925_end_0"), val = tensor<int32, [4]>([1, 1, 512, 256])];
            tensor<bool, [4]> var_9925_end_mask_0 = const()[name = string("op_9925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 512, 256]> var_9925_cast_fp16 = slice_by_index(begin = var_9925_begin_0, end = var_9925_end_0, end_mask = var_9925_end_mask_0, x = k_139_cast_fp16)[name = string("op_9925_cast_fp16")];
            int32 var_9927 = const()[name = string("op_9927"), val = int32(-1)];
            bool var_9928_interleave_0 = const()[name = string("op_9928_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512, 512]> var_9928_cast_fp16 = concat(axis = var_9927, interleave = var_9928_interleave_0, values = (var_9920_cast_fp16, var_9925_cast_fp16))[name = string("op_9928_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> var_9929_cast_fp16 = mul(x = var_9928_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9929_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> k_141_cast_fp16 = add(x = var_9914_cast_fp16, y = var_9929_cast_fp16)[name = string("k_141_cast_fp16")];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_46_reps_0 = const()[name = string("tile_46_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 512, 512]> transpose_92_cast_fp16 = transpose(perm = transpose_92_perm_0, x = k_141_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 1, 512, 512]> tile_46_cast_fp16 = tile(reps = tile_46_reps_0, x = transpose_92_cast_fp16)[name = string("tile_46_cast_fp16")];
            tensor<int32, [5]> concat_92 = const()[name = string("concat_92"), val = tensor<int32, [5]>([4, 1, 1, 512, 512])];
            tensor<fp16, [4, 1, 1, 512, 512]> reshape_92_cast_fp16 = reshape(shape = concat_92, x = tile_46_cast_fp16)[name = string("reshape_92_cast_fp16")];
            tensor<int32, [5]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_93 = const()[name = string("concat_93"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
            tensor<fp16, [1, 4, 1, 512, 512]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = reshape_92_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 1, 512, 512]> reshape_93_cast_fp16 = reshape(shape = concat_93, x = transpose_93_cast_fp16)[name = string("reshape_93_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_47_reps_0 = const()[name = string("tile_47_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 512, 512]> transpose_94_cast_fp16 = transpose(perm = transpose_94_perm_0, x = v_93_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 1, 512, 512]> tile_47_cast_fp16 = tile(reps = tile_47_reps_0, x = transpose_94_cast_fp16)[name = string("tile_47_cast_fp16")];
            tensor<int32, [5]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [5]>([4, 1, 1, 512, 512])];
            tensor<fp16, [4, 1, 1, 512, 512]> reshape_94_cast_fp16 = reshape(shape = concat_94, x = tile_47_cast_fp16)[name = string("reshape_94_cast_fp16")];
            tensor<int32, [5]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [4]>([-1, 1, 512, 512])];
            tensor<fp16, [1, 4, 1, 512, 512]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = reshape_94_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 1, 512, 512]> reshape_95_cast_fp16 = reshape(shape = concat_95, x = transpose_95_cast_fp16)[name = string("reshape_95_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9943_transpose_x_0 = const()[name = string("op_9943_transpose_x_0"), val = bool(false)];
            bool var_9943_transpose_y_0 = const()[name = string("op_9943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = reshape_93_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 4, 512, 512]> var_9943_cast_fp16 = matmul(transpose_x = var_9943_transpose_x_0, transpose_y = var_9943_transpose_y_0, x = q_cast_fp16, y = transpose_119_cast_fp16)[name = string("op_9943_cast_fp16")];
            tensor<fp16, [1, 4, 512, 512]> var_9945_cast_fp16 = add(x = var_9943_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9945_cast_fp16")];
            int32 var_9946 = const()[name = string("op_9946"), val = int32(-1)];
            tensor<fp16, [1, 4, 512, 512]> w_377_cast_fp16 = softmax(axis = var_9946, x = var_9945_cast_fp16)[name = string("w_377_cast_fp16")];
            bool var_9949_transpose_x_0 = const()[name = string("op_9949_transpose_x_0"), val = bool(false)];
            bool var_9949_transpose_y_0 = const()[name = string("op_9949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 512]> v_cast_fp16 = transpose(perm = v_perm_0, x = reshape_95_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 4, 512, 512]> var_9949_cast_fp16 = matmul(transpose_x = var_9949_transpose_x_0, transpose_y = var_9949_transpose_y_0, x = w_377_cast_fp16, y = v_cast_fp16)[name = string("op_9949_cast_fp16")];
            tensor<int32, [4]> var_9952_perm_0 = const()[name = string("op_9952_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9956 = const()[name = string("op_9956"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 512, 4, 512]> var_9952_cast_fp16 = transpose(perm = var_9952_perm_0, x = var_9949_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 512, 2048]> input_279_cast_fp16 = reshape(shape = var_9956, x = var_9952_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132218816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133267456))))[name = string("tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9961_cast_fp16 = abs(x = linear_211_cast_fp16)[name = string("op_9961_cast_fp16")];
            bool var_9964 = const()[name = string("op_9964"), val = bool(true)];
            tensor<int32, [1]> var_9965_axes_0 = const()[name = string("op_9965_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9965_cast_fp16 = reduce_max(axes = var_9965_axes_0, keep_dims = var_9964, x = var_9961_cast_fp16)[name = string("op_9965_cast_fp16")];
            fp16 var_9966_to_fp16 = const()[name = string("op_9966_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_379_cast_fp16 = maximum(x = var_9965_cast_fp16, y = var_9966_to_fp16)[name = string("s_379_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_379_cast_fp16 = real_div(x = linear_211_cast_fp16, y = s_379_cast_fp16)[name = string("y_379_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9969_cast_fp16 = mul(x = y_379_cast_fp16, y = y_379_cast_fp16)[name = string("op_9969_cast_fp16")];
            tensor<int32, [1]> var_9974_axes_0 = const()[name = string("op_9974_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9974_keep_dims_0 = const()[name = string("op_9974_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9974_cast_fp16 = reduce_mean(axes = var_9974_axes_0, keep_dims = var_9974_keep_dims_0, x = var_9969_cast_fp16)[name = string("op_9974_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9975_cast_fp16 = mul(x = s_379_cast_fp16, y = s_379_cast_fp16)[name = string("op_9975_cast_fp16")];
            fp32 var_9976_epsilon_0 = const()[name = string("op_9976_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_9976_cast_fp16 = inverse(epsilon = var_9976_epsilon_0, x = var_9975_cast_fp16)[name = string("op_9976_cast_fp16")];
            fp16 var_9977_to_fp16 = const()[name = string("op_9977_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_9978_cast_fp16 = mul(x = var_9976_cast_fp16, y = var_9977_to_fp16)[name = string("op_9978_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_9980_cast_fp16 = add(x = var_9974_cast_fp16, y = var_9978_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp32 var_9981_epsilon_0 = const()[name = string("op_9981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_9981_cast_fp16 = rsqrt(epsilon = var_9981_epsilon_0, x = var_9980_cast_fp16)[name = string("op_9981_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_331_cast_fp16 = mul(x = y_379_cast_fp16, y = var_9981_cast_fp16)[name = string("out_331_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133268544)))];
            tensor<fp16, [1, 512, 512]> var_9983_cast_fp16 = mul(x = out_331_cast_fp16, y = tm_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("op_9983_cast_fp16")];
            tensor<fp16, [1, 512, 512]> x_425_cast_fp16 = add(x = x_415_cast_fp16, y = var_9983_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9986_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_9986_cast_fp16")];
            bool var_9989 = const()[name = string("op_9989"), val = bool(true)];
            tensor<int32, [1]> var_9990_axes_0 = const()[name = string("op_9990_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_9990_cast_fp16 = reduce_max(axes = var_9990_axes_0, keep_dims = var_9989, x = var_9986_cast_fp16)[name = string("op_9990_cast_fp16")];
            fp16 var_9991_to_fp16 = const()[name = string("op_9991_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_381_cast_fp16 = maximum(x = var_9990_cast_fp16, y = var_9991_to_fp16)[name = string("s_381_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_381_cast_fp16 = real_div(x = x_425_cast_fp16, y = s_381_cast_fp16)[name = string("y_381_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_9994_cast_fp16 = mul(x = y_381_cast_fp16, y = y_381_cast_fp16)[name = string("op_9994_cast_fp16")];
            tensor<int32, [1]> var_9999_axes_0 = const()[name = string("op_9999_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9999_keep_dims_0 = const()[name = string("op_9999_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9999_cast_fp16 = reduce_mean(axes = var_9999_axes_0, keep_dims = var_9999_keep_dims_0, x = var_9994_cast_fp16)[name = string("op_9999_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_10000_cast_fp16 = mul(x = s_381_cast_fp16, y = s_381_cast_fp16)[name = string("op_10000_cast_fp16")];
            fp32 var_10001_epsilon_0 = const()[name = string("op_10001_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_10001_cast_fp16 = inverse(epsilon = var_10001_epsilon_0, x = var_10000_cast_fp16)[name = string("op_10001_cast_fp16")];
            fp16 var_10002_to_fp16 = const()[name = string("op_10002_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_10003_cast_fp16 = mul(x = var_10001_cast_fp16, y = var_10002_to_fp16)[name = string("op_10003_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_10005_cast_fp16 = add(x = var_9999_cast_fp16, y = var_10003_cast_fp16)[name = string("op_10005_cast_fp16")];
            fp32 var_10006_epsilon_0 = const()[name = string("op_10006_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_10006_cast_fp16 = rsqrt(epsilon = var_10006_epsilon_0, x = var_10005_cast_fp16)[name = string("op_10006_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_333_cast_fp16 = mul(x = y_381_cast_fp16, y = var_10006_cast_fp16)[name = string("out_333_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133269632)))];
            tensor<fp16, [1, 512, 512]> input_281_cast_fp16 = mul(x = out_333_cast_fp16, y = tm_layers_23_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133270720))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134319360))))[name = string("tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_212_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_212_cast_fp16")];
            string var_10013_mode_0 = const()[name = string("op_10013_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 2048]> var_10013_cast_fp16 = gelu(mode = var_10013_mode_0, x = linear_212_cast_fp16)[name = string("op_10013_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134323520))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135372160))))[name = string("tm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 2048]> linear_213_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_213_cast_fp16")];
            tensor<fp16, [1, 512, 2048]> input_283_cast_fp16 = mul(x = var_10013_cast_fp16, y = linear_213_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135376320))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136424960))))[name = string("tm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_283_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_10021_cast_fp16 = abs(x = linear_214_cast_fp16)[name = string("op_10021_cast_fp16")];
            bool var_10024 = const()[name = string("op_10024"), val = bool(true)];
            tensor<int32, [1]> var_10025_axes_0 = const()[name = string("op_10025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_10025_cast_fp16 = reduce_max(axes = var_10025_axes_0, keep_dims = var_10024, x = var_10021_cast_fp16)[name = string("op_10025_cast_fp16")];
            fp16 var_10026_to_fp16 = const()[name = string("op_10026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_383_cast_fp16 = maximum(x = var_10025_cast_fp16, y = var_10026_to_fp16)[name = string("s_383_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_383_cast_fp16 = real_div(x = linear_214_cast_fp16, y = s_383_cast_fp16)[name = string("y_383_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_10029_cast_fp16 = mul(x = y_383_cast_fp16, y = y_383_cast_fp16)[name = string("op_10029_cast_fp16")];
            tensor<int32, [1]> var_10034_axes_0 = const()[name = string("op_10034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10034_keep_dims_0 = const()[name = string("op_10034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_10034_cast_fp16 = reduce_mean(axes = var_10034_axes_0, keep_dims = var_10034_keep_dims_0, x = var_10029_cast_fp16)[name = string("op_10034_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_10035_cast_fp16 = mul(x = s_383_cast_fp16, y = s_383_cast_fp16)[name = string("op_10035_cast_fp16")];
            fp32 var_10036_epsilon_0 = const()[name = string("op_10036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_10036_cast_fp16 = inverse(epsilon = var_10036_epsilon_0, x = var_10035_cast_fp16)[name = string("op_10036_cast_fp16")];
            fp16 var_10037_to_fp16 = const()[name = string("op_10037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_10038_cast_fp16 = mul(x = var_10036_cast_fp16, y = var_10037_to_fp16)[name = string("op_10038_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_10040_cast_fp16 = add(x = var_10034_cast_fp16, y = var_10038_cast_fp16)[name = string("op_10040_cast_fp16")];
            fp32 var_10041_epsilon_0 = const()[name = string("op_10041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_10041_cast_fp16 = rsqrt(epsilon = var_10041_epsilon_0, x = var_10040_cast_fp16)[name = string("op_10041_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_335_cast_fp16 = mul(x = y_383_cast_fp16, y = var_10041_cast_fp16)[name = string("out_335_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136426048)))];
            tensor<fp16, [1, 512, 512]> var_10043_cast_fp16 = mul(x = out_335_cast_fp16, y = tm_layers_23_post_feedforward_layernorm_weight_to_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_285_cast_fp16 = add(x = x_425_cast_fp16, y = var_10043_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136427136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136689344))))[name = string("tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_215_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = string("linear_215_cast_fp16")];
            string x_429_mode_0 = const()[name = string("x_429_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 512, 512]> x_429_cast_fp16 = gelu(mode = x_429_mode_0, x = linear_215_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [4]> var_10063_begin_0 = const()[name = string("op_10063_begin_0"), val = tensor<int32, [4]>([0, 0, 23, 0])];
            tensor<int32, [4]> var_10063_end_0 = const()[name = string("op_10063_end_0"), val = tensor<int32, [4]>([1, 512, 24, 512])];
            tensor<bool, [4]> var_10063_end_mask_0 = const()[name = string("op_10063_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_10063_squeeze_mask_0 = const()[name = string("op_10063_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 512, 512]> var_10063_cast_fp16 = slice_by_index(begin = var_10063_begin_0, end = var_10063_end_0, end_mask = var_10063_end_mask_0, squeeze_mask = var_10063_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 512, 512]> input_287_cast_fp16 = mul(x = x_429_cast_fp16, y = var_10063_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136690432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136952640))))[name = string("tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 512]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_10073_cast_fp16 = abs(x = linear_216_cast_fp16)[name = string("op_10073_cast_fp16")];
            bool var_10076 = const()[name = string("op_10076"), val = bool(true)];
            tensor<int32, [1]> var_10077_axes_0 = const()[name = string("op_10077_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_10077_cast_fp16 = reduce_max(axes = var_10077_axes_0, keep_dims = var_10076, x = var_10073_cast_fp16)[name = string("op_10077_cast_fp16")];
            fp16 var_10078_to_fp16 = const()[name = string("op_10078_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_385_cast_fp16 = maximum(x = var_10077_cast_fp16, y = var_10078_to_fp16)[name = string("s_385_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_385_cast_fp16 = real_div(x = linear_216_cast_fp16, y = s_385_cast_fp16)[name = string("y_385_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_10081_cast_fp16 = mul(x = y_385_cast_fp16, y = y_385_cast_fp16)[name = string("op_10081_cast_fp16")];
            tensor<int32, [1]> var_10086_axes_0 = const()[name = string("op_10086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10086_keep_dims_0 = const()[name = string("op_10086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_10086_cast_fp16 = reduce_mean(axes = var_10086_axes_0, keep_dims = var_10086_keep_dims_0, x = var_10081_cast_fp16)[name = string("op_10086_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_10087_cast_fp16 = mul(x = s_385_cast_fp16, y = s_385_cast_fp16)[name = string("op_10087_cast_fp16")];
            fp32 var_10088_epsilon_0 = const()[name = string("op_10088_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_10088_cast_fp16 = inverse(epsilon = var_10088_epsilon_0, x = var_10087_cast_fp16)[name = string("op_10088_cast_fp16")];
            fp16 var_10089_to_fp16 = const()[name = string("op_10089_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_10090_cast_fp16 = mul(x = var_10088_cast_fp16, y = var_10089_to_fp16)[name = string("op_10090_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_10092_cast_fp16 = add(x = var_10086_cast_fp16, y = var_10090_cast_fp16)[name = string("op_10092_cast_fp16")];
            fp32 var_10093_epsilon_0 = const()[name = string("op_10093_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_10093_cast_fp16 = rsqrt(epsilon = var_10093_epsilon_0, x = var_10092_cast_fp16)[name = string("op_10093_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_337_cast_fp16 = mul(x = y_385_cast_fp16, y = var_10093_cast_fp16)[name = string("out_337_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136953728)))];
            tensor<fp16, [1, 512, 512]> var_10095_cast_fp16 = mul(x = out_337_cast_fp16, y = tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_10095_cast_fp16")];
            tensor<fp16, [1, 512, 512]> h_47_cast_fp16 = add(x = input_285_cast_fp16, y = var_10095_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<fp16, [1]> tm_layers_23_layer_scalar_to_fp16 = const()[name = string("tm_layers_23_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d6p-4])];
            tensor<fp16, [1, 512, 512]> x_cast_fp16 = mul(x = h_47_cast_fp16, y = tm_layers_23_layer_scalar_to_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_10099_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_10099_cast_fp16")];
            bool var_10102 = const()[name = string("op_10102"), val = bool(true)];
            tensor<int32, [1]> var_10103_axes_0 = const()[name = string("op_10103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_10103_cast_fp16 = reduce_max(axes = var_10103_axes_0, keep_dims = var_10102, x = var_10099_cast_fp16)[name = string("op_10103_cast_fp16")];
            fp16 var_10104_to_fp16 = const()[name = string("op_10104_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 512, 1]> s_387_cast_fp16 = maximum(x = var_10103_cast_fp16, y = var_10104_to_fp16)[name = string("s_387_cast_fp16")];
            tensor<fp16, [1, 512, 512]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_387_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 512, 512]> var_10107_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_10107_cast_fp16")];
            tensor<int32, [1]> var_10112_axes_0 = const()[name = string("op_10112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10112_keep_dims_0 = const()[name = string("op_10112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_10112_cast_fp16 = reduce_mean(axes = var_10112_axes_0, keep_dims = var_10112_keep_dims_0, x = var_10107_cast_fp16)[name = string("op_10112_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_10113_cast_fp16 = mul(x = s_387_cast_fp16, y = s_387_cast_fp16)[name = string("op_10113_cast_fp16")];
            fp32 var_10114_epsilon_0 = const()[name = string("op_10114_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_10114_cast_fp16 = inverse(epsilon = var_10114_epsilon_0, x = var_10113_cast_fp16)[name = string("op_10114_cast_fp16")];
            fp16 var_10115_to_fp16 = const()[name = string("op_10115_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_10116_cast_fp16 = mul(x = var_10114_cast_fp16, y = var_10115_to_fp16)[name = string("op_10116_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_10118_cast_fp16 = add(x = var_10112_cast_fp16, y = var_10116_cast_fp16)[name = string("op_10118_cast_fp16")];
            fp32 var_10119_epsilon_0 = const()[name = string("op_10119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_10119_cast_fp16 = rsqrt(epsilon = var_10119_epsilon_0, x = var_10118_cast_fp16)[name = string("op_10119_cast_fp16")];
            tensor<fp16, [1, 512, 512]> out_cast_fp16 = mul(x = y_cast_fp16, y = var_10119_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [512]> tm_norm_weight_to_fp16 = const()[name = string("tm_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136954816)))];
            tensor<fp16, [1, 512, 512]> h_cast_fp16 = mul(x = out_cast_fp16, y = tm_norm_weight_to_fp16)[name = string("h_cast_fp16")];
            tensor<int32, [1]> var_10126_axes_0 = const()[name = string("op_10126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10126_keep_dims_0 = const()[name = string("op_10126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10126_cast_fp16 = reduce_sum(axes = var_10126_axes_0, keep_dims = var_10126_keep_dims_0, x = attention_mask)[name = string("op_10126_cast_fp16")];
            tensor<fp16, [1, 512]> var_10127_cast_fp16 = real_div(x = attention_mask, y = var_10126_cast_fp16)[name = string("op_10127_cast_fp16")];
            tensor<int32, [1]> var_10134_axes_0 = const()[name = string("op_10134_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 512]> var_10134_cast_fp16 = expand_dims(axes = var_10134_axes_0, x = var_10127_cast_fp16)[name = string("op_10134_cast_fp16")];
            bool input_transpose_x_0 = const()[name = string("input_transpose_x_0"), val = bool(false)];
            bool input_transpose_y_0 = const()[name = string("input_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512]> input_cast_fp16 = matmul(transpose_x = input_transpose_x_0, transpose_y = input_transpose_y_0, x = var_10134_cast_fp16, y = h_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [768, 512]> tm_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136955904))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137349184))))[name = string("tm_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_217_bias_0_to_fp16 = const()[name = string("linear_217_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137350784)))];
            tensor<fp16, [1, 1, 768]> linear_217_cast_fp16 = linear(bias = linear_217_bias_0_to_fp16, weight = tm_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<int32, [3]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_10151_squeeze_mask_0 = const()[name = string("op_10151_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, squeeze_mask = var_10151_squeeze_mask_0, x = linear_217_cast_fp16)[name = string("op_10151_cast_fp16")];
            tensor<fp16, [1, 768]> var_10157_cast_fp16 = abs(x = var_10151_cast_fp16)[name = string("op_10157_cast_fp16")];
            bool var_10160 = const()[name = string("op_10160"), val = bool(true)];
            tensor<int32, [1]> s_axes_0 = const()[name = string("s_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> s_cast_fp16 = reduce_max(axes = s_axes_0, keep_dims = var_10160, x = var_10157_cast_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [1, 768]> pooled_cast_fp16 = real_div(x = var_10151_cast_fp16, y = s_cast_fp16)[name = string("pooled_cast_fp16")];
            tensor<fp16, [1, 768]> var_10163_cast_fp16 = mul(x = pooled_cast_fp16, y = pooled_cast_fp16)[name = string("op_10163_cast_fp16")];
            tensor<int32, [1]> var_10168_axes_0 = const()[name = string("op_10168_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10168_keep_dims_0 = const()[name = string("op_10168_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10168_cast_fp16 = reduce_sum(axes = var_10168_axes_0, keep_dims = var_10168_keep_dims_0, x = var_10163_cast_fp16)[name = string("op_10168_cast_fp16")];
            fp32 var_10169_epsilon_0 = const()[name = string("op_10169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_10169_cast_fp16 = rsqrt(epsilon = var_10169_epsilon_0, x = var_10168_cast_fp16)[name = string("op_10169_cast_fp16")];
            tensor<fp16, [1, 768]> var_10170_cast_fp16 = mul(x = pooled_cast_fp16, y = var_10169_cast_fp16)[name = string("op_10170_cast_fp16")];
            string var_10170_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_10170_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 768]> embedding = cast(dtype = var_10170_cast_fp16_to_fp32_dtype_0, x = var_10170_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func embed_64<ios18>(tensor<fp16, [1, 64]> attention_mask, tensor<fp16, [1, 64, 512]> inputs_embeds) {
            tensor<fp16, [12288, 512]> tm_ple_per_layer_model_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [12288, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1728))), scale = tensor<fp16, [12288, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6293248))))[name = string("tm_ple_per_layer_model_projection_weight_to_fp16_quantized")];
            tensor<fp16, [12288]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [12288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6317888)))];
            tensor<fp16, [1, 64, 12288]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = tm_ple_per_layer_model_projection_weight_to_fp16_quantized, x = inputs_embeds)[name = string("linear_0_cast_fp16")];
            fp16 var_2129_to_fp16 = const()[name = string("op_2129_to_fp16"), val = fp16(0x1.6ap-5)];
            tensor<fp16, [1, 64, 12288]> ple_1_cast_fp16 = mul(x = linear_0_cast_fp16, y = var_2129_to_fp16)[name = string("ple_1_cast_fp16")];
            tensor<int32, [4]> var_2135 = const()[name = string("op_2135"), val = tensor<int32, [4]>([1, -1, 24, 512])];
            tensor<fp16, [1, 64, 24, 512]> x_1_cast_fp16 = reshape(shape = var_2135, x = ple_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 64, 24, 512]> var_2137_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_2137_cast_fp16")];
            bool var_2140 = const()[name = string("op_2140"), val = bool(true)];
            tensor<int32, [1]> var_2141_axes_0 = const()[name = string("op_2141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 24, 1]> var_2141_cast_fp16 = reduce_max(axes = var_2141_axes_0, keep_dims = var_2140, x = var_2137_cast_fp16)[name = string("op_2141_cast_fp16")];
            fp16 var_2142_to_fp16 = const()[name = string("op_2142_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 24, 1]> s_1_cast_fp16 = maximum(x = var_2141_cast_fp16, y = var_2142_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 64, 24, 512]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 64, 24, 512]> var_2145_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [1]> var_2150_axes_0 = const()[name = string("op_2150_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2150_keep_dims_0 = const()[name = string("op_2150_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 24, 1]> var_2150_cast_fp16 = reduce_mean(axes = var_2150_axes_0, keep_dims = var_2150_keep_dims_0, x = var_2145_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 64, 24, 1]> var_2151_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 24, 1]> var_2152_cast_fp16 = inverse(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            fp16 var_2153_to_fp16 = const()[name = string("op_2153_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 24, 1]> var_2154_cast_fp16 = mul(x = var_2152_cast_fp16, y = var_2153_to_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 64, 24, 1]> var_2156_cast_fp16 = add(x = var_2150_cast_fp16, y = var_2154_cast_fp16)[name = string("op_2156_cast_fp16")];
            fp32 var_2157_epsilon_0 = const()[name = string("op_2157_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 24, 1]> var_2157_cast_fp16 = rsqrt(epsilon = var_2157_epsilon_0, x = var_2156_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<fp16, [1, 64, 24, 512]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_2157_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [512]> tm_ple_per_layer_projection_norm_weight_to_fp16 = const()[name = string("tm_ple_per_layer_projection_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6342528)))];
            tensor<fp16, [1, 64, 24, 512]> ple_3_cast_fp16 = mul(x = out_1_cast_fp16, y = tm_ple_per_layer_projection_norm_weight_to_fp16)[name = string("ple_3_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 64]> var_2162_cast_fp16 = sub(x = var_2160_to_fp16, y = attention_mask)[name = string("op_2162_cast_fp16")];
            fp16 var_2163_to_fp16 = const()[name = string("op_2163_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 64]> var_2164_cast_fp16 = mul(x = var_2162_cast_fp16, y = var_2163_to_fp16)[name = string("op_2164_cast_fp16")];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 64]> var_2171_cast_fp16 = expand_dims(axes = var_2171_axes_0, x = var_2164_cast_fp16)[name = string("op_2171_cast_fp16")];
            tensor<int32, [1]> var_2173_axes_0 = const()[name = string("op_2173_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 64]> var_2173_cast_fp16 = expand_dims(axes = var_2173_axes_0, x = var_2171_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2179_cast_fp16 = abs(x = inputs_embeds)[name = string("op_2179_cast_fp16")];
            bool var_2182 = const()[name = string("op_2182"), val = bool(true)];
            tensor<int32, [1]> var_2183_axes_0 = const()[name = string("op_2183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2183_cast_fp16 = reduce_max(axes = var_2183_axes_0, keep_dims = var_2182, x = var_2179_cast_fp16)[name = string("op_2183_cast_fp16")];
            fp16 var_2184_to_fp16 = const()[name = string("op_2184_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_3_cast_fp16 = maximum(x = var_2183_cast_fp16, y = var_2184_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_3_cast_fp16 = real_div(x = inputs_embeds, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2187_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> var_2192_axes_0 = const()[name = string("op_2192_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2192_keep_dims_0 = const()[name = string("op_2192_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2192_cast_fp16 = reduce_mean(axes = var_2192_axes_0, keep_dims = var_2192_keep_dims_0, x = var_2187_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2193_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp32 var_2194_epsilon_0 = const()[name = string("op_2194_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2194_cast_fp16 = inverse(epsilon = var_2194_epsilon_0, x = var_2193_cast_fp16)[name = string("op_2194_cast_fp16")];
            fp16 var_2195_to_fp16 = const()[name = string("op_2195_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2196_cast_fp16 = mul(x = var_2194_cast_fp16, y = var_2195_to_fp16)[name = string("op_2196_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2198_cast_fp16 = add(x = var_2192_cast_fp16, y = var_2196_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_2199_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6343616)))];
            tensor<fp16, [1, 64, 512]> input_1_cast_fp16 = mul(x = out_3_cast_fp16, y = tm_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6344704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6869056))))[name = string("tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6871168)))];
            tensor<fp16, [1, 64, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_1_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6873280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7135488))))[name = string("tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7136576)))];
            tensor<fp16, [1, 64, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_2221 = const()[name = string("op_2221"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_2222_cast_fp16 = reshape(shape = var_2221, x = linear_2_cast_fp16)[name = string("op_2222_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7137664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7399872))))[name = string("tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_2234_cast_fp16 = reshape(shape = var_2233, x = linear_3_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<int32, [4]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_2210_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 4, 64, 256]> var_2238_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_2238_cast_fp16")];
            bool var_2241 = const()[name = string("op_2241"), val = bool(true)];
            tensor<int32, [1]> var_2242_axes_0 = const()[name = string("op_2242_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_2242_cast_fp16 = reduce_max(axes = var_2242_axes_0, keep_dims = var_2241, x = var_2238_cast_fp16)[name = string("op_2242_cast_fp16")];
            fp16 var_2243_to_fp16 = const()[name = string("op_2243_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_5_cast_fp16 = maximum(x = var_2242_cast_fp16, y = var_2243_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_2246_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<int32, [1]> var_2251_axes_0 = const()[name = string("op_2251_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2251_keep_dims_0 = const()[name = string("op_2251_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_2251_cast_fp16 = reduce_mean(axes = var_2251_axes_0, keep_dims = var_2251_keep_dims_0, x = var_2246_cast_fp16)[name = string("op_2251_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_2252_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_2252_cast_fp16")];
            fp32 var_2253_epsilon_0 = const()[name = string("op_2253_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_2253_cast_fp16 = inverse(epsilon = var_2253_epsilon_0, x = var_2252_cast_fp16)[name = string("op_2253_cast_fp16")];
            fp16 var_2254_to_fp16 = const()[name = string("op_2254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_2255_cast_fp16 = mul(x = var_2253_cast_fp16, y = var_2254_to_fp16)[name = string("op_2255_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_2257_cast_fp16 = add(x = var_2251_cast_fp16, y = var_2255_cast_fp16)[name = string("op_2257_cast_fp16")];
            fp32 var_2258_epsilon_0 = const()[name = string("op_2258_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_2258_cast_fp16 = rsqrt(epsilon = var_2258_epsilon_0, x = var_2257_cast_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_2258_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7400960)))];
            tensor<fp16, [1, 4, 64, 256]> q_1_cast_fp16 = mul(x = out_5_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_2222_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 2, 64, 256]> var_2261_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_2261_cast_fp16")];
            bool var_2264 = const()[name = string("op_2264"), val = bool(true)];
            tensor<int32, [1]> var_2265_axes_0 = const()[name = string("op_2265_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_2265_cast_fp16 = reduce_max(axes = var_2265_axes_0, keep_dims = var_2264, x = var_2261_cast_fp16)[name = string("op_2265_cast_fp16")];
            fp16 var_2266_to_fp16 = const()[name = string("op_2266_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_7_cast_fp16 = maximum(x = var_2265_cast_fp16, y = var_2266_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2269_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [1]> var_2274_axes_0 = const()[name = string("op_2274_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2274_keep_dims_0 = const()[name = string("op_2274_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_2274_cast_fp16 = reduce_mean(axes = var_2274_axes_0, keep_dims = var_2274_keep_dims_0, x = var_2269_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2275_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_2276_cast_fp16 = inverse(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            fp16 var_2277_to_fp16 = const()[name = string("op_2277_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_2278_cast_fp16 = mul(x = var_2276_cast_fp16, y = var_2277_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2280_cast_fp16 = add(x = var_2274_cast_fp16, y = var_2278_cast_fp16)[name = string("op_2280_cast_fp16")];
            fp32 var_2281_epsilon_0 = const()[name = string("op_2281_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_2281_cast_fp16 = rsqrt(epsilon = var_2281_epsilon_0, x = var_2280_cast_fp16)[name = string("op_2281_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_7_cast_fp16 = mul(x = y_7_cast_fp16, y = var_2281_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7401536)))];
            tensor<fp16, [1, 2, 64, 256]> k_1_cast_fp16 = mul(x = out_7_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = var_2234_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2, 64, 256]> var_2284_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_2284_cast_fp16")];
            bool var_2287 = const()[name = string("op_2287"), val = bool(true)];
            tensor<int32, [1]> var_2288_axes_0 = const()[name = string("op_2288_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_2288_cast_fp16 = reduce_max(axes = var_2288_axes_0, keep_dims = var_2287, x = var_2284_cast_fp16)[name = string("op_2288_cast_fp16")];
            fp16 var_2289_to_fp16 = const()[name = string("op_2289_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_9_cast_fp16 = maximum(x = var_2288_cast_fp16, y = var_2289_to_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_9_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2292_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_2292_cast_fp16")];
            tensor<int32, [1]> var_2297_axes_0 = const()[name = string("op_2297_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2297_keep_dims_0 = const()[name = string("op_2297_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_2297_cast_fp16 = reduce_mean(axes = var_2297_axes_0, keep_dims = var_2297_keep_dims_0, x = var_2292_cast_fp16)[name = string("op_2297_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2298_cast_fp16 = mul(x = s_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_2298_cast_fp16")];
            fp32 var_2299_epsilon_0 = const()[name = string("op_2299_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_2299_cast_fp16 = inverse(epsilon = var_2299_epsilon_0, x = var_2298_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 var_2300_to_fp16 = const()[name = string("op_2300_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_2301_cast_fp16 = mul(x = var_2299_cast_fp16, y = var_2300_to_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2303_cast_fp16 = add(x = var_2297_cast_fp16, y = var_2301_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp32 var_2304_epsilon_0 = const()[name = string("op_2304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_2304_cast_fp16 = rsqrt(epsilon = var_2304_epsilon_0, x = var_2303_cast_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_1_cast_fp16 = mul(x = y_9_cast_fp16, y = var_2304_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<fp16, [64, 256]> cos_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [64, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137476736))), scale = tensor<fp16, [64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137493184))))[name = string("cos_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 64, 256]> var_2306_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2306_cast_fp16")];
            tensor<int32, [4]> var_2311_begin_0 = const()[name = string("op_2311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2311_end_0 = const()[name = string("op_2311_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_2311_end_mask_0 = const()[name = string("op_2311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_2311_cast_fp16 = slice_by_index(begin = var_2311_begin_0, end = var_2311_end_0, end_mask = var_2311_end_mask_0, x = q_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_2312_cast_fp16 = mul(x = var_2311_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_2312_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = q_1_cast_fp16)[name = string("op_2317_cast_fp16")];
            int32 var_2319 = const()[name = string("op_2319"), val = int32(-1)];
            bool var_2320_interleave_0 = const()[name = string("op_2320_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_2320_cast_fp16 = concat(axis = var_2319, interleave = var_2320_interleave_0, values = (var_2312_cast_fp16, var_2317_cast_fp16))[name = string("op_2320_cast_fp16")];
            tensor<fp16, [64, 256]> sin_1_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [64, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137493376))), scale = tensor<fp16, [64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137509824))))[name = string("sin_1_to_fp16_quantized")];
            tensor<fp16, [1, 4, 64, 256]> var_2321_cast_fp16 = mul(x = var_2320_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_3_cast_fp16 = add(x = var_2306_cast_fp16, y = var_2321_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2324_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2324_cast_fp16")];
            tensor<int32, [4]> var_2329_begin_0 = const()[name = string("op_2329_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2329_end_0 = const()[name = string("op_2329_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_2329_end_mask_0 = const()[name = string("op_2329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_2329_cast_fp16 = slice_by_index(begin = var_2329_begin_0, end = var_2329_end_0, end_mask = var_2329_end_mask_0, x = k_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2330_cast_fp16 = mul(x = var_2329_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_2330_cast_fp16")];
            tensor<int32, [4]> var_2335_begin_0 = const()[name = string("op_2335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2335_end_0 = const()[name = string("op_2335_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_2335_end_mask_0 = const()[name = string("op_2335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_2335_cast_fp16 = slice_by_index(begin = var_2335_begin_0, end = var_2335_end_0, end_mask = var_2335_end_mask_0, x = k_1_cast_fp16)[name = string("op_2335_cast_fp16")];
            int32 var_2337 = const()[name = string("op_2337"), val = int32(-1)];
            bool var_2338_interleave_0 = const()[name = string("op_2338_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_2338_cast_fp16 = concat(axis = var_2337, interleave = var_2338_interleave_0, values = (var_2330_cast_fp16, var_2335_cast_fp16))[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2339_cast_fp16 = mul(x = var_2338_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2339_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_3_cast_fp16 = add(x = var_2324_cast_fp16, y = var_2339_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = k_3_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [4, 1, 64, 256]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
            tensor<int32, [5]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_0_cast_fp16 = reshape(shape = concat_0, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [4, 1, 64, 256]> reshape_1_cast_fp16 = reshape(shape = concat_1, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = v_1_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [4, 1, 64, 256]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
            tensor<int32, [5]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_2_cast_fp16 = reshape(shape = concat_2, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [4, 1, 64, 256]> reshape_3_cast_fp16 = reshape(shape = concat_3, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2353_transpose_x_0 = const()[name = string("op_2353_transpose_x_0"), val = bool(false)];
            bool var_2353_transpose_y_0 = const()[name = string("op_2353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 4, 64, 64]> var_2353_cast_fp16 = matmul(transpose_x = var_2353_transpose_x_0, transpose_y = var_2353_transpose_y_0, x = q_3_cast_fp16, y = transpose_96_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_2355_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2356 = const()[name = string("op_2356"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_9_cast_fp16 = softmax(axis = var_2356, x = var_2355_cast_fp16)[name = string("w_9_cast_fp16")];
            bool var_2359_transpose_x_0 = const()[name = string("op_2359_transpose_x_0"), val = bool(false)];
            bool var_2359_transpose_y_0 = const()[name = string("op_2359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 4, 64, 256]> var_2359_cast_fp16 = matmul(transpose_x = var_2359_transpose_x_0, transpose_y = var_2359_transpose_y_0, x = w_9_cast_fp16, y = v_3_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<int32, [4]> var_2362_perm_0 = const()[name = string("op_2362_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2366 = const()[name = string("op_2366"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_2362_cast_fp16 = transpose(perm = var_2362_perm_0, x = var_2359_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 64, 1024]> input_3_cast_fp16 = reshape(shape = var_2366, x = var_2362_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7402112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7926464))))[name = string("tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2371_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_2371_cast_fp16")];
            bool var_2374 = const()[name = string("op_2374"), val = bool(true)];
            tensor<int32, [1]> var_2375_axes_0 = const()[name = string("op_2375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2375_cast_fp16 = reduce_max(axes = var_2375_axes_0, keep_dims = var_2374, x = var_2371_cast_fp16)[name = string("op_2375_cast_fp16")];
            fp16 var_2376_to_fp16 = const()[name = string("op_2376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_11_cast_fp16 = maximum(x = var_2375_cast_fp16, y = var_2376_to_fp16)[name = string("s_11_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_11_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_11_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2379_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_2379_cast_fp16")];
            tensor<int32, [1]> var_2384_axes_0 = const()[name = string("op_2384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2384_keep_dims_0 = const()[name = string("op_2384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2384_cast_fp16 = reduce_mean(axes = var_2384_axes_0, keep_dims = var_2384_keep_dims_0, x = var_2379_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2385_cast_fp16 = mul(x = s_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2385_cast_fp16")];
            fp32 var_2386_epsilon_0 = const()[name = string("op_2386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2386_cast_fp16 = inverse(epsilon = var_2386_epsilon_0, x = var_2385_cast_fp16)[name = string("op_2386_cast_fp16")];
            fp16 var_2387_to_fp16 = const()[name = string("op_2387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2388_cast_fp16 = mul(x = var_2386_cast_fp16, y = var_2387_to_fp16)[name = string("op_2388_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2390_cast_fp16 = add(x = var_2384_cast_fp16, y = var_2388_cast_fp16)[name = string("op_2390_cast_fp16")];
            fp32 var_2391_epsilon_0 = const()[name = string("op_2391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2391_cast_fp16 = rsqrt(epsilon = var_2391_epsilon_0, x = var_2390_cast_fp16)[name = string("op_2391_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_2391_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7927552)))];
            tensor<fp16, [1, 64, 512]> var_2393_cast_fp16 = mul(x = out_9_cast_fp16, y = tm_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_11_cast_fp16 = add(x = inputs_embeds, y = var_2393_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2396_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_2396_cast_fp16")];
            bool var_2399 = const()[name = string("op_2399"), val = bool(true)];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_2399, x = var_2396_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 var_2401_to_fp16 = const()[name = string("op_2401_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_13_cast_fp16 = maximum(x = var_2400_cast_fp16, y = var_2401_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = s_13_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2404_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<int32, [1]> var_2409_axes_0 = const()[name = string("op_2409_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2409_keep_dims_0 = const()[name = string("op_2409_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2409_cast_fp16 = reduce_mean(axes = var_2409_axes_0, keep_dims = var_2409_keep_dims_0, x = var_2404_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2410_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2411_cast_fp16 = inverse(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            fp16 var_2412_to_fp16 = const()[name = string("op_2412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2413_cast_fp16 = mul(x = var_2411_cast_fp16, y = var_2412_to_fp16)[name = string("op_2413_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2415_cast_fp16 = add(x = var_2409_cast_fp16, y = var_2413_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp32 var_2416_epsilon_0 = const()[name = string("op_2416_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2416_cast_fp16 = rsqrt(epsilon = var_2416_epsilon_0, x = var_2415_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_2416_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7928640)))];
            tensor<fp16, [1, 64, 512]> input_5_cast_fp16 = mul(x = out_11_cast_fp16, y = tm_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7929728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8978368))))[name = string("tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8982528)))];
            tensor<fp16, [1, 64, 2048]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_2423_mode_0 = const()[name = string("op_2423_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_2423_cast_fp16 = gelu(mode = var_2423_mode_0, x = linear_5_cast_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8986688))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10035328))))[name = string("tm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_7_cast_fp16 = mul(x = var_2423_cast_fp16, y = linear_6_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10039488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11088128))))[name = string("tm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2431_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_2431_cast_fp16")];
            bool var_2434 = const()[name = string("op_2434"), val = bool(true)];
            tensor<int32, [1]> var_2435_axes_0 = const()[name = string("op_2435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2435_cast_fp16 = reduce_max(axes = var_2435_axes_0, keep_dims = var_2434, x = var_2431_cast_fp16)[name = string("op_2435_cast_fp16")];
            fp16 var_2436_to_fp16 = const()[name = string("op_2436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_15_cast_fp16 = maximum(x = var_2435_cast_fp16, y = var_2436_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_15_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_15_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2439_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_2439_cast_fp16")];
            tensor<int32, [1]> var_2444_axes_0 = const()[name = string("op_2444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2444_keep_dims_0 = const()[name = string("op_2444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2444_cast_fp16 = reduce_mean(axes = var_2444_axes_0, keep_dims = var_2444_keep_dims_0, x = var_2439_cast_fp16)[name = string("op_2444_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2445_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_2445_cast_fp16")];
            fp32 var_2446_epsilon_0 = const()[name = string("op_2446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2446_cast_fp16 = inverse(epsilon = var_2446_epsilon_0, x = var_2445_cast_fp16)[name = string("op_2446_cast_fp16")];
            fp16 var_2447_to_fp16 = const()[name = string("op_2447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2448_cast_fp16 = mul(x = var_2446_cast_fp16, y = var_2447_to_fp16)[name = string("op_2448_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2450_cast_fp16 = add(x = var_2444_cast_fp16, y = var_2448_cast_fp16)[name = string("op_2450_cast_fp16")];
            fp32 var_2451_epsilon_0 = const()[name = string("op_2451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2451_cast_fp16 = rsqrt(epsilon = var_2451_epsilon_0, x = var_2450_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_2451_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11089216)))];
            tensor<fp16, [1, 64, 512]> var_2453_cast_fp16 = mul(x = out_13_cast_fp16, y = tm_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_9_cast_fp16 = add(x = x_11_cast_fp16, y = var_2453_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11090304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11352512))))[name = string("tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_8_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_8_cast_fp16")];
            string x_15_mode_0 = const()[name = string("x_15_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_15_cast_fp16 = gelu(mode = x_15_mode_0, x = linear_8_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [4]> var_2473_begin_0 = const()[name = string("op_2473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2473_end_0 = const()[name = string("op_2473_end_0"), val = tensor<int32, [4]>([1, 64, 1, 512])];
            tensor<bool, [4]> var_2473_end_mask_0 = const()[name = string("op_2473_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2473_squeeze_mask_0 = const()[name = string("op_2473_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_2473_cast_fp16 = slice_by_index(begin = var_2473_begin_0, end = var_2473_end_0, end_mask = var_2473_end_mask_0, squeeze_mask = var_2473_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2473_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_11_cast_fp16 = mul(x = x_15_cast_fp16, y = var_2473_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11353600))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11615808))))[name = string("tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2483_cast_fp16 = abs(x = linear_9_cast_fp16)[name = string("op_2483_cast_fp16")];
            bool var_2486 = const()[name = string("op_2486"), val = bool(true)];
            tensor<int32, [1]> var_2487_axes_0 = const()[name = string("op_2487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2487_cast_fp16 = reduce_max(axes = var_2487_axes_0, keep_dims = var_2486, x = var_2483_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp16 var_2488_to_fp16 = const()[name = string("op_2488_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_17_cast_fp16 = maximum(x = var_2487_cast_fp16, y = var_2488_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_17_cast_fp16 = real_div(x = linear_9_cast_fp16, y = s_17_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2491_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_2491_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2496_keep_dims_0 = const()[name = string("op_2496_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2496_cast_fp16 = reduce_mean(axes = var_2496_axes_0, keep_dims = var_2496_keep_dims_0, x = var_2491_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2497_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_2497_cast_fp16")];
            fp32 var_2498_epsilon_0 = const()[name = string("op_2498_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2498_cast_fp16 = inverse(epsilon = var_2498_epsilon_0, x = var_2497_cast_fp16)[name = string("op_2498_cast_fp16")];
            fp16 var_2499_to_fp16 = const()[name = string("op_2499_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2500_cast_fp16 = mul(x = var_2498_cast_fp16, y = var_2499_to_fp16)[name = string("op_2500_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2502_cast_fp16 = add(x = var_2496_cast_fp16, y = var_2500_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp32 var_2503_epsilon_0 = const()[name = string("op_2503_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2503_cast_fp16 = rsqrt(epsilon = var_2503_epsilon_0, x = var_2502_cast_fp16)[name = string("op_2503_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_2503_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11616896)))];
            tensor<fp16, [1, 64, 512]> var_2505_cast_fp16 = mul(x = out_15_cast_fp16, y = tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2505_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_1_cast_fp16 = add(x = input_9_cast_fp16, y = var_2505_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<fp16, [1]> tm_layers_0_layer_scalar_to_fp16 = const()[name = string("tm_layers_0_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9p-2])];
            tensor<fp16, [1, 64, 512]> x_19_cast_fp16 = mul(x = h_1_cast_fp16, y = tm_layers_0_layer_scalar_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2509_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_2509_cast_fp16")];
            bool var_2512 = const()[name = string("op_2512"), val = bool(true)];
            tensor<int32, [1]> var_2513_axes_0 = const()[name = string("op_2513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2513_cast_fp16 = reduce_max(axes = var_2513_axes_0, keep_dims = var_2512, x = var_2509_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_19_cast_fp16 = maximum(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_19_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2517_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_2517_cast_fp16")];
            tensor<int32, [1]> var_2522_axes_0 = const()[name = string("op_2522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2522_keep_dims_0 = const()[name = string("op_2522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2522_cast_fp16 = reduce_mean(axes = var_2522_axes_0, keep_dims = var_2522_keep_dims_0, x = var_2517_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2523_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_2523_cast_fp16")];
            fp32 var_2524_epsilon_0 = const()[name = string("op_2524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2524_cast_fp16 = inverse(epsilon = var_2524_epsilon_0, x = var_2523_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp16 var_2525_to_fp16 = const()[name = string("op_2525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2526_cast_fp16 = mul(x = var_2524_cast_fp16, y = var_2525_to_fp16)[name = string("op_2526_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2528_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2526_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_2529_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11617984)))];
            tensor<fp16, [1, 64, 512]> input_13_cast_fp16 = mul(x = out_17_cast_fp16, y = tm_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11619072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12143424))))[name = string("tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_2540_cast_fp16 = reshape(shape = var_2539, x = linear_10_cast_fp16)[name = string("op_2540_cast_fp16")];
            tensor<int32, [4]> x_21_perm_0 = const()[name = string("x_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12145536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12407744))))[name = string("tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<int32, [4]> var_2551 = const()[name = string("op_2551"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_2552_cast_fp16 = reshape(shape = var_2551, x = linear_11_cast_fp16)[name = string("op_2552_cast_fp16")];
            tensor<int32, [4]> x_23_perm_0 = const()[name = string("x_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12408832))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671040))))[name = string("tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [4]> var_2563 = const()[name = string("op_2563"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_2564_cast_fp16 = reshape(shape = var_2563, x = linear_12_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = var_2540_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 4, 64, 256]> var_2568_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_2568_cast_fp16")];
            bool var_2571 = const()[name = string("op_2571"), val = bool(true)];
            tensor<int32, [1]> var_2572_axes_0 = const()[name = string("op_2572_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_2572_cast_fp16 = reduce_max(axes = var_2572_axes_0, keep_dims = var_2571, x = var_2568_cast_fp16)[name = string("op_2572_cast_fp16")];
            fp16 var_2573_to_fp16 = const()[name = string("op_2573_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_21_cast_fp16 = maximum(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_21_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_2576_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> var_2581_axes_0 = const()[name = string("op_2581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2581_keep_dims_0 = const()[name = string("op_2581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_2581_cast_fp16 = reduce_mean(axes = var_2581_axes_0, keep_dims = var_2581_keep_dims_0, x = var_2576_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_2582_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp32 var_2583_epsilon_0 = const()[name = string("op_2583_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_2583_cast_fp16 = inverse(epsilon = var_2583_epsilon_0, x = var_2582_cast_fp16)[name = string("op_2583_cast_fp16")];
            fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_2585_cast_fp16 = mul(x = var_2583_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_2587_cast_fp16 = add(x = var_2581_cast_fp16, y = var_2585_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_19_cast_fp16 = mul(x = y_21_cast_fp16, y = var_2588_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672128)))];
            tensor<fp16, [1, 4, 64, 256]> q_5_cast_fp16 = mul(x = out_19_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_23_cast_fp16 = transpose(perm = x_23_perm_0, x = var_2552_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2, 64, 256]> var_2591_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_2591_cast_fp16")];
            bool var_2594 = const()[name = string("op_2594"), val = bool(true)];
            tensor<int32, [1]> var_2595_axes_0 = const()[name = string("op_2595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_2595_cast_fp16 = reduce_max(axes = var_2595_axes_0, keep_dims = var_2594, x = var_2591_cast_fp16)[name = string("op_2595_cast_fp16")];
            fp16 var_2596_to_fp16 = const()[name = string("op_2596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_23_cast_fp16 = maximum(x = var_2595_cast_fp16, y = var_2596_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_23_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2599_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2599_cast_fp16")];
            tensor<int32, [1]> var_2604_axes_0 = const()[name = string("op_2604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2604_keep_dims_0 = const()[name = string("op_2604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_2604_cast_fp16 = reduce_mean(axes = var_2604_axes_0, keep_dims = var_2604_keep_dims_0, x = var_2599_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2605_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_2605_cast_fp16")];
            fp32 var_2606_epsilon_0 = const()[name = string("op_2606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_2606_cast_fp16 = inverse(epsilon = var_2606_epsilon_0, x = var_2605_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp16 var_2607_to_fp16 = const()[name = string("op_2607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_2608_cast_fp16 = mul(x = var_2606_cast_fp16, y = var_2607_to_fp16)[name = string("op_2608_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2610_cast_fp16 = add(x = var_2604_cast_fp16, y = var_2608_cast_fp16)[name = string("op_2610_cast_fp16")];
            fp32 var_2611_epsilon_0 = const()[name = string("op_2611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_2611_cast_fp16 = rsqrt(epsilon = var_2611_epsilon_0, x = var_2610_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_21_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2611_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672704)))];
            tensor<fp16, [1, 2, 64, 256]> k_7_cast_fp16 = mul(x = out_21_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_2564_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 2, 64, 256]> var_2614_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_2614_cast_fp16")];
            bool var_2617 = const()[name = string("op_2617"), val = bool(true)];
            tensor<int32, [1]> var_2618_axes_0 = const()[name = string("op_2618_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_2618_cast_fp16 = reduce_max(axes = var_2618_axes_0, keep_dims = var_2617, x = var_2614_cast_fp16)[name = string("op_2618_cast_fp16")];
            fp16 var_2619_to_fp16 = const()[name = string("op_2619_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_25_cast_fp16 = maximum(x = var_2618_cast_fp16, y = var_2619_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_25_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2622_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2622_cast_fp16")];
            tensor<int32, [1]> var_2627_axes_0 = const()[name = string("op_2627_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2627_keep_dims_0 = const()[name = string("op_2627_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_2627_cast_fp16 = reduce_mean(axes = var_2627_axes_0, keep_dims = var_2627_keep_dims_0, x = var_2622_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2628_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_2628_cast_fp16")];
            fp32 var_2629_epsilon_0 = const()[name = string("op_2629_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_2629_cast_fp16 = inverse(epsilon = var_2629_epsilon_0, x = var_2628_cast_fp16)[name = string("op_2629_cast_fp16")];
            fp16 var_2630_to_fp16 = const()[name = string("op_2630_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_2631_cast_fp16 = mul(x = var_2629_cast_fp16, y = var_2630_to_fp16)[name = string("op_2631_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2633_cast_fp16 = add(x = var_2627_cast_fp16, y = var_2631_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_2634_cast_fp16 = rsqrt(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_5_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2634_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_2636_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2636_cast_fp16")];
            tensor<int32, [4]> var_2641_begin_0 = const()[name = string("op_2641_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2641_end_0 = const()[name = string("op_2641_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_2641_end_mask_0 = const()[name = string("op_2641_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_2641_cast_fp16 = slice_by_index(begin = var_2641_begin_0, end = var_2641_end_0, end_mask = var_2641_end_mask_0, x = q_5_cast_fp16)[name = string("op_2641_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_2642_cast_fp16 = mul(x = var_2641_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_2642_cast_fp16")];
            tensor<int32, [4]> var_2647_begin_0 = const()[name = string("op_2647_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2647_end_0 = const()[name = string("op_2647_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_2647_end_mask_0 = const()[name = string("op_2647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_2647_cast_fp16 = slice_by_index(begin = var_2647_begin_0, end = var_2647_end_0, end_mask = var_2647_end_mask_0, x = q_5_cast_fp16)[name = string("op_2647_cast_fp16")];
            int32 var_2649 = const()[name = string("op_2649"), val = int32(-1)];
            bool var_2650_interleave_0 = const()[name = string("op_2650_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_2650_cast_fp16 = concat(axis = var_2649, interleave = var_2650_interleave_0, values = (var_2642_cast_fp16, var_2647_cast_fp16))[name = string("op_2650_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2651_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_7_cast_fp16 = add(x = var_2636_cast_fp16, y = var_2651_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2654_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = k_7_cast_fp16)[name = string("op_2659_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2660_cast_fp16 = mul(x = var_2659_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2660_cast_fp16")];
            tensor<int32, [4]> var_2665_begin_0 = const()[name = string("op_2665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2665_end_0 = const()[name = string("op_2665_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_2665_end_mask_0 = const()[name = string("op_2665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_2665_cast_fp16 = slice_by_index(begin = var_2665_begin_0, end = var_2665_end_0, end_mask = var_2665_end_mask_0, x = k_7_cast_fp16)[name = string("op_2665_cast_fp16")];
            int32 var_2667 = const()[name = string("op_2667"), val = int32(-1)];
            bool var_2668_interleave_0 = const()[name = string("op_2668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_2668_cast_fp16 = concat(axis = var_2667, interleave = var_2668_interleave_0, values = (var_2660_cast_fp16, var_2665_cast_fp16))[name = string("op_2668_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2669_cast_fp16 = mul(x = var_2668_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_9_cast_fp16 = add(x = var_2654_cast_fp16, y = var_2669_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = k_9_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [4, 1, 64, 256]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_4_cast_fp16 = reshape(shape = concat_4, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [4, 1, 64, 256]> reshape_5_cast_fp16 = reshape(shape = concat_5, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = v_5_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [4, 1, 64, 256]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_6_cast_fp16 = reshape(shape = concat_6, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [4, 1, 64, 256]> reshape_7_cast_fp16 = reshape(shape = concat_7, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2683_transpose_x_0 = const()[name = string("op_2683_transpose_x_0"), val = bool(false)];
            bool var_2683_transpose_y_0 = const()[name = string("op_2683_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_97_cast_fp16 = transpose(perm = transpose_97_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 4, 64, 64]> var_2683_cast_fp16 = matmul(transpose_x = var_2683_transpose_x_0, transpose_y = var_2683_transpose_y_0, x = q_7_cast_fp16, y = transpose_97_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_2685_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2173_cast_fp16)[name = string("op_2685_cast_fp16")];
            int32 var_2686 = const()[name = string("op_2686"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_25_cast_fp16 = softmax(axis = var_2686, x = var_2685_cast_fp16)[name = string("w_25_cast_fp16")];
            bool var_2689_transpose_x_0 = const()[name = string("op_2689_transpose_x_0"), val = bool(false)];
            bool var_2689_transpose_y_0 = const()[name = string("op_2689_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 4, 64, 256]> var_2689_cast_fp16 = matmul(transpose_x = var_2689_transpose_x_0, transpose_y = var_2689_transpose_y_0, x = w_25_cast_fp16, y = v_7_cast_fp16)[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> var_2692_perm_0 = const()[name = string("op_2692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2696 = const()[name = string("op_2696"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_2692_cast_fp16 = transpose(perm = var_2692_perm_0, x = var_2689_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 64, 1024]> input_15_cast_fp16 = reshape(shape = var_2696, x = var_2692_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13197632))))[name = string("tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2701_cast_fp16 = abs(x = linear_13_cast_fp16)[name = string("op_2701_cast_fp16")];
            bool var_2704 = const()[name = string("op_2704"), val = bool(true)];
            tensor<int32, [1]> var_2705_axes_0 = const()[name = string("op_2705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2705_cast_fp16 = reduce_max(axes = var_2705_axes_0, keep_dims = var_2704, x = var_2701_cast_fp16)[name = string("op_2705_cast_fp16")];
            fp16 var_2706_to_fp16 = const()[name = string("op_2706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_27_cast_fp16 = maximum(x = var_2705_cast_fp16, y = var_2706_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_27_cast_fp16 = real_div(x = linear_13_cast_fp16, y = s_27_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2709_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2709_cast_fp16")];
            tensor<int32, [1]> var_2714_axes_0 = const()[name = string("op_2714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2714_keep_dims_0 = const()[name = string("op_2714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2714_cast_fp16 = reduce_mean(axes = var_2714_axes_0, keep_dims = var_2714_keep_dims_0, x = var_2709_cast_fp16)[name = string("op_2714_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2715_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2715_cast_fp16")];
            fp32 var_2716_epsilon_0 = const()[name = string("op_2716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2716_cast_fp16 = inverse(epsilon = var_2716_epsilon_0, x = var_2715_cast_fp16)[name = string("op_2716_cast_fp16")];
            fp16 var_2717_to_fp16 = const()[name = string("op_2717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2718_cast_fp16 = mul(x = var_2716_cast_fp16, y = var_2717_to_fp16)[name = string("op_2718_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2720_cast_fp16 = add(x = var_2714_cast_fp16, y = var_2718_cast_fp16)[name = string("op_2720_cast_fp16")];
            fp32 var_2721_epsilon_0 = const()[name = string("op_2721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2721_cast_fp16 = rsqrt(epsilon = var_2721_epsilon_0, x = var_2720_cast_fp16)[name = string("op_2721_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2721_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13198720)))];
            tensor<fp16, [1, 64, 512]> var_2723_cast_fp16 = mul(x = out_23_cast_fp16, y = tm_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2723_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_29_cast_fp16 = add(x = x_19_cast_fp16, y = var_2723_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2726_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_2726_cast_fp16")];
            bool var_2729 = const()[name = string("op_2729"), val = bool(true)];
            tensor<int32, [1]> var_2730_axes_0 = const()[name = string("op_2730_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2730_cast_fp16 = reduce_max(axes = var_2730_axes_0, keep_dims = var_2729, x = var_2726_cast_fp16)[name = string("op_2730_cast_fp16")];
            fp16 var_2731_to_fp16 = const()[name = string("op_2731_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_29_cast_fp16 = maximum(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_29_cast_fp16 = real_div(x = x_29_cast_fp16, y = s_29_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2734_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2734_cast_fp16")];
            tensor<int32, [1]> var_2739_axes_0 = const()[name = string("op_2739_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2739_keep_dims_0 = const()[name = string("op_2739_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2739_cast_fp16 = reduce_mean(axes = var_2739_axes_0, keep_dims = var_2739_keep_dims_0, x = var_2734_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2740_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2740_cast_fp16")];
            fp32 var_2741_epsilon_0 = const()[name = string("op_2741_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2741_cast_fp16 = inverse(epsilon = var_2741_epsilon_0, x = var_2740_cast_fp16)[name = string("op_2741_cast_fp16")];
            fp16 var_2742_to_fp16 = const()[name = string("op_2742_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2743_cast_fp16 = mul(x = var_2741_cast_fp16, y = var_2742_to_fp16)[name = string("op_2743_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2745_cast_fp16 = add(x = var_2739_cast_fp16, y = var_2743_cast_fp16)[name = string("op_2745_cast_fp16")];
            fp32 var_2746_epsilon_0 = const()[name = string("op_2746_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2746_cast_fp16 = rsqrt(epsilon = var_2746_epsilon_0, x = var_2745_cast_fp16)[name = string("op_2746_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2746_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13199808)))];
            tensor<fp16, [1, 64, 512]> input_17_cast_fp16 = mul(x = out_25_cast_fp16, y = tm_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13200896))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14249536))))[name = string("tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_14_cast_fp16")];
            string var_2753_mode_0 = const()[name = string("op_2753_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_2753_cast_fp16 = gelu(mode = var_2753_mode_0, x = linear_14_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14253696))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15302336))))[name = string("tm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_15_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_19_cast_fp16 = mul(x = var_2753_cast_fp16, y = linear_15_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15306496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16355136))))[name = string("tm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2761_cast_fp16 = abs(x = linear_16_cast_fp16)[name = string("op_2761_cast_fp16")];
            bool var_2764 = const()[name = string("op_2764"), val = bool(true)];
            tensor<int32, [1]> var_2765_axes_0 = const()[name = string("op_2765_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2765_cast_fp16 = reduce_max(axes = var_2765_axes_0, keep_dims = var_2764, x = var_2761_cast_fp16)[name = string("op_2765_cast_fp16")];
            fp16 var_2766_to_fp16 = const()[name = string("op_2766_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_31_cast_fp16 = maximum(x = var_2765_cast_fp16, y = var_2766_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_31_cast_fp16 = real_div(x = linear_16_cast_fp16, y = s_31_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2769_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2769_cast_fp16")];
            tensor<int32, [1]> var_2774_axes_0 = const()[name = string("op_2774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2774_keep_dims_0 = const()[name = string("op_2774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2774_cast_fp16 = reduce_mean(axes = var_2774_axes_0, keep_dims = var_2774_keep_dims_0, x = var_2769_cast_fp16)[name = string("op_2774_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2775_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp32 var_2776_epsilon_0 = const()[name = string("op_2776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2776_cast_fp16 = inverse(epsilon = var_2776_epsilon_0, x = var_2775_cast_fp16)[name = string("op_2776_cast_fp16")];
            fp16 var_2777_to_fp16 = const()[name = string("op_2777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2778_cast_fp16 = mul(x = var_2776_cast_fp16, y = var_2777_to_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2780_cast_fp16 = add(x = var_2774_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2780_cast_fp16")];
            fp32 var_2781_epsilon_0 = const()[name = string("op_2781_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2781_cast_fp16 = rsqrt(epsilon = var_2781_epsilon_0, x = var_2780_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2781_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16356224)))];
            tensor<fp16, [1, 64, 512]> var_2783_cast_fp16 = mul(x = out_27_cast_fp16, y = tm_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2783_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_21_cast_fp16 = add(x = x_29_cast_fp16, y = var_2783_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16357312))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16619520))))[name = string("tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_17_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            string x_33_mode_0 = const()[name = string("x_33_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_33_cast_fp16 = gelu(mode = x_33_mode_0, x = linear_17_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> var_2803_begin_0 = const()[name = string("op_2803_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2803_end_0 = const()[name = string("op_2803_end_0"), val = tensor<int32, [4]>([1, 64, 2, 512])];
            tensor<bool, [4]> var_2803_end_mask_0 = const()[name = string("op_2803_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2803_squeeze_mask_0 = const()[name = string("op_2803_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_2803_cast_fp16 = slice_by_index(begin = var_2803_begin_0, end = var_2803_end_0, end_mask = var_2803_end_mask_0, squeeze_mask = var_2803_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_23_cast_fp16 = mul(x = x_33_cast_fp16, y = var_2803_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16620608))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16882816))))[name = string("tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2813_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2813_cast_fp16")];
            bool var_2816 = const()[name = string("op_2816"), val = bool(true)];
            tensor<int32, [1]> var_2817_axes_0 = const()[name = string("op_2817_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2817_cast_fp16 = reduce_max(axes = var_2817_axes_0, keep_dims = var_2816, x = var_2813_cast_fp16)[name = string("op_2817_cast_fp16")];
            fp16 var_2818_to_fp16 = const()[name = string("op_2818_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_33_cast_fp16 = maximum(x = var_2817_cast_fp16, y = var_2818_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_33_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_33_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2821_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<int32, [1]> var_2826_axes_0 = const()[name = string("op_2826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2826_keep_dims_0 = const()[name = string("op_2826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2826_cast_fp16 = reduce_mean(axes = var_2826_axes_0, keep_dims = var_2826_keep_dims_0, x = var_2821_cast_fp16)[name = string("op_2826_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2827_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp32 var_2828_epsilon_0 = const()[name = string("op_2828_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2828_cast_fp16 = inverse(epsilon = var_2828_epsilon_0, x = var_2827_cast_fp16)[name = string("op_2828_cast_fp16")];
            fp16 var_2829_to_fp16 = const()[name = string("op_2829_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2830_cast_fp16 = mul(x = var_2828_cast_fp16, y = var_2829_to_fp16)[name = string("op_2830_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2832_cast_fp16 = add(x = var_2826_cast_fp16, y = var_2830_cast_fp16)[name = string("op_2832_cast_fp16")];
            fp32 var_2833_epsilon_0 = const()[name = string("op_2833_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2833_cast_fp16 = rsqrt(epsilon = var_2833_epsilon_0, x = var_2832_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2833_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16883904)))];
            tensor<fp16, [1, 64, 512]> var_2835_cast_fp16 = mul(x = out_29_cast_fp16, y = tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2835_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_3_cast_fp16 = add(x = input_21_cast_fp16, y = var_2835_cast_fp16)[name = string("h_3_cast_fp16")];
            tensor<fp16, [1]> tm_layers_1_layer_scalar_to_fp16 = const()[name = string("tm_layers_1_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.5ap-3])];
            tensor<fp16, [1, 64, 512]> x_37_cast_fp16 = mul(x = h_3_cast_fp16, y = tm_layers_1_layer_scalar_to_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2839_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2839_cast_fp16")];
            bool var_2842 = const()[name = string("op_2842"), val = bool(true)];
            tensor<int32, [1]> var_2843_axes_0 = const()[name = string("op_2843_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2843_cast_fp16 = reduce_max(axes = var_2843_axes_0, keep_dims = var_2842, x = var_2839_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp16 var_2844_to_fp16 = const()[name = string("op_2844_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_35_cast_fp16 = maximum(x = var_2843_cast_fp16, y = var_2844_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_35_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_35_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_2847_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2847_cast_fp16")];
            tensor<int32, [1]> var_2852_axes_0 = const()[name = string("op_2852_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2852_keep_dims_0 = const()[name = string("op_2852_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_2852_cast_fp16 = reduce_mean(axes = var_2852_axes_0, keep_dims = var_2852_keep_dims_0, x = var_2847_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2853_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp32 var_2854_epsilon_0 = const()[name = string("op_2854_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2854_cast_fp16 = inverse(epsilon = var_2854_epsilon_0, x = var_2853_cast_fp16)[name = string("op_2854_cast_fp16")];
            fp16 var_2855_to_fp16 = const()[name = string("op_2855_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2856_cast_fp16 = mul(x = var_2854_cast_fp16, y = var_2855_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2858_cast_fp16 = add(x = var_2852_cast_fp16, y = var_2856_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp32 var_2859_epsilon_0 = const()[name = string("op_2859_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2859_cast_fp16 = rsqrt(epsilon = var_2859_epsilon_0, x = var_2858_cast_fp16)[name = string("op_2859_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_31_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2859_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16884992)))];
            tensor<fp16, [1, 64, 512]> input_25_cast_fp16 = mul(x = out_31_cast_fp16, y = tm_layers_2_input_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16886080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17410432))))[name = string("tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [4]> var_2869 = const()[name = string("op_2869"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_2870_cast_fp16 = reshape(shape = var_2869, x = linear_19_cast_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17412544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17674752))))[name = string("tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [4]> var_2881 = const()[name = string("op_2881"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_2882_cast_fp16 = reshape(shape = var_2881, x = linear_20_cast_fp16)[name = string("op_2882_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17675840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17938048))))[name = string("tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_2894_cast_fp16 = reshape(shape = var_2893, x = linear_21_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [4]> x_43_perm_0 = const()[name = string("x_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2870_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 4, 64, 256]> var_2898_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2898_cast_fp16")];
            bool var_2901 = const()[name = string("op_2901"), val = bool(true)];
            tensor<int32, [1]> var_2902_axes_0 = const()[name = string("op_2902_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_2902_cast_fp16 = reduce_max(axes = var_2902_axes_0, keep_dims = var_2901, x = var_2898_cast_fp16)[name = string("op_2902_cast_fp16")];
            fp16 var_2903_to_fp16 = const()[name = string("op_2903_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_37_cast_fp16 = maximum(x = var_2902_cast_fp16, y = var_2903_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_37_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_37_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_2906_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<int32, [1]> var_2911_axes_0 = const()[name = string("op_2911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2911_keep_dims_0 = const()[name = string("op_2911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_2911_cast_fp16 = reduce_mean(axes = var_2911_axes_0, keep_dims = var_2911_keep_dims_0, x = var_2906_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_2912_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2912_cast_fp16")];
            fp32 var_2913_epsilon_0 = const()[name = string("op_2913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_2913_cast_fp16 = inverse(epsilon = var_2913_epsilon_0, x = var_2912_cast_fp16)[name = string("op_2913_cast_fp16")];
            fp16 var_2914_to_fp16 = const()[name = string("op_2914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_2915_cast_fp16 = mul(x = var_2913_cast_fp16, y = var_2914_to_fp16)[name = string("op_2915_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_2917_cast_fp16 = add(x = var_2911_cast_fp16, y = var_2915_cast_fp16)[name = string("op_2917_cast_fp16")];
            fp32 var_2918_epsilon_0 = const()[name = string("op_2918_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_2918_cast_fp16 = rsqrt(epsilon = var_2918_epsilon_0, x = var_2917_cast_fp16)[name = string("op_2918_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_33_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2918_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_9_cast_fp16 = mul(x = out_33_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2882_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 2, 64, 256]> var_2921_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2921_cast_fp16")];
            bool var_2924 = const()[name = string("op_2924"), val = bool(true)];
            tensor<int32, [1]> var_2925_axes_0 = const()[name = string("op_2925_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_2925_cast_fp16 = reduce_max(axes = var_2925_axes_0, keep_dims = var_2924, x = var_2921_cast_fp16)[name = string("op_2925_cast_fp16")];
            fp16 var_2926_to_fp16 = const()[name = string("op_2926_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_39_cast_fp16 = maximum(x = var_2925_cast_fp16, y = var_2926_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_39_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_39_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2929_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2929_cast_fp16")];
            tensor<int32, [1]> var_2934_axes_0 = const()[name = string("op_2934_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2934_keep_dims_0 = const()[name = string("op_2934_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_2934_cast_fp16 = reduce_mean(axes = var_2934_axes_0, keep_dims = var_2934_keep_dims_0, x = var_2929_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2935_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2935_cast_fp16")];
            fp32 var_2936_epsilon_0 = const()[name = string("op_2936_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_2936_cast_fp16 = inverse(epsilon = var_2936_epsilon_0, x = var_2935_cast_fp16)[name = string("op_2936_cast_fp16")];
            fp16 var_2937_to_fp16 = const()[name = string("op_2937_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_2938_cast_fp16 = mul(x = var_2936_cast_fp16, y = var_2937_to_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2940_cast_fp16 = add(x = var_2934_cast_fp16, y = var_2938_cast_fp16)[name = string("op_2940_cast_fp16")];
            fp32 var_2941_epsilon_0 = const()[name = string("op_2941_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_2941_cast_fp16 = rsqrt(epsilon = var_2941_epsilon_0, x = var_2940_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_35_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2941_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_13_cast_fp16 = mul(x = out_35_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_43_cast_fp16 = transpose(perm = x_43_perm_0, x = var_2894_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 2, 64, 256]> var_2944_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2944_cast_fp16")];
            bool var_2947 = const()[name = string("op_2947"), val = bool(true)];
            tensor<int32, [1]> var_2948_axes_0 = const()[name = string("op_2948_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_2948_cast_fp16 = reduce_max(axes = var_2948_axes_0, keep_dims = var_2947, x = var_2944_cast_fp16)[name = string("op_2948_cast_fp16")];
            fp16 var_2949_to_fp16 = const()[name = string("op_2949_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_41_cast_fp16 = maximum(x = var_2948_cast_fp16, y = var_2949_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_41_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_41_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2952_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2952_cast_fp16")];
            tensor<int32, [1]> var_2957_axes_0 = const()[name = string("op_2957_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2957_keep_dims_0 = const()[name = string("op_2957_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_2957_cast_fp16 = reduce_mean(axes = var_2957_axes_0, keep_dims = var_2957_keep_dims_0, x = var_2952_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2958_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2958_cast_fp16")];
            fp32 var_2959_epsilon_0 = const()[name = string("op_2959_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_2959_cast_fp16 = inverse(epsilon = var_2959_epsilon_0, x = var_2958_cast_fp16)[name = string("op_2959_cast_fp16")];
            fp16 var_2960_to_fp16 = const()[name = string("op_2960_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_2961_cast_fp16 = mul(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_2963_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2961_cast_fp16)[name = string("op_2963_cast_fp16")];
            fp32 var_2964_epsilon_0 = const()[name = string("op_2964_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_2964_cast_fp16 = rsqrt(epsilon = var_2964_epsilon_0, x = var_2963_cast_fp16)[name = string("op_2964_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_9_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2964_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_2966_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2966_cast_fp16")];
            tensor<int32, [4]> var_2971_begin_0 = const()[name = string("op_2971_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2971_end_0 = const()[name = string("op_2971_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_2971_end_mask_0 = const()[name = string("op_2971_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_2971_cast_fp16 = slice_by_index(begin = var_2971_begin_0, end = var_2971_end_0, end_mask = var_2971_end_mask_0, x = q_9_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_2972_cast_fp16 = mul(x = var_2971_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2972_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            int32 var_2979 = const()[name = string("op_2979"), val = int32(-1)];
            bool var_2980_interleave_0 = const()[name = string("op_2980_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_2980_cast_fp16 = concat(axis = var_2979, interleave = var_2980_interleave_0, values = (var_2972_cast_fp16, var_2977_cast_fp16))[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_2981_cast_fp16 = mul(x = var_2980_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2981_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_11_cast_fp16 = add(x = var_2966_cast_fp16, y = var_2981_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2984_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2989_begin_0 = const()[name = string("op_2989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2989_end_0 = const()[name = string("op_2989_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_2989_end_mask_0 = const()[name = string("op_2989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_2989_cast_fp16 = slice_by_index(begin = var_2989_begin_0, end = var_2989_end_0, end_mask = var_2989_end_mask_0, x = k_13_cast_fp16)[name = string("op_2989_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_2990_cast_fp16 = mul(x = var_2989_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2990_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = k_13_cast_fp16)[name = string("op_2995_cast_fp16")];
            int32 var_2997 = const()[name = string("op_2997"), val = int32(-1)];
            bool var_2998_interleave_0 = const()[name = string("op_2998_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_2998_cast_fp16 = concat(axis = var_2997, interleave = var_2998_interleave_0, values = (var_2990_cast_fp16, var_2995_cast_fp16))[name = string("op_2998_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_2999_cast_fp16 = mul(x = var_2998_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_2999_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_15_cast_fp16 = add(x = var_2984_cast_fp16, y = var_2999_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = k_15_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [4, 1, 64, 256]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
            tensor<int32, [5]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_8_cast_fp16 = reshape(shape = concat_8, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [4, 1, 64, 256]> reshape_9_cast_fp16 = reshape(shape = concat_9, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = v_9_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [4, 1, 64, 256]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
            tensor<int32, [5]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_10_cast_fp16 = reshape(shape = concat_10, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [4, 1, 64, 256]> reshape_11_cast_fp16 = reshape(shape = concat_11, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3013_transpose_x_0 = const()[name = string("op_3013_transpose_x_0"), val = bool(false)];
            bool var_3013_transpose_y_0 = const()[name = string("op_3013_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 4, 64, 64]> var_3013_cast_fp16 = matmul(transpose_x = var_3013_transpose_x_0, transpose_y = var_3013_transpose_y_0, x = q_11_cast_fp16, y = transpose_98_cast_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_3015_cast_fp16 = add(x = var_3013_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3015_cast_fp16")];
            int32 var_3016 = const()[name = string("op_3016"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_41_cast_fp16 = softmax(axis = var_3016, x = var_3015_cast_fp16)[name = string("w_41_cast_fp16")];
            bool var_3019_transpose_x_0 = const()[name = string("op_3019_transpose_x_0"), val = bool(false)];
            bool var_3019_transpose_y_0 = const()[name = string("op_3019_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 4, 64, 256]> var_3019_cast_fp16 = matmul(transpose_x = var_3019_transpose_x_0, transpose_y = var_3019_transpose_y_0, x = w_41_cast_fp16, y = v_11_cast_fp16)[name = string("op_3019_cast_fp16")];
            tensor<int32, [4]> var_3022_perm_0 = const()[name = string("op_3022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3026 = const()[name = string("op_3026"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_3022_cast_fp16 = transpose(perm = var_3022_perm_0, x = var_3019_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 64, 1024]> input_27_cast_fp16 = reshape(shape = var_3026, x = var_3022_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17939136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18463488))))[name = string("tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3031_cast_fp16 = abs(x = linear_22_cast_fp16)[name = string("op_3031_cast_fp16")];
            bool var_3034 = const()[name = string("op_3034"), val = bool(true)];
            tensor<int32, [1]> var_3035_axes_0 = const()[name = string("op_3035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3035_cast_fp16 = reduce_max(axes = var_3035_axes_0, keep_dims = var_3034, x = var_3031_cast_fp16)[name = string("op_3035_cast_fp16")];
            fp16 var_3036_to_fp16 = const()[name = string("op_3036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_43_cast_fp16 = maximum(x = var_3035_cast_fp16, y = var_3036_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_43_cast_fp16 = real_div(x = linear_22_cast_fp16, y = s_43_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3039_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3039_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3045_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_3045_cast_fp16")];
            fp32 var_3046_epsilon_0 = const()[name = string("op_3046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3046_cast_fp16 = inverse(epsilon = var_3046_epsilon_0, x = var_3045_cast_fp16)[name = string("op_3046_cast_fp16")];
            fp16 var_3047_to_fp16 = const()[name = string("op_3047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3048_cast_fp16 = mul(x = var_3046_cast_fp16, y = var_3047_to_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3050_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3048_cast_fp16)[name = string("op_3050_cast_fp16")];
            fp32 var_3051_epsilon_0 = const()[name = string("op_3051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3051_cast_fp16 = rsqrt(epsilon = var_3051_epsilon_0, x = var_3050_cast_fp16)[name = string("op_3051_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_3051_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18464576)))];
            tensor<fp16, [1, 64, 512]> var_3053_cast_fp16 = mul(x = out_37_cast_fp16, y = tm_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = var_3053_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3056_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_3056_cast_fp16")];
            bool var_3059 = const()[name = string("op_3059"), val = bool(true)];
            tensor<int32, [1]> var_3060_axes_0 = const()[name = string("op_3060_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3060_cast_fp16 = reduce_max(axes = var_3060_axes_0, keep_dims = var_3059, x = var_3056_cast_fp16)[name = string("op_3060_cast_fp16")];
            fp16 var_3061_to_fp16 = const()[name = string("op_3061_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_45_cast_fp16 = maximum(x = var_3060_cast_fp16, y = var_3061_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_45_cast_fp16 = real_div(x = x_47_cast_fp16, y = s_45_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3064_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_3064_cast_fp16")];
            tensor<int32, [1]> var_3069_axes_0 = const()[name = string("op_3069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3069_keep_dims_0 = const()[name = string("op_3069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3069_cast_fp16 = reduce_mean(axes = var_3069_axes_0, keep_dims = var_3069_keep_dims_0, x = var_3064_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3070_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp32 var_3071_epsilon_0 = const()[name = string("op_3071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3071_cast_fp16 = inverse(epsilon = var_3071_epsilon_0, x = var_3070_cast_fp16)[name = string("op_3071_cast_fp16")];
            fp16 var_3072_to_fp16 = const()[name = string("op_3072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3073_cast_fp16 = mul(x = var_3071_cast_fp16, y = var_3072_to_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3075_cast_fp16 = add(x = var_3069_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3075_cast_fp16")];
            fp32 var_3076_epsilon_0 = const()[name = string("op_3076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3076_cast_fp16 = rsqrt(epsilon = var_3076_epsilon_0, x = var_3075_cast_fp16)[name = string("op_3076_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_3076_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18465664)))];
            tensor<fp16, [1, 64, 512]> input_29_cast_fp16 = mul(x = out_39_cast_fp16, y = tm_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18466752))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19515392))))[name = string("tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_23_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_23_cast_fp16")];
            string var_3083_mode_0 = const()[name = string("op_3083_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_3083_cast_fp16 = gelu(mode = var_3083_mode_0, x = linear_23_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19519552))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20568192))))[name = string("tm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_24_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_31_cast_fp16 = mul(x = var_3083_cast_fp16, y = linear_24_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20572352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21620992))))[name = string("tm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3091_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_3091_cast_fp16")];
            bool var_3094 = const()[name = string("op_3094"), val = bool(true)];
            tensor<int32, [1]> var_3095_axes_0 = const()[name = string("op_3095_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3095_cast_fp16 = reduce_max(axes = var_3095_axes_0, keep_dims = var_3094, x = var_3091_cast_fp16)[name = string("op_3095_cast_fp16")];
            fp16 var_3096_to_fp16 = const()[name = string("op_3096_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_47_cast_fp16 = maximum(x = var_3095_cast_fp16, y = var_3096_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_47_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_47_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3099_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3104_keep_dims_0 = const()[name = string("op_3104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3104_cast_fp16 = reduce_mean(axes = var_3104_axes_0, keep_dims = var_3104_keep_dims_0, x = var_3099_cast_fp16)[name = string("op_3104_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3105_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_3105_cast_fp16")];
            fp32 var_3106_epsilon_0 = const()[name = string("op_3106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3106_cast_fp16 = inverse(epsilon = var_3106_epsilon_0, x = var_3105_cast_fp16)[name = string("op_3106_cast_fp16")];
            fp16 var_3107_to_fp16 = const()[name = string("op_3107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3108_cast_fp16 = mul(x = var_3106_cast_fp16, y = var_3107_to_fp16)[name = string("op_3108_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3110_cast_fp16 = add(x = var_3104_cast_fp16, y = var_3108_cast_fp16)[name = string("op_3110_cast_fp16")];
            fp32 var_3111_epsilon_0 = const()[name = string("op_3111_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3111_cast_fp16 = rsqrt(epsilon = var_3111_epsilon_0, x = var_3110_cast_fp16)[name = string("op_3111_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_3111_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21622080)))];
            tensor<fp16, [1, 64, 512]> var_3113_cast_fp16 = mul(x = out_41_cast_fp16, y = tm_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_33_cast_fp16 = add(x = x_47_cast_fp16, y = var_3113_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21623168))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21885376))))[name = string("tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_26_cast_fp16")];
            string x_51_mode_0 = const()[name = string("x_51_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_51_cast_fp16 = gelu(mode = x_51_mode_0, x = linear_26_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<int32, [4]> var_3133_begin_0 = const()[name = string("op_3133_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_3133_end_0 = const()[name = string("op_3133_end_0"), val = tensor<int32, [4]>([1, 64, 3, 512])];
            tensor<bool, [4]> var_3133_end_mask_0 = const()[name = string("op_3133_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3133_squeeze_mask_0 = const()[name = string("op_3133_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_3133_cast_fp16 = slice_by_index(begin = var_3133_begin_0, end = var_3133_end_0, end_mask = var_3133_end_mask_0, squeeze_mask = var_3133_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_35_cast_fp16 = mul(x = x_51_cast_fp16, y = var_3133_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21886464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22148672))))[name = string("tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3143_cast_fp16 = abs(x = linear_27_cast_fp16)[name = string("op_3143_cast_fp16")];
            bool var_3146 = const()[name = string("op_3146"), val = bool(true)];
            tensor<int32, [1]> var_3147_axes_0 = const()[name = string("op_3147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3147_cast_fp16 = reduce_max(axes = var_3147_axes_0, keep_dims = var_3146, x = var_3143_cast_fp16)[name = string("op_3147_cast_fp16")];
            fp16 var_3148_to_fp16 = const()[name = string("op_3148_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_49_cast_fp16 = maximum(x = var_3147_cast_fp16, y = var_3148_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_49_cast_fp16 = real_div(x = linear_27_cast_fp16, y = s_49_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3151_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_3151_cast_fp16")];
            tensor<int32, [1]> var_3156_axes_0 = const()[name = string("op_3156_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3156_keep_dims_0 = const()[name = string("op_3156_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3156_cast_fp16 = reduce_mean(axes = var_3156_axes_0, keep_dims = var_3156_keep_dims_0, x = var_3151_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3157_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_3157_cast_fp16")];
            fp32 var_3158_epsilon_0 = const()[name = string("op_3158_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3158_cast_fp16 = inverse(epsilon = var_3158_epsilon_0, x = var_3157_cast_fp16)[name = string("op_3158_cast_fp16")];
            fp16 var_3159_to_fp16 = const()[name = string("op_3159_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3160_cast_fp16 = mul(x = var_3158_cast_fp16, y = var_3159_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3162_cast_fp16 = add(x = var_3156_cast_fp16, y = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp32 var_3163_epsilon_0 = const()[name = string("op_3163_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3163_cast_fp16 = rsqrt(epsilon = var_3163_epsilon_0, x = var_3162_cast_fp16)[name = string("op_3163_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_43_cast_fp16 = mul(x = y_49_cast_fp16, y = var_3163_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22149760)))];
            tensor<fp16, [1, 64, 512]> var_3165_cast_fp16 = mul(x = out_43_cast_fp16, y = tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_5_cast_fp16 = add(x = input_33_cast_fp16, y = var_3165_cast_fp16)[name = string("h_5_cast_fp16")];
            tensor<fp16, [1]> tm_layers_2_layer_scalar_to_fp16 = const()[name = string("tm_layers_2_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ep-1])];
            tensor<fp16, [1, 64, 512]> x_55_cast_fp16 = mul(x = h_5_cast_fp16, y = tm_layers_2_layer_scalar_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3169_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_3169_cast_fp16")];
            bool var_3172 = const()[name = string("op_3172"), val = bool(true)];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3173_cast_fp16 = reduce_max(axes = var_3173_axes_0, keep_dims = var_3172, x = var_3169_cast_fp16)[name = string("op_3173_cast_fp16")];
            fp16 var_3174_to_fp16 = const()[name = string("op_3174_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_51_cast_fp16 = maximum(x = var_3173_cast_fp16, y = var_3174_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_51_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_51_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3177_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_3177_cast_fp16")];
            tensor<int32, [1]> var_3182_axes_0 = const()[name = string("op_3182_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3182_keep_dims_0 = const()[name = string("op_3182_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3182_cast_fp16 = reduce_mean(axes = var_3182_axes_0, keep_dims = var_3182_keep_dims_0, x = var_3177_cast_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3183_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_3183_cast_fp16")];
            fp32 var_3184_epsilon_0 = const()[name = string("op_3184_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3184_cast_fp16 = inverse(epsilon = var_3184_epsilon_0, x = var_3183_cast_fp16)[name = string("op_3184_cast_fp16")];
            fp16 var_3185_to_fp16 = const()[name = string("op_3185_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3186_cast_fp16 = mul(x = var_3184_cast_fp16, y = var_3185_to_fp16)[name = string("op_3186_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3188_cast_fp16 = add(x = var_3182_cast_fp16, y = var_3186_cast_fp16)[name = string("op_3188_cast_fp16")];
            fp32 var_3189_epsilon_0 = const()[name = string("op_3189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3189_cast_fp16 = rsqrt(epsilon = var_3189_epsilon_0, x = var_3188_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_45_cast_fp16 = mul(x = y_51_cast_fp16, y = var_3189_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22150848)))];
            tensor<fp16, [1, 64, 512]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = tm_layers_3_input_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22151936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22676288))))[name = string("tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_3200_cast_fp16 = reshape(shape = var_3199, x = linear_28_cast_fp16)[name = string("op_3200_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22678400))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22940608))))[name = string("tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_29_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_59_perm_0 = const()[name = string("x_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22941696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23203904))))[name = string("tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_3223 = const()[name = string("op_3223"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_3224_cast_fp16 = reshape(shape = var_3223, x = linear_30_cast_fp16)[name = string("op_3224_cast_fp16")];
            tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_3200_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 4, 64, 256]> var_3228_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3231 = const()[name = string("op_3231"), val = bool(true)];
            tensor<int32, [1]> var_3232_axes_0 = const()[name = string("op_3232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_3232_cast_fp16 = reduce_max(axes = var_3232_axes_0, keep_dims = var_3231, x = var_3228_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_53_cast_fp16 = maximum(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_53_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_53_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_3236_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_3236_cast_fp16")];
            tensor<int32, [1]> var_3241_axes_0 = const()[name = string("op_3241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3241_keep_dims_0 = const()[name = string("op_3241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_3241_cast_fp16 = reduce_mean(axes = var_3241_axes_0, keep_dims = var_3241_keep_dims_0, x = var_3236_cast_fp16)[name = string("op_3241_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_3242_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_3242_cast_fp16")];
            fp32 var_3243_epsilon_0 = const()[name = string("op_3243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_3243_cast_fp16 = inverse(epsilon = var_3243_epsilon_0, x = var_3242_cast_fp16)[name = string("op_3243_cast_fp16")];
            fp16 var_3244_to_fp16 = const()[name = string("op_3244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_3245_cast_fp16 = mul(x = var_3243_cast_fp16, y = var_3244_to_fp16)[name = string("op_3245_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_3247_cast_fp16 = add(x = var_3241_cast_fp16, y = var_3245_cast_fp16)[name = string("op_3247_cast_fp16")];
            fp32 var_3248_epsilon_0 = const()[name = string("op_3248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_3248_cast_fp16 = rsqrt(epsilon = var_3248_epsilon_0, x = var_3247_cast_fp16)[name = string("op_3248_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_47_cast_fp16 = mul(x = y_53_cast_fp16, y = var_3248_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23204992)))];
            tensor<fp16, [1, 4, 64, 256]> q_13_cast_fp16 = mul(x = out_47_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_59_cast_fp16 = transpose(perm = x_59_perm_0, x = var_3212_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 2, 64, 256]> var_3251_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_3251_cast_fp16")];
            bool var_3254 = const()[name = string("op_3254"), val = bool(true)];
            tensor<int32, [1]> var_3255_axes_0 = const()[name = string("op_3255_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_3255_cast_fp16 = reduce_max(axes = var_3255_axes_0, keep_dims = var_3254, x = var_3251_cast_fp16)[name = string("op_3255_cast_fp16")];
            fp16 var_3256_to_fp16 = const()[name = string("op_3256_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_55_cast_fp16 = maximum(x = var_3255_cast_fp16, y = var_3256_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_55_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_55_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_3259_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<int32, [1]> var_3264_axes_0 = const()[name = string("op_3264_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3264_keep_dims_0 = const()[name = string("op_3264_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_3264_cast_fp16 = reduce_mean(axes = var_3264_axes_0, keep_dims = var_3264_keep_dims_0, x = var_3259_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_3265_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_3265_cast_fp16")];
            fp32 var_3266_epsilon_0 = const()[name = string("op_3266_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_3266_cast_fp16 = inverse(epsilon = var_3266_epsilon_0, x = var_3265_cast_fp16)[name = string("op_3266_cast_fp16")];
            fp16 var_3267_to_fp16 = const()[name = string("op_3267_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_3268_cast_fp16 = mul(x = var_3266_cast_fp16, y = var_3267_to_fp16)[name = string("op_3268_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_3270_cast_fp16 = add(x = var_3264_cast_fp16, y = var_3268_cast_fp16)[name = string("op_3270_cast_fp16")];
            fp32 var_3271_epsilon_0 = const()[name = string("op_3271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_3271_cast_fp16 = rsqrt(epsilon = var_3271_epsilon_0, x = var_3270_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_49_cast_fp16 = mul(x = y_55_cast_fp16, y = var_3271_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23205568)))];
            tensor<fp16, [1, 2, 64, 256]> k_19_cast_fp16 = mul(x = out_49_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_3224_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2, 64, 256]> var_3274_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_3274_cast_fp16")];
            bool var_3277 = const()[name = string("op_3277"), val = bool(true)];
            tensor<int32, [1]> var_3278_axes_0 = const()[name = string("op_3278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_3278_cast_fp16 = reduce_max(axes = var_3278_axes_0, keep_dims = var_3277, x = var_3274_cast_fp16)[name = string("op_3278_cast_fp16")];
            fp16 var_3279_to_fp16 = const()[name = string("op_3279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_57_cast_fp16 = maximum(x = var_3278_cast_fp16, y = var_3279_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_57_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_57_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_3282_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_3282_cast_fp16")];
            tensor<int32, [1]> var_3287_axes_0 = const()[name = string("op_3287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3287_keep_dims_0 = const()[name = string("op_3287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_3287_cast_fp16 = reduce_mean(axes = var_3287_axes_0, keep_dims = var_3287_keep_dims_0, x = var_3282_cast_fp16)[name = string("op_3287_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_3288_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_3289_cast_fp16 = inverse(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            fp16 var_3290_to_fp16 = const()[name = string("op_3290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_3291_cast_fp16 = mul(x = var_3289_cast_fp16, y = var_3290_to_fp16)[name = string("op_3291_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_3293_cast_fp16 = add(x = var_3287_cast_fp16, y = var_3291_cast_fp16)[name = string("op_3293_cast_fp16")];
            fp32 var_3294_epsilon_0 = const()[name = string("op_3294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_3294_cast_fp16 = rsqrt(epsilon = var_3294_epsilon_0, x = var_3293_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_13_cast_fp16 = mul(x = y_57_cast_fp16, y = var_3294_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_3296_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3296_cast_fp16")];
            tensor<int32, [4]> var_3301_begin_0 = const()[name = string("op_3301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3301_end_0 = const()[name = string("op_3301_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_3301_end_mask_0 = const()[name = string("op_3301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_3301_cast_fp16 = slice_by_index(begin = var_3301_begin_0, end = var_3301_end_0, end_mask = var_3301_end_mask_0, x = q_13_cast_fp16)[name = string("op_3301_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_3302_cast_fp16 = mul(x = var_3301_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3307_begin_0 = const()[name = string("op_3307_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3307_end_0 = const()[name = string("op_3307_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_3307_end_mask_0 = const()[name = string("op_3307_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_3307_cast_fp16 = slice_by_index(begin = var_3307_begin_0, end = var_3307_end_0, end_mask = var_3307_end_mask_0, x = q_13_cast_fp16)[name = string("op_3307_cast_fp16")];
            int32 var_3309 = const()[name = string("op_3309"), val = int32(-1)];
            bool var_3310_interleave_0 = const()[name = string("op_3310_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_3310_cast_fp16 = concat(axis = var_3309, interleave = var_3310_interleave_0, values = (var_3302_cast_fp16, var_3307_cast_fp16))[name = string("op_3310_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_3311_cast_fp16 = mul(x = var_3310_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3311_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_15_cast_fp16 = add(x = var_3296_cast_fp16, y = var_3311_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_3314_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3314_cast_fp16")];
            tensor<int32, [4]> var_3319_begin_0 = const()[name = string("op_3319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3319_end_0 = const()[name = string("op_3319_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_3319_end_mask_0 = const()[name = string("op_3319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_3319_cast_fp16 = slice_by_index(begin = var_3319_begin_0, end = var_3319_end_0, end_mask = var_3319_end_mask_0, x = k_19_cast_fp16)[name = string("op_3319_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_3320_cast_fp16 = mul(x = var_3319_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_3320_cast_fp16")];
            tensor<int32, [4]> var_3325_begin_0 = const()[name = string("op_3325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3325_end_0 = const()[name = string("op_3325_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_3325_end_mask_0 = const()[name = string("op_3325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_3325_cast_fp16 = slice_by_index(begin = var_3325_begin_0, end = var_3325_end_0, end_mask = var_3325_end_mask_0, x = k_19_cast_fp16)[name = string("op_3325_cast_fp16")];
            int32 var_3327 = const()[name = string("op_3327"), val = int32(-1)];
            bool var_3328_interleave_0 = const()[name = string("op_3328_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_3328_cast_fp16 = concat(axis = var_3327, interleave = var_3328_interleave_0, values = (var_3320_cast_fp16, var_3325_cast_fp16))[name = string("op_3328_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_3329_cast_fp16 = mul(x = var_3328_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3329_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_21_cast_fp16 = add(x = var_3314_cast_fp16, y = var_3329_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = k_21_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [4, 1, 64, 256]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_12_cast_fp16 = reshape(shape = concat_12, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [4, 1, 64, 256]> reshape_13_cast_fp16 = reshape(shape = concat_13, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = v_13_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [4, 1, 64, 256]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_14_cast_fp16 = reshape(shape = concat_14, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [4, 1, 64, 256]> reshape_15_cast_fp16 = reshape(shape = concat_15, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3343_transpose_x_0 = const()[name = string("op_3343_transpose_x_0"), val = bool(false)];
            bool var_3343_transpose_y_0 = const()[name = string("op_3343_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 4, 64, 64]> var_3343_cast_fp16 = matmul(transpose_x = var_3343_transpose_x_0, transpose_y = var_3343_transpose_y_0, x = q_15_cast_fp16, y = transpose_99_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_3345_cast_fp16 = add(x = var_3343_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3345_cast_fp16")];
            int32 var_3346 = const()[name = string("op_3346"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_57_cast_fp16 = softmax(axis = var_3346, x = var_3345_cast_fp16)[name = string("w_57_cast_fp16")];
            bool var_3349_transpose_x_0 = const()[name = string("op_3349_transpose_x_0"), val = bool(false)];
            bool var_3349_transpose_y_0 = const()[name = string("op_3349_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 4, 64, 256]> var_3349_cast_fp16 = matmul(transpose_x = var_3349_transpose_x_0, transpose_y = var_3349_transpose_y_0, x = w_57_cast_fp16, y = v_15_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<int32, [4]> var_3352_perm_0 = const()[name = string("op_3352_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3356 = const()[name = string("op_3356"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_3352_cast_fp16 = transpose(perm = var_3352_perm_0, x = var_3349_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 64, 1024]> input_39_cast_fp16 = reshape(shape = var_3356, x = var_3352_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23206144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23730496))))[name = string("tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3361_cast_fp16 = abs(x = linear_31_cast_fp16)[name = string("op_3361_cast_fp16")];
            bool var_3364 = const()[name = string("op_3364"), val = bool(true)];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3365_cast_fp16 = reduce_max(axes = var_3365_axes_0, keep_dims = var_3364, x = var_3361_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_59_cast_fp16 = maximum(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_59_cast_fp16 = real_div(x = linear_31_cast_fp16, y = s_59_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3369_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<int32, [1]> var_3374_axes_0 = const()[name = string("op_3374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3374_keep_dims_0 = const()[name = string("op_3374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3374_cast_fp16 = reduce_mean(axes = var_3374_axes_0, keep_dims = var_3374_keep_dims_0, x = var_3369_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3375_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_3375_cast_fp16")];
            fp32 var_3376_epsilon_0 = const()[name = string("op_3376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3376_cast_fp16 = inverse(epsilon = var_3376_epsilon_0, x = var_3375_cast_fp16)[name = string("op_3376_cast_fp16")];
            fp16 var_3377_to_fp16 = const()[name = string("op_3377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3378_cast_fp16 = mul(x = var_3376_cast_fp16, y = var_3377_to_fp16)[name = string("op_3378_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3380_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3380_cast_fp16")];
            fp32 var_3381_epsilon_0 = const()[name = string("op_3381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3381_cast_fp16 = rsqrt(epsilon = var_3381_epsilon_0, x = var_3380_cast_fp16)[name = string("op_3381_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_3381_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23731584)))];
            tensor<fp16, [1, 64, 512]> var_3383_cast_fp16 = mul(x = out_51_cast_fp16, y = tm_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_3383_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_65_cast_fp16 = add(x = x_55_cast_fp16, y = var_3383_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3386_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_3386_cast_fp16")];
            bool var_3389 = const()[name = string("op_3389"), val = bool(true)];
            tensor<int32, [1]> var_3390_axes_0 = const()[name = string("op_3390_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3390_cast_fp16 = reduce_max(axes = var_3390_axes_0, keep_dims = var_3389, x = var_3386_cast_fp16)[name = string("op_3390_cast_fp16")];
            fp16 var_3391_to_fp16 = const()[name = string("op_3391_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_61_cast_fp16 = maximum(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_61_cast_fp16 = real_div(x = x_65_cast_fp16, y = s_61_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3394_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<int32, [1]> var_3399_axes_0 = const()[name = string("op_3399_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3399_keep_dims_0 = const()[name = string("op_3399_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3399_cast_fp16 = reduce_mean(axes = var_3399_axes_0, keep_dims = var_3399_keep_dims_0, x = var_3394_cast_fp16)[name = string("op_3399_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3400_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_3400_cast_fp16")];
            fp32 var_3401_epsilon_0 = const()[name = string("op_3401_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3401_cast_fp16 = inverse(epsilon = var_3401_epsilon_0, x = var_3400_cast_fp16)[name = string("op_3401_cast_fp16")];
            fp16 var_3402_to_fp16 = const()[name = string("op_3402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3403_cast_fp16 = mul(x = var_3401_cast_fp16, y = var_3402_to_fp16)[name = string("op_3403_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3405_cast_fp16 = add(x = var_3399_cast_fp16, y = var_3403_cast_fp16)[name = string("op_3405_cast_fp16")];
            fp32 var_3406_epsilon_0 = const()[name = string("op_3406_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3406_cast_fp16 = rsqrt(epsilon = var_3406_epsilon_0, x = var_3405_cast_fp16)[name = string("op_3406_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_3406_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23732672)))];
            tensor<fp16, [1, 64, 512]> input_41_cast_fp16 = mul(x = out_53_cast_fp16, y = tm_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23733760))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24782400))))[name = string("tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_32_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_32_cast_fp16")];
            string var_3413_mode_0 = const()[name = string("op_3413_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_3413_cast_fp16 = gelu(mode = var_3413_mode_0, x = linear_32_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24786560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25835200))))[name = string("tm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_43_cast_fp16 = mul(x = var_3413_cast_fp16, y = linear_33_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25839360))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26888000))))[name = string("tm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3421_cast_fp16 = abs(x = linear_34_cast_fp16)[name = string("op_3421_cast_fp16")];
            bool var_3424 = const()[name = string("op_3424"), val = bool(true)];
            tensor<int32, [1]> var_3425_axes_0 = const()[name = string("op_3425_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3425_cast_fp16 = reduce_max(axes = var_3425_axes_0, keep_dims = var_3424, x = var_3421_cast_fp16)[name = string("op_3425_cast_fp16")];
            fp16 var_3426_to_fp16 = const()[name = string("op_3426_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_63_cast_fp16 = maximum(x = var_3425_cast_fp16, y = var_3426_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_63_cast_fp16 = real_div(x = linear_34_cast_fp16, y = s_63_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3429_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_3429_cast_fp16")];
            tensor<int32, [1]> var_3434_axes_0 = const()[name = string("op_3434_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3434_keep_dims_0 = const()[name = string("op_3434_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3434_cast_fp16 = reduce_mean(axes = var_3434_axes_0, keep_dims = var_3434_keep_dims_0, x = var_3429_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3435_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp32 var_3436_epsilon_0 = const()[name = string("op_3436_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3436_cast_fp16 = inverse(epsilon = var_3436_epsilon_0, x = var_3435_cast_fp16)[name = string("op_3436_cast_fp16")];
            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3438_cast_fp16 = mul(x = var_3436_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3440_cast_fp16 = add(x = var_3434_cast_fp16, y = var_3438_cast_fp16)[name = string("op_3440_cast_fp16")];
            fp32 var_3441_epsilon_0 = const()[name = string("op_3441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3441_cast_fp16 = rsqrt(epsilon = var_3441_epsilon_0, x = var_3440_cast_fp16)[name = string("op_3441_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_55_cast_fp16 = mul(x = y_63_cast_fp16, y = var_3441_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26889088)))];
            tensor<fp16, [1, 64, 512]> var_3443_cast_fp16 = mul(x = out_55_cast_fp16, y = tm_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_45_cast_fp16 = add(x = x_65_cast_fp16, y = var_3443_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26890176))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27152384))))[name = string("tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_35_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_35_cast_fp16")];
            string x_69_mode_0 = const()[name = string("x_69_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_69_cast_fp16 = gelu(mode = x_69_mode_0, x = linear_35_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [4]> var_3463_begin_0 = const()[name = string("op_3463_begin_0"), val = tensor<int32, [4]>([0, 0, 3, 0])];
            tensor<int32, [4]> var_3463_end_0 = const()[name = string("op_3463_end_0"), val = tensor<int32, [4]>([1, 64, 4, 512])];
            tensor<bool, [4]> var_3463_end_mask_0 = const()[name = string("op_3463_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3463_squeeze_mask_0 = const()[name = string("op_3463_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_3463_cast_fp16 = slice_by_index(begin = var_3463_begin_0, end = var_3463_end_0, end_mask = var_3463_end_mask_0, squeeze_mask = var_3463_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3463_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_47_cast_fp16 = mul(x = x_69_cast_fp16, y = var_3463_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27153472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27415680))))[name = string("tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3473_cast_fp16 = abs(x = linear_36_cast_fp16)[name = string("op_3473_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<int32, [1]> var_3477_axes_0 = const()[name = string("op_3477_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3477_cast_fp16 = reduce_max(axes = var_3477_axes_0, keep_dims = var_3476, x = var_3473_cast_fp16)[name = string("op_3477_cast_fp16")];
            fp16 var_3478_to_fp16 = const()[name = string("op_3478_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_65_cast_fp16 = maximum(x = var_3477_cast_fp16, y = var_3478_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_65_cast_fp16 = real_div(x = linear_36_cast_fp16, y = s_65_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3481_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [1]> var_3486_axes_0 = const()[name = string("op_3486_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3486_keep_dims_0 = const()[name = string("op_3486_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3486_cast_fp16 = reduce_mean(axes = var_3486_axes_0, keep_dims = var_3486_keep_dims_0, x = var_3481_cast_fp16)[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3487_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp32 var_3488_epsilon_0 = const()[name = string("op_3488_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3488_cast_fp16 = inverse(epsilon = var_3488_epsilon_0, x = var_3487_cast_fp16)[name = string("op_3488_cast_fp16")];
            fp16 var_3489_to_fp16 = const()[name = string("op_3489_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3490_cast_fp16 = mul(x = var_3488_cast_fp16, y = var_3489_to_fp16)[name = string("op_3490_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3492_cast_fp16 = add(x = var_3486_cast_fp16, y = var_3490_cast_fp16)[name = string("op_3492_cast_fp16")];
            fp32 var_3493_epsilon_0 = const()[name = string("op_3493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3493_cast_fp16 = rsqrt(epsilon = var_3493_epsilon_0, x = var_3492_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_57_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3493_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27416768)))];
            tensor<fp16, [1, 64, 512]> var_3495_cast_fp16 = mul(x = out_57_cast_fp16, y = tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3495_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_7_cast_fp16 = add(x = input_45_cast_fp16, y = var_3495_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<fp16, [1]> tm_layers_3_layer_scalar_to_fp16 = const()[name = string("tm_layers_3_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 64, 512]> x_73_cast_fp16 = mul(x = h_7_cast_fp16, y = tm_layers_3_layer_scalar_to_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3499_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_3499_cast_fp16")];
            bool var_3502 = const()[name = string("op_3502"), val = bool(true)];
            tensor<int32, [1]> var_3503_axes_0 = const()[name = string("op_3503_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3503_cast_fp16 = reduce_max(axes = var_3503_axes_0, keep_dims = var_3502, x = var_3499_cast_fp16)[name = string("op_3503_cast_fp16")];
            fp16 var_3504_to_fp16 = const()[name = string("op_3504_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_67_cast_fp16 = maximum(x = var_3503_cast_fp16, y = var_3504_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_67_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_67_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3507_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<int32, [1]> var_3512_axes_0 = const()[name = string("op_3512_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3512_keep_dims_0 = const()[name = string("op_3512_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3512_cast_fp16 = reduce_mean(axes = var_3512_axes_0, keep_dims = var_3512_keep_dims_0, x = var_3507_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3513_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_3513_cast_fp16")];
            fp32 var_3514_epsilon_0 = const()[name = string("op_3514_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3514_cast_fp16 = inverse(epsilon = var_3514_epsilon_0, x = var_3513_cast_fp16)[name = string("op_3514_cast_fp16")];
            fp16 var_3515_to_fp16 = const()[name = string("op_3515_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3516_cast_fp16 = mul(x = var_3514_cast_fp16, y = var_3515_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3518_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp32 var_3519_epsilon_0 = const()[name = string("op_3519_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3519_cast_fp16 = rsqrt(epsilon = var_3519_epsilon_0, x = var_3518_cast_fp16)[name = string("op_3519_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_59_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3519_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27417856)))];
            tensor<fp16, [1, 64, 512]> input_49_cast_fp16 = mul(x = out_59_cast_fp16, y = tm_layers_4_input_layernorm_weight_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27418944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27943296))))[name = string("tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3529 = const()[name = string("op_3529"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_3530_cast_fp16 = reshape(shape = var_3529, x = linear_37_cast_fp16)[name = string("op_3530_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27945408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28207616))))[name = string("tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3541 = const()[name = string("op_3541"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_3542_cast_fp16 = reshape(shape = var_3541, x = linear_38_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28208704))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28470912))))[name = string("tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_3553 = const()[name = string("op_3553"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_3554_cast_fp16 = reshape(shape = var_3553, x = linear_39_cast_fp16)[name = string("op_3554_cast_fp16")];
            tensor<int32, [4]> x_79_perm_0 = const()[name = string("x_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3530_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 4, 64, 256]> var_3558_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3558_cast_fp16")];
            bool var_3561 = const()[name = string("op_3561"), val = bool(true)];
            tensor<int32, [1]> var_3562_axes_0 = const()[name = string("op_3562_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_3562_cast_fp16 = reduce_max(axes = var_3562_axes_0, keep_dims = var_3561, x = var_3558_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp16 var_3563_to_fp16 = const()[name = string("op_3563_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_69_cast_fp16 = maximum(x = var_3562_cast_fp16, y = var_3563_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_69_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_69_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_3566_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3566_cast_fp16")];
            tensor<int32, [1]> var_3571_axes_0 = const()[name = string("op_3571_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3571_keep_dims_0 = const()[name = string("op_3571_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_3571_cast_fp16 = reduce_mean(axes = var_3571_axes_0, keep_dims = var_3571_keep_dims_0, x = var_3566_cast_fp16)[name = string("op_3571_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_3572_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3572_cast_fp16")];
            fp32 var_3573_epsilon_0 = const()[name = string("op_3573_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_3573_cast_fp16 = inverse(epsilon = var_3573_epsilon_0, x = var_3572_cast_fp16)[name = string("op_3573_cast_fp16")];
            fp16 var_3574_to_fp16 = const()[name = string("op_3574_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_3575_cast_fp16 = mul(x = var_3573_cast_fp16, y = var_3574_to_fp16)[name = string("op_3575_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_3577_cast_fp16 = add(x = var_3571_cast_fp16, y = var_3575_cast_fp16)[name = string("op_3577_cast_fp16")];
            fp32 var_3578_epsilon_0 = const()[name = string("op_3578_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_3578_cast_fp16 = rsqrt(epsilon = var_3578_epsilon_0, x = var_3577_cast_fp16)[name = string("op_3578_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_61_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3578_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [256]> tm_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472000)))];
            tensor<fp16, [1, 4, 64, 256]> q_17_cast_fp16 = mul(x = out_61_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3542_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2, 64, 256]> var_3581_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3581_cast_fp16")];
            bool var_3584 = const()[name = string("op_3584"), val = bool(true)];
            tensor<int32, [1]> var_3585_axes_0 = const()[name = string("op_3585_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_3585_cast_fp16 = reduce_max(axes = var_3585_axes_0, keep_dims = var_3584, x = var_3581_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp16 var_3586_to_fp16 = const()[name = string("op_3586_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_71_cast_fp16 = maximum(x = var_3585_cast_fp16, y = var_3586_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_71_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_71_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_3589_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3589_cast_fp16")];
            tensor<int32, [1]> var_3594_axes_0 = const()[name = string("op_3594_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3594_keep_dims_0 = const()[name = string("op_3594_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_3594_cast_fp16 = reduce_mean(axes = var_3594_axes_0, keep_dims = var_3594_keep_dims_0, x = var_3589_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_3595_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp32 var_3596_epsilon_0 = const()[name = string("op_3596_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_3596_cast_fp16 = inverse(epsilon = var_3596_epsilon_0, x = var_3595_cast_fp16)[name = string("op_3596_cast_fp16")];
            fp16 var_3597_to_fp16 = const()[name = string("op_3597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_3598_cast_fp16 = mul(x = var_3596_cast_fp16, y = var_3597_to_fp16)[name = string("op_3598_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_3600_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3598_cast_fp16)[name = string("op_3600_cast_fp16")];
            fp32 var_3601_epsilon_0 = const()[name = string("op_3601_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_3601_cast_fp16 = rsqrt(epsilon = var_3601_epsilon_0, x = var_3600_cast_fp16)[name = string("op_3601_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_63_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3601_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472576)))];
            tensor<fp16, [1, 2, 64, 256]> k_25_cast_fp16 = mul(x = out_63_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_79_cast_fp16 = transpose(perm = x_79_perm_0, x = var_3554_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 2, 64, 256]> var_3604_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3604_cast_fp16")];
            bool var_3607 = const()[name = string("op_3607"), val = bool(true)];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_3607, x = var_3604_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_73_cast_fp16 = maximum(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_73_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_73_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_3612_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<int32, [1]> var_3617_axes_0 = const()[name = string("op_3617_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3617_keep_dims_0 = const()[name = string("op_3617_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_3617_cast_fp16 = reduce_mean(axes = var_3617_axes_0, keep_dims = var_3617_keep_dims_0, x = var_3612_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_3618_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_3619_cast_fp16 = inverse(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            fp16 var_3620_to_fp16 = const()[name = string("op_3620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_3621_cast_fp16 = mul(x = var_3619_cast_fp16, y = var_3620_to_fp16)[name = string("op_3621_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_3623_cast_fp16 = add(x = var_3617_cast_fp16, y = var_3621_cast_fp16)[name = string("op_3623_cast_fp16")];
            fp32 var_3624_epsilon_0 = const()[name = string("op_3624_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_3624_cast_fp16 = rsqrt(epsilon = var_3624_epsilon_0, x = var_3623_cast_fp16)[name = string("op_3624_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_17_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3624_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_3626_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3626_cast_fp16")];
            tensor<int32, [4]> var_3631_begin_0 = const()[name = string("op_3631_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3631_end_0 = const()[name = string("op_3631_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_3631_end_mask_0 = const()[name = string("op_3631_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_3631_cast_fp16 = slice_by_index(begin = var_3631_begin_0, end = var_3631_end_0, end_mask = var_3631_end_mask_0, x = q_17_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_3632_cast_fp16 = mul(x = var_3631_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> var_3637_begin_0 = const()[name = string("op_3637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3637_end_0 = const()[name = string("op_3637_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_3637_end_mask_0 = const()[name = string("op_3637_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_3637_cast_fp16 = slice_by_index(begin = var_3637_begin_0, end = var_3637_end_0, end_mask = var_3637_end_mask_0, x = q_17_cast_fp16)[name = string("op_3637_cast_fp16")];
            int32 var_3639 = const()[name = string("op_3639"), val = int32(-1)];
            bool var_3640_interleave_0 = const()[name = string("op_3640_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_3640_cast_fp16 = concat(axis = var_3639, interleave = var_3640_interleave_0, values = (var_3632_cast_fp16, var_3637_cast_fp16))[name = string("op_3640_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_3641_cast_fp16 = mul(x = var_3640_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3641_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_19_cast_fp16 = add(x = var_3626_cast_fp16, y = var_3641_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_3644_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = k_25_cast_fp16)[name = string("op_3649_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_3650_cast_fp16 = mul(x = var_3649_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_3650_cast_fp16")];
            tensor<int32, [4]> var_3655_begin_0 = const()[name = string("op_3655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3655_end_0 = const()[name = string("op_3655_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_3655_end_mask_0 = const()[name = string("op_3655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_3655_cast_fp16 = slice_by_index(begin = var_3655_begin_0, end = var_3655_end_0, end_mask = var_3655_end_mask_0, x = k_25_cast_fp16)[name = string("op_3655_cast_fp16")];
            int32 var_3657 = const()[name = string("op_3657"), val = int32(-1)];
            bool var_3658_interleave_0 = const()[name = string("op_3658_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_3658_cast_fp16 = concat(axis = var_3657, interleave = var_3658_interleave_0, values = (var_3650_cast_fp16, var_3655_cast_fp16))[name = string("op_3658_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_3659_cast_fp16 = mul(x = var_3658_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_3659_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_27_cast_fp16 = add(x = var_3644_cast_fp16, y = var_3659_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = k_27_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [4, 1, 64, 256]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
            tensor<int32, [5]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_16_cast_fp16 = reshape(shape = concat_16, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [4, 1, 64, 256]> reshape_17_cast_fp16 = reshape(shape = concat_17, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = v_17_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [4, 1, 64, 256]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
            tensor<int32, [5]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_18_cast_fp16 = reshape(shape = concat_18, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [4, 1, 64, 256]> reshape_19_cast_fp16 = reshape(shape = concat_19, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3673_transpose_x_0 = const()[name = string("op_3673_transpose_x_0"), val = bool(false)];
            bool var_3673_transpose_y_0 = const()[name = string("op_3673_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 4, 64, 64]> var_3673_cast_fp16 = matmul(transpose_x = var_3673_transpose_x_0, transpose_y = var_3673_transpose_y_0, x = q_19_cast_fp16, y = transpose_100_cast_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_3675_cast_fp16 = add(x = var_3673_cast_fp16, y = var_2173_cast_fp16)[name = string("op_3675_cast_fp16")];
            int32 var_3676 = const()[name = string("op_3676"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_73_cast_fp16 = softmax(axis = var_3676, x = var_3675_cast_fp16)[name = string("w_73_cast_fp16")];
            bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)];
            bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 4, 64, 256]> var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = w_73_cast_fp16, y = v_19_cast_fp16)[name = string("op_3679_cast_fp16")];
            tensor<int32, [4]> var_3682_perm_0 = const()[name = string("op_3682_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3686 = const()[name = string("op_3686"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_3682_cast_fp16 = transpose(perm = var_3682_perm_0, x = var_3679_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 64, 1024]> input_51_cast_fp16 = reshape(shape = var_3686, x = var_3682_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28473152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28997504))))[name = string("tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3691_cast_fp16 = abs(x = linear_40_cast_fp16)[name = string("op_3691_cast_fp16")];
            bool var_3694 = const()[name = string("op_3694"), val = bool(true)];
            tensor<int32, [1]> var_3695_axes_0 = const()[name = string("op_3695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3695_cast_fp16 = reduce_max(axes = var_3695_axes_0, keep_dims = var_3694, x = var_3691_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 var_3696_to_fp16 = const()[name = string("op_3696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_75_cast_fp16 = maximum(x = var_3695_cast_fp16, y = var_3696_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_75_cast_fp16 = real_div(x = linear_40_cast_fp16, y = s_75_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3699_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3699_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3704_keep_dims_0 = const()[name = string("op_3704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3704_cast_fp16 = reduce_mean(axes = var_3704_axes_0, keep_dims = var_3704_keep_dims_0, x = var_3699_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3705_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3705_cast_fp16")];
            fp32 var_3706_epsilon_0 = const()[name = string("op_3706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3706_cast_fp16 = inverse(epsilon = var_3706_epsilon_0, x = var_3705_cast_fp16)[name = string("op_3706_cast_fp16")];
            fp16 var_3707_to_fp16 = const()[name = string("op_3707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3708_cast_fp16 = mul(x = var_3706_cast_fp16, y = var_3707_to_fp16)[name = string("op_3708_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3710_cast_fp16 = add(x = var_3704_cast_fp16, y = var_3708_cast_fp16)[name = string("op_3710_cast_fp16")];
            fp32 var_3711_epsilon_0 = const()[name = string("op_3711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3711_cast_fp16 = rsqrt(epsilon = var_3711_epsilon_0, x = var_3710_cast_fp16)[name = string("op_3711_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3711_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28998592)))];
            tensor<fp16, [1, 64, 512]> var_3713_cast_fp16 = mul(x = out_65_cast_fp16, y = tm_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = var_3713_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3716_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3716_cast_fp16")];
            bool var_3719 = const()[name = string("op_3719"), val = bool(true)];
            tensor<int32, [1]> var_3720_axes_0 = const()[name = string("op_3720_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3720_cast_fp16 = reduce_max(axes = var_3720_axes_0, keep_dims = var_3719, x = var_3716_cast_fp16)[name = string("op_3720_cast_fp16")];
            fp16 var_3721_to_fp16 = const()[name = string("op_3721_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_77_cast_fp16 = maximum(x = var_3720_cast_fp16, y = var_3721_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_77_cast_fp16 = real_div(x = x_83_cast_fp16, y = s_77_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3724_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3729_axes_0 = const()[name = string("op_3729_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3729_keep_dims_0 = const()[name = string("op_3729_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3729_cast_fp16 = reduce_mean(axes = var_3729_axes_0, keep_dims = var_3729_keep_dims_0, x = var_3724_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3730_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3730_cast_fp16")];
            fp32 var_3731_epsilon_0 = const()[name = string("op_3731_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3731_cast_fp16 = inverse(epsilon = var_3731_epsilon_0, x = var_3730_cast_fp16)[name = string("op_3731_cast_fp16")];
            fp16 var_3732_to_fp16 = const()[name = string("op_3732_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3733_cast_fp16 = mul(x = var_3731_cast_fp16, y = var_3732_to_fp16)[name = string("op_3733_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3735_cast_fp16 = add(x = var_3729_cast_fp16, y = var_3733_cast_fp16)[name = string("op_3735_cast_fp16")];
            fp32 var_3736_epsilon_0 = const()[name = string("op_3736_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3736_cast_fp16 = rsqrt(epsilon = var_3736_epsilon_0, x = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_67_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3736_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28999680)))];
            tensor<fp16, [1, 64, 512]> input_53_cast_fp16 = mul(x = out_67_cast_fp16, y = tm_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29000768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30049408))))[name = string("tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_41_cast_fp16")];
            string var_3743_mode_0 = const()[name = string("op_3743_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_3743_cast_fp16 = gelu(mode = var_3743_mode_0, x = linear_41_cast_fp16)[name = string("op_3743_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30053568))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31102208))))[name = string("tm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_42_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_55_cast_fp16 = mul(x = var_3743_cast_fp16, y = linear_42_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31106368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32155008))))[name = string("tm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3751_cast_fp16 = abs(x = linear_43_cast_fp16)[name = string("op_3751_cast_fp16")];
            bool var_3754 = const()[name = string("op_3754"), val = bool(true)];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_3754, x = var_3751_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 var_3756_to_fp16 = const()[name = string("op_3756_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_79_cast_fp16 = maximum(x = var_3755_cast_fp16, y = var_3756_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_79_cast_fp16 = real_div(x = linear_43_cast_fp16, y = s_79_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3759_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3759_cast_fp16")];
            tensor<int32, [1]> var_3764_axes_0 = const()[name = string("op_3764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3764_keep_dims_0 = const()[name = string("op_3764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3764_cast_fp16 = reduce_mean(axes = var_3764_axes_0, keep_dims = var_3764_keep_dims_0, x = var_3759_cast_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3765_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3766_cast_fp16 = inverse(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3768_cast_fp16 = mul(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("op_3768_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3770_cast_fp16 = add(x = var_3764_cast_fp16, y = var_3768_cast_fp16)[name = string("op_3770_cast_fp16")];
            fp32 var_3771_epsilon_0 = const()[name = string("op_3771_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3771_cast_fp16 = rsqrt(epsilon = var_3771_epsilon_0, x = var_3770_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_69_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3771_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32156096)))];
            tensor<fp16, [1, 64, 512]> var_3773_cast_fp16 = mul(x = out_69_cast_fp16, y = tm_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3773_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_57_cast_fp16 = add(x = x_83_cast_fp16, y = var_3773_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32157184))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32419392))))[name = string("tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_44_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_44_cast_fp16")];
            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_87_cast_fp16 = gelu(mode = x_87_mode_0, x = linear_44_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [4]> var_3793_begin_0 = const()[name = string("op_3793_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_3793_end_0 = const()[name = string("op_3793_end_0"), val = tensor<int32, [4]>([1, 64, 5, 512])];
            tensor<bool, [4]> var_3793_end_mask_0 = const()[name = string("op_3793_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3793_squeeze_mask_0 = const()[name = string("op_3793_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_3793_cast_fp16 = slice_by_index(begin = var_3793_begin_0, end = var_3793_end_0, end_mask = var_3793_end_mask_0, squeeze_mask = var_3793_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_59_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3793_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32420480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32682688))))[name = string("tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3803_cast_fp16 = abs(x = linear_45_cast_fp16)[name = string("op_3803_cast_fp16")];
            bool var_3806 = const()[name = string("op_3806"), val = bool(true)];
            tensor<int32, [1]> var_3807_axes_0 = const()[name = string("op_3807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3807_cast_fp16 = reduce_max(axes = var_3807_axes_0, keep_dims = var_3806, x = var_3803_cast_fp16)[name = string("op_3807_cast_fp16")];
            fp16 var_3808_to_fp16 = const()[name = string("op_3808_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_81_cast_fp16 = maximum(x = var_3807_cast_fp16, y = var_3808_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_81_cast_fp16 = real_div(x = linear_45_cast_fp16, y = s_81_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3811_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [1]> var_3816_axes_0 = const()[name = string("op_3816_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3816_keep_dims_0 = const()[name = string("op_3816_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3816_cast_fp16 = reduce_mean(axes = var_3816_axes_0, keep_dims = var_3816_keep_dims_0, x = var_3811_cast_fp16)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3817_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3817_cast_fp16")];
            fp32 var_3818_epsilon_0 = const()[name = string("op_3818_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3818_cast_fp16 = inverse(epsilon = var_3818_epsilon_0, x = var_3817_cast_fp16)[name = string("op_3818_cast_fp16")];
            fp16 var_3819_to_fp16 = const()[name = string("op_3819_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3820_cast_fp16 = mul(x = var_3818_cast_fp16, y = var_3819_to_fp16)[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3822_cast_fp16 = add(x = var_3816_cast_fp16, y = var_3820_cast_fp16)[name = string("op_3822_cast_fp16")];
            fp32 var_3823_epsilon_0 = const()[name = string("op_3823_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3823_cast_fp16 = rsqrt(epsilon = var_3823_epsilon_0, x = var_3822_cast_fp16)[name = string("op_3823_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_71_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3823_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32683776)))];
            tensor<fp16, [1, 64, 512]> var_3825_cast_fp16 = mul(x = out_71_cast_fp16, y = tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3825_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_9_cast_fp16 = add(x = input_57_cast_fp16, y = var_3825_cast_fp16)[name = string("h_9_cast_fp16")];
            tensor<fp16, [1]> tm_layers_4_layer_scalar_to_fp16 = const()[name = string("tm_layers_4_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.62p-1])];
            tensor<fp16, [1, 64, 512]> x_91_cast_fp16 = mul(x = h_9_cast_fp16, y = tm_layers_4_layer_scalar_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3829_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3829_cast_fp16")];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<int32, [1]> var_3833_axes_0 = const()[name = string("op_3833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3833_cast_fp16 = reduce_max(axes = var_3833_axes_0, keep_dims = var_3832, x = var_3829_cast_fp16)[name = string("op_3833_cast_fp16")];
            fp16 var_3834_to_fp16 = const()[name = string("op_3834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_83_cast_fp16 = maximum(x = var_3833_cast_fp16, y = var_3834_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_83_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_83_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_3837_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3837_cast_fp16")];
            tensor<int32, [1]> var_3842_axes_0 = const()[name = string("op_3842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3842_keep_dims_0 = const()[name = string("op_3842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3842_cast_fp16 = reduce_mean(axes = var_3842_axes_0, keep_dims = var_3842_keep_dims_0, x = var_3837_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3843_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp32 var_3844_epsilon_0 = const()[name = string("op_3844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3844_cast_fp16 = inverse(epsilon = var_3844_epsilon_0, x = var_3843_cast_fp16)[name = string("op_3844_cast_fp16")];
            fp16 var_3845_to_fp16 = const()[name = string("op_3845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3846_cast_fp16 = mul(x = var_3844_cast_fp16, y = var_3845_to_fp16)[name = string("op_3846_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3848_cast_fp16 = add(x = var_3842_cast_fp16, y = var_3846_cast_fp16)[name = string("op_3848_cast_fp16")];
            fp32 var_3849_epsilon_0 = const()[name = string("op_3849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3849_cast_fp16 = rsqrt(epsilon = var_3849_epsilon_0, x = var_3848_cast_fp16)[name = string("op_3849_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_73_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3849_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32684864)))];
            tensor<fp16, [1, 64, 512]> input_61_cast_fp16 = mul(x = out_73_cast_fp16, y = tm_layers_5_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32685952))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33734592))))[name = string("tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_46_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [4]> var_3859 = const()[name = string("op_3859"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 64, 4, 512]> var_3860_cast_fp16 = reshape(shape = var_3859, x = linear_46_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<int32, [4]> x_93_perm_0 = const()[name = string("x_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33738752))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34000960))))[name = string("tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<int32, [4]> var_3871 = const()[name = string("op_3871"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 64, 1, 512]> var_3872_cast_fp16 = reshape(shape = var_3871, x = linear_47_cast_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [4]> x_95_perm_0 = const()[name = string("x_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34002048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34264256))))[name = string("tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [4]> var_3883 = const()[name = string("op_3883"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 64, 1, 512]> var_3884_cast_fp16 = reshape(shape = var_3883, x = linear_48_cast_fp16)[name = string("op_3884_cast_fp16")];
            tensor<int32, [4]> x_97_perm_0 = const()[name = string("x_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 512]> x_93_cast_fp16 = transpose(perm = x_93_perm_0, x = var_3860_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 4, 64, 512]> var_3888_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3888_cast_fp16")];
            bool var_3891 = const()[name = string("op_3891"), val = bool(true)];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_3892_cast_fp16 = reduce_max(axes = var_3892_axes_0, keep_dims = var_3891, x = var_3888_cast_fp16)[name = string("op_3892_cast_fp16")];
            fp16 var_3893_to_fp16 = const()[name = string("op_3893_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_85_cast_fp16 = maximum(x = var_3892_cast_fp16, y = var_3893_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> y_85_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_85_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_3896_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3896_cast_fp16")];
            tensor<int32, [1]> var_3901_axes_0 = const()[name = string("op_3901_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3901_keep_dims_0 = const()[name = string("op_3901_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_3901_cast_fp16 = reduce_mean(axes = var_3901_axes_0, keep_dims = var_3901_keep_dims_0, x = var_3896_cast_fp16)[name = string("op_3901_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_3902_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3902_cast_fp16")];
            fp32 var_3903_epsilon_0 = const()[name = string("op_3903_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_3903_cast_fp16 = inverse(epsilon = var_3903_epsilon_0, x = var_3902_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_3905_cast_fp16 = mul(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_3907_cast_fp16 = add(x = var_3901_cast_fp16, y = var_3905_cast_fp16)[name = string("op_3907_cast_fp16")];
            fp32 var_3908_epsilon_0 = const()[name = string("op_3908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_3908_cast_fp16 = rsqrt(epsilon = var_3908_epsilon_0, x = var_3907_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> out_75_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3908_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34265344)))];
            tensor<fp16, [1, 4, 64, 512]> q_21_cast_fp16 = mul(x = out_75_cast_fp16, y = tm_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> x_95_cast_fp16 = transpose(perm = x_95_perm_0, x = var_3872_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1, 64, 512]> var_3911_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 64, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 64, 1]> s_87_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> y_87_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_87_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_3919_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 64, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_3925_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 64, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 64, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 64, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> out_77_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3931_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34266432)))];
            tensor<fp16, [1, 1, 64, 512]> k_31_cast_fp16 = mul(x = out_77_cast_fp16, y = tm_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = var_3884_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 1, 64, 512]> var_3934_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3934_cast_fp16")];
            bool var_3937 = const()[name = string("op_3937"), val = bool(true)];
            tensor<int32, [1]> var_3938_axes_0 = const()[name = string("op_3938_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 64, 1]> var_3938_cast_fp16 = reduce_max(axes = var_3938_axes_0, keep_dims = var_3937, x = var_3934_cast_fp16)[name = string("op_3938_cast_fp16")];
            fp16 var_3939_to_fp16 = const()[name = string("op_3939_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 64, 1]> s_89_cast_fp16 = maximum(x = var_3938_cast_fp16, y = var_3939_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> y_89_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_89_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_3942_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3942_cast_fp16")];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3947_keep_dims_0 = const()[name = string("op_3947_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 64, 1]> var_3947_cast_fp16 = reduce_mean(axes = var_3947_axes_0, keep_dims = var_3947_keep_dims_0, x = var_3942_cast_fp16)[name = string("op_3947_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_3948_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3948_cast_fp16")];
            fp32 var_3949_epsilon_0 = const()[name = string("op_3949_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 64, 1]> var_3949_cast_fp16 = inverse(epsilon = var_3949_epsilon_0, x = var_3948_cast_fp16)[name = string("op_3949_cast_fp16")];
            fp16 var_3950_to_fp16 = const()[name = string("op_3950_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 64, 1]> var_3951_cast_fp16 = mul(x = var_3949_cast_fp16, y = var_3950_to_fp16)[name = string("op_3951_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_3953_cast_fp16 = add(x = var_3947_cast_fp16, y = var_3951_cast_fp16)[name = string("op_3953_cast_fp16")];
            fp32 var_3954_epsilon_0 = const()[name = string("op_3954_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 64, 1]> var_3954_cast_fp16 = rsqrt(epsilon = var_3954_epsilon_0, x = var_3953_cast_fp16)[name = string("op_3954_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> v_21_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3954_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<fp16, [64, 512]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137510016))), scale = tensor<fp16, [64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137493184))))[name = string("cos_to_fp16_quantized")];
            tensor<fp16, [1, 4, 64, 512]> var_3956_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3956_cast_fp16")];
            tensor<int32, [4]> var_3961_begin_0 = const()[name = string("op_3961_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3961_end_0 = const()[name = string("op_3961_end_0"), val = tensor<int32, [4]>([1, 4, 64, 512])];
            tensor<bool, [4]> var_3961_end_mask_0 = const()[name = string("op_3961_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 256]> var_3961_cast_fp16 = slice_by_index(begin = var_3961_begin_0, end = var_3961_end_0, end_mask = var_3961_end_mask_0, x = q_21_cast_fp16)[name = string("op_3961_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 256]> var_3962_cast_fp16 = mul(x = var_3961_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_3962_cast_fp16")];
            tensor<int32, [4]> var_3967_begin_0 = const()[name = string("op_3967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3967_end_0 = const()[name = string("op_3967_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_3967_end_mask_0 = const()[name = string("op_3967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 256]> var_3967_cast_fp16 = slice_by_index(begin = var_3967_begin_0, end = var_3967_end_0, end_mask = var_3967_end_mask_0, x = q_21_cast_fp16)[name = string("op_3967_cast_fp16")];
            int32 var_3969 = const()[name = string("op_3969"), val = int32(-1)];
            bool var_3970_interleave_0 = const()[name = string("op_3970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 512]> var_3970_cast_fp16 = concat(axis = var_3969, interleave = var_3970_interleave_0, values = (var_3962_cast_fp16, var_3967_cast_fp16))[name = string("op_3970_cast_fp16")];
            tensor<fp16, [64, 512]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137542848))), scale = tensor<fp16, [64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137575680))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [1, 4, 64, 512]> var_3971_cast_fp16 = mul(x = var_3970_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3971_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> q_23_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3971_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_3974_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_3974_cast_fp16")];
            tensor<int32, [4]> var_3979_begin_0 = const()[name = string("op_3979_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3979_end_0 = const()[name = string("op_3979_end_0"), val = tensor<int32, [4]>([1, 1, 64, 512])];
            tensor<bool, [4]> var_3979_end_mask_0 = const()[name = string("op_3979_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 64, 256]> var_3979_cast_fp16 = slice_by_index(begin = var_3979_begin_0, end = var_3979_end_0, end_mask = var_3979_end_mask_0, x = k_31_cast_fp16)[name = string("op_3979_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 64, 256]> var_3980_cast_fp16 = mul(x = var_3979_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_3980_cast_fp16")];
            tensor<int32, [4]> var_3985_begin_0 = const()[name = string("op_3985_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3985_end_0 = const()[name = string("op_3985_end_0"), val = tensor<int32, [4]>([1, 1, 64, 256])];
            tensor<bool, [4]> var_3985_end_mask_0 = const()[name = string("op_3985_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 64, 256]> var_3985_cast_fp16 = slice_by_index(begin = var_3985_begin_0, end = var_3985_end_0, end_mask = var_3985_end_mask_0, x = k_31_cast_fp16)[name = string("op_3985_cast_fp16")];
            int32 var_3987 = const()[name = string("op_3987"), val = int32(-1)];
            bool var_3988_interleave_0 = const()[name = string("op_3988_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 64, 512]> var_3988_cast_fp16 = concat(axis = var_3987, interleave = var_3988_interleave_0, values = (var_3980_cast_fp16, var_3985_cast_fp16))[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_3989_cast_fp16 = mul(x = var_3988_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_3989_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> k_33_cast_fp16 = add(x = var_3974_cast_fp16, y = var_3989_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 64, 512]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = k_33_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [4, 1, 64, 512]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([4, 1, 1, 64, 512])];
            tensor<fp16, [4, 1, 1, 64, 512]> reshape_20_cast_fp16 = reshape(shape = concat_20, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 64, 512])];
            tensor<fp16, [1, 4, 1, 64, 512]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [4, 1, 64, 512]> reshape_21_cast_fp16 = reshape(shape = concat_21, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 64, 512]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = v_21_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [4, 1, 64, 512]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([4, 1, 1, 64, 512])];
            tensor<fp16, [4, 1, 1, 64, 512]> reshape_22_cast_fp16 = reshape(shape = concat_22, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 64, 512])];
            tensor<fp16, [1, 4, 1, 64, 512]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [4, 1, 64, 512]> reshape_23_cast_fp16 = reshape(shape = concat_23, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4003_transpose_x_0 = const()[name = string("op_4003_transpose_x_0"), val = bool(false)];
            bool var_4003_transpose_y_0 = const()[name = string("op_4003_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 64]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 4, 64, 64]> var_4003_cast_fp16 = matmul(transpose_x = var_4003_transpose_x_0, transpose_y = var_4003_transpose_y_0, x = q_23_cast_fp16, y = transpose_101_cast_fp16)[name = string("op_4003_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_4005_cast_fp16 = add(x = var_4003_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4005_cast_fp16")];
            int32 var_4006 = const()[name = string("op_4006"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_89_cast_fp16 = softmax(axis = var_4006, x = var_4005_cast_fp16)[name = string("w_89_cast_fp16")];
            bool var_4009_transpose_x_0 = const()[name = string("op_4009_transpose_x_0"), val = bool(false)];
            bool var_4009_transpose_y_0 = const()[name = string("op_4009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 512]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 4, 64, 512]> var_4009_cast_fp16 = matmul(transpose_x = var_4009_transpose_x_0, transpose_y = var_4009_transpose_y_0, x = w_89_cast_fp16, y = v_23_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4012_perm_0 = const()[name = string("op_4012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4016 = const()[name = string("op_4016"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 64, 4, 512]> var_4012_cast_fp16 = transpose(perm = var_4012_perm_0, x = var_4009_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 64, 2048]> input_63_cast_fp16 = reshape(shape = var_4016, x = var_4012_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34267520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35316160))))[name = string("tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4021_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_4021_cast_fp16")];
            bool var_4024 = const()[name = string("op_4024"), val = bool(true)];
            tensor<int32, [1]> var_4025_axes_0 = const()[name = string("op_4025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4025_cast_fp16 = reduce_max(axes = var_4025_axes_0, keep_dims = var_4024, x = var_4021_cast_fp16)[name = string("op_4025_cast_fp16")];
            fp16 var_4026_to_fp16 = const()[name = string("op_4026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_91_cast_fp16 = maximum(x = var_4025_cast_fp16, y = var_4026_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_91_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_91_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4029_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<int32, [1]> var_4034_axes_0 = const()[name = string("op_4034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4034_keep_dims_0 = const()[name = string("op_4034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4034_cast_fp16 = reduce_mean(axes = var_4034_axes_0, keep_dims = var_4034_keep_dims_0, x = var_4029_cast_fp16)[name = string("op_4034_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4035_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_4035_cast_fp16")];
            fp32 var_4036_epsilon_0 = const()[name = string("op_4036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4036_cast_fp16 = inverse(epsilon = var_4036_epsilon_0, x = var_4035_cast_fp16)[name = string("op_4036_cast_fp16")];
            fp16 var_4037_to_fp16 = const()[name = string("op_4037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4038_cast_fp16 = mul(x = var_4036_cast_fp16, y = var_4037_to_fp16)[name = string("op_4038_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4040_cast_fp16 = add(x = var_4034_cast_fp16, y = var_4038_cast_fp16)[name = string("op_4040_cast_fp16")];
            fp32 var_4041_epsilon_0 = const()[name = string("op_4041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4041_cast_fp16 = rsqrt(epsilon = var_4041_epsilon_0, x = var_4040_cast_fp16)[name = string("op_4041_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_79_cast_fp16 = mul(x = y_91_cast_fp16, y = var_4041_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35317248)))];
            tensor<fp16, [1, 64, 512]> var_4043_cast_fp16 = mul(x = out_79_cast_fp16, y = tm_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_4043_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_101_cast_fp16 = add(x = x_91_cast_fp16, y = var_4043_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4046_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_4046_cast_fp16")];
            bool var_4049 = const()[name = string("op_4049"), val = bool(true)];
            tensor<int32, [1]> var_4050_axes_0 = const()[name = string("op_4050_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4050_cast_fp16 = reduce_max(axes = var_4050_axes_0, keep_dims = var_4049, x = var_4046_cast_fp16)[name = string("op_4050_cast_fp16")];
            fp16 var_4051_to_fp16 = const()[name = string("op_4051_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_93_cast_fp16 = maximum(x = var_4050_cast_fp16, y = var_4051_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_93_cast_fp16 = real_div(x = x_101_cast_fp16, y = s_93_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4054_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [1]> var_4059_axes_0 = const()[name = string("op_4059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4059_keep_dims_0 = const()[name = string("op_4059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4059_cast_fp16 = reduce_mean(axes = var_4059_axes_0, keep_dims = var_4059_keep_dims_0, x = var_4054_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4060_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_4060_cast_fp16")];
            fp32 var_4061_epsilon_0 = const()[name = string("op_4061_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4061_cast_fp16 = inverse(epsilon = var_4061_epsilon_0, x = var_4060_cast_fp16)[name = string("op_4061_cast_fp16")];
            fp16 var_4062_to_fp16 = const()[name = string("op_4062_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4063_cast_fp16 = mul(x = var_4061_cast_fp16, y = var_4062_to_fp16)[name = string("op_4063_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4065_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4063_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp32 var_4066_epsilon_0 = const()[name = string("op_4066_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4066_cast_fp16 = rsqrt(epsilon = var_4066_epsilon_0, x = var_4065_cast_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_81_cast_fp16 = mul(x = y_93_cast_fp16, y = var_4066_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35318336)))];
            tensor<fp16, [1, 64, 512]> input_65_cast_fp16 = mul(x = out_81_cast_fp16, y = tm_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35319424))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36368064))))[name = string("tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_50_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_50_cast_fp16")];
            string var_4073_mode_0 = const()[name = string("op_4073_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_4073_cast_fp16 = gelu(mode = var_4073_mode_0, x = linear_50_cast_fp16)[name = string("op_4073_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36372224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37420864))))[name = string("tm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_51_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_67_cast_fp16 = mul(x = var_4073_cast_fp16, y = linear_51_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37425024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38473664))))[name = string("tm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4081_cast_fp16 = abs(x = linear_52_cast_fp16)[name = string("op_4081_cast_fp16")];
            bool var_4084 = const()[name = string("op_4084"), val = bool(true)];
            tensor<int32, [1]> var_4085_axes_0 = const()[name = string("op_4085_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4085_cast_fp16 = reduce_max(axes = var_4085_axes_0, keep_dims = var_4084, x = var_4081_cast_fp16)[name = string("op_4085_cast_fp16")];
            fp16 var_4086_to_fp16 = const()[name = string("op_4086_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_95_cast_fp16 = maximum(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_95_cast_fp16 = real_div(x = linear_52_cast_fp16, y = s_95_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4089_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_4089_cast_fp16")];
            tensor<int32, [1]> var_4094_axes_0 = const()[name = string("op_4094_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4094_keep_dims_0 = const()[name = string("op_4094_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4094_cast_fp16 = reduce_mean(axes = var_4094_axes_0, keep_dims = var_4094_keep_dims_0, x = var_4089_cast_fp16)[name = string("op_4094_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4095_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_4095_cast_fp16")];
            fp32 var_4096_epsilon_0 = const()[name = string("op_4096_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4096_cast_fp16 = inverse(epsilon = var_4096_epsilon_0, x = var_4095_cast_fp16)[name = string("op_4096_cast_fp16")];
            fp16 var_4097_to_fp16 = const()[name = string("op_4097_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4098_cast_fp16 = mul(x = var_4096_cast_fp16, y = var_4097_to_fp16)[name = string("op_4098_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4100_cast_fp16 = add(x = var_4094_cast_fp16, y = var_4098_cast_fp16)[name = string("op_4100_cast_fp16")];
            fp32 var_4101_epsilon_0 = const()[name = string("op_4101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4101_cast_fp16 = rsqrt(epsilon = var_4101_epsilon_0, x = var_4100_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_83_cast_fp16 = mul(x = y_95_cast_fp16, y = var_4101_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38474752)))];
            tensor<fp16, [1, 64, 512]> var_4103_cast_fp16 = mul(x = out_83_cast_fp16, y = tm_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_69_cast_fp16 = add(x = x_101_cast_fp16, y = var_4103_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38475840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38738048))))[name = string("tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_53_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_53_cast_fp16")];
            string x_105_mode_0 = const()[name = string("x_105_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_105_cast_fp16 = gelu(mode = x_105_mode_0, x = linear_53_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [4]> var_4123_begin_0 = const()[name = string("op_4123_begin_0"), val = tensor<int32, [4]>([0, 0, 5, 0])];
            tensor<int32, [4]> var_4123_end_0 = const()[name = string("op_4123_end_0"), val = tensor<int32, [4]>([1, 64, 6, 512])];
            tensor<bool, [4]> var_4123_end_mask_0 = const()[name = string("op_4123_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4123_squeeze_mask_0 = const()[name = string("op_4123_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_4123_cast_fp16 = slice_by_index(begin = var_4123_begin_0, end = var_4123_end_0, end_mask = var_4123_end_mask_0, squeeze_mask = var_4123_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_71_cast_fp16 = mul(x = x_105_cast_fp16, y = var_4123_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38739136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39001344))))[name = string("tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4133_cast_fp16 = abs(x = linear_54_cast_fp16)[name = string("op_4133_cast_fp16")];
            bool var_4136 = const()[name = string("op_4136"), val = bool(true)];
            tensor<int32, [1]> var_4137_axes_0 = const()[name = string("op_4137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4137_cast_fp16 = reduce_max(axes = var_4137_axes_0, keep_dims = var_4136, x = var_4133_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp16 var_4138_to_fp16 = const()[name = string("op_4138_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_97_cast_fp16 = maximum(x = var_4137_cast_fp16, y = var_4138_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_97_cast_fp16 = real_div(x = linear_54_cast_fp16, y = s_97_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4141_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_4141_cast_fp16")];
            tensor<int32, [1]> var_4146_axes_0 = const()[name = string("op_4146_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4146_keep_dims_0 = const()[name = string("op_4146_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4146_cast_fp16 = reduce_mean(axes = var_4146_axes_0, keep_dims = var_4146_keep_dims_0, x = var_4141_cast_fp16)[name = string("op_4146_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4147_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_4147_cast_fp16")];
            fp32 var_4148_epsilon_0 = const()[name = string("op_4148_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4148_cast_fp16 = inverse(epsilon = var_4148_epsilon_0, x = var_4147_cast_fp16)[name = string("op_4148_cast_fp16")];
            fp16 var_4149_to_fp16 = const()[name = string("op_4149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4150_cast_fp16 = mul(x = var_4148_cast_fp16, y = var_4149_to_fp16)[name = string("op_4150_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4152_cast_fp16 = add(x = var_4146_cast_fp16, y = var_4150_cast_fp16)[name = string("op_4152_cast_fp16")];
            fp32 var_4153_epsilon_0 = const()[name = string("op_4153_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4153_cast_fp16 = rsqrt(epsilon = var_4153_epsilon_0, x = var_4152_cast_fp16)[name = string("op_4153_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_85_cast_fp16 = mul(x = y_97_cast_fp16, y = var_4153_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39002432)))];
            tensor<fp16, [1, 64, 512]> var_4155_cast_fp16 = mul(x = out_85_cast_fp16, y = tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4155_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_11_cast_fp16 = add(x = input_69_cast_fp16, y = var_4155_cast_fp16)[name = string("h_11_cast_fp16")];
            tensor<fp16, [1]> tm_layers_5_layer_scalar_to_fp16 = const()[name = string("tm_layers_5_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-2])];
            tensor<fp16, [1, 64, 512]> x_109_cast_fp16 = mul(x = h_11_cast_fp16, y = tm_layers_5_layer_scalar_to_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4159_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_4159_cast_fp16")];
            bool var_4162 = const()[name = string("op_4162"), val = bool(true)];
            tensor<int32, [1]> var_4163_axes_0 = const()[name = string("op_4163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4163_cast_fp16 = reduce_max(axes = var_4163_axes_0, keep_dims = var_4162, x = var_4159_cast_fp16)[name = string("op_4163_cast_fp16")];
            fp16 var_4164_to_fp16 = const()[name = string("op_4164_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_99_cast_fp16 = maximum(x = var_4163_cast_fp16, y = var_4164_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_99_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_99_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4167_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_4167_cast_fp16")];
            tensor<int32, [1]> var_4172_axes_0 = const()[name = string("op_4172_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4172_keep_dims_0 = const()[name = string("op_4172_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4172_cast_fp16 = reduce_mean(axes = var_4172_axes_0, keep_dims = var_4172_keep_dims_0, x = var_4167_cast_fp16)[name = string("op_4172_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4173_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4174_cast_fp16 = inverse(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            fp16 var_4175_to_fp16 = const()[name = string("op_4175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4176_cast_fp16 = mul(x = var_4174_cast_fp16, y = var_4175_to_fp16)[name = string("op_4176_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4178_cast_fp16 = add(x = var_4172_cast_fp16, y = var_4176_cast_fp16)[name = string("op_4178_cast_fp16")];
            fp32 var_4179_epsilon_0 = const()[name = string("op_4179_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4179_cast_fp16 = rsqrt(epsilon = var_4179_epsilon_0, x = var_4178_cast_fp16)[name = string("op_4179_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_87_cast_fp16 = mul(x = y_99_cast_fp16, y = var_4179_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39003520)))];
            tensor<fp16, [1, 64, 512]> input_73_cast_fp16 = mul(x = out_87_cast_fp16, y = tm_layers_6_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39004608))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39528960))))[name = string("tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [4]> var_4189 = const()[name = string("op_4189"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_4190_cast_fp16 = reshape(shape = var_4189, x = linear_55_cast_fp16)[name = string("op_4190_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39531072))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39793280))))[name = string("tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_4201 = const()[name = string("op_4201"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_4202_cast_fp16 = reshape(shape = var_4201, x = linear_56_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39794368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40056576))))[name = string("tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_57_cast_fp16)[name = string("op_4214_cast_fp16")];
            tensor<int32, [4]> x_115_perm_0 = const()[name = string("x_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4190_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 4, 64, 256]> var_4218_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_4218_cast_fp16")];
            bool var_4221 = const()[name = string("op_4221"), val = bool(true)];
            tensor<int32, [1]> var_4222_axes_0 = const()[name = string("op_4222_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_4222_cast_fp16 = reduce_max(axes = var_4222_axes_0, keep_dims = var_4221, x = var_4218_cast_fp16)[name = string("op_4222_cast_fp16")];
            fp16 var_4223_to_fp16 = const()[name = string("op_4223_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_101_cast_fp16 = maximum(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_101_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_101_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4226_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_4226_cast_fp16")];
            tensor<int32, [1]> var_4231_axes_0 = const()[name = string("op_4231_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4231_keep_dims_0 = const()[name = string("op_4231_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_4231_cast_fp16 = reduce_mean(axes = var_4231_axes_0, keep_dims = var_4231_keep_dims_0, x = var_4226_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_4232_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_4232_cast_fp16")];
            fp32 var_4233_epsilon_0 = const()[name = string("op_4233_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_4233_cast_fp16 = inverse(epsilon = var_4233_epsilon_0, x = var_4232_cast_fp16)[name = string("op_4233_cast_fp16")];
            fp16 var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_4235_cast_fp16 = mul(x = var_4233_cast_fp16, y = var_4234_to_fp16)[name = string("op_4235_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_4237_cast_fp16 = add(x = var_4231_cast_fp16, y = var_4235_cast_fp16)[name = string("op_4237_cast_fp16")];
            fp32 var_4238_epsilon_0 = const()[name = string("op_4238_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_4238_cast_fp16 = rsqrt(epsilon = var_4238_epsilon_0, x = var_4237_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_89_cast_fp16 = mul(x = y_101_cast_fp16, y = var_4238_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40057664)))];
            tensor<fp16, [1, 4, 64, 256]> q_25_cast_fp16 = mul(x = out_89_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4202_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 2, 64, 256]> var_4241_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_4241_cast_fp16")];
            bool var_4244 = const()[name = string("op_4244"), val = bool(true)];
            tensor<int32, [1]> var_4245_axes_0 = const()[name = string("op_4245_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_4245_cast_fp16 = reduce_max(axes = var_4245_axes_0, keep_dims = var_4244, x = var_4241_cast_fp16)[name = string("op_4245_cast_fp16")];
            fp16 var_4246_to_fp16 = const()[name = string("op_4246_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_103_cast_fp16 = maximum(x = var_4245_cast_fp16, y = var_4246_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_103_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_103_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4249_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<int32, [1]> var_4254_axes_0 = const()[name = string("op_4254_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4254_keep_dims_0 = const()[name = string("op_4254_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_4254_cast_fp16 = reduce_mean(axes = var_4254_axes_0, keep_dims = var_4254_keep_dims_0, x = var_4249_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4255_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_4255_cast_fp16")];
            fp32 var_4256_epsilon_0 = const()[name = string("op_4256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_4256_cast_fp16 = inverse(epsilon = var_4256_epsilon_0, x = var_4255_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp16 var_4257_to_fp16 = const()[name = string("op_4257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_4258_cast_fp16 = mul(x = var_4256_cast_fp16, y = var_4257_to_fp16)[name = string("op_4258_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4260_cast_fp16 = add(x = var_4254_cast_fp16, y = var_4258_cast_fp16)[name = string("op_4260_cast_fp16")];
            fp32 var_4261_epsilon_0 = const()[name = string("op_4261_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_4261_cast_fp16 = rsqrt(epsilon = var_4261_epsilon_0, x = var_4260_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_91_cast_fp16 = mul(x = y_103_cast_fp16, y = var_4261_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058240)))];
            tensor<fp16, [1, 2, 64, 256]> k_37_cast_fp16 = mul(x = out_91_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_37_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_115_cast_fp16 = transpose(perm = x_115_perm_0, x = var_4214_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 2, 64, 256]> var_4264_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_4264_cast_fp16")];
            bool var_4267 = const()[name = string("op_4267"), val = bool(true)];
            tensor<int32, [1]> var_4268_axes_0 = const()[name = string("op_4268_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_4268_cast_fp16 = reduce_max(axes = var_4268_axes_0, keep_dims = var_4267, x = var_4264_cast_fp16)[name = string("op_4268_cast_fp16")];
            fp16 var_4269_to_fp16 = const()[name = string("op_4269_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_105_cast_fp16 = maximum(x = var_4268_cast_fp16, y = var_4269_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_105_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_105_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4272_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_4272_cast_fp16")];
            tensor<int32, [1]> var_4277_axes_0 = const()[name = string("op_4277_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4277_keep_dims_0 = const()[name = string("op_4277_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_4277_cast_fp16 = reduce_mean(axes = var_4277_axes_0, keep_dims = var_4277_keep_dims_0, x = var_4272_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4278_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp32 var_4279_epsilon_0 = const()[name = string("op_4279_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_4279_cast_fp16 = inverse(epsilon = var_4279_epsilon_0, x = var_4278_cast_fp16)[name = string("op_4279_cast_fp16")];
            fp16 var_4280_to_fp16 = const()[name = string("op_4280_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_4281_cast_fp16 = mul(x = var_4279_cast_fp16, y = var_4280_to_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4283_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4281_cast_fp16)[name = string("op_4283_cast_fp16")];
            fp32 var_4284_epsilon_0 = const()[name = string("op_4284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_4284_cast_fp16 = rsqrt(epsilon = var_4284_epsilon_0, x = var_4283_cast_fp16)[name = string("op_4284_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_25_cast_fp16 = mul(x = y_105_cast_fp16, y = var_4284_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4286_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4286_cast_fp16")];
            tensor<int32, [4]> var_4291_begin_0 = const()[name = string("op_4291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4291_end_0 = const()[name = string("op_4291_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_4291_end_mask_0 = const()[name = string("op_4291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_4291_cast_fp16 = slice_by_index(begin = var_4291_begin_0, end = var_4291_end_0, end_mask = var_4291_end_mask_0, x = q_25_cast_fp16)[name = string("op_4291_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_4292_cast_fp16 = mul(x = var_4291_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_4292_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = q_25_cast_fp16)[name = string("op_4297_cast_fp16")];
            int32 var_4299 = const()[name = string("op_4299"), val = int32(-1)];
            bool var_4300_interleave_0 = const()[name = string("op_4300_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_4300_cast_fp16 = concat(axis = var_4299, interleave = var_4300_interleave_0, values = (var_4292_cast_fp16, var_4297_cast_fp16))[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4301_cast_fp16 = mul(x = var_4300_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4301_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_27_cast_fp16 = add(x = var_4286_cast_fp16, y = var_4301_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4304_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4304_cast_fp16")];
            tensor<int32, [4]> var_4309_begin_0 = const()[name = string("op_4309_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4309_end_0 = const()[name = string("op_4309_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_4309_end_mask_0 = const()[name = string("op_4309_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_4309_cast_fp16 = slice_by_index(begin = var_4309_begin_0, end = var_4309_end_0, end_mask = var_4309_end_mask_0, x = k_37_cast_fp16)[name = string("op_4309_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4310_cast_fp16 = mul(x = var_4309_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_4310_cast_fp16")];
            tensor<int32, [4]> var_4315_begin_0 = const()[name = string("op_4315_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4315_end_0 = const()[name = string("op_4315_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_4315_end_mask_0 = const()[name = string("op_4315_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_4315_cast_fp16 = slice_by_index(begin = var_4315_begin_0, end = var_4315_end_0, end_mask = var_4315_end_mask_0, x = k_37_cast_fp16)[name = string("op_4315_cast_fp16")];
            int32 var_4317 = const()[name = string("op_4317"), val = int32(-1)];
            bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_4318_cast_fp16 = concat(axis = var_4317, interleave = var_4318_interleave_0, values = (var_4310_cast_fp16, var_4315_cast_fp16))[name = string("op_4318_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4319_cast_fp16 = mul(x = var_4318_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_39_cast_fp16 = add(x = var_4304_cast_fp16, y = var_4319_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = k_39_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [4, 1, 64, 256]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
            tensor<int32, [5]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_24_cast_fp16 = reshape(shape = concat_24, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [4, 1, 64, 256]> reshape_25_cast_fp16 = reshape(shape = concat_25, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = v_25_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [4, 1, 64, 256]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
            tensor<int32, [5]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_26_cast_fp16 = reshape(shape = concat_26, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [4, 1, 64, 256]> reshape_27_cast_fp16 = reshape(shape = concat_27, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4333_transpose_x_0 = const()[name = string("op_4333_transpose_x_0"), val = bool(false)];
            bool var_4333_transpose_y_0 = const()[name = string("op_4333_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_102_cast_fp16 = transpose(perm = transpose_102_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 4, 64, 64]> var_4333_cast_fp16 = matmul(transpose_x = var_4333_transpose_x_0, transpose_y = var_4333_transpose_y_0, x = q_27_cast_fp16, y = transpose_102_cast_fp16)[name = string("op_4333_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_4335_cast_fp16 = add(x = var_4333_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4335_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_105_cast_fp16 = softmax(axis = var_4336, x = var_4335_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4339_transpose_x_0 = const()[name = string("op_4339_transpose_x_0"), val = bool(false)];
            bool var_4339_transpose_y_0 = const()[name = string("op_4339_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 4, 64, 256]> var_4339_cast_fp16 = matmul(transpose_x = var_4339_transpose_x_0, transpose_y = var_4339_transpose_y_0, x = w_105_cast_fp16, y = v_27_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<int32, [4]> var_4342_perm_0 = const()[name = string("op_4342_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4346 = const()[name = string("op_4346"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_4342_cast_fp16 = transpose(perm = var_4342_perm_0, x = var_4339_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 64, 1024]> input_75_cast_fp16 = reshape(shape = var_4346, x = var_4342_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40583168))))[name = string("tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4351_cast_fp16 = abs(x = linear_58_cast_fp16)[name = string("op_4351_cast_fp16")];
            bool var_4354 = const()[name = string("op_4354"), val = bool(true)];
            tensor<int32, [1]> var_4355_axes_0 = const()[name = string("op_4355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4355_cast_fp16 = reduce_max(axes = var_4355_axes_0, keep_dims = var_4354, x = var_4351_cast_fp16)[name = string("op_4355_cast_fp16")];
            fp16 var_4356_to_fp16 = const()[name = string("op_4356_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_107_cast_fp16 = maximum(x = var_4355_cast_fp16, y = var_4356_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_107_cast_fp16 = real_div(x = linear_58_cast_fp16, y = s_107_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4359_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<int32, [1]> var_4364_axes_0 = const()[name = string("op_4364_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4364_keep_dims_0 = const()[name = string("op_4364_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4364_cast_fp16 = reduce_mean(axes = var_4364_axes_0, keep_dims = var_4364_keep_dims_0, x = var_4359_cast_fp16)[name = string("op_4364_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4365_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_4365_cast_fp16")];
            fp32 var_4366_epsilon_0 = const()[name = string("op_4366_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4366_cast_fp16 = inverse(epsilon = var_4366_epsilon_0, x = var_4365_cast_fp16)[name = string("op_4366_cast_fp16")];
            fp16 var_4367_to_fp16 = const()[name = string("op_4367_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4368_cast_fp16 = mul(x = var_4366_cast_fp16, y = var_4367_to_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4370_cast_fp16 = add(x = var_4364_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            fp32 var_4371_epsilon_0 = const()[name = string("op_4371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4371_cast_fp16 = rsqrt(epsilon = var_4371_epsilon_0, x = var_4370_cast_fp16)[name = string("op_4371_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_93_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4371_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40584256)))];
            tensor<fp16, [1, 64, 512]> var_4373_cast_fp16 = mul(x = out_93_cast_fp16, y = tm_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_4373_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = var_4373_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4376_cast_fp16 = abs(x = x_119_cast_fp16)[name = string("op_4376_cast_fp16")];
            bool var_4379 = const()[name = string("op_4379"), val = bool(true)];
            tensor<int32, [1]> var_4380_axes_0 = const()[name = string("op_4380_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4380_cast_fp16 = reduce_max(axes = var_4380_axes_0, keep_dims = var_4379, x = var_4376_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 var_4381_to_fp16 = const()[name = string("op_4381_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_109_cast_fp16 = maximum(x = var_4380_cast_fp16, y = var_4381_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_109_cast_fp16 = real_div(x = x_119_cast_fp16, y = s_109_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4384_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4384_cast_fp16")];
            tensor<int32, [1]> var_4389_axes_0 = const()[name = string("op_4389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4389_keep_dims_0 = const()[name = string("op_4389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4389_cast_fp16 = reduce_mean(axes = var_4389_axes_0, keep_dims = var_4389_keep_dims_0, x = var_4384_cast_fp16)[name = string("op_4389_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4390_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_4390_cast_fp16")];
            fp32 var_4391_epsilon_0 = const()[name = string("op_4391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4391_cast_fp16 = inverse(epsilon = var_4391_epsilon_0, x = var_4390_cast_fp16)[name = string("op_4391_cast_fp16")];
            fp16 var_4392_to_fp16 = const()[name = string("op_4392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = string("op_4393_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4395_cast_fp16 = add(x = var_4389_cast_fp16, y = var_4393_cast_fp16)[name = string("op_4395_cast_fp16")];
            fp32 var_4396_epsilon_0 = const()[name = string("op_4396_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4396_cast_fp16 = rsqrt(epsilon = var_4396_epsilon_0, x = var_4395_cast_fp16)[name = string("op_4396_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_95_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4396_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40585344)))];
            tensor<fp16, [1, 64, 512]> input_77_cast_fp16 = mul(x = out_95_cast_fp16, y = tm_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40586432))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41635072))))[name = string("tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_59_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_59_cast_fp16")];
            string var_4403_mode_0 = const()[name = string("op_4403_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_4403_cast_fp16 = gelu(mode = var_4403_mode_0, x = linear_59_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41639232))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42687872))))[name = string("tm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_60_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_79_cast_fp16 = mul(x = var_4403_cast_fp16, y = linear_60_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42692032))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43740672))))[name = string("tm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4411_cast_fp16 = abs(x = linear_61_cast_fp16)[name = string("op_4411_cast_fp16")];
            bool var_4414 = const()[name = string("op_4414"), val = bool(true)];
            tensor<int32, [1]> var_4415_axes_0 = const()[name = string("op_4415_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4415_cast_fp16 = reduce_max(axes = var_4415_axes_0, keep_dims = var_4414, x = var_4411_cast_fp16)[name = string("op_4415_cast_fp16")];
            fp16 var_4416_to_fp16 = const()[name = string("op_4416_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_111_cast_fp16 = maximum(x = var_4415_cast_fp16, y = var_4416_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_111_cast_fp16 = real_div(x = linear_61_cast_fp16, y = s_111_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4419_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4419_cast_fp16")];
            tensor<int32, [1]> var_4424_axes_0 = const()[name = string("op_4424_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4424_keep_dims_0 = const()[name = string("op_4424_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4424_cast_fp16 = reduce_mean(axes = var_4424_axes_0, keep_dims = var_4424_keep_dims_0, x = var_4419_cast_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4425_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4426_cast_fp16 = inverse(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            fp16 var_4427_to_fp16 = const()[name = string("op_4427_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4428_cast_fp16 = mul(x = var_4426_cast_fp16, y = var_4427_to_fp16)[name = string("op_4428_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4430_cast_fp16 = add(x = var_4424_cast_fp16, y = var_4428_cast_fp16)[name = string("op_4430_cast_fp16")];
            fp32 var_4431_epsilon_0 = const()[name = string("op_4431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4431_cast_fp16 = rsqrt(epsilon = var_4431_epsilon_0, x = var_4430_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_97_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4431_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43741760)))];
            tensor<fp16, [1, 64, 512]> var_4433_cast_fp16 = mul(x = out_97_cast_fp16, y = tm_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4433_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_81_cast_fp16 = add(x = x_119_cast_fp16, y = var_4433_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43742848))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44005056))))[name = string("tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_62_cast_fp16")];
            string x_123_mode_0 = const()[name = string("x_123_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_123_cast_fp16 = gelu(mode = x_123_mode_0, x = linear_62_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [4]> var_4453_begin_0 = const()[name = string("op_4453_begin_0"), val = tensor<int32, [4]>([0, 0, 6, 0])];
            tensor<int32, [4]> var_4453_end_0 = const()[name = string("op_4453_end_0"), val = tensor<int32, [4]>([1, 64, 7, 512])];
            tensor<bool, [4]> var_4453_end_mask_0 = const()[name = string("op_4453_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4453_squeeze_mask_0 = const()[name = string("op_4453_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_4453_cast_fp16 = slice_by_index(begin = var_4453_begin_0, end = var_4453_end_0, end_mask = var_4453_end_mask_0, squeeze_mask = var_4453_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_83_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4453_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44006144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44268352))))[name = string("tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4463_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4463_cast_fp16")];
            bool var_4466 = const()[name = string("op_4466"), val = bool(true)];
            tensor<int32, [1]> var_4467_axes_0 = const()[name = string("op_4467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4467_cast_fp16 = reduce_max(axes = var_4467_axes_0, keep_dims = var_4466, x = var_4463_cast_fp16)[name = string("op_4467_cast_fp16")];
            fp16 var_4468_to_fp16 = const()[name = string("op_4468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_113_cast_fp16 = maximum(x = var_4467_cast_fp16, y = var_4468_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_113_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_113_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4471_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [1]> var_4476_axes_0 = const()[name = string("op_4476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4476_keep_dims_0 = const()[name = string("op_4476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4476_cast_fp16 = reduce_mean(axes = var_4476_axes_0, keep_dims = var_4476_keep_dims_0, x = var_4471_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4477_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_fp16)[name = string("op_4477_cast_fp16")];
            fp32 var_4478_epsilon_0 = const()[name = string("op_4478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4478_cast_fp16 = inverse(epsilon = var_4478_epsilon_0, x = var_4477_cast_fp16)[name = string("op_4478_cast_fp16")];
            fp16 var_4479_to_fp16 = const()[name = string("op_4479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4480_cast_fp16 = mul(x = var_4478_cast_fp16, y = var_4479_to_fp16)[name = string("op_4480_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4482_cast_fp16 = add(x = var_4476_cast_fp16, y = var_4480_cast_fp16)[name = string("op_4482_cast_fp16")];
            fp32 var_4483_epsilon_0 = const()[name = string("op_4483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4483_cast_fp16 = rsqrt(epsilon = var_4483_epsilon_0, x = var_4482_cast_fp16)[name = string("op_4483_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_99_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4483_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44269440)))];
            tensor<fp16, [1, 64, 512]> var_4485_cast_fp16 = mul(x = out_99_cast_fp16, y = tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_13_cast_fp16 = add(x = input_81_cast_fp16, y = var_4485_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<fp16, [1]> tm_layers_6_layer_scalar_to_fp16 = const()[name = string("tm_layers_6_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.cep-3])];
            tensor<fp16, [1, 64, 512]> x_127_cast_fp16 = mul(x = h_13_cast_fp16, y = tm_layers_6_layer_scalar_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4489_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_4489_cast_fp16")];
            bool var_4492 = const()[name = string("op_4492"), val = bool(true)];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4493_cast_fp16 = reduce_max(axes = var_4493_axes_0, keep_dims = var_4492, x = var_4489_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_115_cast_fp16 = maximum(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_115_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_115_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4497_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<int32, [1]> var_4502_axes_0 = const()[name = string("op_4502_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4502_keep_dims_0 = const()[name = string("op_4502_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4502_cast_fp16 = reduce_mean(axes = var_4502_axes_0, keep_dims = var_4502_keep_dims_0, x = var_4497_cast_fp16)[name = string("op_4502_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4503_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4504_cast_fp16 = inverse(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            fp16 var_4505_to_fp16 = const()[name = string("op_4505_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4506_cast_fp16 = mul(x = var_4504_cast_fp16, y = var_4505_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4508_cast_fp16 = add(x = var_4502_cast_fp16, y = var_4506_cast_fp16)[name = string("op_4508_cast_fp16")];
            fp32 var_4509_epsilon_0 = const()[name = string("op_4509_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4509_cast_fp16 = rsqrt(epsilon = var_4509_epsilon_0, x = var_4508_cast_fp16)[name = string("op_4509_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_101_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4509_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44270528)))];
            tensor<fp16, [1, 64, 512]> input_85_cast_fp16 = mul(x = out_101_cast_fp16, y = tm_layers_7_input_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44271616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44795968))))[name = string("tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4519 = const()[name = string("op_4519"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_4520_cast_fp16 = reshape(shape = var_4519, x = linear_64_cast_fp16)[name = string("op_4520_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44798080))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45060288))))[name = string("tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4531 = const()[name = string("op_4531"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_4532_cast_fp16 = reshape(shape = var_4531, x = linear_65_cast_fp16)[name = string("op_4532_cast_fp16")];
            tensor<int32, [4]> x_131_perm_0 = const()[name = string("x_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45061376))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45323584))))[name = string("tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4543 = const()[name = string("op_4543"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_4544_cast_fp16 = reshape(shape = var_4543, x = linear_66_cast_fp16)[name = string("op_4544_cast_fp16")];
            tensor<int32, [4]> x_133_perm_0 = const()[name = string("x_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_4520_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 4, 64, 256]> var_4548_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_4548_cast_fp16")];
            bool var_4551 = const()[name = string("op_4551"), val = bool(true)];
            tensor<int32, [1]> var_4552_axes_0 = const()[name = string("op_4552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_4552_cast_fp16 = reduce_max(axes = var_4552_axes_0, keep_dims = var_4551, x = var_4548_cast_fp16)[name = string("op_4552_cast_fp16")];
            fp16 var_4553_to_fp16 = const()[name = string("op_4553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_117_cast_fp16 = maximum(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_117_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_117_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4556_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<int32, [1]> var_4561_axes_0 = const()[name = string("op_4561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4561_keep_dims_0 = const()[name = string("op_4561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_4561_cast_fp16 = reduce_mean(axes = var_4561_axes_0, keep_dims = var_4561_keep_dims_0, x = var_4556_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_4562_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4562_cast_fp16")];
            fp32 var_4563_epsilon_0 = const()[name = string("op_4563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_4563_cast_fp16 = inverse(epsilon = var_4563_epsilon_0, x = var_4562_cast_fp16)[name = string("op_4563_cast_fp16")];
            fp16 var_4564_to_fp16 = const()[name = string("op_4564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_4565_cast_fp16 = mul(x = var_4563_cast_fp16, y = var_4564_to_fp16)[name = string("op_4565_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_4567_cast_fp16 = add(x = var_4561_cast_fp16, y = var_4565_cast_fp16)[name = string("op_4567_cast_fp16")];
            fp32 var_4568_epsilon_0 = const()[name = string("op_4568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_4568_cast_fp16 = rsqrt(epsilon = var_4568_epsilon_0, x = var_4567_cast_fp16)[name = string("op_4568_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_103_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4568_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_29_cast_fp16 = mul(x = out_103_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_131_cast_fp16 = transpose(perm = x_131_perm_0, x = var_4532_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2, 64, 256]> var_4571_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_4571_cast_fp16")];
            bool var_4574 = const()[name = string("op_4574"), val = bool(true)];
            tensor<int32, [1]> var_4575_axes_0 = const()[name = string("op_4575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_4575_cast_fp16 = reduce_max(axes = var_4575_axes_0, keep_dims = var_4574, x = var_4571_cast_fp16)[name = string("op_4575_cast_fp16")];
            fp16 var_4576_to_fp16 = const()[name = string("op_4576_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_119_cast_fp16 = maximum(x = var_4575_cast_fp16, y = var_4576_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_119_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_119_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4579_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4579_cast_fp16")];
            tensor<int32, [1]> var_4584_axes_0 = const()[name = string("op_4584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4584_keep_dims_0 = const()[name = string("op_4584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_4584_cast_fp16 = reduce_mean(axes = var_4584_axes_0, keep_dims = var_4584_keep_dims_0, x = var_4579_cast_fp16)[name = string("op_4584_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4585_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4585_cast_fp16")];
            fp32 var_4586_epsilon_0 = const()[name = string("op_4586_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_4586_cast_fp16 = inverse(epsilon = var_4586_epsilon_0, x = var_4585_cast_fp16)[name = string("op_4586_cast_fp16")];
            fp16 var_4587_to_fp16 = const()[name = string("op_4587_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_4588_cast_fp16 = mul(x = var_4586_cast_fp16, y = var_4587_to_fp16)[name = string("op_4588_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4590_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4588_cast_fp16)[name = string("op_4590_cast_fp16")];
            fp32 var_4591_epsilon_0 = const()[name = string("op_4591_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_4591_cast_fp16 = rsqrt(epsilon = var_4591_epsilon_0, x = var_4590_cast_fp16)[name = string("op_4591_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_105_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4591_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [256]> tm_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45324672)))];
            tensor<fp16, [1, 2, 64, 256]> k_43_cast_fp16 = mul(x = out_105_cast_fp16, y = tm_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_133_cast_fp16 = transpose(perm = x_133_perm_0, x = var_4544_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 2, 64, 256]> var_4594_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_4594_cast_fp16")];
            bool var_4597 = const()[name = string("op_4597"), val = bool(true)];
            tensor<int32, [1]> var_4598_axes_0 = const()[name = string("op_4598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_4598_cast_fp16 = reduce_max(axes = var_4598_axes_0, keep_dims = var_4597, x = var_4594_cast_fp16)[name = string("op_4598_cast_fp16")];
            fp16 var_4599_to_fp16 = const()[name = string("op_4599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_121_cast_fp16 = maximum(x = var_4598_cast_fp16, y = var_4599_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_121_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_121_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4602_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4602_cast_fp16")];
            tensor<int32, [1]> var_4607_axes_0 = const()[name = string("op_4607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4607_keep_dims_0 = const()[name = string("op_4607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_4607_cast_fp16 = reduce_mean(axes = var_4607_axes_0, keep_dims = var_4607_keep_dims_0, x = var_4602_cast_fp16)[name = string("op_4607_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4608_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4608_cast_fp16")];
            fp32 var_4609_epsilon_0 = const()[name = string("op_4609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_4609_cast_fp16 = inverse(epsilon = var_4609_epsilon_0, x = var_4608_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_4611_cast_fp16 = mul(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("op_4611_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4613_cast_fp16 = add(x = var_4607_cast_fp16, y = var_4611_cast_fp16)[name = string("op_4613_cast_fp16")];
            fp32 var_4614_epsilon_0 = const()[name = string("op_4614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_4614_cast_fp16 = rsqrt(epsilon = var_4614_epsilon_0, x = var_4613_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_29_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4614_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4616_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4616_cast_fp16")];
            tensor<int32, [4]> var_4621_begin_0 = const()[name = string("op_4621_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4621_end_0 = const()[name = string("op_4621_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_4621_end_mask_0 = const()[name = string("op_4621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_4621_cast_fp16 = slice_by_index(begin = var_4621_begin_0, end = var_4621_end_0, end_mask = var_4621_end_mask_0, x = q_29_cast_fp16)[name = string("op_4621_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_4622_cast_fp16 = mul(x = var_4621_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<int32, [4]> var_4627_begin_0 = const()[name = string("op_4627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4627_end_0 = const()[name = string("op_4627_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_4627_end_mask_0 = const()[name = string("op_4627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_4627_cast_fp16 = slice_by_index(begin = var_4627_begin_0, end = var_4627_end_0, end_mask = var_4627_end_mask_0, x = q_29_cast_fp16)[name = string("op_4627_cast_fp16")];
            int32 var_4629 = const()[name = string("op_4629"), val = int32(-1)];
            bool var_4630_interleave_0 = const()[name = string("op_4630_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_4630_cast_fp16 = concat(axis = var_4629, interleave = var_4630_interleave_0, values = (var_4622_cast_fp16, var_4627_cast_fp16))[name = string("op_4630_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4631_cast_fp16 = mul(x = var_4630_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4631_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_31_cast_fp16 = add(x = var_4616_cast_fp16, y = var_4631_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4634_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4634_cast_fp16")];
            tensor<int32, [4]> var_4639_begin_0 = const()[name = string("op_4639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4639_end_0 = const()[name = string("op_4639_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_4639_end_mask_0 = const()[name = string("op_4639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_4639_cast_fp16 = slice_by_index(begin = var_4639_begin_0, end = var_4639_end_0, end_mask = var_4639_end_mask_0, x = k_43_cast_fp16)[name = string("op_4639_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4640_cast_fp16 = mul(x = var_4639_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4640_cast_fp16")];
            tensor<int32, [4]> var_4645_begin_0 = const()[name = string("op_4645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4645_end_0 = const()[name = string("op_4645_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_4645_end_mask_0 = const()[name = string("op_4645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_4645_cast_fp16 = slice_by_index(begin = var_4645_begin_0, end = var_4645_end_0, end_mask = var_4645_end_mask_0, x = k_43_cast_fp16)[name = string("op_4645_cast_fp16")];
            int32 var_4647 = const()[name = string("op_4647"), val = int32(-1)];
            bool var_4648_interleave_0 = const()[name = string("op_4648_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_4648_cast_fp16 = concat(axis = var_4647, interleave = var_4648_interleave_0, values = (var_4640_cast_fp16, var_4645_cast_fp16))[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4649_cast_fp16 = mul(x = var_4648_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4649_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_45_cast_fp16 = add(x = var_4634_cast_fp16, y = var_4649_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_14_reps_0 = const()[name = string("tile_14_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = k_45_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [4, 1, 64, 256]> tile_14_cast_fp16 = tile(reps = tile_14_reps_0, x = transpose_28_cast_fp16)[name = string("tile_14_cast_fp16")];
            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_28_cast_fp16 = reshape(shape = concat_28, x = tile_14_cast_fp16)[name = string("reshape_28_cast_fp16")];
            tensor<int32, [5]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_28_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [4, 1, 64, 256]> reshape_29_cast_fp16 = reshape(shape = concat_29, x = transpose_29_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_15_reps_0 = const()[name = string("tile_15_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = v_29_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [4, 1, 64, 256]> tile_15_cast_fp16 = tile(reps = tile_15_reps_0, x = transpose_30_cast_fp16)[name = string("tile_15_cast_fp16")];
            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_30_cast_fp16 = reshape(shape = concat_30, x = tile_15_cast_fp16)[name = string("reshape_30_cast_fp16")];
            tensor<int32, [5]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_30_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [4, 1, 64, 256]> reshape_31_cast_fp16 = reshape(shape = concat_31, x = transpose_31_cast_fp16)[name = string("reshape_31_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4663_transpose_x_0 = const()[name = string("op_4663_transpose_x_0"), val = bool(false)];
            bool var_4663_transpose_y_0 = const()[name = string("op_4663_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 4, 64, 64]> var_4663_cast_fp16 = matmul(transpose_x = var_4663_transpose_x_0, transpose_y = var_4663_transpose_y_0, x = q_31_cast_fp16, y = transpose_103_cast_fp16)[name = string("op_4663_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_4665_cast_fp16 = add(x = var_4663_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4666 = const()[name = string("op_4666"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_121_cast_fp16 = softmax(axis = var_4666, x = var_4665_cast_fp16)[name = string("w_121_cast_fp16")];
            bool var_4669_transpose_x_0 = const()[name = string("op_4669_transpose_x_0"), val = bool(false)];
            bool var_4669_transpose_y_0 = const()[name = string("op_4669_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = reshape_31_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 4, 64, 256]> var_4669_cast_fp16 = matmul(transpose_x = var_4669_transpose_x_0, transpose_y = var_4669_transpose_y_0, x = w_121_cast_fp16, y = v_31_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<int32, [4]> var_4672_perm_0 = const()[name = string("op_4672_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4676 = const()[name = string("op_4676"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_4672_cast_fp16 = transpose(perm = var_4672_perm_0, x = var_4669_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 64, 1024]> input_87_cast_fp16 = reshape(shape = var_4676, x = var_4672_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45325248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45849600))))[name = string("tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4681_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4684 = const()[name = string("op_4684"), val = bool(true)];
            tensor<int32, [1]> var_4685_axes_0 = const()[name = string("op_4685_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4685_cast_fp16 = reduce_max(axes = var_4685_axes_0, keep_dims = var_4684, x = var_4681_cast_fp16)[name = string("op_4685_cast_fp16")];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_123_cast_fp16 = maximum(x = var_4685_cast_fp16, y = var_4686_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_123_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_123_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4689_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [1]> var_4694_axes_0 = const()[name = string("op_4694_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4694_keep_dims_0 = const()[name = string("op_4694_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4694_cast_fp16 = reduce_mean(axes = var_4694_axes_0, keep_dims = var_4694_keep_dims_0, x = var_4689_cast_fp16)[name = string("op_4694_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4695_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4695_cast_fp16")];
            fp32 var_4696_epsilon_0 = const()[name = string("op_4696_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4696_cast_fp16 = inverse(epsilon = var_4696_epsilon_0, x = var_4695_cast_fp16)[name = string("op_4696_cast_fp16")];
            fp16 var_4697_to_fp16 = const()[name = string("op_4697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4698_cast_fp16 = mul(x = var_4696_cast_fp16, y = var_4697_to_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4700_cast_fp16 = add(x = var_4694_cast_fp16, y = var_4698_cast_fp16)[name = string("op_4700_cast_fp16")];
            fp32 var_4701_epsilon_0 = const()[name = string("op_4701_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4701_cast_fp16 = rsqrt(epsilon = var_4701_epsilon_0, x = var_4700_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_107_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4701_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45850688)))];
            tensor<fp16, [1, 64, 512]> var_4703_cast_fp16 = mul(x = out_107_cast_fp16, y = tm_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4703_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_137_cast_fp16 = add(x = x_127_cast_fp16, y = var_4703_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4706_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_4706_cast_fp16")];
            bool var_4709 = const()[name = string("op_4709"), val = bool(true)];
            tensor<int32, [1]> var_4710_axes_0 = const()[name = string("op_4710_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4710_cast_fp16 = reduce_max(axes = var_4710_axes_0, keep_dims = var_4709, x = var_4706_cast_fp16)[name = string("op_4710_cast_fp16")];
            fp16 var_4711_to_fp16 = const()[name = string("op_4711_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_125_cast_fp16 = maximum(x = var_4710_cast_fp16, y = var_4711_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_125_cast_fp16 = real_div(x = x_137_cast_fp16, y = s_125_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4714_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4714_cast_fp16")];
            tensor<int32, [1]> var_4719_axes_0 = const()[name = string("op_4719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4719_keep_dims_0 = const()[name = string("op_4719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4719_cast_fp16 = reduce_mean(axes = var_4719_axes_0, keep_dims = var_4719_keep_dims_0, x = var_4714_cast_fp16)[name = string("op_4719_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4720_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4720_cast_fp16")];
            fp32 var_4721_epsilon_0 = const()[name = string("op_4721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4721_cast_fp16 = inverse(epsilon = var_4721_epsilon_0, x = var_4720_cast_fp16)[name = string("op_4721_cast_fp16")];
            fp16 var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4723_cast_fp16 = mul(x = var_4721_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4725_cast_fp16 = add(x = var_4719_cast_fp16, y = var_4723_cast_fp16)[name = string("op_4725_cast_fp16")];
            fp32 var_4726_epsilon_0 = const()[name = string("op_4726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4726_cast_fp16 = rsqrt(epsilon = var_4726_epsilon_0, x = var_4725_cast_fp16)[name = string("op_4726_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_109_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4726_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45851776)))];
            tensor<fp16, [1, 64, 512]> input_89_cast_fp16 = mul(x = out_109_cast_fp16, y = tm_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45852864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46901504))))[name = string("tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4733_mode_0 = const()[name = string("op_4733_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_4733_cast_fp16 = gelu(mode = var_4733_mode_0, x = linear_68_cast_fp16)[name = string("op_4733_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46905664))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47954304))))[name = string("tm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_91_cast_fp16 = mul(x = var_4733_cast_fp16, y = linear_69_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47958464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49007104))))[name = string("tm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4741_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4741_cast_fp16")];
            bool var_4744 = const()[name = string("op_4744"), val = bool(true)];
            tensor<int32, [1]> var_4745_axes_0 = const()[name = string("op_4745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4745_cast_fp16 = reduce_max(axes = var_4745_axes_0, keep_dims = var_4744, x = var_4741_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_127_cast_fp16 = maximum(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_127_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_127_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4749_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<int32, [1]> var_4754_axes_0 = const()[name = string("op_4754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4754_keep_dims_0 = const()[name = string("op_4754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4754_cast_fp16 = reduce_mean(axes = var_4754_axes_0, keep_dims = var_4754_keep_dims_0, x = var_4749_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4755_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4755_cast_fp16")];
            fp32 var_4756_epsilon_0 = const()[name = string("op_4756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4756_cast_fp16 = inverse(epsilon = var_4756_epsilon_0, x = var_4755_cast_fp16)[name = string("op_4756_cast_fp16")];
            fp16 var_4757_to_fp16 = const()[name = string("op_4757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4758_cast_fp16 = mul(x = var_4756_cast_fp16, y = var_4757_to_fp16)[name = string("op_4758_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4760_cast_fp16 = add(x = var_4754_cast_fp16, y = var_4758_cast_fp16)[name = string("op_4760_cast_fp16")];
            fp32 var_4761_epsilon_0 = const()[name = string("op_4761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4761_cast_fp16 = rsqrt(epsilon = var_4761_epsilon_0, x = var_4760_cast_fp16)[name = string("op_4761_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_111_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4761_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49008192)))];
            tensor<fp16, [1, 64, 512]> var_4763_cast_fp16 = mul(x = out_111_cast_fp16, y = tm_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4763_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_93_cast_fp16 = add(x = x_137_cast_fp16, y = var_4763_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49009280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49271488))))[name = string("tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_71_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_71_cast_fp16")];
            string x_141_mode_0 = const()[name = string("x_141_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_141_cast_fp16 = gelu(mode = x_141_mode_0, x = linear_71_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_4783_begin_0 = const()[name = string("op_4783_begin_0"), val = tensor<int32, [4]>([0, 0, 7, 0])];
            tensor<int32, [4]> var_4783_end_0 = const()[name = string("op_4783_end_0"), val = tensor<int32, [4]>([1, 64, 8, 512])];
            tensor<bool, [4]> var_4783_end_mask_0 = const()[name = string("op_4783_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4783_squeeze_mask_0 = const()[name = string("op_4783_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_4783_cast_fp16 = slice_by_index(begin = var_4783_begin_0, end = var_4783_end_0, end_mask = var_4783_end_mask_0, squeeze_mask = var_4783_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_95_cast_fp16 = mul(x = x_141_cast_fp16, y = var_4783_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49272576))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49534784))))[name = string("tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4793_cast_fp16 = abs(x = linear_72_cast_fp16)[name = string("op_4793_cast_fp16")];
            bool var_4796 = const()[name = string("op_4796"), val = bool(true)];
            tensor<int32, [1]> var_4797_axes_0 = const()[name = string("op_4797_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4797_cast_fp16 = reduce_max(axes = var_4797_axes_0, keep_dims = var_4796, x = var_4793_cast_fp16)[name = string("op_4797_cast_fp16")];
            fp16 var_4798_to_fp16 = const()[name = string("op_4798_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_129_cast_fp16 = maximum(x = var_4797_cast_fp16, y = var_4798_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_129_cast_fp16 = real_div(x = linear_72_cast_fp16, y = s_129_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4801_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4801_cast_fp16")];
            tensor<int32, [1]> var_4806_axes_0 = const()[name = string("op_4806_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4806_keep_dims_0 = const()[name = string("op_4806_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4806_cast_fp16 = reduce_mean(axes = var_4806_axes_0, keep_dims = var_4806_keep_dims_0, x = var_4801_cast_fp16)[name = string("op_4806_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4807_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4807_cast_fp16")];
            fp32 var_4808_epsilon_0 = const()[name = string("op_4808_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4808_cast_fp16 = inverse(epsilon = var_4808_epsilon_0, x = var_4807_cast_fp16)[name = string("op_4808_cast_fp16")];
            fp16 var_4809_to_fp16 = const()[name = string("op_4809_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4810_cast_fp16 = mul(x = var_4808_cast_fp16, y = var_4809_to_fp16)[name = string("op_4810_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4812_cast_fp16 = add(x = var_4806_cast_fp16, y = var_4810_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp32 var_4813_epsilon_0 = const()[name = string("op_4813_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4813_cast_fp16 = rsqrt(epsilon = var_4813_epsilon_0, x = var_4812_cast_fp16)[name = string("op_4813_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_113_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4813_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49535872)))];
            tensor<fp16, [1, 64, 512]> var_4815_cast_fp16 = mul(x = out_113_cast_fp16, y = tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4815_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_15_cast_fp16 = add(x = input_93_cast_fp16, y = var_4815_cast_fp16)[name = string("h_15_cast_fp16")];
            tensor<fp16, [1]> tm_layers_7_layer_scalar_to_fp16 = const()[name = string("tm_layers_7_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-1])];
            tensor<fp16, [1, 64, 512]> x_145_cast_fp16 = mul(x = h_15_cast_fp16, y = tm_layers_7_layer_scalar_to_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4819_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4819_cast_fp16")];
            bool var_4822 = const()[name = string("op_4822"), val = bool(true)];
            tensor<int32, [1]> var_4823_axes_0 = const()[name = string("op_4823_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_4823_cast_fp16 = reduce_max(axes = var_4823_axes_0, keep_dims = var_4822, x = var_4819_cast_fp16)[name = string("op_4823_cast_fp16")];
            fp16 var_4824_to_fp16 = const()[name = string("op_4824_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_131_cast_fp16 = maximum(x = var_4823_cast_fp16, y = var_4824_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_131_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_131_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_4827_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [1]> var_4832_axes_0 = const()[name = string("op_4832_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4832_keep_dims_0 = const()[name = string("op_4832_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_4832_cast_fp16 = reduce_mean(axes = var_4832_axes_0, keep_dims = var_4832_keep_dims_0, x = var_4827_cast_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4833_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4833_cast_fp16")];
            fp32 var_4834_epsilon_0 = const()[name = string("op_4834_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_4834_cast_fp16 = inverse(epsilon = var_4834_epsilon_0, x = var_4833_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp16 var_4835_to_fp16 = const()[name = string("op_4835_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_4836_cast_fp16 = mul(x = var_4834_cast_fp16, y = var_4835_to_fp16)[name = string("op_4836_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_4838_cast_fp16 = add(x = var_4832_cast_fp16, y = var_4836_cast_fp16)[name = string("op_4838_cast_fp16")];
            fp32 var_4839_epsilon_0 = const()[name = string("op_4839_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_4839_cast_fp16 = rsqrt(epsilon = var_4839_epsilon_0, x = var_4838_cast_fp16)[name = string("op_4839_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_115_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4839_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49536960)))];
            tensor<fp16, [1, 64, 512]> input_97_cast_fp16 = mul(x = out_115_cast_fp16, y = tm_layers_8_input_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49538048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50062400))))[name = string("tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4849 = const()[name = string("op_4849"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_4850_cast_fp16 = reshape(shape = var_4849, x = linear_73_cast_fp16)[name = string("op_4850_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50064512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50326720))))[name = string("tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [4]> var_4861 = const()[name = string("op_4861"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_4862_cast_fp16 = reshape(shape = var_4861, x = linear_74_cast_fp16)[name = string("op_4862_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50327808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50590016))))[name = string("tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [4]> var_4873 = const()[name = string("op_4873"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_4874_cast_fp16 = reshape(shape = var_4873, x = linear_75_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<int32, [4]> x_151_perm_0 = const()[name = string("x_151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_4850_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 4, 64, 256]> var_4878_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4878_cast_fp16")];
            bool var_4881 = const()[name = string("op_4881"), val = bool(true)];
            tensor<int32, [1]> var_4882_axes_0 = const()[name = string("op_4882_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_4882_cast_fp16 = reduce_max(axes = var_4882_axes_0, keep_dims = var_4881, x = var_4878_cast_fp16)[name = string("op_4882_cast_fp16")];
            fp16 var_4883_to_fp16 = const()[name = string("op_4883_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_133_cast_fp16 = maximum(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_133_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_133_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4886_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4886_cast_fp16")];
            tensor<int32, [1]> var_4891_axes_0 = const()[name = string("op_4891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4891_keep_dims_0 = const()[name = string("op_4891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_4891_cast_fp16 = reduce_mean(axes = var_4891_axes_0, keep_dims = var_4891_keep_dims_0, x = var_4886_cast_fp16)[name = string("op_4891_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_4892_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4892_cast_fp16")];
            fp32 var_4893_epsilon_0 = const()[name = string("op_4893_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_4893_cast_fp16 = inverse(epsilon = var_4893_epsilon_0, x = var_4892_cast_fp16)[name = string("op_4893_cast_fp16")];
            fp16 var_4894_to_fp16 = const()[name = string("op_4894_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_4895_cast_fp16 = mul(x = var_4893_cast_fp16, y = var_4894_to_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_4897_cast_fp16 = add(x = var_4891_cast_fp16, y = var_4895_cast_fp16)[name = string("op_4897_cast_fp16")];
            fp32 var_4898_epsilon_0 = const()[name = string("op_4898_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_4898_cast_fp16 = rsqrt(epsilon = var_4898_epsilon_0, x = var_4897_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_117_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4898_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591104)))];
            tensor<fp16, [1, 4, 64, 256]> q_33_cast_fp16 = mul(x = out_117_cast_fp16, y = tm_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_4862_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 2, 64, 256]> var_4901_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4901_cast_fp16")];
            bool var_4904 = const()[name = string("op_4904"), val = bool(true)];
            tensor<int32, [1]> var_4905_axes_0 = const()[name = string("op_4905_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_4905_cast_fp16 = reduce_max(axes = var_4905_axes_0, keep_dims = var_4904, x = var_4901_cast_fp16)[name = string("op_4905_cast_fp16")];
            fp16 var_4906_to_fp16 = const()[name = string("op_4906_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_135_cast_fp16 = maximum(x = var_4905_cast_fp16, y = var_4906_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_135_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_135_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4909_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [1]> var_4914_axes_0 = const()[name = string("op_4914_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4914_keep_dims_0 = const()[name = string("op_4914_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_4914_cast_fp16 = reduce_mean(axes = var_4914_axes_0, keep_dims = var_4914_keep_dims_0, x = var_4909_cast_fp16)[name = string("op_4914_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4915_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4915_cast_fp16")];
            fp32 var_4916_epsilon_0 = const()[name = string("op_4916_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_4916_cast_fp16 = inverse(epsilon = var_4916_epsilon_0, x = var_4915_cast_fp16)[name = string("op_4916_cast_fp16")];
            fp16 var_4917_to_fp16 = const()[name = string("op_4917_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_4918_cast_fp16 = mul(x = var_4916_cast_fp16, y = var_4917_to_fp16)[name = string("op_4918_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4920_cast_fp16 = add(x = var_4914_cast_fp16, y = var_4918_cast_fp16)[name = string("op_4920_cast_fp16")];
            fp32 var_4921_epsilon_0 = const()[name = string("op_4921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_4921_cast_fp16 = rsqrt(epsilon = var_4921_epsilon_0, x = var_4920_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_119_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4921_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591680)))];
            tensor<fp16, [1, 2, 64, 256]> k_49_cast_fp16 = mul(x = out_119_cast_fp16, y = tm_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_151_cast_fp16 = transpose(perm = x_151_perm_0, x = var_4874_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 2, 64, 256]> var_4924_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4924_cast_fp16")];
            bool var_4927 = const()[name = string("op_4927"), val = bool(true)];
            tensor<int32, [1]> var_4928_axes_0 = const()[name = string("op_4928_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_4928_cast_fp16 = reduce_max(axes = var_4928_axes_0, keep_dims = var_4927, x = var_4924_cast_fp16)[name = string("op_4928_cast_fp16")];
            fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_137_cast_fp16 = maximum(x = var_4928_cast_fp16, y = var_4929_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_137_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_137_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4932_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4932_cast_fp16")];
            tensor<int32, [1]> var_4937_axes_0 = const()[name = string("op_4937_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4937_keep_dims_0 = const()[name = string("op_4937_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_4937_cast_fp16 = reduce_mean(axes = var_4937_axes_0, keep_dims = var_4937_keep_dims_0, x = var_4932_cast_fp16)[name = string("op_4937_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4938_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4938_cast_fp16")];
            fp32 var_4939_epsilon_0 = const()[name = string("op_4939_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_4939_cast_fp16 = inverse(epsilon = var_4939_epsilon_0, x = var_4938_cast_fp16)[name = string("op_4939_cast_fp16")];
            fp16 var_4940_to_fp16 = const()[name = string("op_4940_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_4941_cast_fp16 = mul(x = var_4939_cast_fp16, y = var_4940_to_fp16)[name = string("op_4941_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_4943_cast_fp16 = add(x = var_4937_cast_fp16, y = var_4941_cast_fp16)[name = string("op_4943_cast_fp16")];
            fp32 var_4944_epsilon_0 = const()[name = string("op_4944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_4944_cast_fp16 = rsqrt(epsilon = var_4944_epsilon_0, x = var_4943_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_33_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4944_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4946_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = q_33_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_4952_cast_fp16 = mul(x = var_4951_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4957_begin_0 = const()[name = string("op_4957_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4957_end_0 = const()[name = string("op_4957_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_4957_end_mask_0 = const()[name = string("op_4957_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_4957_cast_fp16 = slice_by_index(begin = var_4957_begin_0, end = var_4957_end_0, end_mask = var_4957_end_mask_0, x = q_33_cast_fp16)[name = string("op_4957_cast_fp16")];
            int32 var_4959 = const()[name = string("op_4959"), val = int32(-1)];
            bool var_4960_interleave_0 = const()[name = string("op_4960_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_4960_cast_fp16 = concat(axis = var_4959, interleave = var_4960_interleave_0, values = (var_4952_cast_fp16, var_4957_cast_fp16))[name = string("op_4960_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_4961_cast_fp16 = mul(x = var_4960_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4961_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_35_cast_fp16 = add(x = var_4946_cast_fp16, y = var_4961_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4964_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_4964_cast_fp16")];
            tensor<int32, [4]> var_4969_begin_0 = const()[name = string("op_4969_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4969_end_0 = const()[name = string("op_4969_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_4969_end_mask_0 = const()[name = string("op_4969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_4969_cast_fp16 = slice_by_index(begin = var_4969_begin_0, end = var_4969_end_0, end_mask = var_4969_end_mask_0, x = k_49_cast_fp16)[name = string("op_4969_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_4970_cast_fp16 = mul(x = var_4969_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_4970_cast_fp16")];
            tensor<int32, [4]> var_4975_begin_0 = const()[name = string("op_4975_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4975_end_0 = const()[name = string("op_4975_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_4975_end_mask_0 = const()[name = string("op_4975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_4975_cast_fp16 = slice_by_index(begin = var_4975_begin_0, end = var_4975_end_0, end_mask = var_4975_end_mask_0, x = k_49_cast_fp16)[name = string("op_4975_cast_fp16")];
            int32 var_4977 = const()[name = string("op_4977"), val = int32(-1)];
            bool var_4978_interleave_0 = const()[name = string("op_4978_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_4978_cast_fp16 = concat(axis = var_4977, interleave = var_4978_interleave_0, values = (var_4970_cast_fp16, var_4975_cast_fp16))[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_4979_cast_fp16 = mul(x = var_4978_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_4979_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_51_cast_fp16 = add(x = var_4964_cast_fp16, y = var_4979_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_16_reps_0 = const()[name = string("tile_16_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = k_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [4, 1, 64, 256]> tile_16_cast_fp16 = tile(reps = tile_16_reps_0, x = transpose_32_cast_fp16)[name = string("tile_16_cast_fp16")];
            tensor<int32, [5]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_32_cast_fp16 = reshape(shape = concat_32, x = tile_16_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [5]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [4, 1, 64, 256]> reshape_33_cast_fp16 = reshape(shape = concat_33, x = transpose_33_cast_fp16)[name = string("reshape_33_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_17_reps_0 = const()[name = string("tile_17_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = v_33_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [4, 1, 64, 256]> tile_17_cast_fp16 = tile(reps = tile_17_reps_0, x = transpose_34_cast_fp16)[name = string("tile_17_cast_fp16")];
            tensor<int32, [5]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_34_cast_fp16 = reshape(shape = concat_34, x = tile_17_cast_fp16)[name = string("reshape_34_cast_fp16")];
            tensor<int32, [5]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = reshape_34_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [4, 1, 64, 256]> reshape_35_cast_fp16 = reshape(shape = concat_35, x = transpose_35_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = reshape_33_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 4, 64, 64]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = q_35_cast_fp16, y = transpose_104_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_4995_cast_fp16 = add(x = var_4993_cast_fp16, y = var_2173_cast_fp16)[name = string("op_4995_cast_fp16")];
            int32 var_4996 = const()[name = string("op_4996"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_137_cast_fp16 = softmax(axis = var_4996, x = var_4995_cast_fp16)[name = string("w_137_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 4, 64, 256]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = w_137_cast_fp16, y = v_35_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5002_perm_0 = const()[name = string("op_5002_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5006 = const()[name = string("op_5006"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_5002_cast_fp16 = transpose(perm = var_5002_perm_0, x = var_4999_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 64, 1024]> input_99_cast_fp16 = reshape(shape = var_5006, x = var_5002_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50592256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51116608))))[name = string("tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5011_cast_fp16 = abs(x = linear_76_cast_fp16)[name = string("op_5011_cast_fp16")];
            bool var_5014 = const()[name = string("op_5014"), val = bool(true)];
            tensor<int32, [1]> var_5015_axes_0 = const()[name = string("op_5015_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5015_cast_fp16 = reduce_max(axes = var_5015_axes_0, keep_dims = var_5014, x = var_5011_cast_fp16)[name = string("op_5015_cast_fp16")];
            fp16 var_5016_to_fp16 = const()[name = string("op_5016_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_139_cast_fp16 = maximum(x = var_5015_cast_fp16, y = var_5016_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_139_cast_fp16 = real_div(x = linear_76_cast_fp16, y = s_139_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5019_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_5019_cast_fp16")];
            tensor<int32, [1]> var_5024_axes_0 = const()[name = string("op_5024_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5024_keep_dims_0 = const()[name = string("op_5024_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5024_cast_fp16 = reduce_mean(axes = var_5024_axes_0, keep_dims = var_5024_keep_dims_0, x = var_5019_cast_fp16)[name = string("op_5024_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5025_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_5025_cast_fp16")];
            fp32 var_5026_epsilon_0 = const()[name = string("op_5026_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5026_cast_fp16 = inverse(epsilon = var_5026_epsilon_0, x = var_5025_cast_fp16)[name = string("op_5026_cast_fp16")];
            fp16 var_5027_to_fp16 = const()[name = string("op_5027_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5028_cast_fp16 = mul(x = var_5026_cast_fp16, y = var_5027_to_fp16)[name = string("op_5028_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5030_cast_fp16 = add(x = var_5024_cast_fp16, y = var_5028_cast_fp16)[name = string("op_5030_cast_fp16")];
            fp32 var_5031_epsilon_0 = const()[name = string("op_5031_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5031_cast_fp16 = rsqrt(epsilon = var_5031_epsilon_0, x = var_5030_cast_fp16)[name = string("op_5031_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_121_cast_fp16 = mul(x = y_139_cast_fp16, y = var_5031_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51117696)))];
            tensor<fp16, [1, 64, 512]> var_5033_cast_fp16 = mul(x = out_121_cast_fp16, y = tm_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_5033_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = var_5033_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5036_cast_fp16 = abs(x = x_155_cast_fp16)[name = string("op_5036_cast_fp16")];
            bool var_5039 = const()[name = string("op_5039"), val = bool(true)];
            tensor<int32, [1]> var_5040_axes_0 = const()[name = string("op_5040_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5040_cast_fp16 = reduce_max(axes = var_5040_axes_0, keep_dims = var_5039, x = var_5036_cast_fp16)[name = string("op_5040_cast_fp16")];
            fp16 var_5041_to_fp16 = const()[name = string("op_5041_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_141_cast_fp16 = maximum(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_141_cast_fp16 = real_div(x = x_155_cast_fp16, y = s_141_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5044_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_5044_cast_fp16")];
            tensor<int32, [1]> var_5049_axes_0 = const()[name = string("op_5049_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5049_keep_dims_0 = const()[name = string("op_5049_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5049_cast_fp16 = reduce_mean(axes = var_5049_axes_0, keep_dims = var_5049_keep_dims_0, x = var_5044_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5050_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_5050_cast_fp16")];
            fp32 var_5051_epsilon_0 = const()[name = string("op_5051_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5051_cast_fp16 = inverse(epsilon = var_5051_epsilon_0, x = var_5050_cast_fp16)[name = string("op_5051_cast_fp16")];
            fp16 var_5052_to_fp16 = const()[name = string("op_5052_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5053_cast_fp16 = mul(x = var_5051_cast_fp16, y = var_5052_to_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5055_cast_fp16 = add(x = var_5049_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            fp32 var_5056_epsilon_0 = const()[name = string("op_5056_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5056_cast_fp16 = rsqrt(epsilon = var_5056_epsilon_0, x = var_5055_cast_fp16)[name = string("op_5056_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_123_cast_fp16 = mul(x = y_141_cast_fp16, y = var_5056_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51118784)))];
            tensor<fp16, [1, 64, 512]> input_101_cast_fp16 = mul(x = out_123_cast_fp16, y = tm_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51119872))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52168512))))[name = string("tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_77_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_77_cast_fp16")];
            string var_5063_mode_0 = const()[name = string("op_5063_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_5063_cast_fp16 = gelu(mode = var_5063_mode_0, x = linear_77_cast_fp16)[name = string("op_5063_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52172672))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53221312))))[name = string("tm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_78_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_103_cast_fp16 = mul(x = var_5063_cast_fp16, y = linear_78_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53225472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54274112))))[name = string("tm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5071_cast_fp16 = abs(x = linear_79_cast_fp16)[name = string("op_5071_cast_fp16")];
            bool var_5074 = const()[name = string("op_5074"), val = bool(true)];
            tensor<int32, [1]> var_5075_axes_0 = const()[name = string("op_5075_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5075_cast_fp16 = reduce_max(axes = var_5075_axes_0, keep_dims = var_5074, x = var_5071_cast_fp16)[name = string("op_5075_cast_fp16")];
            fp16 var_5076_to_fp16 = const()[name = string("op_5076_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_143_cast_fp16 = maximum(x = var_5075_cast_fp16, y = var_5076_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_143_cast_fp16 = real_div(x = linear_79_cast_fp16, y = s_143_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5079_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_5079_cast_fp16")];
            tensor<int32, [1]> var_5084_axes_0 = const()[name = string("op_5084_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5084_keep_dims_0 = const()[name = string("op_5084_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5084_cast_fp16 = reduce_mean(axes = var_5084_axes_0, keep_dims = var_5084_keep_dims_0, x = var_5079_cast_fp16)[name = string("op_5084_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5085_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5086_cast_fp16 = inverse(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            fp16 var_5087_to_fp16 = const()[name = string("op_5087_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5088_cast_fp16 = mul(x = var_5086_cast_fp16, y = var_5087_to_fp16)[name = string("op_5088_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5090_cast_fp16 = add(x = var_5084_cast_fp16, y = var_5088_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp32 var_5091_epsilon_0 = const()[name = string("op_5091_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5091_cast_fp16 = rsqrt(epsilon = var_5091_epsilon_0, x = var_5090_cast_fp16)[name = string("op_5091_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_125_cast_fp16 = mul(x = y_143_cast_fp16, y = var_5091_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54275200)))];
            tensor<fp16, [1, 64, 512]> var_5093_cast_fp16 = mul(x = out_125_cast_fp16, y = tm_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5093_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_105_cast_fp16 = add(x = x_155_cast_fp16, y = var_5093_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54276288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54538496))))[name = string("tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_80_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_80_cast_fp16")];
            string x_159_mode_0 = const()[name = string("x_159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_159_cast_fp16 = gelu(mode = x_159_mode_0, x = linear_80_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [4]> var_5113_begin_0 = const()[name = string("op_5113_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_5113_end_0 = const()[name = string("op_5113_end_0"), val = tensor<int32, [4]>([1, 64, 9, 512])];
            tensor<bool, [4]> var_5113_end_mask_0 = const()[name = string("op_5113_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5113_squeeze_mask_0 = const()[name = string("op_5113_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_5113_cast_fp16 = slice_by_index(begin = var_5113_begin_0, end = var_5113_end_0, end_mask = var_5113_end_mask_0, squeeze_mask = var_5113_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_107_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5113_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54539584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54801792))))[name = string("tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5123_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5123_cast_fp16")];
            bool var_5126 = const()[name = string("op_5126"), val = bool(true)];
            tensor<int32, [1]> var_5127_axes_0 = const()[name = string("op_5127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5127_cast_fp16 = reduce_max(axes = var_5127_axes_0, keep_dims = var_5126, x = var_5123_cast_fp16)[name = string("op_5127_cast_fp16")];
            fp16 var_5128_to_fp16 = const()[name = string("op_5128_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_145_cast_fp16 = maximum(x = var_5127_cast_fp16, y = var_5128_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_145_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_145_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5131_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5136_axes_0 = const()[name = string("op_5136_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5136_keep_dims_0 = const()[name = string("op_5136_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5136_cast_fp16 = reduce_mean(axes = var_5136_axes_0, keep_dims = var_5136_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5136_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5137_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5138_cast_fp16 = inverse(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp16 var_5139_to_fp16 = const()[name = string("op_5139_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5140_cast_fp16 = mul(x = var_5138_cast_fp16, y = var_5139_to_fp16)[name = string("op_5140_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5142_cast_fp16 = add(x = var_5136_cast_fp16, y = var_5140_cast_fp16)[name = string("op_5142_cast_fp16")];
            fp32 var_5143_epsilon_0 = const()[name = string("op_5143_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5143_cast_fp16 = rsqrt(epsilon = var_5143_epsilon_0, x = var_5142_cast_fp16)[name = string("op_5143_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_127_cast_fp16 = mul(x = y_145_cast_fp16, y = var_5143_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54802880)))];
            tensor<fp16, [1, 64, 512]> var_5145_cast_fp16 = mul(x = out_127_cast_fp16, y = tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5145_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_17_cast_fp16 = add(x = input_105_cast_fp16, y = var_5145_cast_fp16)[name = string("h_17_cast_fp16")];
            tensor<fp16, [1]> tm_layers_8_layer_scalar_to_fp16 = const()[name = string("tm_layers_8_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bcp-1])];
            tensor<fp16, [1, 64, 512]> x_163_cast_fp16 = mul(x = h_17_cast_fp16, y = tm_layers_8_layer_scalar_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5149_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_5149_cast_fp16")];
            bool var_5152 = const()[name = string("op_5152"), val = bool(true)];
            tensor<int32, [1]> var_5153_axes_0 = const()[name = string("op_5153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5153_cast_fp16 = reduce_max(axes = var_5153_axes_0, keep_dims = var_5152, x = var_5149_cast_fp16)[name = string("op_5153_cast_fp16")];
            fp16 var_5154_to_fp16 = const()[name = string("op_5154_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_147_cast_fp16 = maximum(x = var_5153_cast_fp16, y = var_5154_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_147_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_147_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5157_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_5157_cast_fp16")];
            tensor<int32, [1]> var_5162_axes_0 = const()[name = string("op_5162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5162_keep_dims_0 = const()[name = string("op_5162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5162_cast_fp16 = reduce_mean(axes = var_5162_axes_0, keep_dims = var_5162_keep_dims_0, x = var_5157_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5163_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_5163_cast_fp16")];
            fp32 var_5164_epsilon_0 = const()[name = string("op_5164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5164_cast_fp16 = inverse(epsilon = var_5164_epsilon_0, x = var_5163_cast_fp16)[name = string("op_5164_cast_fp16")];
            fp16 var_5165_to_fp16 = const()[name = string("op_5165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5166_cast_fp16 = mul(x = var_5164_cast_fp16, y = var_5165_to_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5168_cast_fp16 = add(x = var_5162_cast_fp16, y = var_5166_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp32 var_5169_epsilon_0 = const()[name = string("op_5169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5169_cast_fp16 = rsqrt(epsilon = var_5169_epsilon_0, x = var_5168_cast_fp16)[name = string("op_5169_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_129_cast_fp16 = mul(x = y_147_cast_fp16, y = var_5169_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54803968)))];
            tensor<fp16, [1, 64, 512]> input_109_cast_fp16 = mul(x = out_129_cast_fp16, y = tm_layers_9_input_layernorm_weight_to_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54805056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55329408))))[name = string("tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [4]> var_5179 = const()[name = string("op_5179"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_5180_cast_fp16 = reshape(shape = var_5179, x = linear_82_cast_fp16)[name = string("op_5180_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55331520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55593728))))[name = string("tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [4]> var_5191 = const()[name = string("op_5191"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_5192_cast_fp16 = reshape(shape = var_5191, x = linear_83_cast_fp16)[name = string("op_5192_cast_fp16")];
            tensor<int32, [4]> x_167_perm_0 = const()[name = string("x_167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55594816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55857024))))[name = string("tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5203 = const()[name = string("op_5203"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_5204_cast_fp16 = reshape(shape = var_5203, x = linear_84_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_5180_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 4, 64, 256]> var_5208_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_5208_cast_fp16")];
            bool var_5211 = const()[name = string("op_5211"), val = bool(true)];
            tensor<int32, [1]> var_5212_axes_0 = const()[name = string("op_5212_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_5212_cast_fp16 = reduce_max(axes = var_5212_axes_0, keep_dims = var_5211, x = var_5208_cast_fp16)[name = string("op_5212_cast_fp16")];
            fp16 var_5213_to_fp16 = const()[name = string("op_5213_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_149_cast_fp16 = maximum(x = var_5212_cast_fp16, y = var_5213_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_149_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_149_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_5216_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<int32, [1]> var_5221_axes_0 = const()[name = string("op_5221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5221_keep_dims_0 = const()[name = string("op_5221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_5221_cast_fp16 = reduce_mean(axes = var_5221_axes_0, keep_dims = var_5221_keep_dims_0, x = var_5216_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_5222_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_5223_cast_fp16 = inverse(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            fp16 var_5224_to_fp16 = const()[name = string("op_5224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_5225_cast_fp16 = mul(x = var_5223_cast_fp16, y = var_5224_to_fp16)[name = string("op_5225_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_5227_cast_fp16 = add(x = var_5221_cast_fp16, y = var_5225_cast_fp16)[name = string("op_5227_cast_fp16")];
            fp32 var_5228_epsilon_0 = const()[name = string("op_5228_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_5228_cast_fp16 = rsqrt(epsilon = var_5228_epsilon_0, x = var_5227_cast_fp16)[name = string("op_5228_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_131_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5228_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = var_5192_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 2, 64, 256]> var_5231_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_5231_cast_fp16")];
            bool var_5234 = const()[name = string("op_5234"), val = bool(true)];
            tensor<int32, [1]> var_5235_axes_0 = const()[name = string("op_5235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_5235_cast_fp16 = reduce_max(axes = var_5235_axes_0, keep_dims = var_5234, x = var_5231_cast_fp16)[name = string("op_5235_cast_fp16")];
            fp16 var_5236_to_fp16 = const()[name = string("op_5236_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_151_cast_fp16 = maximum(x = var_5235_cast_fp16, y = var_5236_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_151_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_151_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_5239_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5239_cast_fp16")];
            tensor<int32, [1]> var_5244_axes_0 = const()[name = string("op_5244_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5244_keep_dims_0 = const()[name = string("op_5244_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_5244_cast_fp16 = reduce_mean(axes = var_5244_axes_0, keep_dims = var_5244_keep_dims_0, x = var_5239_cast_fp16)[name = string("op_5244_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_5245_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_5245_cast_fp16")];
            fp32 var_5246_epsilon_0 = const()[name = string("op_5246_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_5246_cast_fp16 = inverse(epsilon = var_5246_epsilon_0, x = var_5245_cast_fp16)[name = string("op_5246_cast_fp16")];
            fp16 var_5247_to_fp16 = const()[name = string("op_5247_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_5248_cast_fp16 = mul(x = var_5246_cast_fp16, y = var_5247_to_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_5250_cast_fp16 = add(x = var_5244_cast_fp16, y = var_5248_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp32 var_5251_epsilon_0 = const()[name = string("op_5251_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_5251_cast_fp16 = rsqrt(epsilon = var_5251_epsilon_0, x = var_5250_cast_fp16)[name = string("op_5251_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_133_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5251_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [256]> tm_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858112)))];
            tensor<fp16, [1, 2, 64, 256]> k_55_cast_fp16 = mul(x = out_133_cast_fp16, y = tm_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_5204_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2, 64, 256]> var_5254_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_5254_cast_fp16")];
            bool var_5257 = const()[name = string("op_5257"), val = bool(true)];
            tensor<int32, [1]> var_5258_axes_0 = const()[name = string("op_5258_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_5258_cast_fp16 = reduce_max(axes = var_5258_axes_0, keep_dims = var_5257, x = var_5254_cast_fp16)[name = string("op_5258_cast_fp16")];
            fp16 var_5259_to_fp16 = const()[name = string("op_5259_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_153_cast_fp16 = maximum(x = var_5258_cast_fp16, y = var_5259_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_153_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_153_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_5262_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5262_cast_fp16")];
            tensor<int32, [1]> var_5267_axes_0 = const()[name = string("op_5267_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5267_keep_dims_0 = const()[name = string("op_5267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_5267_cast_fp16 = reduce_mean(axes = var_5267_axes_0, keep_dims = var_5267_keep_dims_0, x = var_5262_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_5268_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5268_cast_fp16")];
            fp32 var_5269_epsilon_0 = const()[name = string("op_5269_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_5269_cast_fp16 = inverse(epsilon = var_5269_epsilon_0, x = var_5268_cast_fp16)[name = string("op_5269_cast_fp16")];
            fp16 var_5270_to_fp16 = const()[name = string("op_5270_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_5271_cast_fp16 = mul(x = var_5269_cast_fp16, y = var_5270_to_fp16)[name = string("op_5271_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_5273_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5271_cast_fp16)[name = string("op_5273_cast_fp16")];
            fp32 var_5274_epsilon_0 = const()[name = string("op_5274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_5274_cast_fp16 = rsqrt(epsilon = var_5274_epsilon_0, x = var_5273_cast_fp16)[name = string("op_5274_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_37_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5274_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_5276_cast_fp16 = mul(x = out_131_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5276_cast_fp16")];
            tensor<int32, [4]> var_5281_begin_0 = const()[name = string("op_5281_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5281_end_0 = const()[name = string("op_5281_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_5281_end_mask_0 = const()[name = string("op_5281_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_5281_cast_fp16 = slice_by_index(begin = var_5281_begin_0, end = var_5281_end_0, end_mask = var_5281_end_mask_0, x = out_131_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_5282_cast_fp16 = mul(x = var_5281_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_5282_cast_fp16")];
            tensor<int32, [4]> var_5287_begin_0 = const()[name = string("op_5287_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5287_end_0 = const()[name = string("op_5287_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_5287_end_mask_0 = const()[name = string("op_5287_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_5287_cast_fp16 = slice_by_index(begin = var_5287_begin_0, end = var_5287_end_0, end_mask = var_5287_end_mask_0, x = out_131_cast_fp16)[name = string("op_5287_cast_fp16")];
            int32 var_5289 = const()[name = string("op_5289"), val = int32(-1)];
            bool var_5290_interleave_0 = const()[name = string("op_5290_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_5290_cast_fp16 = concat(axis = var_5289, interleave = var_5290_interleave_0, values = (var_5282_cast_fp16, var_5287_cast_fp16))[name = string("op_5290_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_5291_cast_fp16 = mul(x = var_5290_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5291_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_39_cast_fp16 = add(x = var_5276_cast_fp16, y = var_5291_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_5294_cast_fp16 = mul(x = k_55_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5294_cast_fp16")];
            tensor<int32, [4]> var_5299_begin_0 = const()[name = string("op_5299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5299_end_0 = const()[name = string("op_5299_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_5299_end_mask_0 = const()[name = string("op_5299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_5299_cast_fp16 = slice_by_index(begin = var_5299_begin_0, end = var_5299_end_0, end_mask = var_5299_end_mask_0, x = k_55_cast_fp16)[name = string("op_5299_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_5300_cast_fp16 = mul(x = var_5299_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_5300_cast_fp16")];
            tensor<int32, [4]> var_5305_begin_0 = const()[name = string("op_5305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5305_end_0 = const()[name = string("op_5305_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_5305_end_mask_0 = const()[name = string("op_5305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_5305_cast_fp16 = slice_by_index(begin = var_5305_begin_0, end = var_5305_end_0, end_mask = var_5305_end_mask_0, x = k_55_cast_fp16)[name = string("op_5305_cast_fp16")];
            int32 var_5307 = const()[name = string("op_5307"), val = int32(-1)];
            bool var_5308_interleave_0 = const()[name = string("op_5308_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_5308_cast_fp16 = concat(axis = var_5307, interleave = var_5308_interleave_0, values = (var_5300_cast_fp16, var_5305_cast_fp16))[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_5309_cast_fp16 = mul(x = var_5308_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_57_cast_fp16 = add(x = var_5294_cast_fp16, y = var_5309_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_18_reps_0 = const()[name = string("tile_18_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = k_57_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [4, 1, 64, 256]> tile_18_cast_fp16 = tile(reps = tile_18_reps_0, x = transpose_36_cast_fp16)[name = string("tile_18_cast_fp16")];
            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_36_cast_fp16 = reshape(shape = concat_36, x = tile_18_cast_fp16)[name = string("reshape_36_cast_fp16")];
            tensor<int32, [5]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = reshape_36_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [4, 1, 64, 256]> reshape_37_cast_fp16 = reshape(shape = concat_37, x = transpose_37_cast_fp16)[name = string("reshape_37_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_19_reps_0 = const()[name = string("tile_19_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = v_37_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [4, 1, 64, 256]> tile_19_cast_fp16 = tile(reps = tile_19_reps_0, x = transpose_38_cast_fp16)[name = string("tile_19_cast_fp16")];
            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_38_cast_fp16 = reshape(shape = concat_38, x = tile_19_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [5]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [4, 1, 64, 256]> reshape_39_cast_fp16 = reshape(shape = concat_39, x = transpose_39_cast_fp16)[name = string("reshape_39_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5323_transpose_x_0 = const()[name = string("op_5323_transpose_x_0"), val = bool(false)];
            bool var_5323_transpose_y_0 = const()[name = string("op_5323_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = reshape_37_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 4, 64, 64]> var_5323_cast_fp16 = matmul(transpose_x = var_5323_transpose_x_0, transpose_y = var_5323_transpose_y_0, x = q_39_cast_fp16, y = transpose_105_cast_fp16)[name = string("op_5323_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_5325_cast_fp16 = add(x = var_5323_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5325_cast_fp16")];
            int32 var_5326 = const()[name = string("op_5326"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_153_cast_fp16 = softmax(axis = var_5326, x = var_5325_cast_fp16)[name = string("w_153_cast_fp16")];
            bool var_5329_transpose_x_0 = const()[name = string("op_5329_transpose_x_0"), val = bool(false)];
            bool var_5329_transpose_y_0 = const()[name = string("op_5329_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = reshape_39_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 4, 64, 256]> var_5329_cast_fp16 = matmul(transpose_x = var_5329_transpose_x_0, transpose_y = var_5329_transpose_y_0, x = w_153_cast_fp16, y = v_39_cast_fp16)[name = string("op_5329_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = var_5329_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 64, 1024]> input_111_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56383040))))[name = string("tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5341_cast_fp16 = abs(x = linear_85_cast_fp16)[name = string("op_5341_cast_fp16")];
            bool var_5344 = const()[name = string("op_5344"), val = bool(true)];
            tensor<int32, [1]> var_5345_axes_0 = const()[name = string("op_5345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5345_cast_fp16 = reduce_max(axes = var_5345_axes_0, keep_dims = var_5344, x = var_5341_cast_fp16)[name = string("op_5345_cast_fp16")];
            fp16 var_5346_to_fp16 = const()[name = string("op_5346_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_155_cast_fp16 = maximum(x = var_5345_cast_fp16, y = var_5346_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_155_cast_fp16 = real_div(x = linear_85_cast_fp16, y = s_155_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5349_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5354_axes_0 = const()[name = string("op_5354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5354_keep_dims_0 = const()[name = string("op_5354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5354_cast_fp16 = reduce_mean(axes = var_5354_axes_0, keep_dims = var_5354_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5355_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5355_cast_fp16")];
            fp32 var_5356_epsilon_0 = const()[name = string("op_5356_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5356_cast_fp16 = inverse(epsilon = var_5356_epsilon_0, x = var_5355_cast_fp16)[name = string("op_5356_cast_fp16")];
            fp16 var_5357_to_fp16 = const()[name = string("op_5357_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5358_cast_fp16 = mul(x = var_5356_cast_fp16, y = var_5357_to_fp16)[name = string("op_5358_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5360_cast_fp16 = add(x = var_5354_cast_fp16, y = var_5358_cast_fp16)[name = string("op_5360_cast_fp16")];
            fp32 var_5361_epsilon_0 = const()[name = string("op_5361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5361_cast_fp16 = rsqrt(epsilon = var_5361_epsilon_0, x = var_5360_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_135_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5361_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56384128)))];
            tensor<fp16, [1, 64, 512]> var_5363_cast_fp16 = mul(x = out_135_cast_fp16, y = tm_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_5363_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_173_cast_fp16 = add(x = x_163_cast_fp16, y = var_5363_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5366_cast_fp16 = abs(x = x_173_cast_fp16)[name = string("op_5366_cast_fp16")];
            bool var_5369 = const()[name = string("op_5369"), val = bool(true)];
            tensor<int32, [1]> var_5370_axes_0 = const()[name = string("op_5370_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5370_cast_fp16 = reduce_max(axes = var_5370_axes_0, keep_dims = var_5369, x = var_5366_cast_fp16)[name = string("op_5370_cast_fp16")];
            fp16 var_5371_to_fp16 = const()[name = string("op_5371_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_157_cast_fp16 = maximum(x = var_5370_cast_fp16, y = var_5371_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_157_cast_fp16 = real_div(x = x_173_cast_fp16, y = s_157_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5374_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [1]> var_5379_axes_0 = const()[name = string("op_5379_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5379_keep_dims_0 = const()[name = string("op_5379_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5379_cast_fp16 = reduce_mean(axes = var_5379_axes_0, keep_dims = var_5379_keep_dims_0, x = var_5374_cast_fp16)[name = string("op_5379_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5380_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5381_cast_fp16 = inverse(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            fp16 var_5382_to_fp16 = const()[name = string("op_5382_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5383_cast_fp16 = mul(x = var_5381_cast_fp16, y = var_5382_to_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5385_cast_fp16 = add(x = var_5379_cast_fp16, y = var_5383_cast_fp16)[name = string("op_5385_cast_fp16")];
            fp32 var_5386_epsilon_0 = const()[name = string("op_5386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5386_cast_fp16 = rsqrt(epsilon = var_5386_epsilon_0, x = var_5385_cast_fp16)[name = string("op_5386_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_137_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5386_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56385216)))];
            tensor<fp16, [1, 64, 512]> input_113_cast_fp16 = mul(x = out_137_cast_fp16, y = tm_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56386304))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57434944))))[name = string("tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_86_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_86_cast_fp16")];
            string var_5393_mode_0 = const()[name = string("op_5393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_5393_cast_fp16 = gelu(mode = var_5393_mode_0, x = linear_86_cast_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57439104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58487744))))[name = string("tm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_87_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_115_cast_fp16 = mul(x = var_5393_cast_fp16, y = linear_87_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58491904))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59540544))))[name = string("tm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5401_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5401_cast_fp16")];
            bool var_5404 = const()[name = string("op_5404"), val = bool(true)];
            tensor<int32, [1]> var_5405_axes_0 = const()[name = string("op_5405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5405_cast_fp16 = reduce_max(axes = var_5405_axes_0, keep_dims = var_5404, x = var_5401_cast_fp16)[name = string("op_5405_cast_fp16")];
            fp16 var_5406_to_fp16 = const()[name = string("op_5406_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_159_cast_fp16 = maximum(x = var_5405_cast_fp16, y = var_5406_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_159_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_159_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5409_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5409_cast_fp16")];
            tensor<int32, [1]> var_5414_axes_0 = const()[name = string("op_5414_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5414_keep_dims_0 = const()[name = string("op_5414_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5414_cast_fp16 = reduce_mean(axes = var_5414_axes_0, keep_dims = var_5414_keep_dims_0, x = var_5409_cast_fp16)[name = string("op_5414_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5415_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5415_cast_fp16")];
            fp32 var_5416_epsilon_0 = const()[name = string("op_5416_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5416_cast_fp16 = inverse(epsilon = var_5416_epsilon_0, x = var_5415_cast_fp16)[name = string("op_5416_cast_fp16")];
            fp16 var_5417_to_fp16 = const()[name = string("op_5417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5418_cast_fp16 = mul(x = var_5416_cast_fp16, y = var_5417_to_fp16)[name = string("op_5418_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5420_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5418_cast_fp16)[name = string("op_5420_cast_fp16")];
            fp32 var_5421_epsilon_0 = const()[name = string("op_5421_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5421_cast_fp16 = rsqrt(epsilon = var_5421_epsilon_0, x = var_5420_cast_fp16)[name = string("op_5421_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_139_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5421_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59541632)))];
            tensor<fp16, [1, 64, 512]> var_5423_cast_fp16 = mul(x = out_139_cast_fp16, y = tm_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5423_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_117_cast_fp16 = add(x = x_173_cast_fp16, y = var_5423_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59542720))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59804928))))[name = string("tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_89_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_89_cast_fp16")];
            string x_177_mode_0 = const()[name = string("x_177_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_177_cast_fp16 = gelu(mode = x_177_mode_0, x = linear_89_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [4]> var_5443_begin_0 = const()[name = string("op_5443_begin_0"), val = tensor<int32, [4]>([0, 0, 9, 0])];
            tensor<int32, [4]> var_5443_end_0 = const()[name = string("op_5443_end_0"), val = tensor<int32, [4]>([1, 64, 10, 512])];
            tensor<bool, [4]> var_5443_end_mask_0 = const()[name = string("op_5443_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5443_squeeze_mask_0 = const()[name = string("op_5443_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_5443_cast_fp16 = slice_by_index(begin = var_5443_begin_0, end = var_5443_end_0, end_mask = var_5443_end_mask_0, squeeze_mask = var_5443_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5443_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_119_cast_fp16 = mul(x = x_177_cast_fp16, y = var_5443_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59806016))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60068224))))[name = string("tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5453_cast_fp16 = abs(x = linear_90_cast_fp16)[name = string("op_5453_cast_fp16")];
            bool var_5456 = const()[name = string("op_5456"), val = bool(true)];
            tensor<int32, [1]> var_5457_axes_0 = const()[name = string("op_5457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5457_cast_fp16 = reduce_max(axes = var_5457_axes_0, keep_dims = var_5456, x = var_5453_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_161_cast_fp16 = maximum(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_161_cast_fp16 = real_div(x = linear_90_cast_fp16, y = s_161_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5461_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<int32, [1]> var_5466_axes_0 = const()[name = string("op_5466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5466_keep_dims_0 = const()[name = string("op_5466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5466_cast_fp16 = reduce_mean(axes = var_5466_axes_0, keep_dims = var_5466_keep_dims_0, x = var_5461_cast_fp16)[name = string("op_5466_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5467_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5467_cast_fp16")];
            fp32 var_5468_epsilon_0 = const()[name = string("op_5468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5468_cast_fp16 = inverse(epsilon = var_5468_epsilon_0, x = var_5467_cast_fp16)[name = string("op_5468_cast_fp16")];
            fp16 var_5469_to_fp16 = const()[name = string("op_5469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5470_cast_fp16 = mul(x = var_5468_cast_fp16, y = var_5469_to_fp16)[name = string("op_5470_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5472_cast_fp16 = add(x = var_5466_cast_fp16, y = var_5470_cast_fp16)[name = string("op_5472_cast_fp16")];
            fp32 var_5473_epsilon_0 = const()[name = string("op_5473_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5473_cast_fp16 = rsqrt(epsilon = var_5473_epsilon_0, x = var_5472_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_141_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5473_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60069312)))];
            tensor<fp16, [1, 64, 512]> var_5475_cast_fp16 = mul(x = out_141_cast_fp16, y = tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5475_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_19_cast_fp16 = add(x = input_117_cast_fp16, y = var_5475_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<fp16, [1]> tm_layers_9_layer_scalar_to_fp16 = const()[name = string("tm_layers_9_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9ep-1])];
            tensor<fp16, [1, 64, 512]> x_181_cast_fp16 = mul(x = h_19_cast_fp16, y = tm_layers_9_layer_scalar_to_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5479_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_5479_cast_fp16")];
            bool var_5482 = const()[name = string("op_5482"), val = bool(true)];
            tensor<int32, [1]> var_5483_axes_0 = const()[name = string("op_5483_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5483_cast_fp16 = reduce_max(axes = var_5483_axes_0, keep_dims = var_5482, x = var_5479_cast_fp16)[name = string("op_5483_cast_fp16")];
            fp16 var_5484_to_fp16 = const()[name = string("op_5484_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_163_cast_fp16 = maximum(x = var_5483_cast_fp16, y = var_5484_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_163_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_163_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5487_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5492_axes_0 = const()[name = string("op_5492_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5492_keep_dims_0 = const()[name = string("op_5492_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5492_cast_fp16 = reduce_mean(axes = var_5492_axes_0, keep_dims = var_5492_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5493_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5493_cast_fp16")];
            fp32 var_5494_epsilon_0 = const()[name = string("op_5494_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5494_cast_fp16 = inverse(epsilon = var_5494_epsilon_0, x = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp16 var_5495_to_fp16 = const()[name = string("op_5495_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5496_cast_fp16 = mul(x = var_5494_cast_fp16, y = var_5495_to_fp16)[name = string("op_5496_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5498_cast_fp16 = add(x = var_5492_cast_fp16, y = var_5496_cast_fp16)[name = string("op_5498_cast_fp16")];
            fp32 var_5499_epsilon_0 = const()[name = string("op_5499_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5499_cast_fp16 = rsqrt(epsilon = var_5499_epsilon_0, x = var_5498_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_143_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5499_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60070400)))];
            tensor<fp16, [1, 64, 512]> input_121_cast_fp16 = mul(x = out_143_cast_fp16, y = tm_layers_10_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60071488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60595840))))[name = string("tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5509 = const()[name = string("op_5509"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_5510_cast_fp16 = reshape(shape = var_5509, x = linear_91_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60597952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60860160))))[name = string("tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5521 = const()[name = string("op_5521"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_5522_cast_fp16 = reshape(shape = var_5521, x = linear_92_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60861248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61123456))))[name = string("tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5533 = const()[name = string("op_5533"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_5534_cast_fp16 = reshape(shape = var_5533, x = linear_93_cast_fp16)[name = string("op_5534_cast_fp16")];
            tensor<int32, [4]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5510_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 4, 64, 256]> var_5538_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_5538_cast_fp16")];
            bool var_5541 = const()[name = string("op_5541"), val = bool(true)];
            tensor<int32, [1]> var_5542_axes_0 = const()[name = string("op_5542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_5542_cast_fp16 = reduce_max(axes = var_5542_axes_0, keep_dims = var_5541, x = var_5538_cast_fp16)[name = string("op_5542_cast_fp16")];
            fp16 var_5543_to_fp16 = const()[name = string("op_5543_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_165_cast_fp16 = maximum(x = var_5542_cast_fp16, y = var_5543_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_165_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_165_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_5546_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<int32, [1]> var_5551_axes_0 = const()[name = string("op_5551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5551_keep_dims_0 = const()[name = string("op_5551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_5551_cast_fp16 = reduce_mean(axes = var_5551_axes_0, keep_dims = var_5551_keep_dims_0, x = var_5546_cast_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_5552_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5552_cast_fp16")];
            fp32 var_5553_epsilon_0 = const()[name = string("op_5553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_5553_cast_fp16 = inverse(epsilon = var_5553_epsilon_0, x = var_5552_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp16 var_5554_to_fp16 = const()[name = string("op_5554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_5555_cast_fp16 = mul(x = var_5553_cast_fp16, y = var_5554_to_fp16)[name = string("op_5555_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_5557_cast_fp16 = add(x = var_5551_cast_fp16, y = var_5555_cast_fp16)[name = string("op_5557_cast_fp16")];
            fp32 var_5558_epsilon_0 = const()[name = string("op_5558_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_5558_cast_fp16 = rsqrt(epsilon = var_5558_epsilon_0, x = var_5557_cast_fp16)[name = string("op_5558_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_145_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5558_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61124544)))];
            tensor<fp16, [1, 4, 64, 256]> q_41_cast_fp16 = mul(x = out_145_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_41_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_5522_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2, 64, 256]> var_5561_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_5561_cast_fp16")];
            bool var_5564 = const()[name = string("op_5564"), val = bool(true)];
            tensor<int32, [1]> var_5565_axes_0 = const()[name = string("op_5565_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_5565_cast_fp16 = reduce_max(axes = var_5565_axes_0, keep_dims = var_5564, x = var_5561_cast_fp16)[name = string("op_5565_cast_fp16")];
            fp16 var_5566_to_fp16 = const()[name = string("op_5566_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_167_cast_fp16 = maximum(x = var_5565_cast_fp16, y = var_5566_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_167_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_167_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_5569_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5569_cast_fp16")];
            tensor<int32, [1]> var_5574_axes_0 = const()[name = string("op_5574_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5574_keep_dims_0 = const()[name = string("op_5574_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_5574_cast_fp16 = reduce_mean(axes = var_5574_axes_0, keep_dims = var_5574_keep_dims_0, x = var_5569_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_5575_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5575_cast_fp16")];
            fp32 var_5576_epsilon_0 = const()[name = string("op_5576_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_5576_cast_fp16 = inverse(epsilon = var_5576_epsilon_0, x = var_5575_cast_fp16)[name = string("op_5576_cast_fp16")];
            fp16 var_5577_to_fp16 = const()[name = string("op_5577_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_5578_cast_fp16 = mul(x = var_5576_cast_fp16, y = var_5577_to_fp16)[name = string("op_5578_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_5580_cast_fp16 = add(x = var_5574_cast_fp16, y = var_5578_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_5581_cast_fp16 = rsqrt(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_147_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5581_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [256]> tm_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125120)))];
            tensor<fp16, [1, 2, 64, 256]> k_61_cast_fp16 = mul(x = out_147_cast_fp16, y = tm_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = var_5534_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 2, 64, 256]> var_5584_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_5584_cast_fp16")];
            bool var_5587 = const()[name = string("op_5587"), val = bool(true)];
            tensor<int32, [1]> var_5588_axes_0 = const()[name = string("op_5588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_5588_cast_fp16 = reduce_max(axes = var_5588_axes_0, keep_dims = var_5587, x = var_5584_cast_fp16)[name = string("op_5588_cast_fp16")];
            fp16 var_5589_to_fp16 = const()[name = string("op_5589_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_169_cast_fp16 = maximum(x = var_5588_cast_fp16, y = var_5589_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_169_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_169_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_5592_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5592_cast_fp16")];
            tensor<int32, [1]> var_5597_axes_0 = const()[name = string("op_5597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5597_keep_dims_0 = const()[name = string("op_5597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_5597_cast_fp16 = reduce_mean(axes = var_5597_axes_0, keep_dims = var_5597_keep_dims_0, x = var_5592_cast_fp16)[name = string("op_5597_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_5598_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5598_cast_fp16")];
            fp32 var_5599_epsilon_0 = const()[name = string("op_5599_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_5599_cast_fp16 = inverse(epsilon = var_5599_epsilon_0, x = var_5598_cast_fp16)[name = string("op_5599_cast_fp16")];
            fp16 var_5600_to_fp16 = const()[name = string("op_5600_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_5601_cast_fp16 = mul(x = var_5599_cast_fp16, y = var_5600_to_fp16)[name = string("op_5601_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_5603_cast_fp16 = add(x = var_5597_cast_fp16, y = var_5601_cast_fp16)[name = string("op_5603_cast_fp16")];
            fp32 var_5604_epsilon_0 = const()[name = string("op_5604_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_5604_cast_fp16 = rsqrt(epsilon = var_5604_epsilon_0, x = var_5603_cast_fp16)[name = string("op_5604_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_41_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5604_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_5606_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5611_begin_0 = const()[name = string("op_5611_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5611_end_0 = const()[name = string("op_5611_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_5611_end_mask_0 = const()[name = string("op_5611_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_5611_cast_fp16 = slice_by_index(begin = var_5611_begin_0, end = var_5611_end_0, end_mask = var_5611_end_mask_0, x = q_41_cast_fp16)[name = string("op_5611_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_5612_cast_fp16 = mul(x = var_5611_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_5612_cast_fp16")];
            tensor<int32, [4]> var_5617_begin_0 = const()[name = string("op_5617_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5617_end_0 = const()[name = string("op_5617_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_5617_end_mask_0 = const()[name = string("op_5617_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_5617_cast_fp16 = slice_by_index(begin = var_5617_begin_0, end = var_5617_end_0, end_mask = var_5617_end_mask_0, x = q_41_cast_fp16)[name = string("op_5617_cast_fp16")];
            int32 var_5619 = const()[name = string("op_5619"), val = int32(-1)];
            bool var_5620_interleave_0 = const()[name = string("op_5620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_5620_cast_fp16 = concat(axis = var_5619, interleave = var_5620_interleave_0, values = (var_5612_cast_fp16, var_5617_cast_fp16))[name = string("op_5620_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_5621_cast_fp16 = mul(x = var_5620_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5621_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_43_cast_fp16 = add(x = var_5606_cast_fp16, y = var_5621_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_5624_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = k_61_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = k_61_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool var_5638_interleave_0 = const()[name = string("op_5638_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_5638_cast_fp16 = concat(axis = var_5637, interleave = var_5638_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("op_5638_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_5639_cast_fp16 = mul(x = var_5638_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_63_cast_fp16 = add(x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_20_reps_0 = const()[name = string("tile_20_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = k_63_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [4, 1, 64, 256]> tile_20_cast_fp16 = tile(reps = tile_20_reps_0, x = transpose_40_cast_fp16)[name = string("tile_20_cast_fp16")];
            tensor<int32, [5]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_40_cast_fp16 = reshape(shape = concat_40, x = tile_20_cast_fp16)[name = string("reshape_40_cast_fp16")];
            tensor<int32, [5]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = reshape_40_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [4, 1, 64, 256]> reshape_41_cast_fp16 = reshape(shape = concat_41, x = transpose_41_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_21_reps_0 = const()[name = string("tile_21_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = v_41_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [4, 1, 64, 256]> tile_21_cast_fp16 = tile(reps = tile_21_reps_0, x = transpose_42_cast_fp16)[name = string("tile_21_cast_fp16")];
            tensor<int32, [5]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_42_cast_fp16 = reshape(shape = concat_42, x = tile_21_cast_fp16)[name = string("reshape_42_cast_fp16")];
            tensor<int32, [5]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = reshape_42_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [4, 1, 64, 256]> reshape_43_cast_fp16 = reshape(shape = concat_43, x = transpose_43_cast_fp16)[name = string("reshape_43_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5653_transpose_x_0 = const()[name = string("op_5653_transpose_x_0"), val = bool(false)];
            bool var_5653_transpose_y_0 = const()[name = string("op_5653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 4, 64, 64]> var_5653_cast_fp16 = matmul(transpose_x = var_5653_transpose_x_0, transpose_y = var_5653_transpose_y_0, x = q_43_cast_fp16, y = transpose_106_cast_fp16)[name = string("op_5653_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_5655_cast_fp16 = add(x = var_5653_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5655_cast_fp16")];
            int32 var_5656 = const()[name = string("op_5656"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_169_cast_fp16 = softmax(axis = var_5656, x = var_5655_cast_fp16)[name = string("w_169_cast_fp16")];
            bool var_5659_transpose_x_0 = const()[name = string("op_5659_transpose_x_0"), val = bool(false)];
            bool var_5659_transpose_y_0 = const()[name = string("op_5659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = reshape_43_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 4, 64, 256]> var_5659_cast_fp16 = matmul(transpose_x = var_5659_transpose_x_0, transpose_y = var_5659_transpose_y_0, x = w_169_cast_fp16, y = v_43_cast_fp16)[name = string("op_5659_cast_fp16")];
            tensor<int32, [4]> var_5662_perm_0 = const()[name = string("op_5662_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5666 = const()[name = string("op_5666"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_5662_cast_fp16 = transpose(perm = var_5662_perm_0, x = var_5659_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 64, 1024]> input_123_cast_fp16 = reshape(shape = var_5666, x = var_5662_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61650048))))[name = string("tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5671_cast_fp16 = abs(x = linear_94_cast_fp16)[name = string("op_5671_cast_fp16")];
            bool var_5674 = const()[name = string("op_5674"), val = bool(true)];
            tensor<int32, [1]> var_5675_axes_0 = const()[name = string("op_5675_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5675_cast_fp16 = reduce_max(axes = var_5675_axes_0, keep_dims = var_5674, x = var_5671_cast_fp16)[name = string("op_5675_cast_fp16")];
            fp16 var_5676_to_fp16 = const()[name = string("op_5676_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_171_cast_fp16 = maximum(x = var_5675_cast_fp16, y = var_5676_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_171_cast_fp16 = real_div(x = linear_94_cast_fp16, y = s_171_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5679_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5679_cast_fp16")];
            tensor<int32, [1]> var_5684_axes_0 = const()[name = string("op_5684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5684_keep_dims_0 = const()[name = string("op_5684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5684_cast_fp16 = reduce_mean(axes = var_5684_axes_0, keep_dims = var_5684_keep_dims_0, x = var_5679_cast_fp16)[name = string("op_5684_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5685_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5685_cast_fp16")];
            fp32 var_5686_epsilon_0 = const()[name = string("op_5686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5686_cast_fp16 = inverse(epsilon = var_5686_epsilon_0, x = var_5685_cast_fp16)[name = string("op_5686_cast_fp16")];
            fp16 var_5687_to_fp16 = const()[name = string("op_5687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5688_cast_fp16 = mul(x = var_5686_cast_fp16, y = var_5687_to_fp16)[name = string("op_5688_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5690_cast_fp16 = add(x = var_5684_cast_fp16, y = var_5688_cast_fp16)[name = string("op_5690_cast_fp16")];
            fp32 var_5691_epsilon_0 = const()[name = string("op_5691_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5691_cast_fp16 = rsqrt(epsilon = var_5691_epsilon_0, x = var_5690_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_149_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5691_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61651136)))];
            tensor<fp16, [1, 64, 512]> var_5693_cast_fp16 = mul(x = out_149_cast_fp16, y = tm_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5693_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = var_5693_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5696_cast_fp16 = abs(x = x_191_cast_fp16)[name = string("op_5696_cast_fp16")];
            bool var_5699 = const()[name = string("op_5699"), val = bool(true)];
            tensor<int32, [1]> var_5700_axes_0 = const()[name = string("op_5700_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5700_cast_fp16 = reduce_max(axes = var_5700_axes_0, keep_dims = var_5699, x = var_5696_cast_fp16)[name = string("op_5700_cast_fp16")];
            fp16 var_5701_to_fp16 = const()[name = string("op_5701_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_173_cast_fp16 = maximum(x = var_5700_cast_fp16, y = var_5701_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_173_cast_fp16 = real_div(x = x_191_cast_fp16, y = s_173_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5704_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5704_cast_fp16")];
            tensor<int32, [1]> var_5709_axes_0 = const()[name = string("op_5709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5709_keep_dims_0 = const()[name = string("op_5709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5709_cast_fp16 = reduce_mean(axes = var_5709_axes_0, keep_dims = var_5709_keep_dims_0, x = var_5704_cast_fp16)[name = string("op_5709_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5710_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5710_cast_fp16")];
            fp32 var_5711_epsilon_0 = const()[name = string("op_5711_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5711_cast_fp16 = inverse(epsilon = var_5711_epsilon_0, x = var_5710_cast_fp16)[name = string("op_5711_cast_fp16")];
            fp16 var_5712_to_fp16 = const()[name = string("op_5712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = string("op_5713_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5715_cast_fp16 = add(x = var_5709_cast_fp16, y = var_5713_cast_fp16)[name = string("op_5715_cast_fp16")];
            fp32 var_5716_epsilon_0 = const()[name = string("op_5716_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5716_cast_fp16 = rsqrt(epsilon = var_5716_epsilon_0, x = var_5715_cast_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_151_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5716_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61652224)))];
            tensor<fp16, [1, 64, 512]> input_125_cast_fp16 = mul(x = out_151_cast_fp16, y = tm_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61653312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62701952))))[name = string("tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_95_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_95_cast_fp16")];
            string var_5723_mode_0 = const()[name = string("op_5723_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_5723_cast_fp16 = gelu(mode = var_5723_mode_0, x = linear_95_cast_fp16)[name = string("op_5723_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62706112))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63754752))))[name = string("tm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_127_cast_fp16 = mul(x = var_5723_cast_fp16, y = linear_96_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63758912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64807552))))[name = string("tm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5731_cast_fp16 = abs(x = linear_97_cast_fp16)[name = string("op_5731_cast_fp16")];
            bool var_5734 = const()[name = string("op_5734"), val = bool(true)];
            tensor<int32, [1]> var_5735_axes_0 = const()[name = string("op_5735_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5735_cast_fp16 = reduce_max(axes = var_5735_axes_0, keep_dims = var_5734, x = var_5731_cast_fp16)[name = string("op_5735_cast_fp16")];
            fp16 var_5736_to_fp16 = const()[name = string("op_5736_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_175_cast_fp16 = maximum(x = var_5735_cast_fp16, y = var_5736_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_175_cast_fp16 = real_div(x = linear_97_cast_fp16, y = s_175_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5739_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [1]> var_5744_axes_0 = const()[name = string("op_5744_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5744_keep_dims_0 = const()[name = string("op_5744_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5744_cast_fp16 = reduce_mean(axes = var_5744_axes_0, keep_dims = var_5744_keep_dims_0, x = var_5739_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5745_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5745_cast_fp16")];
            fp32 var_5746_epsilon_0 = const()[name = string("op_5746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5746_cast_fp16 = inverse(epsilon = var_5746_epsilon_0, x = var_5745_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5747_to_fp16 = const()[name = string("op_5747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5748_cast_fp16 = mul(x = var_5746_cast_fp16, y = var_5747_to_fp16)[name = string("op_5748_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5750_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5748_cast_fp16)[name = string("op_5750_cast_fp16")];
            fp32 var_5751_epsilon_0 = const()[name = string("op_5751_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5751_cast_fp16 = rsqrt(epsilon = var_5751_epsilon_0, x = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_153_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5751_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64808640)))];
            tensor<fp16, [1, 64, 512]> var_5753_cast_fp16 = mul(x = out_153_cast_fp16, y = tm_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_129_cast_fp16 = add(x = x_191_cast_fp16, y = var_5753_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64809728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65071936))))[name = string("tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_98_cast_fp16")];
            string x_195_mode_0 = const()[name = string("x_195_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_195_cast_fp16 = gelu(mode = x_195_mode_0, x = linear_98_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<int32, [4]> var_5773_begin_0 = const()[name = string("op_5773_begin_0"), val = tensor<int32, [4]>([0, 0, 10, 0])];
            tensor<int32, [4]> var_5773_end_0 = const()[name = string("op_5773_end_0"), val = tensor<int32, [4]>([1, 64, 11, 512])];
            tensor<bool, [4]> var_5773_end_mask_0 = const()[name = string("op_5773_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5773_squeeze_mask_0 = const()[name = string("op_5773_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_5773_cast_fp16 = slice_by_index(begin = var_5773_begin_0, end = var_5773_end_0, end_mask = var_5773_end_mask_0, squeeze_mask = var_5773_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_131_cast_fp16 = mul(x = x_195_cast_fp16, y = var_5773_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65073024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65335232))))[name = string("tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5783_cast_fp16 = abs(x = linear_99_cast_fp16)[name = string("op_5783_cast_fp16")];
            bool var_5786 = const()[name = string("op_5786"), val = bool(true)];
            tensor<int32, [1]> var_5787_axes_0 = const()[name = string("op_5787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5787_cast_fp16 = reduce_max(axes = var_5787_axes_0, keep_dims = var_5786, x = var_5783_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp16 var_5788_to_fp16 = const()[name = string("op_5788_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_177_cast_fp16 = maximum(x = var_5787_cast_fp16, y = var_5788_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_177_cast_fp16 = real_div(x = linear_99_cast_fp16, y = s_177_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5791_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5791_cast_fp16")];
            tensor<int32, [1]> var_5796_axes_0 = const()[name = string("op_5796_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5796_keep_dims_0 = const()[name = string("op_5796_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5796_cast_fp16 = reduce_mean(axes = var_5796_axes_0, keep_dims = var_5796_keep_dims_0, x = var_5791_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5797_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5797_cast_fp16")];
            fp32 var_5798_epsilon_0 = const()[name = string("op_5798_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5798_cast_fp16 = inverse(epsilon = var_5798_epsilon_0, x = var_5797_cast_fp16)[name = string("op_5798_cast_fp16")];
            fp16 var_5799_to_fp16 = const()[name = string("op_5799_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5800_cast_fp16 = mul(x = var_5798_cast_fp16, y = var_5799_to_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5802_cast_fp16 = add(x = var_5796_cast_fp16, y = var_5800_cast_fp16)[name = string("op_5802_cast_fp16")];
            fp32 var_5803_epsilon_0 = const()[name = string("op_5803_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5803_cast_fp16 = rsqrt(epsilon = var_5803_epsilon_0, x = var_5802_cast_fp16)[name = string("op_5803_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_155_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5803_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65336320)))];
            tensor<fp16, [1, 64, 512]> var_5805_cast_fp16 = mul(x = out_155_cast_fp16, y = tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5805_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_21_cast_fp16 = add(x = input_129_cast_fp16, y = var_5805_cast_fp16)[name = string("h_21_cast_fp16")];
            tensor<fp16, [1]> tm_layers_10_layer_scalar_to_fp16 = const()[name = string("tm_layers_10_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.1cp-1])];
            tensor<fp16, [1, 64, 512]> x_199_cast_fp16 = mul(x = h_21_cast_fp16, y = tm_layers_10_layer_scalar_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5809_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5809_cast_fp16")];
            bool var_5812 = const()[name = string("op_5812"), val = bool(true)];
            tensor<int32, [1]> var_5813_axes_0 = const()[name = string("op_5813_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_5813_cast_fp16 = reduce_max(axes = var_5813_axes_0, keep_dims = var_5812, x = var_5809_cast_fp16)[name = string("op_5813_cast_fp16")];
            fp16 var_5814_to_fp16 = const()[name = string("op_5814_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_179_cast_fp16 = maximum(x = var_5813_cast_fp16, y = var_5814_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_179_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_179_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_5817_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5817_cast_fp16")];
            tensor<int32, [1]> var_5822_axes_0 = const()[name = string("op_5822_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5822_keep_dims_0 = const()[name = string("op_5822_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5822_cast_fp16 = reduce_mean(axes = var_5822_axes_0, keep_dims = var_5822_keep_dims_0, x = var_5817_cast_fp16)[name = string("op_5822_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5823_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp32 var_5824_epsilon_0 = const()[name = string("op_5824_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_5824_cast_fp16 = inverse(epsilon = var_5824_epsilon_0, x = var_5823_cast_fp16)[name = string("op_5824_cast_fp16")];
            fp16 var_5825_to_fp16 = const()[name = string("op_5825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_5826_cast_fp16 = mul(x = var_5824_cast_fp16, y = var_5825_to_fp16)[name = string("op_5826_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_5828_cast_fp16 = add(x = var_5822_cast_fp16, y = var_5826_cast_fp16)[name = string("op_5828_cast_fp16")];
            fp32 var_5829_epsilon_0 = const()[name = string("op_5829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_5829_cast_fp16 = rsqrt(epsilon = var_5829_epsilon_0, x = var_5828_cast_fp16)[name = string("op_5829_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_157_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5829_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65337408)))];
            tensor<fp16, [1, 64, 512]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = tm_layers_11_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65338496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66387136))))[name = string("tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_100_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5839 = const()[name = string("op_5839"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 64, 4, 512]> var_5840_cast_fp16 = reshape(shape = var_5839, x = linear_100_cast_fp16)[name = string("op_5840_cast_fp16")];
            tensor<int32, [4]> x_201_perm_0 = const()[name = string("x_201_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66391296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66653504))))[name = string("tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_5851 = const()[name = string("op_5851"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 64, 1, 512]> var_5852_cast_fp16 = reshape(shape = var_5851, x = linear_101_cast_fp16)[name = string("op_5852_cast_fp16")];
            tensor<int32, [4]> x_203_perm_0 = const()[name = string("x_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66654592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66916800))))[name = string("tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 64, 1, 512]> var_5864_cast_fp16 = reshape(shape = var_5863, x = linear_102_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 512]> x_201_cast_fp16 = transpose(perm = x_201_perm_0, x = var_5840_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 4, 64, 512]> var_5868_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5868_cast_fp16")];
            bool var_5871 = const()[name = string("op_5871"), val = bool(true)];
            tensor<int32, [1]> var_5872_axes_0 = const()[name = string("op_5872_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_5872_cast_fp16 = reduce_max(axes = var_5872_axes_0, keep_dims = var_5871, x = var_5868_cast_fp16)[name = string("op_5872_cast_fp16")];
            fp16 var_5873_to_fp16 = const()[name = string("op_5873_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_181_cast_fp16 = maximum(x = var_5872_cast_fp16, y = var_5873_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> y_181_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_181_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_5876_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5876_cast_fp16")];
            tensor<int32, [1]> var_5881_axes_0 = const()[name = string("op_5881_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5881_keep_dims_0 = const()[name = string("op_5881_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_5881_cast_fp16 = reduce_mean(axes = var_5881_axes_0, keep_dims = var_5881_keep_dims_0, x = var_5876_cast_fp16)[name = string("op_5881_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_5882_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5882_cast_fp16")];
            fp32 var_5883_epsilon_0 = const()[name = string("op_5883_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_5883_cast_fp16 = inverse(epsilon = var_5883_epsilon_0, x = var_5882_cast_fp16)[name = string("op_5883_cast_fp16")];
            fp16 var_5884_to_fp16 = const()[name = string("op_5884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_5885_cast_fp16 = mul(x = var_5883_cast_fp16, y = var_5884_to_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_5887_cast_fp16 = add(x = var_5881_cast_fp16, y = var_5885_cast_fp16)[name = string("op_5887_cast_fp16")];
            fp32 var_5888_epsilon_0 = const()[name = string("op_5888_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_5888_cast_fp16 = rsqrt(epsilon = var_5888_epsilon_0, x = var_5887_cast_fp16)[name = string("op_5888_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> out_159_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5888_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66917888)))];
            tensor<fp16, [1, 4, 64, 512]> q_45_cast_fp16 = mul(x = out_159_cast_fp16, y = tm_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> x_203_cast_fp16 = transpose(perm = x_203_perm_0, x = var_5852_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1, 64, 512]> var_5891_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5891_cast_fp16")];
            bool var_5894 = const()[name = string("op_5894"), val = bool(true)];
            tensor<int32, [1]> var_5895_axes_0 = const()[name = string("op_5895_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 64, 1]> var_5895_cast_fp16 = reduce_max(axes = var_5895_axes_0, keep_dims = var_5894, x = var_5891_cast_fp16)[name = string("op_5895_cast_fp16")];
            fp16 var_5896_to_fp16 = const()[name = string("op_5896_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 64, 1]> s_183_cast_fp16 = maximum(x = var_5895_cast_fp16, y = var_5896_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> y_183_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_183_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_5899_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5899_cast_fp16")];
            tensor<int32, [1]> var_5904_axes_0 = const()[name = string("op_5904_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5904_keep_dims_0 = const()[name = string("op_5904_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 64, 1]> var_5904_cast_fp16 = reduce_mean(axes = var_5904_axes_0, keep_dims = var_5904_keep_dims_0, x = var_5899_cast_fp16)[name = string("op_5904_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_5905_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5905_cast_fp16")];
            fp32 var_5906_epsilon_0 = const()[name = string("op_5906_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 64, 1]> var_5906_cast_fp16 = inverse(epsilon = var_5906_epsilon_0, x = var_5905_cast_fp16)[name = string("op_5906_cast_fp16")];
            fp16 var_5907_to_fp16 = const()[name = string("op_5907_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 64, 1]> var_5908_cast_fp16 = mul(x = var_5906_cast_fp16, y = var_5907_to_fp16)[name = string("op_5908_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_5910_cast_fp16 = add(x = var_5904_cast_fp16, y = var_5908_cast_fp16)[name = string("op_5910_cast_fp16")];
            fp32 var_5911_epsilon_0 = const()[name = string("op_5911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 64, 1]> var_5911_cast_fp16 = rsqrt(epsilon = var_5911_epsilon_0, x = var_5910_cast_fp16)[name = string("op_5911_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> out_161_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5911_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66918976)))];
            tensor<fp16, [1, 1, 64, 512]> k_67_cast_fp16 = mul(x = out_161_cast_fp16, y = tm_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_5864_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1, 64, 512]> var_5914_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5914_cast_fp16")];
            bool var_5917 = const()[name = string("op_5917"), val = bool(true)];
            tensor<int32, [1]> var_5918_axes_0 = const()[name = string("op_5918_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 64, 1]> var_5918_cast_fp16 = reduce_max(axes = var_5918_axes_0, keep_dims = var_5917, x = var_5914_cast_fp16)[name = string("op_5918_cast_fp16")];
            fp16 var_5919_to_fp16 = const()[name = string("op_5919_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 64, 1]> s_185_cast_fp16 = maximum(x = var_5918_cast_fp16, y = var_5919_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> y_185_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_185_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_5922_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5922_cast_fp16")];
            tensor<int32, [1]> var_5927_axes_0 = const()[name = string("op_5927_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5927_keep_dims_0 = const()[name = string("op_5927_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 64, 1]> var_5927_cast_fp16 = reduce_mean(axes = var_5927_axes_0, keep_dims = var_5927_keep_dims_0, x = var_5922_cast_fp16)[name = string("op_5927_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_5928_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5928_cast_fp16")];
            fp32 var_5929_epsilon_0 = const()[name = string("op_5929_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 64, 1]> var_5929_cast_fp16 = inverse(epsilon = var_5929_epsilon_0, x = var_5928_cast_fp16)[name = string("op_5929_cast_fp16")];
            fp16 var_5930_to_fp16 = const()[name = string("op_5930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 64, 1]> var_5931_cast_fp16 = mul(x = var_5929_cast_fp16, y = var_5930_to_fp16)[name = string("op_5931_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_5933_cast_fp16 = add(x = var_5927_cast_fp16, y = var_5931_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp32 var_5934_epsilon_0 = const()[name = string("op_5934_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 64, 1]> var_5934_cast_fp16 = rsqrt(epsilon = var_5934_epsilon_0, x = var_5933_cast_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> v_45_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5934_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_5936_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5936_cast_fp16")];
            tensor<int32, [4]> var_5941_begin_0 = const()[name = string("op_5941_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5941_end_0 = const()[name = string("op_5941_end_0"), val = tensor<int32, [4]>([1, 4, 64, 512])];
            tensor<bool, [4]> var_5941_end_mask_0 = const()[name = string("op_5941_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 256]> var_5941_cast_fp16 = slice_by_index(begin = var_5941_begin_0, end = var_5941_end_0, end_mask = var_5941_end_mask_0, x = q_45_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 256]> var_5942_cast_fp16 = mul(x = var_5941_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_5942_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 256]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = q_45_cast_fp16)[name = string("op_5947_cast_fp16")];
            int32 var_5949 = const()[name = string("op_5949"), val = int32(-1)];
            bool var_5950_interleave_0 = const()[name = string("op_5950_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 512]> var_5950_cast_fp16 = concat(axis = var_5949, interleave = var_5950_interleave_0, values = (var_5942_cast_fp16, var_5947_cast_fp16))[name = string("op_5950_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_5951_cast_fp16 = mul(x = var_5950_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5951_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> q_47_cast_fp16 = add(x = var_5936_cast_fp16, y = var_5951_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_5954_cast_fp16 = mul(x = k_67_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_5954_cast_fp16")];
            tensor<int32, [4]> var_5959_begin_0 = const()[name = string("op_5959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5959_end_0 = const()[name = string("op_5959_end_0"), val = tensor<int32, [4]>([1, 1, 64, 512])];
            tensor<bool, [4]> var_5959_end_mask_0 = const()[name = string("op_5959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 64, 256]> var_5959_cast_fp16 = slice_by_index(begin = var_5959_begin_0, end = var_5959_end_0, end_mask = var_5959_end_mask_0, x = k_67_cast_fp16)[name = string("op_5959_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 64, 256]> var_5960_cast_fp16 = mul(x = var_5959_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_5960_cast_fp16")];
            tensor<int32, [4]> var_5965_begin_0 = const()[name = string("op_5965_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5965_end_0 = const()[name = string("op_5965_end_0"), val = tensor<int32, [4]>([1, 1, 64, 256])];
            tensor<bool, [4]> var_5965_end_mask_0 = const()[name = string("op_5965_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 64, 256]> var_5965_cast_fp16 = slice_by_index(begin = var_5965_begin_0, end = var_5965_end_0, end_mask = var_5965_end_mask_0, x = k_67_cast_fp16)[name = string("op_5965_cast_fp16")];
            int32 var_5967 = const()[name = string("op_5967"), val = int32(-1)];
            bool var_5968_interleave_0 = const()[name = string("op_5968_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 64, 512]> var_5968_cast_fp16 = concat(axis = var_5967, interleave = var_5968_interleave_0, values = (var_5960_cast_fp16, var_5965_cast_fp16))[name = string("op_5968_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_5969_cast_fp16 = mul(x = var_5968_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_5969_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> k_69_cast_fp16 = add(x = var_5954_cast_fp16, y = var_5969_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_22_reps_0 = const()[name = string("tile_22_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 64, 512]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = k_69_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [4, 1, 64, 512]> tile_22_cast_fp16 = tile(reps = tile_22_reps_0, x = transpose_44_cast_fp16)[name = string("tile_22_cast_fp16")];
            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([4, 1, 1, 64, 512])];
            tensor<fp16, [4, 1, 1, 64, 512]> reshape_44_cast_fp16 = reshape(shape = concat_44, x = tile_22_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [5]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 64, 512])];
            tensor<fp16, [1, 4, 1, 64, 512]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [4, 1, 64, 512]> reshape_45_cast_fp16 = reshape(shape = concat_45, x = transpose_45_cast_fp16)[name = string("reshape_45_cast_fp16")];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_23_reps_0 = const()[name = string("tile_23_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 64, 512]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = v_45_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [4, 1, 64, 512]> tile_23_cast_fp16 = tile(reps = tile_23_reps_0, x = transpose_46_cast_fp16)[name = string("tile_23_cast_fp16")];
            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([4, 1, 1, 64, 512])];
            tensor<fp16, [4, 1, 1, 64, 512]> reshape_46_cast_fp16 = reshape(shape = concat_46, x = tile_23_cast_fp16)[name = string("reshape_46_cast_fp16")];
            tensor<int32, [5]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 64, 512])];
            tensor<fp16, [1, 4, 1, 64, 512]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = reshape_46_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [4, 1, 64, 512]> reshape_47_cast_fp16 = reshape(shape = concat_47, x = transpose_47_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5983_transpose_x_0 = const()[name = string("op_5983_transpose_x_0"), val = bool(false)];
            bool var_5983_transpose_y_0 = const()[name = string("op_5983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 64]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = reshape_45_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 4, 64, 64]> var_5983_cast_fp16 = matmul(transpose_x = var_5983_transpose_x_0, transpose_y = var_5983_transpose_y_0, x = q_47_cast_fp16, y = transpose_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_5985_cast_fp16 = add(x = var_5983_cast_fp16, y = var_2173_cast_fp16)[name = string("op_5985_cast_fp16")];
            int32 var_5986 = const()[name = string("op_5986"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_185_cast_fp16 = softmax(axis = var_5986, x = var_5985_cast_fp16)[name = string("w_185_cast_fp16")];
            bool var_5989_transpose_x_0 = const()[name = string("op_5989_transpose_x_0"), val = bool(false)];
            bool var_5989_transpose_y_0 = const()[name = string("op_5989_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 512]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 4, 64, 512]> var_5989_cast_fp16 = matmul(transpose_x = var_5989_transpose_x_0, transpose_y = var_5989_transpose_y_0, x = w_185_cast_fp16, y = v_47_cast_fp16)[name = string("op_5989_cast_fp16")];
            tensor<int32, [4]> var_5992_perm_0 = const()[name = string("op_5992_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5996 = const()[name = string("op_5996"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 64, 4, 512]> var_5992_cast_fp16 = transpose(perm = var_5992_perm_0, x = var_5989_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 64, 2048]> input_135_cast_fp16 = reshape(shape = var_5996, x = var_5992_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66920064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67968704))))[name = string("tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6001_cast_fp16 = abs(x = linear_103_cast_fp16)[name = string("op_6001_cast_fp16")];
            bool var_6004 = const()[name = string("op_6004"), val = bool(true)];
            tensor<int32, [1]> var_6005_axes_0 = const()[name = string("op_6005_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6005_cast_fp16 = reduce_max(axes = var_6005_axes_0, keep_dims = var_6004, x = var_6001_cast_fp16)[name = string("op_6005_cast_fp16")];
            fp16 var_6006_to_fp16 = const()[name = string("op_6006_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_187_cast_fp16 = maximum(x = var_6005_cast_fp16, y = var_6006_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_187_cast_fp16 = real_div(x = linear_103_cast_fp16, y = s_187_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6009_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_6009_cast_fp16")];
            tensor<int32, [1]> var_6014_axes_0 = const()[name = string("op_6014_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6014_keep_dims_0 = const()[name = string("op_6014_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6014_cast_fp16 = reduce_mean(axes = var_6014_axes_0, keep_dims = var_6014_keep_dims_0, x = var_6009_cast_fp16)[name = string("op_6014_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6015_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_6015_cast_fp16")];
            fp32 var_6016_epsilon_0 = const()[name = string("op_6016_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6016_cast_fp16 = inverse(epsilon = var_6016_epsilon_0, x = var_6015_cast_fp16)[name = string("op_6016_cast_fp16")];
            fp16 var_6017_to_fp16 = const()[name = string("op_6017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6018_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6020_cast_fp16 = add(x = var_6014_cast_fp16, y = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp32 var_6021_epsilon_0 = const()[name = string("op_6021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6021_cast_fp16 = rsqrt(epsilon = var_6021_epsilon_0, x = var_6020_cast_fp16)[name = string("op_6021_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_163_cast_fp16 = mul(x = y_187_cast_fp16, y = var_6021_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969792)))];
            tensor<fp16, [1, 64, 512]> var_6023_cast_fp16 = mul(x = out_163_cast_fp16, y = tm_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_209_cast_fp16 = add(x = x_199_cast_fp16, y = var_6023_cast_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6026_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_6026_cast_fp16")];
            bool var_6029 = const()[name = string("op_6029"), val = bool(true)];
            tensor<int32, [1]> var_6030_axes_0 = const()[name = string("op_6030_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6030_cast_fp16 = reduce_max(axes = var_6030_axes_0, keep_dims = var_6029, x = var_6026_cast_fp16)[name = string("op_6030_cast_fp16")];
            fp16 var_6031_to_fp16 = const()[name = string("op_6031_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_189_cast_fp16 = maximum(x = var_6030_cast_fp16, y = var_6031_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_189_cast_fp16 = real_div(x = x_209_cast_fp16, y = s_189_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6034_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [1]> var_6039_axes_0 = const()[name = string("op_6039_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6039_keep_dims_0 = const()[name = string("op_6039_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6039_cast_fp16 = reduce_mean(axes = var_6039_axes_0, keep_dims = var_6039_keep_dims_0, x = var_6034_cast_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6040_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6041_cast_fp16 = inverse(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            fp16 var_6042_to_fp16 = const()[name = string("op_6042_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6043_cast_fp16 = mul(x = var_6041_cast_fp16, y = var_6042_to_fp16)[name = string("op_6043_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6045_cast_fp16 = add(x = var_6039_cast_fp16, y = var_6043_cast_fp16)[name = string("op_6045_cast_fp16")];
            fp32 var_6046_epsilon_0 = const()[name = string("op_6046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6046_cast_fp16 = rsqrt(epsilon = var_6046_epsilon_0, x = var_6045_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_165_cast_fp16 = mul(x = y_189_cast_fp16, y = var_6046_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67970880)))];
            tensor<fp16, [1, 64, 512]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = tm_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971968))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69020608))))[name = string("tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_104_cast_fp16")];
            string var_6053_mode_0 = const()[name = string("op_6053_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_6053_cast_fp16 = gelu(mode = var_6053_mode_0, x = linear_104_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70073408))))[name = string("tm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_105_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_139_cast_fp16 = mul(x = var_6053_cast_fp16, y = linear_105_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70077568))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71126208))))[name = string("tm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6061_cast_fp16 = abs(x = linear_106_cast_fp16)[name = string("op_6061_cast_fp16")];
            bool var_6064 = const()[name = string("op_6064"), val = bool(true)];
            tensor<int32, [1]> var_6065_axes_0 = const()[name = string("op_6065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6065_cast_fp16 = reduce_max(axes = var_6065_axes_0, keep_dims = var_6064, x = var_6061_cast_fp16)[name = string("op_6065_cast_fp16")];
            fp16 var_6066_to_fp16 = const()[name = string("op_6066_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_191_cast_fp16 = maximum(x = var_6065_cast_fp16, y = var_6066_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_191_cast_fp16 = real_div(x = linear_106_cast_fp16, y = s_191_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6069_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<int32, [1]> var_6074_axes_0 = const()[name = string("op_6074_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6074_keep_dims_0 = const()[name = string("op_6074_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6074_cast_fp16 = reduce_mean(axes = var_6074_axes_0, keep_dims = var_6074_keep_dims_0, x = var_6069_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6075_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_6075_cast_fp16")];
            fp32 var_6076_epsilon_0 = const()[name = string("op_6076_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6076_cast_fp16 = inverse(epsilon = var_6076_epsilon_0, x = var_6075_cast_fp16)[name = string("op_6076_cast_fp16")];
            fp16 var_6077_to_fp16 = const()[name = string("op_6077_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6078_cast_fp16 = mul(x = var_6076_cast_fp16, y = var_6077_to_fp16)[name = string("op_6078_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6080_cast_fp16 = add(x = var_6074_cast_fp16, y = var_6078_cast_fp16)[name = string("op_6080_cast_fp16")];
            fp32 var_6081_epsilon_0 = const()[name = string("op_6081_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6081_cast_fp16 = rsqrt(epsilon = var_6081_epsilon_0, x = var_6080_cast_fp16)[name = string("op_6081_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_167_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6081_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71127296)))];
            tensor<fp16, [1, 64, 512]> var_6083_cast_fp16 = mul(x = out_167_cast_fp16, y = tm_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6083_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_141_cast_fp16 = add(x = x_209_cast_fp16, y = var_6083_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71128384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71390592))))[name = string("tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_107_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_107_cast_fp16")];
            string x_213_mode_0 = const()[name = string("x_213_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_213_cast_fp16 = gelu(mode = x_213_mode_0, x = linear_107_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [4]> var_6103_begin_0 = const()[name = string("op_6103_begin_0"), val = tensor<int32, [4]>([0, 0, 11, 0])];
            tensor<int32, [4]> var_6103_end_0 = const()[name = string("op_6103_end_0"), val = tensor<int32, [4]>([1, 64, 12, 512])];
            tensor<bool, [4]> var_6103_end_mask_0 = const()[name = string("op_6103_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6103_squeeze_mask_0 = const()[name = string("op_6103_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_6103_cast_fp16 = slice_by_index(begin = var_6103_begin_0, end = var_6103_end_0, end_mask = var_6103_end_mask_0, squeeze_mask = var_6103_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6103_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_143_cast_fp16 = mul(x = x_213_cast_fp16, y = var_6103_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71391680))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71653888))))[name = string("tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6113_cast_fp16 = abs(x = linear_108_cast_fp16)[name = string("op_6113_cast_fp16")];
            bool var_6116 = const()[name = string("op_6116"), val = bool(true)];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6117_cast_fp16 = reduce_max(axes = var_6117_axes_0, keep_dims = var_6116, x = var_6113_cast_fp16)[name = string("op_6117_cast_fp16")];
            fp16 var_6118_to_fp16 = const()[name = string("op_6118_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_193_cast_fp16 = maximum(x = var_6117_cast_fp16, y = var_6118_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_193_cast_fp16 = real_div(x = linear_108_cast_fp16, y = s_193_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6121_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6121_cast_fp16")];
            tensor<int32, [1]> var_6126_axes_0 = const()[name = string("op_6126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6126_keep_dims_0 = const()[name = string("op_6126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6126_cast_fp16 = reduce_mean(axes = var_6126_axes_0, keep_dims = var_6126_keep_dims_0, x = var_6121_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6127_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_6127_cast_fp16")];
            fp32 var_6128_epsilon_0 = const()[name = string("op_6128_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6128_cast_fp16 = inverse(epsilon = var_6128_epsilon_0, x = var_6127_cast_fp16)[name = string("op_6128_cast_fp16")];
            fp16 var_6129_to_fp16 = const()[name = string("op_6129_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6130_cast_fp16 = mul(x = var_6128_cast_fp16, y = var_6129_to_fp16)[name = string("op_6130_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6132_cast_fp16 = add(x = var_6126_cast_fp16, y = var_6130_cast_fp16)[name = string("op_6132_cast_fp16")];
            fp32 var_6133_epsilon_0 = const()[name = string("op_6133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6133_cast_fp16 = rsqrt(epsilon = var_6133_epsilon_0, x = var_6132_cast_fp16)[name = string("op_6133_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_169_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6133_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71654976)))];
            tensor<fp16, [1, 64, 512]> var_6135_cast_fp16 = mul(x = out_169_cast_fp16, y = tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6135_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_23_cast_fp16 = add(x = input_141_cast_fp16, y = var_6135_cast_fp16)[name = string("h_23_cast_fp16")];
            tensor<fp16, [1]> tm_layers_11_layer_scalar_to_fp16 = const()[name = string("tm_layers_11_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.86p-2])];
            tensor<fp16, [1, 64, 512]> x_217_cast_fp16 = mul(x = h_23_cast_fp16, y = tm_layers_11_layer_scalar_to_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6139_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_195_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_195_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_195_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6153_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_171_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71656064)))];
            tensor<fp16, [1, 64, 512]> input_145_cast_fp16 = mul(x = out_171_cast_fp16, y = tm_layers_12_input_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71657152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72181504))))[name = string("tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<int32, [4]> var_6169 = const()[name = string("op_6169"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_6170_cast_fp16 = reshape(shape = var_6169, x = linear_109_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72183616))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72445824))))[name = string("tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [4]> var_6181 = const()[name = string("op_6181"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_6182_cast_fp16 = reshape(shape = var_6181, x = linear_110_cast_fp16)[name = string("op_6182_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72446912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72709120))))[name = string("tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<int32, [4]> var_6193 = const()[name = string("op_6193"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_6194_cast_fp16 = reshape(shape = var_6193, x = linear_111_cast_fp16)[name = string("op_6194_cast_fp16")];
            tensor<int32, [4]> x_223_perm_0 = const()[name = string("x_223_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6170_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 4, 64, 256]> var_6198_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_6198_cast_fp16")];
            bool var_6201 = const()[name = string("op_6201"), val = bool(true)];
            tensor<int32, [1]> var_6202_axes_0 = const()[name = string("op_6202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_6202_cast_fp16 = reduce_max(axes = var_6202_axes_0, keep_dims = var_6201, x = var_6198_cast_fp16)[name = string("op_6202_cast_fp16")];
            fp16 var_6203_to_fp16 = const()[name = string("op_6203_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_197_cast_fp16 = maximum(x = var_6202_cast_fp16, y = var_6203_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_197_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_197_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6206_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6206_cast_fp16")];
            tensor<int32, [1]> var_6211_axes_0 = const()[name = string("op_6211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6211_keep_dims_0 = const()[name = string("op_6211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_6211_cast_fp16 = reduce_mean(axes = var_6211_axes_0, keep_dims = var_6211_keep_dims_0, x = var_6206_cast_fp16)[name = string("op_6211_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_6212_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6212_cast_fp16")];
            fp32 var_6213_epsilon_0 = const()[name = string("op_6213_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_6213_cast_fp16 = inverse(epsilon = var_6213_epsilon_0, x = var_6212_cast_fp16)[name = string("op_6213_cast_fp16")];
            fp16 var_6214_to_fp16 = const()[name = string("op_6214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_6215_cast_fp16 = mul(x = var_6213_cast_fp16, y = var_6214_to_fp16)[name = string("op_6215_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_6217_cast_fp16 = add(x = var_6211_cast_fp16, y = var_6215_cast_fp16)[name = string("op_6217_cast_fp16")];
            fp32 var_6218_epsilon_0 = const()[name = string("op_6218_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_6218_cast_fp16 = rsqrt(epsilon = var_6218_epsilon_0, x = var_6217_cast_fp16)[name = string("op_6218_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_173_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6218_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_49_cast_fp16 = mul(x = out_173_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_49_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6182_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 2, 64, 256]> var_6221_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_6221_cast_fp16")];
            bool var_6224 = const()[name = string("op_6224"), val = bool(true)];
            tensor<int32, [1]> var_6225_axes_0 = const()[name = string("op_6225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_6225_cast_fp16 = reduce_max(axes = var_6225_axes_0, keep_dims = var_6224, x = var_6221_cast_fp16)[name = string("op_6225_cast_fp16")];
            fp16 var_6226_to_fp16 = const()[name = string("op_6226_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_199_cast_fp16 = maximum(x = var_6225_cast_fp16, y = var_6226_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_199_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_199_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6229_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6229_cast_fp16")];
            tensor<int32, [1]> var_6234_axes_0 = const()[name = string("op_6234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6234_keep_dims_0 = const()[name = string("op_6234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_6234_cast_fp16 = reduce_mean(axes = var_6234_axes_0, keep_dims = var_6234_keep_dims_0, x = var_6229_cast_fp16)[name = string("op_6234_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6235_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6235_cast_fp16")];
            fp32 var_6236_epsilon_0 = const()[name = string("op_6236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_6236_cast_fp16 = inverse(epsilon = var_6236_epsilon_0, x = var_6235_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_6238_cast_fp16 = mul(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6240_cast_fp16 = add(x = var_6234_cast_fp16, y = var_6238_cast_fp16)[name = string("op_6240_cast_fp16")];
            fp32 var_6241_epsilon_0 = const()[name = string("op_6241_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_6241_cast_fp16 = rsqrt(epsilon = var_6241_epsilon_0, x = var_6240_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_175_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6241_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_73_cast_fp16 = mul(x = out_175_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_223_cast_fp16 = transpose(perm = x_223_perm_0, x = var_6194_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2, 64, 256]> var_6244_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_6244_cast_fp16")];
            bool var_6247 = const()[name = string("op_6247"), val = bool(true)];
            tensor<int32, [1]> var_6248_axes_0 = const()[name = string("op_6248_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_6248_cast_fp16 = reduce_max(axes = var_6248_axes_0, keep_dims = var_6247, x = var_6244_cast_fp16)[name = string("op_6248_cast_fp16")];
            fp16 var_6249_to_fp16 = const()[name = string("op_6249_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_201_cast_fp16 = maximum(x = var_6248_cast_fp16, y = var_6249_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_201_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_201_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6252_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6252_cast_fp16")];
            tensor<int32, [1]> var_6257_axes_0 = const()[name = string("op_6257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6257_keep_dims_0 = const()[name = string("op_6257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_6257_cast_fp16 = reduce_mean(axes = var_6257_axes_0, keep_dims = var_6257_keep_dims_0, x = var_6252_cast_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6258_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6258_cast_fp16")];
            fp32 var_6259_epsilon_0 = const()[name = string("op_6259_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_6259_cast_fp16 = inverse(epsilon = var_6259_epsilon_0, x = var_6258_cast_fp16)[name = string("op_6259_cast_fp16")];
            fp16 var_6260_to_fp16 = const()[name = string("op_6260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_6261_cast_fp16 = mul(x = var_6259_cast_fp16, y = var_6260_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6263_cast_fp16 = add(x = var_6257_cast_fp16, y = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp32 var_6264_epsilon_0 = const()[name = string("op_6264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_6264_cast_fp16 = rsqrt(epsilon = var_6264_epsilon_0, x = var_6263_cast_fp16)[name = string("op_6264_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_49_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6264_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6266_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6266_cast_fp16")];
            tensor<int32, [4]> var_6271_begin_0 = const()[name = string("op_6271_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6271_end_0 = const()[name = string("op_6271_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_6271_end_mask_0 = const()[name = string("op_6271_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_6271_cast_fp16 = slice_by_index(begin = var_6271_begin_0, end = var_6271_end_0, end_mask = var_6271_end_mask_0, x = q_49_cast_fp16)[name = string("op_6271_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_6272_cast_fp16 = mul(x = var_6271_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_6272_cast_fp16")];
            tensor<int32, [4]> var_6277_begin_0 = const()[name = string("op_6277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6277_end_0 = const()[name = string("op_6277_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_6277_end_mask_0 = const()[name = string("op_6277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_6277_cast_fp16 = slice_by_index(begin = var_6277_begin_0, end = var_6277_end_0, end_mask = var_6277_end_mask_0, x = q_49_cast_fp16)[name = string("op_6277_cast_fp16")];
            int32 var_6279 = const()[name = string("op_6279"), val = int32(-1)];
            bool var_6280_interleave_0 = const()[name = string("op_6280_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_6280_cast_fp16 = concat(axis = var_6279, interleave = var_6280_interleave_0, values = (var_6272_cast_fp16, var_6277_cast_fp16))[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6281_cast_fp16 = mul(x = var_6280_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6281_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_51_cast_fp16 = add(x = var_6266_cast_fp16, y = var_6281_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6284_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6284_cast_fp16")];
            tensor<int32, [4]> var_6289_begin_0 = const()[name = string("op_6289_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6289_end_0 = const()[name = string("op_6289_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_6289_end_mask_0 = const()[name = string("op_6289_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_6289_cast_fp16 = slice_by_index(begin = var_6289_begin_0, end = var_6289_end_0, end_mask = var_6289_end_mask_0, x = k_73_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_6290_cast_fp16 = mul(x = var_6289_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_6290_cast_fp16")];
            tensor<int32, [4]> var_6295_begin_0 = const()[name = string("op_6295_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6295_end_0 = const()[name = string("op_6295_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_6295_end_mask_0 = const()[name = string("op_6295_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_6295_cast_fp16 = slice_by_index(begin = var_6295_begin_0, end = var_6295_end_0, end_mask = var_6295_end_mask_0, x = k_73_cast_fp16)[name = string("op_6295_cast_fp16")];
            int32 var_6297 = const()[name = string("op_6297"), val = int32(-1)];
            bool var_6298_interleave_0 = const()[name = string("op_6298_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_6298_cast_fp16 = concat(axis = var_6297, interleave = var_6298_interleave_0, values = (var_6290_cast_fp16, var_6295_cast_fp16))[name = string("op_6298_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6299_cast_fp16 = mul(x = var_6298_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6299_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_75_cast_fp16 = add(x = var_6284_cast_fp16, y = var_6299_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_24_reps_0 = const()[name = string("tile_24_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = k_75_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [4, 1, 64, 256]> tile_24_cast_fp16 = tile(reps = tile_24_reps_0, x = transpose_48_cast_fp16)[name = string("tile_24_cast_fp16")];
            tensor<int32, [5]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_48_cast_fp16 = reshape(shape = concat_48, x = tile_24_cast_fp16)[name = string("reshape_48_cast_fp16")];
            tensor<int32, [5]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = reshape_48_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [4, 1, 64, 256]> reshape_49_cast_fp16 = reshape(shape = concat_49, x = transpose_49_cast_fp16)[name = string("reshape_49_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_25_reps_0 = const()[name = string("tile_25_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = v_49_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [4, 1, 64, 256]> tile_25_cast_fp16 = tile(reps = tile_25_reps_0, x = transpose_50_cast_fp16)[name = string("tile_25_cast_fp16")];
            tensor<int32, [5]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_50_cast_fp16 = reshape(shape = concat_50, x = tile_25_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [5]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [4, 1, 64, 256]> reshape_51_cast_fp16 = reshape(shape = concat_51, x = transpose_51_cast_fp16)[name = string("reshape_51_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6313_transpose_x_0 = const()[name = string("op_6313_transpose_x_0"), val = bool(false)];
            bool var_6313_transpose_y_0 = const()[name = string("op_6313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = reshape_49_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 4, 64, 64]> var_6313_cast_fp16 = matmul(transpose_x = var_6313_transpose_x_0, transpose_y = var_6313_transpose_y_0, x = q_51_cast_fp16, y = transpose_108_cast_fp16)[name = string("op_6313_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_6315_cast_fp16 = add(x = var_6313_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6315_cast_fp16")];
            int32 var_6316 = const()[name = string("op_6316"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_201_cast_fp16 = softmax(axis = var_6316, x = var_6315_cast_fp16)[name = string("w_201_cast_fp16")];
            bool var_6319_transpose_x_0 = const()[name = string("op_6319_transpose_x_0"), val = bool(false)];
            bool var_6319_transpose_y_0 = const()[name = string("op_6319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = reshape_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 4, 64, 256]> var_6319_cast_fp16 = matmul(transpose_x = var_6319_transpose_x_0, transpose_y = var_6319_transpose_y_0, x = w_201_cast_fp16, y = v_51_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<int32, [4]> var_6322_perm_0 = const()[name = string("op_6322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6326 = const()[name = string("op_6326"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_6322_cast_fp16 = transpose(perm = var_6322_perm_0, x = var_6319_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 64, 1024]> input_147_cast_fp16 = reshape(shape = var_6326, x = var_6322_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72710208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73234560))))[name = string("tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6331_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6331_cast_fp16")];
            bool var_6334 = const()[name = string("op_6334"), val = bool(true)];
            tensor<int32, [1]> var_6335_axes_0 = const()[name = string("op_6335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6335_cast_fp16 = reduce_max(axes = var_6335_axes_0, keep_dims = var_6334, x = var_6331_cast_fp16)[name = string("op_6335_cast_fp16")];
            fp16 var_6336_to_fp16 = const()[name = string("op_6336_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_203_cast_fp16 = maximum(x = var_6335_cast_fp16, y = var_6336_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_203_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_203_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6339_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6339_cast_fp16")];
            tensor<int32, [1]> var_6344_axes_0 = const()[name = string("op_6344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6344_keep_dims_0 = const()[name = string("op_6344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6344_cast_fp16 = reduce_mean(axes = var_6344_axes_0, keep_dims = var_6344_keep_dims_0, x = var_6339_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6345_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6345_cast_fp16")];
            fp32 var_6346_epsilon_0 = const()[name = string("op_6346_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6346_cast_fp16 = inverse(epsilon = var_6346_epsilon_0, x = var_6345_cast_fp16)[name = string("op_6346_cast_fp16")];
            fp16 var_6347_to_fp16 = const()[name = string("op_6347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6348_cast_fp16 = mul(x = var_6346_cast_fp16, y = var_6347_to_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6350_cast_fp16 = add(x = var_6344_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            fp32 var_6351_epsilon_0 = const()[name = string("op_6351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6351_cast_fp16 = rsqrt(epsilon = var_6351_epsilon_0, x = var_6350_cast_fp16)[name = string("op_6351_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_177_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6351_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73235648)))];
            tensor<fp16, [1, 64, 512]> var_6353_cast_fp16 = mul(x = out_177_cast_fp16, y = tm_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = var_6353_cast_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6356_cast_fp16 = abs(x = x_227_cast_fp16)[name = string("op_6356_cast_fp16")];
            bool var_6359 = const()[name = string("op_6359"), val = bool(true)];
            tensor<int32, [1]> var_6360_axes_0 = const()[name = string("op_6360_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6360_cast_fp16 = reduce_max(axes = var_6360_axes_0, keep_dims = var_6359, x = var_6356_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_205_cast_fp16 = maximum(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_205_cast_fp16 = real_div(x = x_227_cast_fp16, y = s_205_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6364_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<int32, [1]> var_6369_axes_0 = const()[name = string("op_6369_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6369_keep_dims_0 = const()[name = string("op_6369_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6369_cast_fp16 = reduce_mean(axes = var_6369_axes_0, keep_dims = var_6369_keep_dims_0, x = var_6364_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6370_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6370_cast_fp16")];
            fp32 var_6371_epsilon_0 = const()[name = string("op_6371_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6371_cast_fp16 = inverse(epsilon = var_6371_epsilon_0, x = var_6370_cast_fp16)[name = string("op_6371_cast_fp16")];
            fp16 var_6372_to_fp16 = const()[name = string("op_6372_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6373_cast_fp16 = mul(x = var_6371_cast_fp16, y = var_6372_to_fp16)[name = string("op_6373_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6375_cast_fp16 = add(x = var_6369_cast_fp16, y = var_6373_cast_fp16)[name = string("op_6375_cast_fp16")];
            fp32 var_6376_epsilon_0 = const()[name = string("op_6376_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6376_cast_fp16 = rsqrt(epsilon = var_6376_epsilon_0, x = var_6375_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_179_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6376_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73236736)))];
            tensor<fp16, [1, 64, 512]> input_149_cast_fp16 = mul(x = out_179_cast_fp16, y = tm_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73237824))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74286464))))[name = string("tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_113_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_113_cast_fp16")];
            string var_6383_mode_0 = const()[name = string("op_6383_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_6383_cast_fp16 = gelu(mode = var_6383_mode_0, x = linear_113_cast_fp16)[name = string("op_6383_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74290624))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339264))))[name = string("tm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_114_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_151_cast_fp16 = mul(x = var_6383_cast_fp16, y = linear_114_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75343424))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76392064))))[name = string("tm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6391_cast_fp16 = abs(x = linear_115_cast_fp16)[name = string("op_6391_cast_fp16")];
            bool var_6394 = const()[name = string("op_6394"), val = bool(true)];
            tensor<int32, [1]> var_6395_axes_0 = const()[name = string("op_6395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6395_cast_fp16 = reduce_max(axes = var_6395_axes_0, keep_dims = var_6394, x = var_6391_cast_fp16)[name = string("op_6395_cast_fp16")];
            fp16 var_6396_to_fp16 = const()[name = string("op_6396_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_207_cast_fp16 = maximum(x = var_6395_cast_fp16, y = var_6396_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_207_cast_fp16 = real_div(x = linear_115_cast_fp16, y = s_207_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6399_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6399_cast_fp16")];
            tensor<int32, [1]> var_6404_axes_0 = const()[name = string("op_6404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6404_keep_dims_0 = const()[name = string("op_6404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6404_cast_fp16 = reduce_mean(axes = var_6404_axes_0, keep_dims = var_6404_keep_dims_0, x = var_6399_cast_fp16)[name = string("op_6404_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6405_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6405_cast_fp16")];
            fp32 var_6406_epsilon_0 = const()[name = string("op_6406_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6406_cast_fp16 = inverse(epsilon = var_6406_epsilon_0, x = var_6405_cast_fp16)[name = string("op_6406_cast_fp16")];
            fp16 var_6407_to_fp16 = const()[name = string("op_6407_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6408_cast_fp16 = mul(x = var_6406_cast_fp16, y = var_6407_to_fp16)[name = string("op_6408_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6410_cast_fp16 = add(x = var_6404_cast_fp16, y = var_6408_cast_fp16)[name = string("op_6410_cast_fp16")];
            fp32 var_6411_epsilon_0 = const()[name = string("op_6411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6411_cast_fp16 = rsqrt(epsilon = var_6411_epsilon_0, x = var_6410_cast_fp16)[name = string("op_6411_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_181_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6411_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76393152)))];
            tensor<fp16, [1, 64, 512]> var_6413_cast_fp16 = mul(x = out_181_cast_fp16, y = tm_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6413_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_153_cast_fp16 = add(x = x_227_cast_fp16, y = var_6413_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76394240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76656448))))[name = string("tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_116_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_116_cast_fp16")];
            string x_231_mode_0 = const()[name = string("x_231_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_231_cast_fp16 = gelu(mode = x_231_mode_0, x = linear_116_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_6433_begin_0 = const()[name = string("op_6433_begin_0"), val = tensor<int32, [4]>([0, 0, 12, 0])];
            tensor<int32, [4]> var_6433_end_0 = const()[name = string("op_6433_end_0"), val = tensor<int32, [4]>([1, 64, 13, 512])];
            tensor<bool, [4]> var_6433_end_mask_0 = const()[name = string("op_6433_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6433_squeeze_mask_0 = const()[name = string("op_6433_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_6433_cast_fp16 = slice_by_index(begin = var_6433_begin_0, end = var_6433_end_0, end_mask = var_6433_end_mask_0, squeeze_mask = var_6433_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6433_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_155_cast_fp16 = mul(x = x_231_cast_fp16, y = var_6433_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76657536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76919744))))[name = string("tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6443_cast_fp16 = abs(x = linear_117_cast_fp16)[name = string("op_6443_cast_fp16")];
            bool var_6446 = const()[name = string("op_6446"), val = bool(true)];
            tensor<int32, [1]> var_6447_axes_0 = const()[name = string("op_6447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6447_cast_fp16 = reduce_max(axes = var_6447_axes_0, keep_dims = var_6446, x = var_6443_cast_fp16)[name = string("op_6447_cast_fp16")];
            fp16 var_6448_to_fp16 = const()[name = string("op_6448_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_209_cast_fp16 = maximum(x = var_6447_cast_fp16, y = var_6448_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_209_cast_fp16 = real_div(x = linear_117_cast_fp16, y = s_209_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6451_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6451_cast_fp16")];
            tensor<int32, [1]> var_6456_axes_0 = const()[name = string("op_6456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6456_keep_dims_0 = const()[name = string("op_6456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6456_cast_fp16 = reduce_mean(axes = var_6456_axes_0, keep_dims = var_6456_keep_dims_0, x = var_6451_cast_fp16)[name = string("op_6456_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6457_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6457_cast_fp16")];
            fp32 var_6458_epsilon_0 = const()[name = string("op_6458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6458_cast_fp16 = inverse(epsilon = var_6458_epsilon_0, x = var_6457_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6459_to_fp16 = const()[name = string("op_6459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6460_cast_fp16 = mul(x = var_6458_cast_fp16, y = var_6459_to_fp16)[name = string("op_6460_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6462_cast_fp16 = add(x = var_6456_cast_fp16, y = var_6460_cast_fp16)[name = string("op_6462_cast_fp16")];
            fp32 var_6463_epsilon_0 = const()[name = string("op_6463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6463_cast_fp16 = rsqrt(epsilon = var_6463_epsilon_0, x = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_183_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6463_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76920832)))];
            tensor<fp16, [1, 64, 512]> var_6465_cast_fp16 = mul(x = out_183_cast_fp16, y = tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6465_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_25_cast_fp16 = add(x = input_153_cast_fp16, y = var_6465_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<fp16, [1]> tm_layers_12_layer_scalar_to_fp16 = const()[name = string("tm_layers_12_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.8cp-1])];
            tensor<fp16, [1, 64, 512]> x_235_cast_fp16 = mul(x = h_25_cast_fp16, y = tm_layers_12_layer_scalar_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6469_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_6469_cast_fp16")];
            bool var_6472 = const()[name = string("op_6472"), val = bool(true)];
            tensor<int32, [1]> var_6473_axes_0 = const()[name = string("op_6473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6473_cast_fp16 = reduce_max(axes = var_6473_axes_0, keep_dims = var_6472, x = var_6469_cast_fp16)[name = string("op_6473_cast_fp16")];
            fp16 var_6474_to_fp16 = const()[name = string("op_6474_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_211_cast_fp16 = maximum(x = var_6473_cast_fp16, y = var_6474_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_211_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_211_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6477_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6477_cast_fp16")];
            tensor<int32, [1]> var_6482_axes_0 = const()[name = string("op_6482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6482_keep_dims_0 = const()[name = string("op_6482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6482_cast_fp16 = reduce_mean(axes = var_6482_axes_0, keep_dims = var_6482_keep_dims_0, x = var_6477_cast_fp16)[name = string("op_6482_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6483_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6483_cast_fp16")];
            fp32 var_6484_epsilon_0 = const()[name = string("op_6484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6484_cast_fp16 = inverse(epsilon = var_6484_epsilon_0, x = var_6483_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp16 var_6485_to_fp16 = const()[name = string("op_6485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6486_cast_fp16 = mul(x = var_6484_cast_fp16, y = var_6485_to_fp16)[name = string("op_6486_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6488_cast_fp16 = add(x = var_6482_cast_fp16, y = var_6486_cast_fp16)[name = string("op_6488_cast_fp16")];
            fp32 var_6489_epsilon_0 = const()[name = string("op_6489_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6489_cast_fp16 = rsqrt(epsilon = var_6489_epsilon_0, x = var_6488_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_185_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6489_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76921920)))];
            tensor<fp16, [1, 64, 512]> input_157_cast_fp16 = mul(x = out_185_cast_fp16, y = tm_layers_13_input_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76923008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77447360))))[name = string("tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<int32, [4]> var_6499 = const()[name = string("op_6499"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_6500_cast_fp16 = reshape(shape = var_6499, x = linear_118_cast_fp16)[name = string("op_6500_cast_fp16")];
            tensor<int32, [4]> x_237_perm_0 = const()[name = string("x_237_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77449472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77711680))))[name = string("tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6511 = const()[name = string("op_6511"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_6512_cast_fp16 = reshape(shape = var_6511, x = linear_119_cast_fp16)[name = string("op_6512_cast_fp16")];
            tensor<int32, [4]> x_239_perm_0 = const()[name = string("x_239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77712768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77974976))))[name = string("tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6523 = const()[name = string("op_6523"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_6524_cast_fp16 = reshape(shape = var_6523, x = linear_120_cast_fp16)[name = string("op_6524_cast_fp16")];
            tensor<int32, [4]> x_241_perm_0 = const()[name = string("x_241_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_237_cast_fp16 = transpose(perm = x_237_perm_0, x = var_6500_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 4, 64, 256]> var_6528_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_6528_cast_fp16")];
            bool var_6531 = const()[name = string("op_6531"), val = bool(true)];
            tensor<int32, [1]> var_6532_axes_0 = const()[name = string("op_6532_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_6532_cast_fp16 = reduce_max(axes = var_6532_axes_0, keep_dims = var_6531, x = var_6528_cast_fp16)[name = string("op_6532_cast_fp16")];
            fp16 var_6533_to_fp16 = const()[name = string("op_6533_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_213_cast_fp16 = maximum(x = var_6532_cast_fp16, y = var_6533_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_213_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_213_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6536_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6536_cast_fp16")];
            tensor<int32, [1]> var_6541_axes_0 = const()[name = string("op_6541_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6541_keep_dims_0 = const()[name = string("op_6541_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_6541_cast_fp16 = reduce_mean(axes = var_6541_axes_0, keep_dims = var_6541_keep_dims_0, x = var_6536_cast_fp16)[name = string("op_6541_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_6542_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp32 var_6543_epsilon_0 = const()[name = string("op_6543_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_6543_cast_fp16 = inverse(epsilon = var_6543_epsilon_0, x = var_6542_cast_fp16)[name = string("op_6543_cast_fp16")];
            fp16 var_6544_to_fp16 = const()[name = string("op_6544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_6545_cast_fp16 = mul(x = var_6543_cast_fp16, y = var_6544_to_fp16)[name = string("op_6545_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_6547_cast_fp16 = add(x = var_6541_cast_fp16, y = var_6545_cast_fp16)[name = string("op_6547_cast_fp16")];
            fp32 var_6548_epsilon_0 = const()[name = string("op_6548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_6548_cast_fp16 = rsqrt(epsilon = var_6548_epsilon_0, x = var_6547_cast_fp16)[name = string("op_6548_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_187_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6548_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976064)))];
            tensor<fp16, [1, 4, 64, 256]> q_53_cast_fp16 = mul(x = out_187_cast_fp16, y = tm_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("q_53_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_239_cast_fp16 = transpose(perm = x_239_perm_0, x = var_6512_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2, 64, 256]> var_6551_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_6551_cast_fp16")];
            bool var_6554 = const()[name = string("op_6554"), val = bool(true)];
            tensor<int32, [1]> var_6555_axes_0 = const()[name = string("op_6555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_6555_cast_fp16 = reduce_max(axes = var_6555_axes_0, keep_dims = var_6554, x = var_6551_cast_fp16)[name = string("op_6555_cast_fp16")];
            fp16 var_6556_to_fp16 = const()[name = string("op_6556_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_215_cast_fp16 = maximum(x = var_6555_cast_fp16, y = var_6556_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_215_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_215_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6559_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6559_cast_fp16")];
            tensor<int32, [1]> var_6564_axes_0 = const()[name = string("op_6564_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6564_keep_dims_0 = const()[name = string("op_6564_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_6564_cast_fp16 = reduce_mean(axes = var_6564_axes_0, keep_dims = var_6564_keep_dims_0, x = var_6559_cast_fp16)[name = string("op_6564_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6565_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6565_cast_fp16")];
            fp32 var_6566_epsilon_0 = const()[name = string("op_6566_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_6566_cast_fp16 = inverse(epsilon = var_6566_epsilon_0, x = var_6565_cast_fp16)[name = string("op_6566_cast_fp16")];
            fp16 var_6567_to_fp16 = const()[name = string("op_6567_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_6568_cast_fp16 = mul(x = var_6566_cast_fp16, y = var_6567_to_fp16)[name = string("op_6568_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6570_cast_fp16 = add(x = var_6564_cast_fp16, y = var_6568_cast_fp16)[name = string("op_6570_cast_fp16")];
            fp32 var_6571_epsilon_0 = const()[name = string("op_6571_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_6571_cast_fp16 = rsqrt(epsilon = var_6571_epsilon_0, x = var_6570_cast_fp16)[name = string("op_6571_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_189_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6571_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976640)))];
            tensor<fp16, [1, 2, 64, 256]> k_79_cast_fp16 = mul(x = out_189_cast_fp16, y = tm_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("k_79_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_241_cast_fp16 = transpose(perm = x_241_perm_0, x = var_6524_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 2, 64, 256]> var_6574_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_6574_cast_fp16")];
            bool var_6577 = const()[name = string("op_6577"), val = bool(true)];
            tensor<int32, [1]> var_6578_axes_0 = const()[name = string("op_6578_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_6578_cast_fp16 = reduce_max(axes = var_6578_axes_0, keep_dims = var_6577, x = var_6574_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_217_cast_fp16 = maximum(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_217_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_217_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6582_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6587_axes_0 = const()[name = string("op_6587_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6587_keep_dims_0 = const()[name = string("op_6587_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_6587_cast_fp16 = reduce_mean(axes = var_6587_axes_0, keep_dims = var_6587_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6588_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6588_cast_fp16")];
            fp32 var_6589_epsilon_0 = const()[name = string("op_6589_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_6589_cast_fp16 = inverse(epsilon = var_6589_epsilon_0, x = var_6588_cast_fp16)[name = string("op_6589_cast_fp16")];
            fp16 var_6590_to_fp16 = const()[name = string("op_6590_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_6591_cast_fp16 = mul(x = var_6589_cast_fp16, y = var_6590_to_fp16)[name = string("op_6591_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6593_cast_fp16 = add(x = var_6587_cast_fp16, y = var_6591_cast_fp16)[name = string("op_6593_cast_fp16")];
            fp32 var_6594_epsilon_0 = const()[name = string("op_6594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_6594_cast_fp16 = rsqrt(epsilon = var_6594_epsilon_0, x = var_6593_cast_fp16)[name = string("op_6594_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_53_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6594_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6596_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6596_cast_fp16")];
            tensor<int32, [4]> var_6601_begin_0 = const()[name = string("op_6601_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6601_end_0 = const()[name = string("op_6601_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_6601_end_mask_0 = const()[name = string("op_6601_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_6601_cast_fp16 = slice_by_index(begin = var_6601_begin_0, end = var_6601_end_0, end_mask = var_6601_end_mask_0, x = q_53_cast_fp16)[name = string("op_6601_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_6602_cast_fp16 = mul(x = var_6601_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_6602_cast_fp16")];
            tensor<int32, [4]> var_6607_begin_0 = const()[name = string("op_6607_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6607_end_0 = const()[name = string("op_6607_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_6607_end_mask_0 = const()[name = string("op_6607_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_6607_cast_fp16 = slice_by_index(begin = var_6607_begin_0, end = var_6607_end_0, end_mask = var_6607_end_mask_0, x = q_53_cast_fp16)[name = string("op_6607_cast_fp16")];
            int32 var_6609 = const()[name = string("op_6609"), val = int32(-1)];
            bool var_6610_interleave_0 = const()[name = string("op_6610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_6610_cast_fp16 = concat(axis = var_6609, interleave = var_6610_interleave_0, values = (var_6602_cast_fp16, var_6607_cast_fp16))[name = string("op_6610_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6611_cast_fp16 = mul(x = var_6610_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_55_cast_fp16 = add(x = var_6596_cast_fp16, y = var_6611_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6614_cast_fp16 = mul(x = k_79_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = k_79_cast_fp16)[name = string("op_6619_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_6620_cast_fp16 = mul(x = var_6619_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_6620_cast_fp16")];
            tensor<int32, [4]> var_6625_begin_0 = const()[name = string("op_6625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6625_end_0 = const()[name = string("op_6625_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_6625_end_mask_0 = const()[name = string("op_6625_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_6625_cast_fp16 = slice_by_index(begin = var_6625_begin_0, end = var_6625_end_0, end_mask = var_6625_end_mask_0, x = k_79_cast_fp16)[name = string("op_6625_cast_fp16")];
            int32 var_6627 = const()[name = string("op_6627"), val = int32(-1)];
            bool var_6628_interleave_0 = const()[name = string("op_6628_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_6628_cast_fp16 = concat(axis = var_6627, interleave = var_6628_interleave_0, values = (var_6620_cast_fp16, var_6625_cast_fp16))[name = string("op_6628_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6629_cast_fp16 = mul(x = var_6628_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6629_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_81_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6629_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_26_reps_0 = const()[name = string("tile_26_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = k_81_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [4, 1, 64, 256]> tile_26_cast_fp16 = tile(reps = tile_26_reps_0, x = transpose_52_cast_fp16)[name = string("tile_26_cast_fp16")];
            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_52_cast_fp16 = reshape(shape = concat_52, x = tile_26_cast_fp16)[name = string("reshape_52_cast_fp16")];
            tensor<int32, [5]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = reshape_52_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [4, 1, 64, 256]> reshape_53_cast_fp16 = reshape(shape = concat_53, x = transpose_53_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_27_reps_0 = const()[name = string("tile_27_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = v_53_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [4, 1, 64, 256]> tile_27_cast_fp16 = tile(reps = tile_27_reps_0, x = transpose_54_cast_fp16)[name = string("tile_27_cast_fp16")];
            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_54_cast_fp16 = reshape(shape = concat_54, x = tile_27_cast_fp16)[name = string("reshape_54_cast_fp16")];
            tensor<int32, [5]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = reshape_54_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [4, 1, 64, 256]> reshape_55_cast_fp16 = reshape(shape = concat_55, x = transpose_55_cast_fp16)[name = string("reshape_55_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6643_transpose_x_0 = const()[name = string("op_6643_transpose_x_0"), val = bool(false)];
            bool var_6643_transpose_y_0 = const()[name = string("op_6643_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 4, 64, 64]> var_6643_cast_fp16 = matmul(transpose_x = var_6643_transpose_x_0, transpose_y = var_6643_transpose_y_0, x = q_55_cast_fp16, y = transpose_109_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_6645_cast_fp16 = add(x = var_6643_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6646 = const()[name = string("op_6646"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_217_cast_fp16 = softmax(axis = var_6646, x = var_6645_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6649_transpose_x_0 = const()[name = string("op_6649_transpose_x_0"), val = bool(false)];
            bool var_6649_transpose_y_0 = const()[name = string("op_6649_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = reshape_55_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 4, 64, 256]> var_6649_cast_fp16 = matmul(transpose_x = var_6649_transpose_x_0, transpose_y = var_6649_transpose_y_0, x = w_217_cast_fp16, y = v_55_cast_fp16)[name = string("op_6649_cast_fp16")];
            tensor<int32, [4]> var_6652_perm_0 = const()[name = string("op_6652_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6656 = const()[name = string("op_6656"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_6652_cast_fp16 = transpose(perm = var_6652_perm_0, x = var_6649_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 64, 1024]> input_159_cast_fp16 = reshape(shape = var_6656, x = var_6652_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77977216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78501568))))[name = string("tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6661_cast_fp16 = abs(x = linear_121_cast_fp16)[name = string("op_6661_cast_fp16")];
            bool var_6664 = const()[name = string("op_6664"), val = bool(true)];
            tensor<int32, [1]> var_6665_axes_0 = const()[name = string("op_6665_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6665_cast_fp16 = reduce_max(axes = var_6665_axes_0, keep_dims = var_6664, x = var_6661_cast_fp16)[name = string("op_6665_cast_fp16")];
            fp16 var_6666_to_fp16 = const()[name = string("op_6666_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_219_cast_fp16 = maximum(x = var_6665_cast_fp16, y = var_6666_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_219_cast_fp16 = real_div(x = linear_121_cast_fp16, y = s_219_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6669_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6669_cast_fp16")];
            tensor<int32, [1]> var_6674_axes_0 = const()[name = string("op_6674_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6674_keep_dims_0 = const()[name = string("op_6674_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6674_cast_fp16 = reduce_mean(axes = var_6674_axes_0, keep_dims = var_6674_keep_dims_0, x = var_6669_cast_fp16)[name = string("op_6674_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6675_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6675_cast_fp16")];
            fp32 var_6676_epsilon_0 = const()[name = string("op_6676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6676_cast_fp16 = inverse(epsilon = var_6676_epsilon_0, x = var_6675_cast_fp16)[name = string("op_6676_cast_fp16")];
            fp16 var_6677_to_fp16 = const()[name = string("op_6677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6678_cast_fp16 = mul(x = var_6676_cast_fp16, y = var_6677_to_fp16)[name = string("op_6678_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6680_cast_fp16 = add(x = var_6674_cast_fp16, y = var_6678_cast_fp16)[name = string("op_6680_cast_fp16")];
            fp32 var_6681_epsilon_0 = const()[name = string("op_6681_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6681_cast_fp16 = rsqrt(epsilon = var_6681_epsilon_0, x = var_6680_cast_fp16)[name = string("op_6681_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_191_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6681_cast_fp16)[name = string("out_191_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78502656)))];
            tensor<fp16, [1, 64, 512]> var_6683_cast_fp16 = mul(x = out_191_cast_fp16, y = tm_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6683_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_245_cast_fp16 = add(x = x_235_cast_fp16, y = var_6683_cast_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6686_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_6686_cast_fp16")];
            bool var_6689 = const()[name = string("op_6689"), val = bool(true)];
            tensor<int32, [1]> var_6690_axes_0 = const()[name = string("op_6690_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6690_cast_fp16 = reduce_max(axes = var_6690_axes_0, keep_dims = var_6689, x = var_6686_cast_fp16)[name = string("op_6690_cast_fp16")];
            fp16 var_6691_to_fp16 = const()[name = string("op_6691_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_221_cast_fp16 = maximum(x = var_6690_cast_fp16, y = var_6691_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_221_cast_fp16 = real_div(x = x_245_cast_fp16, y = s_221_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6694_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6694_cast_fp16")];
            tensor<int32, [1]> var_6699_axes_0 = const()[name = string("op_6699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6699_keep_dims_0 = const()[name = string("op_6699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6699_cast_fp16 = reduce_mean(axes = var_6699_axes_0, keep_dims = var_6699_keep_dims_0, x = var_6694_cast_fp16)[name = string("op_6699_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6700_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6700_cast_fp16")];
            fp32 var_6701_epsilon_0 = const()[name = string("op_6701_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6701_cast_fp16 = inverse(epsilon = var_6701_epsilon_0, x = var_6700_cast_fp16)[name = string("op_6701_cast_fp16")];
            fp16 var_6702_to_fp16 = const()[name = string("op_6702_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6703_cast_fp16 = mul(x = var_6701_cast_fp16, y = var_6702_to_fp16)[name = string("op_6703_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6705_cast_fp16 = add(x = var_6699_cast_fp16, y = var_6703_cast_fp16)[name = string("op_6705_cast_fp16")];
            fp32 var_6706_epsilon_0 = const()[name = string("op_6706_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6706_cast_fp16 = rsqrt(epsilon = var_6706_epsilon_0, x = var_6705_cast_fp16)[name = string("op_6706_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_193_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6706_cast_fp16)[name = string("out_193_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78503744)))];
            tensor<fp16, [1, 64, 512]> input_161_cast_fp16 = mul(x = out_193_cast_fp16, y = tm_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78504832))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79553472))))[name = string("tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_122_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_122_cast_fp16")];
            string var_6713_mode_0 = const()[name = string("op_6713_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_6713_cast_fp16 = gelu(mode = var_6713_mode_0, x = linear_122_cast_fp16)[name = string("op_6713_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79557632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80606272))))[name = string("tm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_123_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_163_cast_fp16 = mul(x = var_6713_cast_fp16, y = linear_123_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80610432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81659072))))[name = string("tm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6721_cast_fp16 = abs(x = linear_124_cast_fp16)[name = string("op_6721_cast_fp16")];
            bool var_6724 = const()[name = string("op_6724"), val = bool(true)];
            tensor<int32, [1]> var_6725_axes_0 = const()[name = string("op_6725_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6725_cast_fp16 = reduce_max(axes = var_6725_axes_0, keep_dims = var_6724, x = var_6721_cast_fp16)[name = string("op_6725_cast_fp16")];
            fp16 var_6726_to_fp16 = const()[name = string("op_6726_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_223_cast_fp16 = maximum(x = var_6725_cast_fp16, y = var_6726_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_223_cast_fp16 = real_div(x = linear_124_cast_fp16, y = s_223_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6729_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [1]> var_6734_axes_0 = const()[name = string("op_6734_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6734_keep_dims_0 = const()[name = string("op_6734_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6734_cast_fp16 = reduce_mean(axes = var_6734_axes_0, keep_dims = var_6734_keep_dims_0, x = var_6729_cast_fp16)[name = string("op_6734_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6735_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6735_cast_fp16")];
            fp32 var_6736_epsilon_0 = const()[name = string("op_6736_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6736_cast_fp16 = inverse(epsilon = var_6736_epsilon_0, x = var_6735_cast_fp16)[name = string("op_6736_cast_fp16")];
            fp16 var_6737_to_fp16 = const()[name = string("op_6737_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6738_cast_fp16 = mul(x = var_6736_cast_fp16, y = var_6737_to_fp16)[name = string("op_6738_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6740_cast_fp16 = add(x = var_6734_cast_fp16, y = var_6738_cast_fp16)[name = string("op_6740_cast_fp16")];
            fp32 var_6741_epsilon_0 = const()[name = string("op_6741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6741_cast_fp16 = rsqrt(epsilon = var_6741_epsilon_0, x = var_6740_cast_fp16)[name = string("op_6741_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_195_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6741_cast_fp16)[name = string("out_195_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81660160)))];
            tensor<fp16, [1, 64, 512]> var_6743_cast_fp16 = mul(x = out_195_cast_fp16, y = tm_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6743_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_165_cast_fp16 = add(x = x_245_cast_fp16, y = var_6743_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81661248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81923456))))[name = string("tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_125_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_125_cast_fp16")];
            string x_249_mode_0 = const()[name = string("x_249_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_249_cast_fp16 = gelu(mode = x_249_mode_0, x = linear_125_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<int32, [4]> var_6763_begin_0 = const()[name = string("op_6763_begin_0"), val = tensor<int32, [4]>([0, 0, 13, 0])];
            tensor<int32, [4]> var_6763_end_0 = const()[name = string("op_6763_end_0"), val = tensor<int32, [4]>([1, 64, 14, 512])];
            tensor<bool, [4]> var_6763_end_mask_0 = const()[name = string("op_6763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6763_squeeze_mask_0 = const()[name = string("op_6763_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_6763_cast_fp16 = slice_by_index(begin = var_6763_begin_0, end = var_6763_end_0, end_mask = var_6763_end_mask_0, squeeze_mask = var_6763_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_167_cast_fp16 = mul(x = x_249_cast_fp16, y = var_6763_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81924544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82186752))))[name = string("tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6773_cast_fp16 = abs(x = linear_126_cast_fp16)[name = string("op_6773_cast_fp16")];
            bool var_6776 = const()[name = string("op_6776"), val = bool(true)];
            tensor<int32, [1]> var_6777_axes_0 = const()[name = string("op_6777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6777_cast_fp16 = reduce_max(axes = var_6777_axes_0, keep_dims = var_6776, x = var_6773_cast_fp16)[name = string("op_6777_cast_fp16")];
            fp16 var_6778_to_fp16 = const()[name = string("op_6778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_225_cast_fp16 = maximum(x = var_6777_cast_fp16, y = var_6778_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_225_cast_fp16 = real_div(x = linear_126_cast_fp16, y = s_225_cast_fp16)[name = string("y_225_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6781_cast_fp16 = mul(x = y_225_cast_fp16, y = y_225_cast_fp16)[name = string("op_6781_cast_fp16")];
            tensor<int32, [1]> var_6786_axes_0 = const()[name = string("op_6786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6786_keep_dims_0 = const()[name = string("op_6786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6786_cast_fp16 = reduce_mean(axes = var_6786_axes_0, keep_dims = var_6786_keep_dims_0, x = var_6781_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6787_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6787_cast_fp16")];
            fp32 var_6788_epsilon_0 = const()[name = string("op_6788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6788_cast_fp16 = inverse(epsilon = var_6788_epsilon_0, x = var_6787_cast_fp16)[name = string("op_6788_cast_fp16")];
            fp16 var_6789_to_fp16 = const()[name = string("op_6789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6790_cast_fp16 = mul(x = var_6788_cast_fp16, y = var_6789_to_fp16)[name = string("op_6790_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6792_cast_fp16 = add(x = var_6786_cast_fp16, y = var_6790_cast_fp16)[name = string("op_6792_cast_fp16")];
            fp32 var_6793_epsilon_0 = const()[name = string("op_6793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6793_cast_fp16 = rsqrt(epsilon = var_6793_epsilon_0, x = var_6792_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_197_cast_fp16 = mul(x = y_225_cast_fp16, y = var_6793_cast_fp16)[name = string("out_197_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82187840)))];
            tensor<fp16, [1, 64, 512]> var_6795_cast_fp16 = mul(x = out_197_cast_fp16, y = tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6795_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_27_cast_fp16 = add(x = input_165_cast_fp16, y = var_6795_cast_fp16)[name = string("h_27_cast_fp16")];
            tensor<fp16, [1]> tm_layers_13_layer_scalar_to_fp16 = const()[name = string("tm_layers_13_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9cp-1])];
            tensor<fp16, [1, 64, 512]> x_253_cast_fp16 = mul(x = h_27_cast_fp16, y = tm_layers_13_layer_scalar_to_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6799_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6799_cast_fp16")];
            bool var_6802 = const()[name = string("op_6802"), val = bool(true)];
            tensor<int32, [1]> var_6803_axes_0 = const()[name = string("op_6803_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6803_cast_fp16 = reduce_max(axes = var_6803_axes_0, keep_dims = var_6802, x = var_6799_cast_fp16)[name = string("op_6803_cast_fp16")];
            fp16 var_6804_to_fp16 = const()[name = string("op_6804_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_227_cast_fp16 = maximum(x = var_6803_cast_fp16, y = var_6804_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_227_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_227_cast_fp16)[name = string("y_227_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6807_cast_fp16 = mul(x = y_227_cast_fp16, y = y_227_cast_fp16)[name = string("op_6807_cast_fp16")];
            tensor<int32, [1]> var_6812_axes_0 = const()[name = string("op_6812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6812_keep_dims_0 = const()[name = string("op_6812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_6812_cast_fp16 = reduce_mean(axes = var_6812_axes_0, keep_dims = var_6812_keep_dims_0, x = var_6807_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6813_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6813_cast_fp16")];
            fp32 var_6814_epsilon_0 = const()[name = string("op_6814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_6814_cast_fp16 = inverse(epsilon = var_6814_epsilon_0, x = var_6813_cast_fp16)[name = string("op_6814_cast_fp16")];
            fp16 var_6815_to_fp16 = const()[name = string("op_6815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_6816_cast_fp16 = mul(x = var_6814_cast_fp16, y = var_6815_to_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_6818_cast_fp16 = add(x = var_6812_cast_fp16, y = var_6816_cast_fp16)[name = string("op_6818_cast_fp16")];
            fp32 var_6819_epsilon_0 = const()[name = string("op_6819_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_6819_cast_fp16 = rsqrt(epsilon = var_6819_epsilon_0, x = var_6818_cast_fp16)[name = string("op_6819_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_199_cast_fp16 = mul(x = y_227_cast_fp16, y = var_6819_cast_fp16)[name = string("out_199_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82188928)))];
            tensor<fp16, [1, 64, 512]> input_169_cast_fp16 = mul(x = out_199_cast_fp16, y = tm_layers_14_input_layernorm_weight_to_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82190016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82714368))))[name = string("tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6829 = const()[name = string("op_6829"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_6830_cast_fp16 = reshape(shape = var_6829, x = linear_127_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82716480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82978688))))[name = string("tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6841 = const()[name = string("op_6841"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_6842_cast_fp16 = reshape(shape = var_6841, x = linear_128_cast_fp16)[name = string("op_6842_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82979776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83241984))))[name = string("tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_6853 = const()[name = string("op_6853"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_6854_cast_fp16 = reshape(shape = var_6853, x = linear_129_cast_fp16)[name = string("op_6854_cast_fp16")];
            tensor<int32, [4]> x_259_perm_0 = const()[name = string("x_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_6830_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 4, 64, 256]> var_6858_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6858_cast_fp16")];
            bool var_6861 = const()[name = string("op_6861"), val = bool(true)];
            tensor<int32, [1]> var_6862_axes_0 = const()[name = string("op_6862_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_6862_cast_fp16 = reduce_max(axes = var_6862_axes_0, keep_dims = var_6861, x = var_6858_cast_fp16)[name = string("op_6862_cast_fp16")];
            fp16 var_6863_to_fp16 = const()[name = string("op_6863_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_229_cast_fp16 = maximum(x = var_6862_cast_fp16, y = var_6863_to_fp16)[name = string("s_229_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_229_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_229_cast_fp16)[name = string("y_229_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6866_cast_fp16 = mul(x = y_229_cast_fp16, y = y_229_cast_fp16)[name = string("op_6866_cast_fp16")];
            tensor<int32, [1]> var_6871_axes_0 = const()[name = string("op_6871_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6871_keep_dims_0 = const()[name = string("op_6871_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_6871_cast_fp16 = reduce_mean(axes = var_6871_axes_0, keep_dims = var_6871_keep_dims_0, x = var_6866_cast_fp16)[name = string("op_6871_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_6872_cast_fp16 = mul(x = s_229_cast_fp16, y = s_229_cast_fp16)[name = string("op_6872_cast_fp16")];
            fp32 var_6873_epsilon_0 = const()[name = string("op_6873_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_6873_cast_fp16 = inverse(epsilon = var_6873_epsilon_0, x = var_6872_cast_fp16)[name = string("op_6873_cast_fp16")];
            fp16 var_6874_to_fp16 = const()[name = string("op_6874_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_6875_cast_fp16 = mul(x = var_6873_cast_fp16, y = var_6874_to_fp16)[name = string("op_6875_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_6877_cast_fp16 = add(x = var_6871_cast_fp16, y = var_6875_cast_fp16)[name = string("op_6877_cast_fp16")];
            fp32 var_6878_epsilon_0 = const()[name = string("op_6878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_6878_cast_fp16 = rsqrt(epsilon = var_6878_epsilon_0, x = var_6877_cast_fp16)[name = string("op_6878_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_201_cast_fp16 = mul(x = y_229_cast_fp16, y = var_6878_cast_fp16)[name = string("out_201_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243072)))];
            tensor<fp16, [1, 4, 64, 256]> q_57_cast_fp16 = mul(x = out_201_cast_fp16, y = tm_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("q_57_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_6842_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, 64, 256]> var_6881_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6884 = const()[name = string("op_6884"), val = bool(true)];
            tensor<int32, [1]> var_6885_axes_0 = const()[name = string("op_6885_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_6885_cast_fp16 = reduce_max(axes = var_6885_axes_0, keep_dims = var_6884, x = var_6881_cast_fp16)[name = string("op_6885_cast_fp16")];
            fp16 var_6886_to_fp16 = const()[name = string("op_6886_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_231_cast_fp16 = maximum(x = var_6885_cast_fp16, y = var_6886_to_fp16)[name = string("s_231_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_231_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_231_cast_fp16)[name = string("y_231_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6889_cast_fp16 = mul(x = y_231_cast_fp16, y = y_231_cast_fp16)[name = string("op_6889_cast_fp16")];
            tensor<int32, [1]> var_6894_axes_0 = const()[name = string("op_6894_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6894_keep_dims_0 = const()[name = string("op_6894_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_6894_cast_fp16 = reduce_mean(axes = var_6894_axes_0, keep_dims = var_6894_keep_dims_0, x = var_6889_cast_fp16)[name = string("op_6894_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6895_cast_fp16 = mul(x = s_231_cast_fp16, y = s_231_cast_fp16)[name = string("op_6895_cast_fp16")];
            fp32 var_6896_epsilon_0 = const()[name = string("op_6896_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_6896_cast_fp16 = inverse(epsilon = var_6896_epsilon_0, x = var_6895_cast_fp16)[name = string("op_6896_cast_fp16")];
            fp16 var_6897_to_fp16 = const()[name = string("op_6897_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_6898_cast_fp16 = mul(x = var_6896_cast_fp16, y = var_6897_to_fp16)[name = string("op_6898_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6900_cast_fp16 = add(x = var_6894_cast_fp16, y = var_6898_cast_fp16)[name = string("op_6900_cast_fp16")];
            fp32 var_6901_epsilon_0 = const()[name = string("op_6901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_6901_cast_fp16 = rsqrt(epsilon = var_6901_epsilon_0, x = var_6900_cast_fp16)[name = string("op_6901_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_203_cast_fp16 = mul(x = y_231_cast_fp16, y = var_6901_cast_fp16)[name = string("out_203_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243648)))];
            tensor<fp16, [1, 2, 64, 256]> k_85_cast_fp16 = mul(x = out_203_cast_fp16, y = tm_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("k_85_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_259_cast_fp16 = transpose(perm = x_259_perm_0, x = var_6854_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 2, 64, 256]> var_6904_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6904_cast_fp16")];
            bool var_6907 = const()[name = string("op_6907"), val = bool(true)];
            tensor<int32, [1]> var_6908_axes_0 = const()[name = string("op_6908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_6908_cast_fp16 = reduce_max(axes = var_6908_axes_0, keep_dims = var_6907, x = var_6904_cast_fp16)[name = string("op_6908_cast_fp16")];
            fp16 var_6909_to_fp16 = const()[name = string("op_6909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_233_cast_fp16 = maximum(x = var_6908_cast_fp16, y = var_6909_to_fp16)[name = string("s_233_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_233_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_233_cast_fp16)[name = string("y_233_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6912_cast_fp16 = mul(x = y_233_cast_fp16, y = y_233_cast_fp16)[name = string("op_6912_cast_fp16")];
            tensor<int32, [1]> var_6917_axes_0 = const()[name = string("op_6917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6917_keep_dims_0 = const()[name = string("op_6917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_6917_cast_fp16 = reduce_mean(axes = var_6917_axes_0, keep_dims = var_6917_keep_dims_0, x = var_6912_cast_fp16)[name = string("op_6917_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6918_cast_fp16 = mul(x = s_233_cast_fp16, y = s_233_cast_fp16)[name = string("op_6918_cast_fp16")];
            fp32 var_6919_epsilon_0 = const()[name = string("op_6919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_6919_cast_fp16 = inverse(epsilon = var_6919_epsilon_0, x = var_6918_cast_fp16)[name = string("op_6919_cast_fp16")];
            fp16 var_6920_to_fp16 = const()[name = string("op_6920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_6921_cast_fp16 = mul(x = var_6919_cast_fp16, y = var_6920_to_fp16)[name = string("op_6921_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_6923_cast_fp16 = add(x = var_6917_cast_fp16, y = var_6921_cast_fp16)[name = string("op_6923_cast_fp16")];
            fp32 var_6924_epsilon_0 = const()[name = string("op_6924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_6924_cast_fp16 = rsqrt(epsilon = var_6924_epsilon_0, x = var_6923_cast_fp16)[name = string("op_6924_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_57_cast_fp16 = mul(x = y_233_cast_fp16, y = var_6924_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6926_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6926_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = q_57_cast_fp16)[name = string("op_6931_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_6932_cast_fp16 = mul(x = var_6931_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_6932_cast_fp16")];
            tensor<int32, [4]> var_6937_begin_0 = const()[name = string("op_6937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6937_end_0 = const()[name = string("op_6937_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_6937_end_mask_0 = const()[name = string("op_6937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_6937_cast_fp16 = slice_by_index(begin = var_6937_begin_0, end = var_6937_end_0, end_mask = var_6937_end_mask_0, x = q_57_cast_fp16)[name = string("op_6937_cast_fp16")];
            int32 var_6939 = const()[name = string("op_6939"), val = int32(-1)];
            bool var_6940_interleave_0 = const()[name = string("op_6940_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_6940_cast_fp16 = concat(axis = var_6939, interleave = var_6940_interleave_0, values = (var_6932_cast_fp16, var_6937_cast_fp16))[name = string("op_6940_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_6941_cast_fp16 = mul(x = var_6940_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6941_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_59_cast_fp16 = add(x = var_6926_cast_fp16, y = var_6941_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6944_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_6944_cast_fp16")];
            tensor<int32, [4]> var_6949_begin_0 = const()[name = string("op_6949_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6949_end_0 = const()[name = string("op_6949_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_6949_end_mask_0 = const()[name = string("op_6949_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_6949_cast_fp16 = slice_by_index(begin = var_6949_begin_0, end = var_6949_end_0, end_mask = var_6949_end_mask_0, x = k_85_cast_fp16)[name = string("op_6949_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_6950_cast_fp16 = mul(x = var_6949_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_6950_cast_fp16")];
            tensor<int32, [4]> var_6955_begin_0 = const()[name = string("op_6955_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6955_end_0 = const()[name = string("op_6955_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_6955_end_mask_0 = const()[name = string("op_6955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_6955_cast_fp16 = slice_by_index(begin = var_6955_begin_0, end = var_6955_end_0, end_mask = var_6955_end_mask_0, x = k_85_cast_fp16)[name = string("op_6955_cast_fp16")];
            int32 var_6957 = const()[name = string("op_6957"), val = int32(-1)];
            bool var_6958_interleave_0 = const()[name = string("op_6958_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_6958_cast_fp16 = concat(axis = var_6957, interleave = var_6958_interleave_0, values = (var_6950_cast_fp16, var_6955_cast_fp16))[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_6959_cast_fp16 = mul(x = var_6958_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_87_cast_fp16 = add(x = var_6944_cast_fp16, y = var_6959_cast_fp16)[name = string("k_87_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_28_reps_0 = const()[name = string("tile_28_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = k_87_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 1, 64, 256]> tile_28_cast_fp16 = tile(reps = tile_28_reps_0, x = transpose_56_cast_fp16)[name = string("tile_28_cast_fp16")];
            tensor<int32, [5]> concat_56 = const()[name = string("concat_56"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_56_cast_fp16 = reshape(shape = concat_56, x = tile_28_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [5]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_57 = const()[name = string("concat_57"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 1, 64, 256]> reshape_57_cast_fp16 = reshape(shape = concat_57, x = transpose_57_cast_fp16)[name = string("reshape_57_cast_fp16")];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_29_reps_0 = const()[name = string("tile_29_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = v_57_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 1, 64, 256]> tile_29_cast_fp16 = tile(reps = tile_29_reps_0, x = transpose_58_cast_fp16)[name = string("tile_29_cast_fp16")];
            tensor<int32, [5]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_58_cast_fp16 = reshape(shape = concat_58, x = tile_29_cast_fp16)[name = string("reshape_58_cast_fp16")];
            tensor<int32, [5]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = reshape_58_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 1, 64, 256]> reshape_59_cast_fp16 = reshape(shape = concat_59, x = transpose_59_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6973_transpose_x_0 = const()[name = string("op_6973_transpose_x_0"), val = bool(false)];
            bool var_6973_transpose_y_0 = const()[name = string("op_6973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_110_cast_fp16 = transpose(perm = transpose_110_perm_0, x = reshape_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 4, 64, 64]> var_6973_cast_fp16 = matmul(transpose_x = var_6973_transpose_x_0, transpose_y = var_6973_transpose_y_0, x = q_59_cast_fp16, y = transpose_110_cast_fp16)[name = string("op_6973_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_6975_cast_fp16 = add(x = var_6973_cast_fp16, y = var_2173_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6976 = const()[name = string("op_6976"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_233_cast_fp16 = softmax(axis = var_6976, x = var_6975_cast_fp16)[name = string("w_233_cast_fp16")];
            bool var_6979_transpose_x_0 = const()[name = string("op_6979_transpose_x_0"), val = bool(false)];
            bool var_6979_transpose_y_0 = const()[name = string("op_6979_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 4, 64, 256]> var_6979_cast_fp16 = matmul(transpose_x = var_6979_transpose_x_0, transpose_y = var_6979_transpose_y_0, x = w_233_cast_fp16, y = v_59_cast_fp16)[name = string("op_6979_cast_fp16")];
            tensor<int32, [4]> var_6982_perm_0 = const()[name = string("op_6982_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6986 = const()[name = string("op_6986"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_6982_cast_fp16 = transpose(perm = var_6982_perm_0, x = var_6979_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 64, 1024]> input_171_cast_fp16 = reshape(shape = var_6986, x = var_6982_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83244224))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83768576))))[name = string("tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6991_cast_fp16 = abs(x = linear_130_cast_fp16)[name = string("op_6991_cast_fp16")];
            bool var_6994 = const()[name = string("op_6994"), val = bool(true)];
            tensor<int32, [1]> var_6995_axes_0 = const()[name = string("op_6995_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_6995_cast_fp16 = reduce_max(axes = var_6995_axes_0, keep_dims = var_6994, x = var_6991_cast_fp16)[name = string("op_6995_cast_fp16")];
            fp16 var_6996_to_fp16 = const()[name = string("op_6996_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_235_cast_fp16 = maximum(x = var_6995_cast_fp16, y = var_6996_to_fp16)[name = string("s_235_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_235_cast_fp16 = real_div(x = linear_130_cast_fp16, y = s_235_cast_fp16)[name = string("y_235_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_6999_cast_fp16 = mul(x = y_235_cast_fp16, y = y_235_cast_fp16)[name = string("op_6999_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_6999_cast_fp16)[name = string("op_7004_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7005_cast_fp16 = mul(x = s_235_cast_fp16, y = s_235_cast_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7010_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7010_cast_fp16")];
            fp32 var_7011_epsilon_0 = const()[name = string("op_7011_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7011_cast_fp16 = rsqrt(epsilon = var_7011_epsilon_0, x = var_7010_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_205_cast_fp16 = mul(x = y_235_cast_fp16, y = var_7011_cast_fp16)[name = string("out_205_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83769664)))];
            tensor<fp16, [1, 64, 512]> var_7013_cast_fp16 = mul(x = out_205_cast_fp16, y = tm_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_7013_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = var_7013_cast_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7016_cast_fp16 = abs(x = x_263_cast_fp16)[name = string("op_7016_cast_fp16")];
            bool var_7019 = const()[name = string("op_7019"), val = bool(true)];
            tensor<int32, [1]> var_7020_axes_0 = const()[name = string("op_7020_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7020_cast_fp16 = reduce_max(axes = var_7020_axes_0, keep_dims = var_7019, x = var_7016_cast_fp16)[name = string("op_7020_cast_fp16")];
            fp16 var_7021_to_fp16 = const()[name = string("op_7021_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_237_cast_fp16 = maximum(x = var_7020_cast_fp16, y = var_7021_to_fp16)[name = string("s_237_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_237_cast_fp16 = real_div(x = x_263_cast_fp16, y = s_237_cast_fp16)[name = string("y_237_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7024_cast_fp16 = mul(x = y_237_cast_fp16, y = y_237_cast_fp16)[name = string("op_7024_cast_fp16")];
            tensor<int32, [1]> var_7029_axes_0 = const()[name = string("op_7029_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7029_keep_dims_0 = const()[name = string("op_7029_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7029_cast_fp16 = reduce_mean(axes = var_7029_axes_0, keep_dims = var_7029_keep_dims_0, x = var_7024_cast_fp16)[name = string("op_7029_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7030_cast_fp16 = mul(x = s_237_cast_fp16, y = s_237_cast_fp16)[name = string("op_7030_cast_fp16")];
            fp32 var_7031_epsilon_0 = const()[name = string("op_7031_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7031_cast_fp16 = inverse(epsilon = var_7031_epsilon_0, x = var_7030_cast_fp16)[name = string("op_7031_cast_fp16")];
            fp16 var_7032_to_fp16 = const()[name = string("op_7032_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7033_cast_fp16 = mul(x = var_7031_cast_fp16, y = var_7032_to_fp16)[name = string("op_7033_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7035_cast_fp16 = add(x = var_7029_cast_fp16, y = var_7033_cast_fp16)[name = string("op_7035_cast_fp16")];
            fp32 var_7036_epsilon_0 = const()[name = string("op_7036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7036_cast_fp16 = rsqrt(epsilon = var_7036_epsilon_0, x = var_7035_cast_fp16)[name = string("op_7036_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_207_cast_fp16 = mul(x = y_237_cast_fp16, y = var_7036_cast_fp16)[name = string("out_207_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83770752)))];
            tensor<fp16, [1, 64, 512]> input_173_cast_fp16 = mul(x = out_207_cast_fp16, y = tm_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83771840))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84820480))))[name = string("tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_131_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_131_cast_fp16")];
            string var_7043_mode_0 = const()[name = string("op_7043_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_7043_cast_fp16 = gelu(mode = var_7043_mode_0, x = linear_131_cast_fp16)[name = string("op_7043_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84824640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85873280))))[name = string("tm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_132_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_175_cast_fp16 = mul(x = var_7043_cast_fp16, y = linear_132_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85877440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86926080))))[name = string("tm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7051_cast_fp16 = abs(x = linear_133_cast_fp16)[name = string("op_7051_cast_fp16")];
            bool var_7054 = const()[name = string("op_7054"), val = bool(true)];
            tensor<int32, [1]> var_7055_axes_0 = const()[name = string("op_7055_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7055_cast_fp16 = reduce_max(axes = var_7055_axes_0, keep_dims = var_7054, x = var_7051_cast_fp16)[name = string("op_7055_cast_fp16")];
            fp16 var_7056_to_fp16 = const()[name = string("op_7056_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_239_cast_fp16 = maximum(x = var_7055_cast_fp16, y = var_7056_to_fp16)[name = string("s_239_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_239_cast_fp16 = real_div(x = linear_133_cast_fp16, y = s_239_cast_fp16)[name = string("y_239_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7059_cast_fp16 = mul(x = y_239_cast_fp16, y = y_239_cast_fp16)[name = string("op_7059_cast_fp16")];
            tensor<int32, [1]> var_7064_axes_0 = const()[name = string("op_7064_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7064_keep_dims_0 = const()[name = string("op_7064_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7064_cast_fp16 = reduce_mean(axes = var_7064_axes_0, keep_dims = var_7064_keep_dims_0, x = var_7059_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7065_cast_fp16 = mul(x = s_239_cast_fp16, y = s_239_cast_fp16)[name = string("op_7065_cast_fp16")];
            fp32 var_7066_epsilon_0 = const()[name = string("op_7066_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7066_cast_fp16 = inverse(epsilon = var_7066_epsilon_0, x = var_7065_cast_fp16)[name = string("op_7066_cast_fp16")];
            fp16 var_7067_to_fp16 = const()[name = string("op_7067_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7068_cast_fp16 = mul(x = var_7066_cast_fp16, y = var_7067_to_fp16)[name = string("op_7068_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7070_cast_fp16 = add(x = var_7064_cast_fp16, y = var_7068_cast_fp16)[name = string("op_7070_cast_fp16")];
            fp32 var_7071_epsilon_0 = const()[name = string("op_7071_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7071_cast_fp16 = rsqrt(epsilon = var_7071_epsilon_0, x = var_7070_cast_fp16)[name = string("op_7071_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_209_cast_fp16 = mul(x = y_239_cast_fp16, y = var_7071_cast_fp16)[name = string("out_209_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86927168)))];
            tensor<fp16, [1, 64, 512]> var_7073_cast_fp16 = mul(x = out_209_cast_fp16, y = tm_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7073_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_177_cast_fp16 = add(x = x_263_cast_fp16, y = var_7073_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86928256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87190464))))[name = string("tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_134_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_134_cast_fp16")];
            string x_267_mode_0 = const()[name = string("x_267_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_267_cast_fp16 = gelu(mode = x_267_mode_0, x = linear_134_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<int32, [4]> var_7093_begin_0 = const()[name = string("op_7093_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
            tensor<int32, [4]> var_7093_end_0 = const()[name = string("op_7093_end_0"), val = tensor<int32, [4]>([1, 64, 15, 512])];
            tensor<bool, [4]> var_7093_end_mask_0 = const()[name = string("op_7093_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7093_squeeze_mask_0 = const()[name = string("op_7093_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_7093_cast_fp16 = slice_by_index(begin = var_7093_begin_0, end = var_7093_end_0, end_mask = var_7093_end_mask_0, squeeze_mask = var_7093_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7093_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_179_cast_fp16 = mul(x = x_267_cast_fp16, y = var_7093_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87191552))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87453760))))[name = string("tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7103_cast_fp16 = abs(x = linear_135_cast_fp16)[name = string("op_7103_cast_fp16")];
            bool var_7106 = const()[name = string("op_7106"), val = bool(true)];
            tensor<int32, [1]> var_7107_axes_0 = const()[name = string("op_7107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7107_cast_fp16 = reduce_max(axes = var_7107_axes_0, keep_dims = var_7106, x = var_7103_cast_fp16)[name = string("op_7107_cast_fp16")];
            fp16 var_7108_to_fp16 = const()[name = string("op_7108_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_241_cast_fp16 = maximum(x = var_7107_cast_fp16, y = var_7108_to_fp16)[name = string("s_241_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_241_cast_fp16 = real_div(x = linear_135_cast_fp16, y = s_241_cast_fp16)[name = string("y_241_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7111_cast_fp16 = mul(x = y_241_cast_fp16, y = y_241_cast_fp16)[name = string("op_7111_cast_fp16")];
            tensor<int32, [1]> var_7116_axes_0 = const()[name = string("op_7116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7116_keep_dims_0 = const()[name = string("op_7116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7116_cast_fp16 = reduce_mean(axes = var_7116_axes_0, keep_dims = var_7116_keep_dims_0, x = var_7111_cast_fp16)[name = string("op_7116_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7117_cast_fp16 = mul(x = s_241_cast_fp16, y = s_241_cast_fp16)[name = string("op_7117_cast_fp16")];
            fp32 var_7118_epsilon_0 = const()[name = string("op_7118_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7118_cast_fp16 = inverse(epsilon = var_7118_epsilon_0, x = var_7117_cast_fp16)[name = string("op_7118_cast_fp16")];
            fp16 var_7119_to_fp16 = const()[name = string("op_7119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7120_cast_fp16 = mul(x = var_7118_cast_fp16, y = var_7119_to_fp16)[name = string("op_7120_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7122_cast_fp16 = add(x = var_7116_cast_fp16, y = var_7120_cast_fp16)[name = string("op_7122_cast_fp16")];
            fp32 var_7123_epsilon_0 = const()[name = string("op_7123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7123_cast_fp16 = rsqrt(epsilon = var_7123_epsilon_0, x = var_7122_cast_fp16)[name = string("op_7123_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_211_cast_fp16 = mul(x = y_241_cast_fp16, y = var_7123_cast_fp16)[name = string("out_211_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87454848)))];
            tensor<fp16, [1, 64, 512]> var_7125_cast_fp16 = mul(x = out_211_cast_fp16, y = tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_29_cast_fp16 = add(x = input_177_cast_fp16, y = var_7125_cast_fp16)[name = string("h_29_cast_fp16")];
            tensor<fp16, [1]> tm_layers_14_layer_scalar_to_fp16 = const()[name = string("tm_layers_14_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1p-1])];
            tensor<fp16, [1, 64, 512]> x_271_cast_fp16 = mul(x = h_29_cast_fp16, y = tm_layers_14_layer_scalar_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7129_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_7129_cast_fp16")];
            bool var_7132 = const()[name = string("op_7132"), val = bool(true)];
            tensor<int32, [1]> var_7133_axes_0 = const()[name = string("op_7133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7133_cast_fp16 = reduce_max(axes = var_7133_axes_0, keep_dims = var_7132, x = var_7129_cast_fp16)[name = string("op_7133_cast_fp16")];
            fp16 var_7134_to_fp16 = const()[name = string("op_7134_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_243_cast_fp16 = maximum(x = var_7133_cast_fp16, y = var_7134_to_fp16)[name = string("s_243_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_243_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_243_cast_fp16)[name = string("y_243_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7137_cast_fp16 = mul(x = y_243_cast_fp16, y = y_243_cast_fp16)[name = string("op_7137_cast_fp16")];
            tensor<int32, [1]> var_7142_axes_0 = const()[name = string("op_7142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7142_keep_dims_0 = const()[name = string("op_7142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7142_cast_fp16 = reduce_mean(axes = var_7142_axes_0, keep_dims = var_7142_keep_dims_0, x = var_7137_cast_fp16)[name = string("op_7142_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7143_cast_fp16 = mul(x = s_243_cast_fp16, y = s_243_cast_fp16)[name = string("op_7143_cast_fp16")];
            fp32 var_7144_epsilon_0 = const()[name = string("op_7144_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7144_cast_fp16 = inverse(epsilon = var_7144_epsilon_0, x = var_7143_cast_fp16)[name = string("op_7144_cast_fp16")];
            fp16 var_7145_to_fp16 = const()[name = string("op_7145_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7146_cast_fp16 = mul(x = var_7144_cast_fp16, y = var_7145_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7148_cast_fp16 = add(x = var_7142_cast_fp16, y = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp32 var_7149_epsilon_0 = const()[name = string("op_7149_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7149_cast_fp16 = rsqrt(epsilon = var_7149_epsilon_0, x = var_7148_cast_fp16)[name = string("op_7149_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_213_cast_fp16 = mul(x = y_243_cast_fp16, y = var_7149_cast_fp16)[name = string("out_213_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87455936)))];
            tensor<fp16, [1, 64, 512]> input_181_cast_fp16 = mul(x = out_213_cast_fp16, y = tm_layers_15_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87457024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87981376))))[name = string("tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<int32, [4]> var_7159 = const()[name = string("op_7159"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_7160_cast_fp16 = reshape(shape = var_7159, x = linear_136_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<int32, [4]> x_273_perm_0 = const()[name = string("x_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87983488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88245696))))[name = string("tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<int32, [4]> var_7171 = const()[name = string("op_7171"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_7172_cast_fp16 = reshape(shape = var_7171, x = linear_137_cast_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [4]> x_275_perm_0 = const()[name = string("x_275_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88246784))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88508992))))[name = string("tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_7183 = const()[name = string("op_7183"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_7184_cast_fp16 = reshape(shape = var_7183, x = linear_138_cast_fp16)[name = string("op_7184_cast_fp16")];
            tensor<int32, [4]> x_277_perm_0 = const()[name = string("x_277_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_273_cast_fp16 = transpose(perm = x_273_perm_0, x = var_7160_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 4, 64, 256]> var_7188_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_7188_cast_fp16")];
            bool var_7191 = const()[name = string("op_7191"), val = bool(true)];
            tensor<int32, [1]> var_7192_axes_0 = const()[name = string("op_7192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_7192_cast_fp16 = reduce_max(axes = var_7192_axes_0, keep_dims = var_7191, x = var_7188_cast_fp16)[name = string("op_7192_cast_fp16")];
            fp16 var_7193_to_fp16 = const()[name = string("op_7193_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_245_cast_fp16 = maximum(x = var_7192_cast_fp16, y = var_7193_to_fp16)[name = string("s_245_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_245_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_245_cast_fp16)[name = string("y_245_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_7196_cast_fp16 = mul(x = y_245_cast_fp16, y = y_245_cast_fp16)[name = string("op_7196_cast_fp16")];
            tensor<int32, [1]> var_7201_axes_0 = const()[name = string("op_7201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7201_keep_dims_0 = const()[name = string("op_7201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_7201_cast_fp16 = reduce_mean(axes = var_7201_axes_0, keep_dims = var_7201_keep_dims_0, x = var_7196_cast_fp16)[name = string("op_7201_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_7202_cast_fp16 = mul(x = s_245_cast_fp16, y = s_245_cast_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_7203_cast_fp16 = inverse(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            fp16 var_7204_to_fp16 = const()[name = string("op_7204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_7205_cast_fp16 = mul(x = var_7203_cast_fp16, y = var_7204_to_fp16)[name = string("op_7205_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_7207_cast_fp16 = add(x = var_7201_cast_fp16, y = var_7205_cast_fp16)[name = string("op_7207_cast_fp16")];
            fp32 var_7208_epsilon_0 = const()[name = string("op_7208_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_7208_cast_fp16 = rsqrt(epsilon = var_7208_epsilon_0, x = var_7207_cast_fp16)[name = string("op_7208_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_215_cast_fp16 = mul(x = y_245_cast_fp16, y = var_7208_cast_fp16)[name = string("out_215_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_61_cast_fp16 = mul(x = out_215_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_61_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_275_cast_fp16 = transpose(perm = x_275_perm_0, x = var_7172_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 2, 64, 256]> var_7211_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_7211_cast_fp16")];
            bool var_7214 = const()[name = string("op_7214"), val = bool(true)];
            tensor<int32, [1]> var_7215_axes_0 = const()[name = string("op_7215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_7215_cast_fp16 = reduce_max(axes = var_7215_axes_0, keep_dims = var_7214, x = var_7211_cast_fp16)[name = string("op_7215_cast_fp16")];
            fp16 var_7216_to_fp16 = const()[name = string("op_7216_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_247_cast_fp16 = maximum(x = var_7215_cast_fp16, y = var_7216_to_fp16)[name = string("s_247_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_247_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_247_cast_fp16)[name = string("y_247_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_7219_cast_fp16 = mul(x = y_247_cast_fp16, y = y_247_cast_fp16)[name = string("op_7219_cast_fp16")];
            tensor<int32, [1]> var_7224_axes_0 = const()[name = string("op_7224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7224_keep_dims_0 = const()[name = string("op_7224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_7224_cast_fp16 = reduce_mean(axes = var_7224_axes_0, keep_dims = var_7224_keep_dims_0, x = var_7219_cast_fp16)[name = string("op_7224_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_7225_cast_fp16 = mul(x = s_247_cast_fp16, y = s_247_cast_fp16)[name = string("op_7225_cast_fp16")];
            fp32 var_7226_epsilon_0 = const()[name = string("op_7226_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_7226_cast_fp16 = inverse(epsilon = var_7226_epsilon_0, x = var_7225_cast_fp16)[name = string("op_7226_cast_fp16")];
            fp16 var_7227_to_fp16 = const()[name = string("op_7227_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_7228_cast_fp16 = mul(x = var_7226_cast_fp16, y = var_7227_to_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_7230_cast_fp16 = add(x = var_7224_cast_fp16, y = var_7228_cast_fp16)[name = string("op_7230_cast_fp16")];
            fp32 var_7231_epsilon_0 = const()[name = string("op_7231_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_7231_cast_fp16 = rsqrt(epsilon = var_7231_epsilon_0, x = var_7230_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_217_cast_fp16 = mul(x = y_247_cast_fp16, y = var_7231_cast_fp16)[name = string("out_217_cast_fp16")];
            tensor<fp16, [256]> tm_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510080)))];
            tensor<fp16, [1, 2, 64, 256]> k_91_cast_fp16 = mul(x = out_217_cast_fp16, y = tm_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_277_cast_fp16 = transpose(perm = x_277_perm_0, x = var_7184_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2, 64, 256]> var_7234_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_7234_cast_fp16")];
            bool var_7237 = const()[name = string("op_7237"), val = bool(true)];
            tensor<int32, [1]> var_7238_axes_0 = const()[name = string("op_7238_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_7238_cast_fp16 = reduce_max(axes = var_7238_axes_0, keep_dims = var_7237, x = var_7234_cast_fp16)[name = string("op_7238_cast_fp16")];
            fp16 var_7239_to_fp16 = const()[name = string("op_7239_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_249_cast_fp16 = maximum(x = var_7238_cast_fp16, y = var_7239_to_fp16)[name = string("s_249_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_249_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_249_cast_fp16)[name = string("y_249_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_7242_cast_fp16 = mul(x = y_249_cast_fp16, y = y_249_cast_fp16)[name = string("op_7242_cast_fp16")];
            tensor<int32, [1]> var_7247_axes_0 = const()[name = string("op_7247_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7247_keep_dims_0 = const()[name = string("op_7247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_7247_cast_fp16 = reduce_mean(axes = var_7247_axes_0, keep_dims = var_7247_keep_dims_0, x = var_7242_cast_fp16)[name = string("op_7247_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_7248_cast_fp16 = mul(x = s_249_cast_fp16, y = s_249_cast_fp16)[name = string("op_7248_cast_fp16")];
            fp32 var_7249_epsilon_0 = const()[name = string("op_7249_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_7249_cast_fp16 = inverse(epsilon = var_7249_epsilon_0, x = var_7248_cast_fp16)[name = string("op_7249_cast_fp16")];
            fp16 var_7250_to_fp16 = const()[name = string("op_7250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_7251_cast_fp16 = mul(x = var_7249_cast_fp16, y = var_7250_to_fp16)[name = string("op_7251_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_7253_cast_fp16 = add(x = var_7247_cast_fp16, y = var_7251_cast_fp16)[name = string("op_7253_cast_fp16")];
            fp32 var_7254_epsilon_0 = const()[name = string("op_7254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_7254_cast_fp16 = rsqrt(epsilon = var_7254_epsilon_0, x = var_7253_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_61_cast_fp16 = mul(x = y_249_cast_fp16, y = var_7254_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_7256_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7256_cast_fp16")];
            tensor<int32, [4]> var_7261_begin_0 = const()[name = string("op_7261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7261_end_0 = const()[name = string("op_7261_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_7261_end_mask_0 = const()[name = string("op_7261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_7261_cast_fp16 = slice_by_index(begin = var_7261_begin_0, end = var_7261_end_0, end_mask = var_7261_end_mask_0, x = q_61_cast_fp16)[name = string("op_7261_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_7262_cast_fp16 = mul(x = var_7261_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_7262_cast_fp16")];
            tensor<int32, [4]> var_7267_begin_0 = const()[name = string("op_7267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7267_end_0 = const()[name = string("op_7267_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_7267_end_mask_0 = const()[name = string("op_7267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_7267_cast_fp16 = slice_by_index(begin = var_7267_begin_0, end = var_7267_end_0, end_mask = var_7267_end_mask_0, x = q_61_cast_fp16)[name = string("op_7267_cast_fp16")];
            int32 var_7269 = const()[name = string("op_7269"), val = int32(-1)];
            bool var_7270_interleave_0 = const()[name = string("op_7270_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_7270_cast_fp16 = concat(axis = var_7269, interleave = var_7270_interleave_0, values = (var_7262_cast_fp16, var_7267_cast_fp16))[name = string("op_7270_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_7271_cast_fp16 = mul(x = var_7270_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7271_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_63_cast_fp16 = add(x = var_7256_cast_fp16, y = var_7271_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_7274_cast_fp16 = mul(x = k_91_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7274_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, x = k_91_cast_fp16)[name = string("op_7279_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_7280_cast_fp16 = mul(x = var_7279_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_7280_cast_fp16")];
            tensor<int32, [4]> var_7285_begin_0 = const()[name = string("op_7285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7285_end_0 = const()[name = string("op_7285_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_7285_end_mask_0 = const()[name = string("op_7285_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, x = k_91_cast_fp16)[name = string("op_7285_cast_fp16")];
            int32 var_7287 = const()[name = string("op_7287"), val = int32(-1)];
            bool var_7288_interleave_0 = const()[name = string("op_7288_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_7288_cast_fp16 = concat(axis = var_7287, interleave = var_7288_interleave_0, values = (var_7280_cast_fp16, var_7285_cast_fp16))[name = string("op_7288_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_7289_cast_fp16 = mul(x = var_7288_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7289_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_93_cast_fp16 = add(x = var_7274_cast_fp16, y = var_7289_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_30_reps_0 = const()[name = string("tile_30_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = k_93_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 1, 64, 256]> tile_30_cast_fp16 = tile(reps = tile_30_reps_0, x = transpose_60_cast_fp16)[name = string("tile_30_cast_fp16")];
            tensor<int32, [5]> concat_60 = const()[name = string("concat_60"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_60_cast_fp16 = reshape(shape = concat_60, x = tile_30_cast_fp16)[name = string("reshape_60_cast_fp16")];
            tensor<int32, [5]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = reshape_60_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 1, 64, 256]> reshape_61_cast_fp16 = reshape(shape = concat_61, x = transpose_61_cast_fp16)[name = string("reshape_61_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_31_reps_0 = const()[name = string("tile_31_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = v_61_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 1, 64, 256]> tile_31_cast_fp16 = tile(reps = tile_31_reps_0, x = transpose_62_cast_fp16)[name = string("tile_31_cast_fp16")];
            tensor<int32, [5]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_62_cast_fp16 = reshape(shape = concat_62, x = tile_31_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [5]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 1, 64, 256]> reshape_63_cast_fp16 = reshape(shape = concat_63, x = transpose_63_cast_fp16)[name = string("reshape_63_cast_fp16")];
            tensor<int32, [4]> v_63_perm_0 = const()[name = string("v_63_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7303_transpose_x_0 = const()[name = string("op_7303_transpose_x_0"), val = bool(false)];
            bool var_7303_transpose_y_0 = const()[name = string("op_7303_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = reshape_61_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 4, 64, 64]> var_7303_cast_fp16 = matmul(transpose_x = var_7303_transpose_x_0, transpose_y = var_7303_transpose_y_0, x = q_63_cast_fp16, y = transpose_111_cast_fp16)[name = string("op_7303_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_7305_cast_fp16 = add(x = var_7303_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7305_cast_fp16")];
            int32 var_7306 = const()[name = string("op_7306"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_249_cast_fp16 = softmax(axis = var_7306, x = var_7305_cast_fp16)[name = string("w_249_cast_fp16")];
            bool var_7309_transpose_x_0 = const()[name = string("op_7309_transpose_x_0"), val = bool(false)];
            bool var_7309_transpose_y_0 = const()[name = string("op_7309_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = reshape_63_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 4, 64, 256]> var_7309_cast_fp16 = matmul(transpose_x = var_7309_transpose_x_0, transpose_y = var_7309_transpose_y_0, x = w_249_cast_fp16, y = v_63_cast_fp16)[name = string("op_7309_cast_fp16")];
            tensor<int32, [4]> var_7312_perm_0 = const()[name = string("op_7312_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7316 = const()[name = string("op_7316"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_7312_cast_fp16 = transpose(perm = var_7312_perm_0, x = var_7309_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 64, 1024]> input_183_cast_fp16 = reshape(shape = var_7316, x = var_7312_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510656))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89035008))))[name = string("tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7321_cast_fp16 = abs(x = linear_139_cast_fp16)[name = string("op_7321_cast_fp16")];
            bool var_7324 = const()[name = string("op_7324"), val = bool(true)];
            tensor<int32, [1]> var_7325_axes_0 = const()[name = string("op_7325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7325_cast_fp16 = reduce_max(axes = var_7325_axes_0, keep_dims = var_7324, x = var_7321_cast_fp16)[name = string("op_7325_cast_fp16")];
            fp16 var_7326_to_fp16 = const()[name = string("op_7326_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_251_cast_fp16 = maximum(x = var_7325_cast_fp16, y = var_7326_to_fp16)[name = string("s_251_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_251_cast_fp16 = real_div(x = linear_139_cast_fp16, y = s_251_cast_fp16)[name = string("y_251_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7329_cast_fp16 = mul(x = y_251_cast_fp16, y = y_251_cast_fp16)[name = string("op_7329_cast_fp16")];
            tensor<int32, [1]> var_7334_axes_0 = const()[name = string("op_7334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7334_keep_dims_0 = const()[name = string("op_7334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7334_cast_fp16 = reduce_mean(axes = var_7334_axes_0, keep_dims = var_7334_keep_dims_0, x = var_7329_cast_fp16)[name = string("op_7334_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7335_cast_fp16 = mul(x = s_251_cast_fp16, y = s_251_cast_fp16)[name = string("op_7335_cast_fp16")];
            fp32 var_7336_epsilon_0 = const()[name = string("op_7336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7336_cast_fp16 = inverse(epsilon = var_7336_epsilon_0, x = var_7335_cast_fp16)[name = string("op_7336_cast_fp16")];
            fp16 var_7337_to_fp16 = const()[name = string("op_7337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7338_cast_fp16 = mul(x = var_7336_cast_fp16, y = var_7337_to_fp16)[name = string("op_7338_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7340_cast_fp16 = add(x = var_7334_cast_fp16, y = var_7338_cast_fp16)[name = string("op_7340_cast_fp16")];
            fp32 var_7341_epsilon_0 = const()[name = string("op_7341_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7341_cast_fp16 = rsqrt(epsilon = var_7341_epsilon_0, x = var_7340_cast_fp16)[name = string("op_7341_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_219_cast_fp16 = mul(x = y_251_cast_fp16, y = var_7341_cast_fp16)[name = string("out_219_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89036096)))];
            tensor<fp16, [1, 64, 512]> var_7343_cast_fp16 = mul(x = out_219_cast_fp16, y = tm_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_7343_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_281_cast_fp16 = add(x = x_271_cast_fp16, y = var_7343_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7346_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_7346_cast_fp16")];
            bool var_7349 = const()[name = string("op_7349"), val = bool(true)];
            tensor<int32, [1]> var_7350_axes_0 = const()[name = string("op_7350_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7350_cast_fp16 = reduce_max(axes = var_7350_axes_0, keep_dims = var_7349, x = var_7346_cast_fp16)[name = string("op_7350_cast_fp16")];
            fp16 var_7351_to_fp16 = const()[name = string("op_7351_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_253_cast_fp16 = maximum(x = var_7350_cast_fp16, y = var_7351_to_fp16)[name = string("s_253_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_253_cast_fp16 = real_div(x = x_281_cast_fp16, y = s_253_cast_fp16)[name = string("y_253_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7354_cast_fp16 = mul(x = y_253_cast_fp16, y = y_253_cast_fp16)[name = string("op_7354_cast_fp16")];
            tensor<int32, [1]> var_7359_axes_0 = const()[name = string("op_7359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7359_keep_dims_0 = const()[name = string("op_7359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7359_cast_fp16 = reduce_mean(axes = var_7359_axes_0, keep_dims = var_7359_keep_dims_0, x = var_7354_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7360_cast_fp16 = mul(x = s_253_cast_fp16, y = s_253_cast_fp16)[name = string("op_7360_cast_fp16")];
            fp32 var_7361_epsilon_0 = const()[name = string("op_7361_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7361_cast_fp16 = inverse(epsilon = var_7361_epsilon_0, x = var_7360_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7362_to_fp16 = const()[name = string("op_7362_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7363_cast_fp16 = mul(x = var_7361_cast_fp16, y = var_7362_to_fp16)[name = string("op_7363_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7365_cast_fp16 = add(x = var_7359_cast_fp16, y = var_7363_cast_fp16)[name = string("op_7365_cast_fp16")];
            fp32 var_7366_epsilon_0 = const()[name = string("op_7366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7366_cast_fp16 = rsqrt(epsilon = var_7366_epsilon_0, x = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_221_cast_fp16 = mul(x = y_253_cast_fp16, y = var_7366_cast_fp16)[name = string("out_221_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89037184)))];
            tensor<fp16, [1, 64, 512]> input_185_cast_fp16 = mul(x = out_221_cast_fp16, y = tm_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89038272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90086912))))[name = string("tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_140_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_140_cast_fp16")];
            string var_7373_mode_0 = const()[name = string("op_7373_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_7373_cast_fp16 = gelu(mode = var_7373_mode_0, x = linear_140_cast_fp16)[name = string("op_7373_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90091072))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91139712))))[name = string("tm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_141_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_187_cast_fp16 = mul(x = var_7373_cast_fp16, y = linear_141_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91143872))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92192512))))[name = string("tm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7381_cast_fp16 = abs(x = linear_142_cast_fp16)[name = string("op_7381_cast_fp16")];
            bool var_7384 = const()[name = string("op_7384"), val = bool(true)];
            tensor<int32, [1]> var_7385_axes_0 = const()[name = string("op_7385_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7385_cast_fp16 = reduce_max(axes = var_7385_axes_0, keep_dims = var_7384, x = var_7381_cast_fp16)[name = string("op_7385_cast_fp16")];
            fp16 var_7386_to_fp16 = const()[name = string("op_7386_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_255_cast_fp16 = maximum(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = string("s_255_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_255_cast_fp16 = real_div(x = linear_142_cast_fp16, y = s_255_cast_fp16)[name = string("y_255_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7389_cast_fp16 = mul(x = y_255_cast_fp16, y = y_255_cast_fp16)[name = string("op_7389_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7389_cast_fp16)[name = string("op_7394_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7395_cast_fp16 = mul(x = s_255_cast_fp16, y = s_255_cast_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7400_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7400_cast_fp16")];
            fp32 var_7401_epsilon_0 = const()[name = string("op_7401_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7401_cast_fp16 = rsqrt(epsilon = var_7401_epsilon_0, x = var_7400_cast_fp16)[name = string("op_7401_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_223_cast_fp16 = mul(x = y_255_cast_fp16, y = var_7401_cast_fp16)[name = string("out_223_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92193600)))];
            tensor<fp16, [1, 64, 512]> var_7403_cast_fp16 = mul(x = out_223_cast_fp16, y = tm_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7403_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_189_cast_fp16 = add(x = x_281_cast_fp16, y = var_7403_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92194688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92456896))))[name = string("tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_143_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_143_cast_fp16")];
            string x_285_mode_0 = const()[name = string("x_285_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_285_cast_fp16 = gelu(mode = x_285_mode_0, x = linear_143_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [4]> var_7423_begin_0 = const()[name = string("op_7423_begin_0"), val = tensor<int32, [4]>([0, 0, 15, 0])];
            tensor<int32, [4]> var_7423_end_0 = const()[name = string("op_7423_end_0"), val = tensor<int32, [4]>([1, 64, 16, 512])];
            tensor<bool, [4]> var_7423_end_mask_0 = const()[name = string("op_7423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7423_squeeze_mask_0 = const()[name = string("op_7423_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_7423_cast_fp16 = slice_by_index(begin = var_7423_begin_0, end = var_7423_end_0, end_mask = var_7423_end_mask_0, squeeze_mask = var_7423_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7423_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_191_cast_fp16 = mul(x = x_285_cast_fp16, y = var_7423_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92457984))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92720192))))[name = string("tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7433_cast_fp16 = abs(x = linear_144_cast_fp16)[name = string("op_7433_cast_fp16")];
            bool var_7436 = const()[name = string("op_7436"), val = bool(true)];
            tensor<int32, [1]> var_7437_axes_0 = const()[name = string("op_7437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7437_cast_fp16 = reduce_max(axes = var_7437_axes_0, keep_dims = var_7436, x = var_7433_cast_fp16)[name = string("op_7437_cast_fp16")];
            fp16 var_7438_to_fp16 = const()[name = string("op_7438_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_257_cast_fp16 = maximum(x = var_7437_cast_fp16, y = var_7438_to_fp16)[name = string("s_257_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_257_cast_fp16 = real_div(x = linear_144_cast_fp16, y = s_257_cast_fp16)[name = string("y_257_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7441_cast_fp16 = mul(x = y_257_cast_fp16, y = y_257_cast_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7446_axes_0 = const()[name = string("op_7446_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7446_keep_dims_0 = const()[name = string("op_7446_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7446_cast_fp16 = reduce_mean(axes = var_7446_axes_0, keep_dims = var_7446_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7447_cast_fp16 = mul(x = s_257_cast_fp16, y = s_257_cast_fp16)[name = string("op_7447_cast_fp16")];
            fp32 var_7448_epsilon_0 = const()[name = string("op_7448_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7448_cast_fp16 = inverse(epsilon = var_7448_epsilon_0, x = var_7447_cast_fp16)[name = string("op_7448_cast_fp16")];
            fp16 var_7449_to_fp16 = const()[name = string("op_7449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7450_cast_fp16 = mul(x = var_7448_cast_fp16, y = var_7449_to_fp16)[name = string("op_7450_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7452_cast_fp16 = add(x = var_7446_cast_fp16, y = var_7450_cast_fp16)[name = string("op_7452_cast_fp16")];
            fp32 var_7453_epsilon_0 = const()[name = string("op_7453_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7453_cast_fp16 = rsqrt(epsilon = var_7453_epsilon_0, x = var_7452_cast_fp16)[name = string("op_7453_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_225_cast_fp16 = mul(x = y_257_cast_fp16, y = var_7453_cast_fp16)[name = string("out_225_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92721280)))];
            tensor<fp16, [1, 64, 512]> var_7455_cast_fp16 = mul(x = out_225_cast_fp16, y = tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7455_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_31_cast_fp16 = add(x = input_189_cast_fp16, y = var_7455_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<fp16, [1]> tm_layers_15_layer_scalar_to_fp16 = const()[name = string("tm_layers_15_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7ap-1])];
            tensor<fp16, [1, 64, 512]> x_289_cast_fp16 = mul(x = h_31_cast_fp16, y = tm_layers_15_layer_scalar_to_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7459_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_7459_cast_fp16")];
            bool var_7462 = const()[name = string("op_7462"), val = bool(true)];
            tensor<int32, [1]> var_7463_axes_0 = const()[name = string("op_7463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7463_cast_fp16 = reduce_max(axes = var_7463_axes_0, keep_dims = var_7462, x = var_7459_cast_fp16)[name = string("op_7463_cast_fp16")];
            fp16 var_7464_to_fp16 = const()[name = string("op_7464_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_259_cast_fp16 = maximum(x = var_7463_cast_fp16, y = var_7464_to_fp16)[name = string("s_259_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_259_cast_fp16 = real_div(x = x_289_cast_fp16, y = s_259_cast_fp16)[name = string("y_259_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7467_cast_fp16 = mul(x = y_259_cast_fp16, y = y_259_cast_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7472_axes_0 = const()[name = string("op_7472_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7472_keep_dims_0 = const()[name = string("op_7472_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7472_cast_fp16 = reduce_mean(axes = var_7472_axes_0, keep_dims = var_7472_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7473_cast_fp16 = mul(x = s_259_cast_fp16, y = s_259_cast_fp16)[name = string("op_7473_cast_fp16")];
            fp32 var_7474_epsilon_0 = const()[name = string("op_7474_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7474_cast_fp16 = inverse(epsilon = var_7474_epsilon_0, x = var_7473_cast_fp16)[name = string("op_7474_cast_fp16")];
            fp16 var_7475_to_fp16 = const()[name = string("op_7475_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7476_cast_fp16 = mul(x = var_7474_cast_fp16, y = var_7475_to_fp16)[name = string("op_7476_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7478_cast_fp16 = add(x = var_7472_cast_fp16, y = var_7476_cast_fp16)[name = string("op_7478_cast_fp16")];
            fp32 var_7479_epsilon_0 = const()[name = string("op_7479_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7479_cast_fp16 = rsqrt(epsilon = var_7479_epsilon_0, x = var_7478_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_227_cast_fp16 = mul(x = y_259_cast_fp16, y = var_7479_cast_fp16)[name = string("out_227_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92722368)))];
            tensor<fp16, [1, 64, 512]> input_193_cast_fp16 = mul(x = out_227_cast_fp16, y = tm_layers_16_input_layernorm_weight_to_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92723456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93247808))))[name = string("tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<int32, [4]> var_7489 = const()[name = string("op_7489"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_7490_cast_fp16 = reshape(shape = var_7489, x = linear_145_cast_fp16)[name = string("op_7490_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93249920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93512128))))[name = string("tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<int32, [4]> var_7501 = const()[name = string("op_7501"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_7502_cast_fp16 = reshape(shape = var_7501, x = linear_146_cast_fp16)[name = string("op_7502_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93513216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93775424))))[name = string("tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7513 = const()[name = string("op_7513"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_7514_cast_fp16 = reshape(shape = var_7513, x = linear_147_cast_fp16)[name = string("op_7514_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_7490_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 4, 64, 256]> var_7518_cast_fp16 = abs(x = x_291_cast_fp16)[name = string("op_7518_cast_fp16")];
            bool var_7521 = const()[name = string("op_7521"), val = bool(true)];
            tensor<int32, [1]> var_7522_axes_0 = const()[name = string("op_7522_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_7522_cast_fp16 = reduce_max(axes = var_7522_axes_0, keep_dims = var_7521, x = var_7518_cast_fp16)[name = string("op_7522_cast_fp16")];
            fp16 var_7523_to_fp16 = const()[name = string("op_7523_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_261_cast_fp16 = maximum(x = var_7522_cast_fp16, y = var_7523_to_fp16)[name = string("s_261_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_261_cast_fp16 = real_div(x = x_291_cast_fp16, y = s_261_cast_fp16)[name = string("y_261_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_7526_cast_fp16 = mul(x = y_261_cast_fp16, y = y_261_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<int32, [1]> var_7531_axes_0 = const()[name = string("op_7531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7531_keep_dims_0 = const()[name = string("op_7531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_7531_cast_fp16 = reduce_mean(axes = var_7531_axes_0, keep_dims = var_7531_keep_dims_0, x = var_7526_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_7532_cast_fp16 = mul(x = s_261_cast_fp16, y = s_261_cast_fp16)[name = string("op_7532_cast_fp16")];
            fp32 var_7533_epsilon_0 = const()[name = string("op_7533_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_7533_cast_fp16 = inverse(epsilon = var_7533_epsilon_0, x = var_7532_cast_fp16)[name = string("op_7533_cast_fp16")];
            fp16 var_7534_to_fp16 = const()[name = string("op_7534_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_7535_cast_fp16 = mul(x = var_7533_cast_fp16, y = var_7534_to_fp16)[name = string("op_7535_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_7537_cast_fp16 = add(x = var_7531_cast_fp16, y = var_7535_cast_fp16)[name = string("op_7537_cast_fp16")];
            fp32 var_7538_epsilon_0 = const()[name = string("op_7538_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_7538_cast_fp16 = rsqrt(epsilon = var_7538_epsilon_0, x = var_7537_cast_fp16)[name = string("op_7538_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_229_cast_fp16 = mul(x = y_261_cast_fp16, y = var_7538_cast_fp16)[name = string("out_229_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_65_cast_fp16 = mul(x = out_229_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_65_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_7502_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 64, 256]> var_7541_cast_fp16 = abs(x = x_293_cast_fp16)[name = string("op_7541_cast_fp16")];
            bool var_7544 = const()[name = string("op_7544"), val = bool(true)];
            tensor<int32, [1]> var_7545_axes_0 = const()[name = string("op_7545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_7545_cast_fp16 = reduce_max(axes = var_7545_axes_0, keep_dims = var_7544, x = var_7541_cast_fp16)[name = string("op_7545_cast_fp16")];
            fp16 var_7546_to_fp16 = const()[name = string("op_7546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_263_cast_fp16 = maximum(x = var_7545_cast_fp16, y = var_7546_to_fp16)[name = string("s_263_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_263_cast_fp16 = real_div(x = x_293_cast_fp16, y = s_263_cast_fp16)[name = string("y_263_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_7549_cast_fp16 = mul(x = y_263_cast_fp16, y = y_263_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<int32, [1]> var_7554_axes_0 = const()[name = string("op_7554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7554_keep_dims_0 = const()[name = string("op_7554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_7554_cast_fp16 = reduce_mean(axes = var_7554_axes_0, keep_dims = var_7554_keep_dims_0, x = var_7549_cast_fp16)[name = string("op_7554_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_7555_cast_fp16 = mul(x = s_263_cast_fp16, y = s_263_cast_fp16)[name = string("op_7555_cast_fp16")];
            fp32 var_7556_epsilon_0 = const()[name = string("op_7556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_7556_cast_fp16 = inverse(epsilon = var_7556_epsilon_0, x = var_7555_cast_fp16)[name = string("op_7556_cast_fp16")];
            fp16 var_7557_to_fp16 = const()[name = string("op_7557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_7558_cast_fp16 = mul(x = var_7556_cast_fp16, y = var_7557_to_fp16)[name = string("op_7558_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_7560_cast_fp16 = add(x = var_7554_cast_fp16, y = var_7558_cast_fp16)[name = string("op_7560_cast_fp16")];
            fp32 var_7561_epsilon_0 = const()[name = string("op_7561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_7561_cast_fp16 = rsqrt(epsilon = var_7561_epsilon_0, x = var_7560_cast_fp16)[name = string("op_7561_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_231_cast_fp16 = mul(x = y_263_cast_fp16, y = var_7561_cast_fp16)[name = string("out_231_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_97_cast_fp16 = mul(x = out_231_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_97_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_7514_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 64, 256]> var_7564_cast_fp16 = abs(x = x_295_cast_fp16)[name = string("op_7564_cast_fp16")];
            bool var_7567 = const()[name = string("op_7567"), val = bool(true)];
            tensor<int32, [1]> var_7568_axes_0 = const()[name = string("op_7568_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_7568_cast_fp16 = reduce_max(axes = var_7568_axes_0, keep_dims = var_7567, x = var_7564_cast_fp16)[name = string("op_7568_cast_fp16")];
            fp16 var_7569_to_fp16 = const()[name = string("op_7569_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_265_cast_fp16 = maximum(x = var_7568_cast_fp16, y = var_7569_to_fp16)[name = string("s_265_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_265_cast_fp16 = real_div(x = x_295_cast_fp16, y = s_265_cast_fp16)[name = string("y_265_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_7572_cast_fp16 = mul(x = y_265_cast_fp16, y = y_265_cast_fp16)[name = string("op_7572_cast_fp16")];
            tensor<int32, [1]> var_7577_axes_0 = const()[name = string("op_7577_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7577_keep_dims_0 = const()[name = string("op_7577_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_7577_cast_fp16 = reduce_mean(axes = var_7577_axes_0, keep_dims = var_7577_keep_dims_0, x = var_7572_cast_fp16)[name = string("op_7577_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_7578_cast_fp16 = mul(x = s_265_cast_fp16, y = s_265_cast_fp16)[name = string("op_7578_cast_fp16")];
            fp32 var_7579_epsilon_0 = const()[name = string("op_7579_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_7579_cast_fp16 = inverse(epsilon = var_7579_epsilon_0, x = var_7578_cast_fp16)[name = string("op_7579_cast_fp16")];
            fp16 var_7580_to_fp16 = const()[name = string("op_7580_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_7581_cast_fp16 = mul(x = var_7579_cast_fp16, y = var_7580_to_fp16)[name = string("op_7581_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_7583_cast_fp16 = add(x = var_7577_cast_fp16, y = var_7581_cast_fp16)[name = string("op_7583_cast_fp16")];
            fp32 var_7584_epsilon_0 = const()[name = string("op_7584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_7584_cast_fp16 = rsqrt(epsilon = var_7584_epsilon_0, x = var_7583_cast_fp16)[name = string("op_7584_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_65_cast_fp16 = mul(x = y_265_cast_fp16, y = var_7584_cast_fp16)[name = string("v_65_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_7586_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7586_cast_fp16")];
            tensor<int32, [4]> var_7591_begin_0 = const()[name = string("op_7591_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7591_end_0 = const()[name = string("op_7591_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_7591_end_mask_0 = const()[name = string("op_7591_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_7591_cast_fp16 = slice_by_index(begin = var_7591_begin_0, end = var_7591_end_0, end_mask = var_7591_end_mask_0, x = q_65_cast_fp16)[name = string("op_7591_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_7592_cast_fp16 = mul(x = var_7591_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_7592_cast_fp16")];
            tensor<int32, [4]> var_7597_begin_0 = const()[name = string("op_7597_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7597_end_0 = const()[name = string("op_7597_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_7597_end_mask_0 = const()[name = string("op_7597_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, x = q_65_cast_fp16)[name = string("op_7597_cast_fp16")];
            int32 var_7599 = const()[name = string("op_7599"), val = int32(-1)];
            bool var_7600_interleave_0 = const()[name = string("op_7600_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_7600_cast_fp16 = concat(axis = var_7599, interleave = var_7600_interleave_0, values = (var_7592_cast_fp16, var_7597_cast_fp16))[name = string("op_7600_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_7601_cast_fp16 = mul(x = var_7600_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7601_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_67_cast_fp16 = add(x = var_7586_cast_fp16, y = var_7601_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_7604_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_7604_cast_fp16")];
            tensor<int32, [4]> var_7609_begin_0 = const()[name = string("op_7609_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7609_end_0 = const()[name = string("op_7609_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_7609_end_mask_0 = const()[name = string("op_7609_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_7609_cast_fp16 = slice_by_index(begin = var_7609_begin_0, end = var_7609_end_0, end_mask = var_7609_end_mask_0, x = k_97_cast_fp16)[name = string("op_7609_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_7610_cast_fp16 = mul(x = var_7609_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_7610_cast_fp16")];
            tensor<int32, [4]> var_7615_begin_0 = const()[name = string("op_7615_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7615_end_0 = const()[name = string("op_7615_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_7615_end_mask_0 = const()[name = string("op_7615_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_7615_cast_fp16 = slice_by_index(begin = var_7615_begin_0, end = var_7615_end_0, end_mask = var_7615_end_mask_0, x = k_97_cast_fp16)[name = string("op_7615_cast_fp16")];
            int32 var_7617 = const()[name = string("op_7617"), val = int32(-1)];
            bool var_7618_interleave_0 = const()[name = string("op_7618_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_7618_cast_fp16 = concat(axis = var_7617, interleave = var_7618_interleave_0, values = (var_7610_cast_fp16, var_7615_cast_fp16))[name = string("op_7618_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_7619_cast_fp16 = mul(x = var_7618_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_7619_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_99_cast_fp16 = add(x = var_7604_cast_fp16, y = var_7619_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_32_reps_0 = const()[name = string("tile_32_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = k_99_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 1, 64, 256]> tile_32_cast_fp16 = tile(reps = tile_32_reps_0, x = transpose_64_cast_fp16)[name = string("tile_32_cast_fp16")];
            tensor<int32, [5]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_64_cast_fp16 = reshape(shape = concat_64, x = tile_32_cast_fp16)[name = string("reshape_64_cast_fp16")];
            tensor<int32, [5]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = reshape_64_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 1, 64, 256]> reshape_65_cast_fp16 = reshape(shape = concat_65, x = transpose_65_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_33_reps_0 = const()[name = string("tile_33_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = v_65_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 1, 64, 256]> tile_33_cast_fp16 = tile(reps = tile_33_reps_0, x = transpose_66_cast_fp16)[name = string("tile_33_cast_fp16")];
            tensor<int32, [5]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_66_cast_fp16 = reshape(shape = concat_66, x = tile_33_cast_fp16)[name = string("reshape_66_cast_fp16")];
            tensor<int32, [5]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = reshape_66_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 1, 64, 256]> reshape_67_cast_fp16 = reshape(shape = concat_67, x = transpose_67_cast_fp16)[name = string("reshape_67_cast_fp16")];
            tensor<int32, [4]> v_67_perm_0 = const()[name = string("v_67_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7633_transpose_x_0 = const()[name = string("op_7633_transpose_x_0"), val = bool(false)];
            bool var_7633_transpose_y_0 = const()[name = string("op_7633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_112_cast_fp16 = transpose(perm = transpose_112_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 4, 64, 64]> var_7633_cast_fp16 = matmul(transpose_x = var_7633_transpose_x_0, transpose_y = var_7633_transpose_y_0, x = q_67_cast_fp16, y = transpose_112_cast_fp16)[name = string("op_7633_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_7635_cast_fp16 = add(x = var_7633_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7635_cast_fp16")];
            int32 var_7636 = const()[name = string("op_7636"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_265_cast_fp16 = softmax(axis = var_7636, x = var_7635_cast_fp16)[name = string("w_265_cast_fp16")];
            bool var_7639_transpose_x_0 = const()[name = string("op_7639_transpose_x_0"), val = bool(false)];
            bool var_7639_transpose_y_0 = const()[name = string("op_7639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = reshape_67_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 4, 64, 256]> var_7639_cast_fp16 = matmul(transpose_x = var_7639_transpose_x_0, transpose_y = var_7639_transpose_y_0, x = w_265_cast_fp16, y = v_67_cast_fp16)[name = string("op_7639_cast_fp16")];
            tensor<int32, [4]> var_7642_perm_0 = const()[name = string("op_7642_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7646 = const()[name = string("op_7646"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_7642_cast_fp16 = transpose(perm = var_7642_perm_0, x = var_7639_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 64, 1024]> input_195_cast_fp16 = reshape(shape = var_7646, x = var_7642_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93776512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94300864))))[name = string("tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7651_cast_fp16 = abs(x = linear_148_cast_fp16)[name = string("op_7651_cast_fp16")];
            bool var_7654 = const()[name = string("op_7654"), val = bool(true)];
            tensor<int32, [1]> var_7655_axes_0 = const()[name = string("op_7655_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7655_cast_fp16 = reduce_max(axes = var_7655_axes_0, keep_dims = var_7654, x = var_7651_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp16 var_7656_to_fp16 = const()[name = string("op_7656_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_267_cast_fp16 = maximum(x = var_7655_cast_fp16, y = var_7656_to_fp16)[name = string("s_267_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_267_cast_fp16 = real_div(x = linear_148_cast_fp16, y = s_267_cast_fp16)[name = string("y_267_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7659_cast_fp16 = mul(x = y_267_cast_fp16, y = y_267_cast_fp16)[name = string("op_7659_cast_fp16")];
            tensor<int32, [1]> var_7664_axes_0 = const()[name = string("op_7664_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7664_keep_dims_0 = const()[name = string("op_7664_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7664_cast_fp16 = reduce_mean(axes = var_7664_axes_0, keep_dims = var_7664_keep_dims_0, x = var_7659_cast_fp16)[name = string("op_7664_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7665_cast_fp16 = mul(x = s_267_cast_fp16, y = s_267_cast_fp16)[name = string("op_7665_cast_fp16")];
            fp32 var_7666_epsilon_0 = const()[name = string("op_7666_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7666_cast_fp16 = inverse(epsilon = var_7666_epsilon_0, x = var_7665_cast_fp16)[name = string("op_7666_cast_fp16")];
            fp16 var_7667_to_fp16 = const()[name = string("op_7667_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7668_cast_fp16 = mul(x = var_7666_cast_fp16, y = var_7667_to_fp16)[name = string("op_7668_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7670_cast_fp16 = add(x = var_7664_cast_fp16, y = var_7668_cast_fp16)[name = string("op_7670_cast_fp16")];
            fp32 var_7671_epsilon_0 = const()[name = string("op_7671_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7671_cast_fp16 = rsqrt(epsilon = var_7671_epsilon_0, x = var_7670_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_233_cast_fp16 = mul(x = y_267_cast_fp16, y = var_7671_cast_fp16)[name = string("out_233_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94301952)))];
            tensor<fp16, [1, 64, 512]> var_7673_cast_fp16 = mul(x = out_233_cast_fp16, y = tm_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("op_7673_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = var_7673_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7676_cast_fp16 = abs(x = x_299_cast_fp16)[name = string("op_7676_cast_fp16")];
            bool var_7679 = const()[name = string("op_7679"), val = bool(true)];
            tensor<int32, [1]> var_7680_axes_0 = const()[name = string("op_7680_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7680_cast_fp16 = reduce_max(axes = var_7680_axes_0, keep_dims = var_7679, x = var_7676_cast_fp16)[name = string("op_7680_cast_fp16")];
            fp16 var_7681_to_fp16 = const()[name = string("op_7681_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_269_cast_fp16 = maximum(x = var_7680_cast_fp16, y = var_7681_to_fp16)[name = string("s_269_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_269_cast_fp16 = real_div(x = x_299_cast_fp16, y = s_269_cast_fp16)[name = string("y_269_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7684_cast_fp16 = mul(x = y_269_cast_fp16, y = y_269_cast_fp16)[name = string("op_7684_cast_fp16")];
            tensor<int32, [1]> var_7689_axes_0 = const()[name = string("op_7689_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7689_keep_dims_0 = const()[name = string("op_7689_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7689_cast_fp16 = reduce_mean(axes = var_7689_axes_0, keep_dims = var_7689_keep_dims_0, x = var_7684_cast_fp16)[name = string("op_7689_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7690_cast_fp16 = mul(x = s_269_cast_fp16, y = s_269_cast_fp16)[name = string("op_7690_cast_fp16")];
            fp32 var_7691_epsilon_0 = const()[name = string("op_7691_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7691_cast_fp16 = inverse(epsilon = var_7691_epsilon_0, x = var_7690_cast_fp16)[name = string("op_7691_cast_fp16")];
            fp16 var_7692_to_fp16 = const()[name = string("op_7692_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7693_cast_fp16 = mul(x = var_7691_cast_fp16, y = var_7692_to_fp16)[name = string("op_7693_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7695_cast_fp16 = add(x = var_7689_cast_fp16, y = var_7693_cast_fp16)[name = string("op_7695_cast_fp16")];
            fp32 var_7696_epsilon_0 = const()[name = string("op_7696_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7696_cast_fp16 = rsqrt(epsilon = var_7696_epsilon_0, x = var_7695_cast_fp16)[name = string("op_7696_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_235_cast_fp16 = mul(x = y_269_cast_fp16, y = var_7696_cast_fp16)[name = string("out_235_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94303040)))];
            tensor<fp16, [1, 64, 512]> input_197_cast_fp16 = mul(x = out_235_cast_fp16, y = tm_layers_16_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94304128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95352768))))[name = string("tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_149_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_149_cast_fp16")];
            string var_7703_mode_0 = const()[name = string("op_7703_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_7703_cast_fp16 = gelu(mode = var_7703_mode_0, x = linear_149_cast_fp16)[name = string("op_7703_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95356928))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96405568))))[name = string("tm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_150_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_199_cast_fp16 = mul(x = var_7703_cast_fp16, y = linear_150_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96409728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97458368))))[name = string("tm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7711_cast_fp16 = abs(x = linear_151_cast_fp16)[name = string("op_7711_cast_fp16")];
            bool var_7714 = const()[name = string("op_7714"), val = bool(true)];
            tensor<int32, [1]> var_7715_axes_0 = const()[name = string("op_7715_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7715_cast_fp16 = reduce_max(axes = var_7715_axes_0, keep_dims = var_7714, x = var_7711_cast_fp16)[name = string("op_7715_cast_fp16")];
            fp16 var_7716_to_fp16 = const()[name = string("op_7716_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_271_cast_fp16 = maximum(x = var_7715_cast_fp16, y = var_7716_to_fp16)[name = string("s_271_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_271_cast_fp16 = real_div(x = linear_151_cast_fp16, y = s_271_cast_fp16)[name = string("y_271_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7719_cast_fp16 = mul(x = y_271_cast_fp16, y = y_271_cast_fp16)[name = string("op_7719_cast_fp16")];
            tensor<int32, [1]> var_7724_axes_0 = const()[name = string("op_7724_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7724_keep_dims_0 = const()[name = string("op_7724_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7724_cast_fp16 = reduce_mean(axes = var_7724_axes_0, keep_dims = var_7724_keep_dims_0, x = var_7719_cast_fp16)[name = string("op_7724_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7725_cast_fp16 = mul(x = s_271_cast_fp16, y = s_271_cast_fp16)[name = string("op_7725_cast_fp16")];
            fp32 var_7726_epsilon_0 = const()[name = string("op_7726_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7726_cast_fp16 = inverse(epsilon = var_7726_epsilon_0, x = var_7725_cast_fp16)[name = string("op_7726_cast_fp16")];
            fp16 var_7727_to_fp16 = const()[name = string("op_7727_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7728_cast_fp16 = mul(x = var_7726_cast_fp16, y = var_7727_to_fp16)[name = string("op_7728_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7730_cast_fp16 = add(x = var_7724_cast_fp16, y = var_7728_cast_fp16)[name = string("op_7730_cast_fp16")];
            fp32 var_7731_epsilon_0 = const()[name = string("op_7731_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7731_cast_fp16 = rsqrt(epsilon = var_7731_epsilon_0, x = var_7730_cast_fp16)[name = string("op_7731_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_237_cast_fp16 = mul(x = y_271_cast_fp16, y = var_7731_cast_fp16)[name = string("out_237_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97459456)))];
            tensor<fp16, [1, 64, 512]> var_7733_cast_fp16 = mul(x = out_237_cast_fp16, y = tm_layers_16_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7733_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_201_cast_fp16 = add(x = x_299_cast_fp16, y = var_7733_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97460544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97722752))))[name = string("tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_152_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_152_cast_fp16")];
            string x_303_mode_0 = const()[name = string("x_303_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_303_cast_fp16 = gelu(mode = x_303_mode_0, x = linear_152_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [4]> var_7753_begin_0 = const()[name = string("op_7753_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7753_end_0 = const()[name = string("op_7753_end_0"), val = tensor<int32, [4]>([1, 64, 17, 512])];
            tensor<bool, [4]> var_7753_end_mask_0 = const()[name = string("op_7753_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7753_squeeze_mask_0 = const()[name = string("op_7753_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_7753_cast_fp16 = slice_by_index(begin = var_7753_begin_0, end = var_7753_end_0, end_mask = var_7753_end_mask_0, squeeze_mask = var_7753_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_7753_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_203_cast_fp16 = mul(x = x_303_cast_fp16, y = var_7753_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97723840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97986048))))[name = string("tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7763_cast_fp16 = abs(x = linear_153_cast_fp16)[name = string("op_7763_cast_fp16")];
            bool var_7766 = const()[name = string("op_7766"), val = bool(true)];
            tensor<int32, [1]> var_7767_axes_0 = const()[name = string("op_7767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7767_cast_fp16 = reduce_max(axes = var_7767_axes_0, keep_dims = var_7766, x = var_7763_cast_fp16)[name = string("op_7767_cast_fp16")];
            fp16 var_7768_to_fp16 = const()[name = string("op_7768_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_273_cast_fp16 = maximum(x = var_7767_cast_fp16, y = var_7768_to_fp16)[name = string("s_273_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_273_cast_fp16 = real_div(x = linear_153_cast_fp16, y = s_273_cast_fp16)[name = string("y_273_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7771_cast_fp16 = mul(x = y_273_cast_fp16, y = y_273_cast_fp16)[name = string("op_7771_cast_fp16")];
            tensor<int32, [1]> var_7776_axes_0 = const()[name = string("op_7776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7776_keep_dims_0 = const()[name = string("op_7776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7776_cast_fp16 = reduce_mean(axes = var_7776_axes_0, keep_dims = var_7776_keep_dims_0, x = var_7771_cast_fp16)[name = string("op_7776_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7777_cast_fp16 = mul(x = s_273_cast_fp16, y = s_273_cast_fp16)[name = string("op_7777_cast_fp16")];
            fp32 var_7778_epsilon_0 = const()[name = string("op_7778_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7778_cast_fp16 = inverse(epsilon = var_7778_epsilon_0, x = var_7777_cast_fp16)[name = string("op_7778_cast_fp16")];
            fp16 var_7779_to_fp16 = const()[name = string("op_7779_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7780_cast_fp16 = mul(x = var_7778_cast_fp16, y = var_7779_to_fp16)[name = string("op_7780_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7782_cast_fp16 = add(x = var_7776_cast_fp16, y = var_7780_cast_fp16)[name = string("op_7782_cast_fp16")];
            fp32 var_7783_epsilon_0 = const()[name = string("op_7783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7783_cast_fp16 = rsqrt(epsilon = var_7783_epsilon_0, x = var_7782_cast_fp16)[name = string("op_7783_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_239_cast_fp16 = mul(x = y_273_cast_fp16, y = var_7783_cast_fp16)[name = string("out_239_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97987136)))];
            tensor<fp16, [1, 64, 512]> var_7785_cast_fp16 = mul(x = out_239_cast_fp16, y = tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_33_cast_fp16 = add(x = input_201_cast_fp16, y = var_7785_cast_fp16)[name = string("h_33_cast_fp16")];
            tensor<fp16, [1]> tm_layers_16_layer_scalar_to_fp16 = const()[name = string("tm_layers_16_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7p-2])];
            tensor<fp16, [1, 64, 512]> x_307_cast_fp16 = mul(x = h_33_cast_fp16, y = tm_layers_16_layer_scalar_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7789_cast_fp16 = abs(x = x_307_cast_fp16)[name = string("op_7789_cast_fp16")];
            bool var_7792 = const()[name = string("op_7792"), val = bool(true)];
            tensor<int32, [1]> var_7793_axes_0 = const()[name = string("op_7793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7793_cast_fp16 = reduce_max(axes = var_7793_axes_0, keep_dims = var_7792, x = var_7789_cast_fp16)[name = string("op_7793_cast_fp16")];
            fp16 var_7794_to_fp16 = const()[name = string("op_7794_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_275_cast_fp16 = maximum(x = var_7793_cast_fp16, y = var_7794_to_fp16)[name = string("s_275_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_275_cast_fp16 = real_div(x = x_307_cast_fp16, y = s_275_cast_fp16)[name = string("y_275_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7797_cast_fp16 = mul(x = y_275_cast_fp16, y = y_275_cast_fp16)[name = string("op_7797_cast_fp16")];
            tensor<int32, [1]> var_7802_axes_0 = const()[name = string("op_7802_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7802_keep_dims_0 = const()[name = string("op_7802_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7802_cast_fp16 = reduce_mean(axes = var_7802_axes_0, keep_dims = var_7802_keep_dims_0, x = var_7797_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7803_cast_fp16 = mul(x = s_275_cast_fp16, y = s_275_cast_fp16)[name = string("op_7803_cast_fp16")];
            fp32 var_7804_epsilon_0 = const()[name = string("op_7804_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7804_cast_fp16 = inverse(epsilon = var_7804_epsilon_0, x = var_7803_cast_fp16)[name = string("op_7804_cast_fp16")];
            fp16 var_7805_to_fp16 = const()[name = string("op_7805_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7806_cast_fp16 = mul(x = var_7804_cast_fp16, y = var_7805_to_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7808_cast_fp16 = add(x = var_7802_cast_fp16, y = var_7806_cast_fp16)[name = string("op_7808_cast_fp16")];
            fp32 var_7809_epsilon_0 = const()[name = string("op_7809_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_7809_cast_fp16 = rsqrt(epsilon = var_7809_epsilon_0, x = var_7808_cast_fp16)[name = string("op_7809_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_241_cast_fp16 = mul(x = y_275_cast_fp16, y = var_7809_cast_fp16)[name = string("out_241_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97988224)))];
            tensor<fp16, [1, 64, 512]> input_205_cast_fp16 = mul(x = out_241_cast_fp16, y = tm_layers_17_input_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97989312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99037952))))[name = string("tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_154_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_7819 = const()[name = string("op_7819"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 64, 4, 512]> var_7820_cast_fp16 = reshape(shape = var_7819, x = linear_154_cast_fp16)[name = string("op_7820_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99042112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99304320))))[name = string("tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_7831 = const()[name = string("op_7831"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 64, 1, 512]> var_7832_cast_fp16 = reshape(shape = var_7831, x = linear_155_cast_fp16)[name = string("op_7832_cast_fp16")];
            tensor<int32, [4]> x_311_perm_0 = const()[name = string("x_311_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99305408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99567616))))[name = string("tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_7843 = const()[name = string("op_7843"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 64, 1, 512]> var_7844_cast_fp16 = reshape(shape = var_7843, x = linear_156_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<int32, [4]> x_313_perm_0 = const()[name = string("x_313_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 512]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_7820_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 4, 64, 512]> var_7848_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_7848_cast_fp16")];
            bool var_7851 = const()[name = string("op_7851"), val = bool(true)];
            tensor<int32, [1]> var_7852_axes_0 = const()[name = string("op_7852_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_7852_cast_fp16 = reduce_max(axes = var_7852_axes_0, keep_dims = var_7851, x = var_7848_cast_fp16)[name = string("op_7852_cast_fp16")];
            fp16 var_7853_to_fp16 = const()[name = string("op_7853_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_277_cast_fp16 = maximum(x = var_7852_cast_fp16, y = var_7853_to_fp16)[name = string("s_277_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> y_277_cast_fp16 = real_div(x = x_309_cast_fp16, y = s_277_cast_fp16)[name = string("y_277_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_7856_cast_fp16 = mul(x = y_277_cast_fp16, y = y_277_cast_fp16)[name = string("op_7856_cast_fp16")];
            tensor<int32, [1]> var_7861_axes_0 = const()[name = string("op_7861_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7861_keep_dims_0 = const()[name = string("op_7861_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_7861_cast_fp16 = reduce_mean(axes = var_7861_axes_0, keep_dims = var_7861_keep_dims_0, x = var_7856_cast_fp16)[name = string("op_7861_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_7862_cast_fp16 = mul(x = s_277_cast_fp16, y = s_277_cast_fp16)[name = string("op_7862_cast_fp16")];
            fp32 var_7863_epsilon_0 = const()[name = string("op_7863_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_7863_cast_fp16 = inverse(epsilon = var_7863_epsilon_0, x = var_7862_cast_fp16)[name = string("op_7863_cast_fp16")];
            fp16 var_7864_to_fp16 = const()[name = string("op_7864_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_7865_cast_fp16 = mul(x = var_7863_cast_fp16, y = var_7864_to_fp16)[name = string("op_7865_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_7867_cast_fp16 = add(x = var_7861_cast_fp16, y = var_7865_cast_fp16)[name = string("op_7867_cast_fp16")];
            fp32 var_7868_epsilon_0 = const()[name = string("op_7868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_7868_cast_fp16 = rsqrt(epsilon = var_7868_epsilon_0, x = var_7867_cast_fp16)[name = string("op_7868_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> out_243_cast_fp16 = mul(x = y_277_cast_fp16, y = var_7868_cast_fp16)[name = string("out_243_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99568704)))];
            tensor<fp16, [1, 4, 64, 512]> q_69_cast_fp16 = mul(x = out_243_cast_fp16, y = tm_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("q_69_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> x_311_cast_fp16 = transpose(perm = x_311_perm_0, x = var_7832_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 1, 64, 512]> var_7871_cast_fp16 = abs(x = x_311_cast_fp16)[name = string("op_7871_cast_fp16")];
            bool var_7874 = const()[name = string("op_7874"), val = bool(true)];
            tensor<int32, [1]> var_7875_axes_0 = const()[name = string("op_7875_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 64, 1]> var_7875_cast_fp16 = reduce_max(axes = var_7875_axes_0, keep_dims = var_7874, x = var_7871_cast_fp16)[name = string("op_7875_cast_fp16")];
            fp16 var_7876_to_fp16 = const()[name = string("op_7876_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 64, 1]> s_279_cast_fp16 = maximum(x = var_7875_cast_fp16, y = var_7876_to_fp16)[name = string("s_279_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> y_279_cast_fp16 = real_div(x = x_311_cast_fp16, y = s_279_cast_fp16)[name = string("y_279_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_7879_cast_fp16 = mul(x = y_279_cast_fp16, y = y_279_cast_fp16)[name = string("op_7879_cast_fp16")];
            tensor<int32, [1]> var_7884_axes_0 = const()[name = string("op_7884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7884_keep_dims_0 = const()[name = string("op_7884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 64, 1]> var_7884_cast_fp16 = reduce_mean(axes = var_7884_axes_0, keep_dims = var_7884_keep_dims_0, x = var_7879_cast_fp16)[name = string("op_7884_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_7885_cast_fp16 = mul(x = s_279_cast_fp16, y = s_279_cast_fp16)[name = string("op_7885_cast_fp16")];
            fp32 var_7886_epsilon_0 = const()[name = string("op_7886_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 64, 1]> var_7886_cast_fp16 = inverse(epsilon = var_7886_epsilon_0, x = var_7885_cast_fp16)[name = string("op_7886_cast_fp16")];
            fp16 var_7887_to_fp16 = const()[name = string("op_7887_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 64, 1]> var_7888_cast_fp16 = mul(x = var_7886_cast_fp16, y = var_7887_to_fp16)[name = string("op_7888_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_7890_cast_fp16 = add(x = var_7884_cast_fp16, y = var_7888_cast_fp16)[name = string("op_7890_cast_fp16")];
            fp32 var_7891_epsilon_0 = const()[name = string("op_7891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 64, 1]> var_7891_cast_fp16 = rsqrt(epsilon = var_7891_epsilon_0, x = var_7890_cast_fp16)[name = string("op_7891_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> out_245_cast_fp16 = mul(x = y_279_cast_fp16, y = var_7891_cast_fp16)[name = string("out_245_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99569792)))];
            tensor<fp16, [1, 1, 64, 512]> k_103_cast_fp16 = mul(x = out_245_cast_fp16, y = tm_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("k_103_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> x_313_cast_fp16 = transpose(perm = x_313_perm_0, x = var_7844_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1, 64, 512]> var_7894_cast_fp16 = abs(x = x_313_cast_fp16)[name = string("op_7894_cast_fp16")];
            bool var_7897 = const()[name = string("op_7897"), val = bool(true)];
            tensor<int32, [1]> var_7898_axes_0 = const()[name = string("op_7898_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 64, 1]> var_7898_cast_fp16 = reduce_max(axes = var_7898_axes_0, keep_dims = var_7897, x = var_7894_cast_fp16)[name = string("op_7898_cast_fp16")];
            fp16 var_7899_to_fp16 = const()[name = string("op_7899_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 64, 1]> s_281_cast_fp16 = maximum(x = var_7898_cast_fp16, y = var_7899_to_fp16)[name = string("s_281_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> y_281_cast_fp16 = real_div(x = x_313_cast_fp16, y = s_281_cast_fp16)[name = string("y_281_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_7902_cast_fp16 = mul(x = y_281_cast_fp16, y = y_281_cast_fp16)[name = string("op_7902_cast_fp16")];
            tensor<int32, [1]> var_7907_axes_0 = const()[name = string("op_7907_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7907_keep_dims_0 = const()[name = string("op_7907_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 64, 1]> var_7907_cast_fp16 = reduce_mean(axes = var_7907_axes_0, keep_dims = var_7907_keep_dims_0, x = var_7902_cast_fp16)[name = string("op_7907_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_7908_cast_fp16 = mul(x = s_281_cast_fp16, y = s_281_cast_fp16)[name = string("op_7908_cast_fp16")];
            fp32 var_7909_epsilon_0 = const()[name = string("op_7909_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 64, 1]> var_7909_cast_fp16 = inverse(epsilon = var_7909_epsilon_0, x = var_7908_cast_fp16)[name = string("op_7909_cast_fp16")];
            fp16 var_7910_to_fp16 = const()[name = string("op_7910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 64, 1]> var_7911_cast_fp16 = mul(x = var_7909_cast_fp16, y = var_7910_to_fp16)[name = string("op_7911_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_7913_cast_fp16 = add(x = var_7907_cast_fp16, y = var_7911_cast_fp16)[name = string("op_7913_cast_fp16")];
            fp32 var_7914_epsilon_0 = const()[name = string("op_7914_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 64, 1]> var_7914_cast_fp16 = rsqrt(epsilon = var_7914_epsilon_0, x = var_7913_cast_fp16)[name = string("op_7914_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> v_69_cast_fp16 = mul(x = y_281_cast_fp16, y = var_7914_cast_fp16)[name = string("v_69_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_7916_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7916_cast_fp16")];
            tensor<int32, [4]> var_7921_begin_0 = const()[name = string("op_7921_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7921_end_0 = const()[name = string("op_7921_end_0"), val = tensor<int32, [4]>([1, 4, 64, 512])];
            tensor<bool, [4]> var_7921_end_mask_0 = const()[name = string("op_7921_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 256]> var_7921_cast_fp16 = slice_by_index(begin = var_7921_begin_0, end = var_7921_end_0, end_mask = var_7921_end_mask_0, x = q_69_cast_fp16)[name = string("op_7921_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 256]> var_7922_cast_fp16 = mul(x = var_7921_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 256]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = q_69_cast_fp16)[name = string("op_7927_cast_fp16")];
            int32 var_7929 = const()[name = string("op_7929"), val = int32(-1)];
            bool var_7930_interleave_0 = const()[name = string("op_7930_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 512]> var_7930_cast_fp16 = concat(axis = var_7929, interleave = var_7930_interleave_0, values = (var_7922_cast_fp16, var_7927_cast_fp16))[name = string("op_7930_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_7931_cast_fp16 = mul(x = var_7930_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7931_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> q_71_cast_fp16 = add(x = var_7916_cast_fp16, y = var_7931_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_7934_cast_fp16 = mul(x = k_103_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_7934_cast_fp16")];
            tensor<int32, [4]> var_7939_begin_0 = const()[name = string("op_7939_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7939_end_0 = const()[name = string("op_7939_end_0"), val = tensor<int32, [4]>([1, 1, 64, 512])];
            tensor<bool, [4]> var_7939_end_mask_0 = const()[name = string("op_7939_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 64, 256]> var_7939_cast_fp16 = slice_by_index(begin = var_7939_begin_0, end = var_7939_end_0, end_mask = var_7939_end_mask_0, x = k_103_cast_fp16)[name = string("op_7939_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 64, 256]> var_7940_cast_fp16 = mul(x = var_7939_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_7940_cast_fp16")];
            tensor<int32, [4]> var_7945_begin_0 = const()[name = string("op_7945_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7945_end_0 = const()[name = string("op_7945_end_0"), val = tensor<int32, [4]>([1, 1, 64, 256])];
            tensor<bool, [4]> var_7945_end_mask_0 = const()[name = string("op_7945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 64, 256]> var_7945_cast_fp16 = slice_by_index(begin = var_7945_begin_0, end = var_7945_end_0, end_mask = var_7945_end_mask_0, x = k_103_cast_fp16)[name = string("op_7945_cast_fp16")];
            int32 var_7947 = const()[name = string("op_7947"), val = int32(-1)];
            bool var_7948_interleave_0 = const()[name = string("op_7948_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 64, 512]> var_7948_cast_fp16 = concat(axis = var_7947, interleave = var_7948_interleave_0, values = (var_7940_cast_fp16, var_7945_cast_fp16))[name = string("op_7948_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_7949_cast_fp16 = mul(x = var_7948_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> k_105_cast_fp16 = add(x = var_7934_cast_fp16, y = var_7949_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_34_reps_0 = const()[name = string("tile_34_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 64, 512]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = k_105_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 1, 64, 512]> tile_34_cast_fp16 = tile(reps = tile_34_reps_0, x = transpose_68_cast_fp16)[name = string("tile_34_cast_fp16")];
            tensor<int32, [5]> concat_68 = const()[name = string("concat_68"), val = tensor<int32, [5]>([4, 1, 1, 64, 512])];
            tensor<fp16, [4, 1, 1, 64, 512]> reshape_68_cast_fp16 = reshape(shape = concat_68, x = tile_34_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [5]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([-1, 1, 64, 512])];
            tensor<fp16, [1, 4, 1, 64, 512]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 1, 64, 512]> reshape_69_cast_fp16 = reshape(shape = concat_69, x = transpose_69_cast_fp16)[name = string("reshape_69_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_35_reps_0 = const()[name = string("tile_35_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 64, 512]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = v_69_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 1, 64, 512]> tile_35_cast_fp16 = tile(reps = tile_35_reps_0, x = transpose_70_cast_fp16)[name = string("tile_35_cast_fp16")];
            tensor<int32, [5]> concat_70 = const()[name = string("concat_70"), val = tensor<int32, [5]>([4, 1, 1, 64, 512])];
            tensor<fp16, [4, 1, 1, 64, 512]> reshape_70_cast_fp16 = reshape(shape = concat_70, x = tile_35_cast_fp16)[name = string("reshape_70_cast_fp16")];
            tensor<int32, [5]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([-1, 1, 64, 512])];
            tensor<fp16, [1, 4, 1, 64, 512]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = reshape_70_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 1, 64, 512]> reshape_71_cast_fp16 = reshape(shape = concat_71, x = transpose_71_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7963_transpose_x_0 = const()[name = string("op_7963_transpose_x_0"), val = bool(false)];
            bool var_7963_transpose_y_0 = const()[name = string("op_7963_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 64]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = reshape_69_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 4, 64, 64]> var_7963_cast_fp16 = matmul(transpose_x = var_7963_transpose_x_0, transpose_y = var_7963_transpose_y_0, x = q_71_cast_fp16, y = transpose_113_cast_fp16)[name = string("op_7963_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_7965_cast_fp16 = add(x = var_7963_cast_fp16, y = var_2173_cast_fp16)[name = string("op_7965_cast_fp16")];
            int32 var_7966 = const()[name = string("op_7966"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_281_cast_fp16 = softmax(axis = var_7966, x = var_7965_cast_fp16)[name = string("w_281_cast_fp16")];
            bool var_7969_transpose_x_0 = const()[name = string("op_7969_transpose_x_0"), val = bool(false)];
            bool var_7969_transpose_y_0 = const()[name = string("op_7969_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 512]> v_71_cast_fp16 = transpose(perm = v_71_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 4, 64, 512]> var_7969_cast_fp16 = matmul(transpose_x = var_7969_transpose_x_0, transpose_y = var_7969_transpose_y_0, x = w_281_cast_fp16, y = v_71_cast_fp16)[name = string("op_7969_cast_fp16")];
            tensor<int32, [4]> var_7972_perm_0 = const()[name = string("op_7972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7976 = const()[name = string("op_7976"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 64, 4, 512]> var_7972_cast_fp16 = transpose(perm = var_7972_perm_0, x = var_7969_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 64, 2048]> input_207_cast_fp16 = reshape(shape = var_7976, x = var_7972_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99570880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619520))))[name = string("tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7981_cast_fp16 = abs(x = linear_157_cast_fp16)[name = string("op_7981_cast_fp16")];
            bool var_7984 = const()[name = string("op_7984"), val = bool(true)];
            tensor<int32, [1]> var_7985_axes_0 = const()[name = string("op_7985_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_7985_cast_fp16 = reduce_max(axes = var_7985_axes_0, keep_dims = var_7984, x = var_7981_cast_fp16)[name = string("op_7985_cast_fp16")];
            fp16 var_7986_to_fp16 = const()[name = string("op_7986_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_283_cast_fp16 = maximum(x = var_7985_cast_fp16, y = var_7986_to_fp16)[name = string("s_283_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_283_cast_fp16 = real_div(x = linear_157_cast_fp16, y = s_283_cast_fp16)[name = string("y_283_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_7989_cast_fp16 = mul(x = y_283_cast_fp16, y = y_283_cast_fp16)[name = string("op_7989_cast_fp16")];
            tensor<int32, [1]> var_7994_axes_0 = const()[name = string("op_7994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7994_keep_dims_0 = const()[name = string("op_7994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7994_cast_fp16 = reduce_mean(axes = var_7994_axes_0, keep_dims = var_7994_keep_dims_0, x = var_7989_cast_fp16)[name = string("op_7994_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_7995_cast_fp16 = mul(x = s_283_cast_fp16, y = s_283_cast_fp16)[name = string("op_7995_cast_fp16")];
            fp32 var_7996_epsilon_0 = const()[name = string("op_7996_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_7996_cast_fp16 = inverse(epsilon = var_7996_epsilon_0, x = var_7995_cast_fp16)[name = string("op_7996_cast_fp16")];
            fp16 var_7997_to_fp16 = const()[name = string("op_7997_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_7998_cast_fp16 = mul(x = var_7996_cast_fp16, y = var_7997_to_fp16)[name = string("op_7998_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8000_cast_fp16 = add(x = var_7994_cast_fp16, y = var_7998_cast_fp16)[name = string("op_8000_cast_fp16")];
            fp32 var_8001_epsilon_0 = const()[name = string("op_8001_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8001_cast_fp16 = rsqrt(epsilon = var_8001_epsilon_0, x = var_8000_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_247_cast_fp16 = mul(x = y_283_cast_fp16, y = var_8001_cast_fp16)[name = string("out_247_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100620608)))];
            tensor<fp16, [1, 64, 512]> var_8003_cast_fp16 = mul(x = out_247_cast_fp16, y = tm_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("op_8003_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_317_cast_fp16 = add(x = x_307_cast_fp16, y = var_8003_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8006_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_8006_cast_fp16")];
            bool var_8009 = const()[name = string("op_8009"), val = bool(true)];
            tensor<int32, [1]> var_8010_axes_0 = const()[name = string("op_8010_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8010_cast_fp16 = reduce_max(axes = var_8010_axes_0, keep_dims = var_8009, x = var_8006_cast_fp16)[name = string("op_8010_cast_fp16")];
            fp16 var_8011_to_fp16 = const()[name = string("op_8011_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_285_cast_fp16 = maximum(x = var_8010_cast_fp16, y = var_8011_to_fp16)[name = string("s_285_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_285_cast_fp16 = real_div(x = x_317_cast_fp16, y = s_285_cast_fp16)[name = string("y_285_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8014_cast_fp16 = mul(x = y_285_cast_fp16, y = y_285_cast_fp16)[name = string("op_8014_cast_fp16")];
            tensor<int32, [1]> var_8019_axes_0 = const()[name = string("op_8019_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8019_keep_dims_0 = const()[name = string("op_8019_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8019_cast_fp16 = reduce_mean(axes = var_8019_axes_0, keep_dims = var_8019_keep_dims_0, x = var_8014_cast_fp16)[name = string("op_8019_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8020_cast_fp16 = mul(x = s_285_cast_fp16, y = s_285_cast_fp16)[name = string("op_8020_cast_fp16")];
            fp32 var_8021_epsilon_0 = const()[name = string("op_8021_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8021_cast_fp16 = inverse(epsilon = var_8021_epsilon_0, x = var_8020_cast_fp16)[name = string("op_8021_cast_fp16")];
            fp16 var_8022_to_fp16 = const()[name = string("op_8022_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8023_cast_fp16 = mul(x = var_8021_cast_fp16, y = var_8022_to_fp16)[name = string("op_8023_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8025_cast_fp16 = add(x = var_8019_cast_fp16, y = var_8023_cast_fp16)[name = string("op_8025_cast_fp16")];
            fp32 var_8026_epsilon_0 = const()[name = string("op_8026_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8026_cast_fp16 = rsqrt(epsilon = var_8026_epsilon_0, x = var_8025_cast_fp16)[name = string("op_8026_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_249_cast_fp16 = mul(x = y_285_cast_fp16, y = var_8026_cast_fp16)[name = string("out_249_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621696)))];
            tensor<fp16, [1, 64, 512]> input_209_cast_fp16 = mul(x = out_249_cast_fp16, y = tm_layers_17_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100622784))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101671424))))[name = string("tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_158_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_158_cast_fp16")];
            string var_8033_mode_0 = const()[name = string("op_8033_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_8033_cast_fp16 = gelu(mode = var_8033_mode_0, x = linear_158_cast_fp16)[name = string("op_8033_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101675584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102724224))))[name = string("tm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_159_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_211_cast_fp16 = mul(x = var_8033_cast_fp16, y = linear_159_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102728384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103777024))))[name = string("tm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8041_cast_fp16 = abs(x = linear_160_cast_fp16)[name = string("op_8041_cast_fp16")];
            bool var_8044 = const()[name = string("op_8044"), val = bool(true)];
            tensor<int32, [1]> var_8045_axes_0 = const()[name = string("op_8045_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8045_cast_fp16 = reduce_max(axes = var_8045_axes_0, keep_dims = var_8044, x = var_8041_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp16 var_8046_to_fp16 = const()[name = string("op_8046_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_287_cast_fp16 = maximum(x = var_8045_cast_fp16, y = var_8046_to_fp16)[name = string("s_287_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_287_cast_fp16 = real_div(x = linear_160_cast_fp16, y = s_287_cast_fp16)[name = string("y_287_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8049_cast_fp16 = mul(x = y_287_cast_fp16, y = y_287_cast_fp16)[name = string("op_8049_cast_fp16")];
            tensor<int32, [1]> var_8054_axes_0 = const()[name = string("op_8054_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8054_keep_dims_0 = const()[name = string("op_8054_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8054_cast_fp16 = reduce_mean(axes = var_8054_axes_0, keep_dims = var_8054_keep_dims_0, x = var_8049_cast_fp16)[name = string("op_8054_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8055_cast_fp16 = mul(x = s_287_cast_fp16, y = s_287_cast_fp16)[name = string("op_8055_cast_fp16")];
            fp32 var_8056_epsilon_0 = const()[name = string("op_8056_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8056_cast_fp16 = inverse(epsilon = var_8056_epsilon_0, x = var_8055_cast_fp16)[name = string("op_8056_cast_fp16")];
            fp16 var_8057_to_fp16 = const()[name = string("op_8057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8058_cast_fp16 = mul(x = var_8056_cast_fp16, y = var_8057_to_fp16)[name = string("op_8058_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8060_cast_fp16 = add(x = var_8054_cast_fp16, y = var_8058_cast_fp16)[name = string("op_8060_cast_fp16")];
            fp32 var_8061_epsilon_0 = const()[name = string("op_8061_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8061_cast_fp16 = rsqrt(epsilon = var_8061_epsilon_0, x = var_8060_cast_fp16)[name = string("op_8061_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_251_cast_fp16 = mul(x = y_287_cast_fp16, y = var_8061_cast_fp16)[name = string("out_251_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103778112)))];
            tensor<fp16, [1, 64, 512]> var_8063_cast_fp16 = mul(x = out_251_cast_fp16, y = tm_layers_17_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_213_cast_fp16 = add(x = x_317_cast_fp16, y = var_8063_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103779200))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041408))))[name = string("tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_161_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_161_cast_fp16")];
            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_321_cast_fp16 = gelu(mode = x_321_mode_0, x = linear_161_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [4]> var_8083_begin_0 = const()[name = string("op_8083_begin_0"), val = tensor<int32, [4]>([0, 0, 17, 0])];
            tensor<int32, [4]> var_8083_end_0 = const()[name = string("op_8083_end_0"), val = tensor<int32, [4]>([1, 64, 18, 512])];
            tensor<bool, [4]> var_8083_end_mask_0 = const()[name = string("op_8083_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8083_squeeze_mask_0 = const()[name = string("op_8083_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_8083_cast_fp16 = slice_by_index(begin = var_8083_begin_0, end = var_8083_end_0, end_mask = var_8083_end_mask_0, squeeze_mask = var_8083_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8083_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_215_cast_fp16 = mul(x = x_321_cast_fp16, y = var_8083_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104042496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104304704))))[name = string("tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8093_cast_fp16 = abs(x = linear_162_cast_fp16)[name = string("op_8093_cast_fp16")];
            bool var_8096 = const()[name = string("op_8096"), val = bool(true)];
            tensor<int32, [1]> var_8097_axes_0 = const()[name = string("op_8097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8097_cast_fp16 = reduce_max(axes = var_8097_axes_0, keep_dims = var_8096, x = var_8093_cast_fp16)[name = string("op_8097_cast_fp16")];
            fp16 var_8098_to_fp16 = const()[name = string("op_8098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_289_cast_fp16 = maximum(x = var_8097_cast_fp16, y = var_8098_to_fp16)[name = string("s_289_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_289_cast_fp16 = real_div(x = linear_162_cast_fp16, y = s_289_cast_fp16)[name = string("y_289_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8101_cast_fp16 = mul(x = y_289_cast_fp16, y = y_289_cast_fp16)[name = string("op_8101_cast_fp16")];
            tensor<int32, [1]> var_8106_axes_0 = const()[name = string("op_8106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8106_keep_dims_0 = const()[name = string("op_8106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8106_cast_fp16 = reduce_mean(axes = var_8106_axes_0, keep_dims = var_8106_keep_dims_0, x = var_8101_cast_fp16)[name = string("op_8106_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8107_cast_fp16 = mul(x = s_289_cast_fp16, y = s_289_cast_fp16)[name = string("op_8107_cast_fp16")];
            fp32 var_8108_epsilon_0 = const()[name = string("op_8108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8108_cast_fp16 = inverse(epsilon = var_8108_epsilon_0, x = var_8107_cast_fp16)[name = string("op_8108_cast_fp16")];
            fp16 var_8109_to_fp16 = const()[name = string("op_8109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8110_cast_fp16 = mul(x = var_8108_cast_fp16, y = var_8109_to_fp16)[name = string("op_8110_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8112_cast_fp16 = add(x = var_8106_cast_fp16, y = var_8110_cast_fp16)[name = string("op_8112_cast_fp16")];
            fp32 var_8113_epsilon_0 = const()[name = string("op_8113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8113_cast_fp16 = rsqrt(epsilon = var_8113_epsilon_0, x = var_8112_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_253_cast_fp16 = mul(x = y_289_cast_fp16, y = var_8113_cast_fp16)[name = string("out_253_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104305792)))];
            tensor<fp16, [1, 64, 512]> var_8115_cast_fp16 = mul(x = out_253_cast_fp16, y = tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8115_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_35_cast_fp16 = add(x = input_213_cast_fp16, y = var_8115_cast_fp16)[name = string("h_35_cast_fp16")];
            tensor<fp16, [1]> tm_layers_17_layer_scalar_to_fp16 = const()[name = string("tm_layers_17_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e6p-1])];
            tensor<fp16, [1, 64, 512]> x_325_cast_fp16 = mul(x = h_35_cast_fp16, y = tm_layers_17_layer_scalar_to_fp16)[name = string("x_325_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8119_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_8119_cast_fp16")];
            bool var_8122 = const()[name = string("op_8122"), val = bool(true)];
            tensor<int32, [1]> var_8123_axes_0 = const()[name = string("op_8123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8123_cast_fp16 = reduce_max(axes = var_8123_axes_0, keep_dims = var_8122, x = var_8119_cast_fp16)[name = string("op_8123_cast_fp16")];
            fp16 var_8124_to_fp16 = const()[name = string("op_8124_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_291_cast_fp16 = maximum(x = var_8123_cast_fp16, y = var_8124_to_fp16)[name = string("s_291_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_291_cast_fp16 = real_div(x = x_325_cast_fp16, y = s_291_cast_fp16)[name = string("y_291_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8127_cast_fp16 = mul(x = y_291_cast_fp16, y = y_291_cast_fp16)[name = string("op_8127_cast_fp16")];
            tensor<int32, [1]> var_8132_axes_0 = const()[name = string("op_8132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8132_keep_dims_0 = const()[name = string("op_8132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8132_cast_fp16 = reduce_mean(axes = var_8132_axes_0, keep_dims = var_8132_keep_dims_0, x = var_8127_cast_fp16)[name = string("op_8132_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8133_cast_fp16 = mul(x = s_291_cast_fp16, y = s_291_cast_fp16)[name = string("op_8133_cast_fp16")];
            fp32 var_8134_epsilon_0 = const()[name = string("op_8134_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8134_cast_fp16 = inverse(epsilon = var_8134_epsilon_0, x = var_8133_cast_fp16)[name = string("op_8134_cast_fp16")];
            fp16 var_8135_to_fp16 = const()[name = string("op_8135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = string("op_8136_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8138_cast_fp16 = add(x = var_8132_cast_fp16, y = var_8136_cast_fp16)[name = string("op_8138_cast_fp16")];
            fp32 var_8139_epsilon_0 = const()[name = string("op_8139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8139_cast_fp16 = rsqrt(epsilon = var_8139_epsilon_0, x = var_8138_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_255_cast_fp16 = mul(x = y_291_cast_fp16, y = var_8139_cast_fp16)[name = string("out_255_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104306880)))];
            tensor<fp16, [1, 64, 512]> input_217_cast_fp16 = mul(x = out_255_cast_fp16, y = tm_layers_18_input_layernorm_weight_to_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104307968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104832320))))[name = string("tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8149 = const()[name = string("op_8149"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_8150_cast_fp16 = reshape(shape = var_8149, x = linear_163_cast_fp16)[name = string("op_8150_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104834432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105096640))))[name = string("tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [4]> var_8161 = const()[name = string("op_8161"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_8162_cast_fp16 = reshape(shape = var_8161, x = linear_164_cast_fp16)[name = string("op_8162_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105097728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105359936))))[name = string("tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<int32, [4]> var_8173 = const()[name = string("op_8173"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_8174_cast_fp16 = reshape(shape = var_8173, x = linear_165_cast_fp16)[name = string("op_8174_cast_fp16")];
            tensor<int32, [4]> x_331_perm_0 = const()[name = string("x_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_8150_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 4, 64, 256]> var_8178_cast_fp16 = abs(x = x_327_cast_fp16)[name = string("op_8178_cast_fp16")];
            bool var_8181 = const()[name = string("op_8181"), val = bool(true)];
            tensor<int32, [1]> var_8182_axes_0 = const()[name = string("op_8182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_8182_cast_fp16 = reduce_max(axes = var_8182_axes_0, keep_dims = var_8181, x = var_8178_cast_fp16)[name = string("op_8182_cast_fp16")];
            fp16 var_8183_to_fp16 = const()[name = string("op_8183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_293_cast_fp16 = maximum(x = var_8182_cast_fp16, y = var_8183_to_fp16)[name = string("s_293_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_293_cast_fp16 = real_div(x = x_327_cast_fp16, y = s_293_cast_fp16)[name = string("y_293_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8186_cast_fp16 = mul(x = y_293_cast_fp16, y = y_293_cast_fp16)[name = string("op_8186_cast_fp16")];
            tensor<int32, [1]> var_8191_axes_0 = const()[name = string("op_8191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8191_keep_dims_0 = const()[name = string("op_8191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_8191_cast_fp16 = reduce_mean(axes = var_8191_axes_0, keep_dims = var_8191_keep_dims_0, x = var_8186_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_8192_cast_fp16 = mul(x = s_293_cast_fp16, y = s_293_cast_fp16)[name = string("op_8192_cast_fp16")];
            fp32 var_8193_epsilon_0 = const()[name = string("op_8193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_8193_cast_fp16 = inverse(epsilon = var_8193_epsilon_0, x = var_8192_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_8197_cast_fp16 = add(x = var_8191_cast_fp16, y = var_8195_cast_fp16)[name = string("op_8197_cast_fp16")];
            fp32 var_8198_epsilon_0 = const()[name = string("op_8198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_8198_cast_fp16 = rsqrt(epsilon = var_8198_epsilon_0, x = var_8197_cast_fp16)[name = string("op_8198_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_257_cast_fp16 = mul(x = y_293_cast_fp16, y = var_8198_cast_fp16)[name = string("out_257_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_73_cast_fp16 = mul(x = out_257_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_73_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_8162_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 64, 256]> var_8201_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_8201_cast_fp16")];
            bool var_8204 = const()[name = string("op_8204"), val = bool(true)];
            tensor<int32, [1]> var_8205_axes_0 = const()[name = string("op_8205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_8205_cast_fp16 = reduce_max(axes = var_8205_axes_0, keep_dims = var_8204, x = var_8201_cast_fp16)[name = string("op_8205_cast_fp16")];
            fp16 var_8206_to_fp16 = const()[name = string("op_8206_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_295_cast_fp16 = maximum(x = var_8205_cast_fp16, y = var_8206_to_fp16)[name = string("s_295_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_295_cast_fp16 = real_div(x = x_329_cast_fp16, y = s_295_cast_fp16)[name = string("y_295_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8209_cast_fp16 = mul(x = y_295_cast_fp16, y = y_295_cast_fp16)[name = string("op_8209_cast_fp16")];
            tensor<int32, [1]> var_8214_axes_0 = const()[name = string("op_8214_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8214_keep_dims_0 = const()[name = string("op_8214_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_8214_cast_fp16 = reduce_mean(axes = var_8214_axes_0, keep_dims = var_8214_keep_dims_0, x = var_8209_cast_fp16)[name = string("op_8214_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8215_cast_fp16 = mul(x = s_295_cast_fp16, y = s_295_cast_fp16)[name = string("op_8215_cast_fp16")];
            fp32 var_8216_epsilon_0 = const()[name = string("op_8216_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_8216_cast_fp16 = inverse(epsilon = var_8216_epsilon_0, x = var_8215_cast_fp16)[name = string("op_8216_cast_fp16")];
            fp16 var_8217_to_fp16 = const()[name = string("op_8217_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_8218_cast_fp16 = mul(x = var_8216_cast_fp16, y = var_8217_to_fp16)[name = string("op_8218_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8220_cast_fp16 = add(x = var_8214_cast_fp16, y = var_8218_cast_fp16)[name = string("op_8220_cast_fp16")];
            fp32 var_8221_epsilon_0 = const()[name = string("op_8221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_8221_cast_fp16 = rsqrt(epsilon = var_8221_epsilon_0, x = var_8220_cast_fp16)[name = string("op_8221_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_259_cast_fp16 = mul(x = y_295_cast_fp16, y = var_8221_cast_fp16)[name = string("out_259_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_109_cast_fp16 = mul(x = out_259_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_109_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_331_cast_fp16 = transpose(perm = x_331_perm_0, x = var_8174_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 64, 256]> var_8224_cast_fp16 = abs(x = x_331_cast_fp16)[name = string("op_8224_cast_fp16")];
            bool var_8227 = const()[name = string("op_8227"), val = bool(true)];
            tensor<int32, [1]> var_8228_axes_0 = const()[name = string("op_8228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_8228_cast_fp16 = reduce_max(axes = var_8228_axes_0, keep_dims = var_8227, x = var_8224_cast_fp16)[name = string("op_8228_cast_fp16")];
            fp16 var_8229_to_fp16 = const()[name = string("op_8229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_297_cast_fp16 = maximum(x = var_8228_cast_fp16, y = var_8229_to_fp16)[name = string("s_297_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_297_cast_fp16 = real_div(x = x_331_cast_fp16, y = s_297_cast_fp16)[name = string("y_297_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8232_cast_fp16 = mul(x = y_297_cast_fp16, y = y_297_cast_fp16)[name = string("op_8232_cast_fp16")];
            tensor<int32, [1]> var_8237_axes_0 = const()[name = string("op_8237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8237_keep_dims_0 = const()[name = string("op_8237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_8237_cast_fp16 = reduce_mean(axes = var_8237_axes_0, keep_dims = var_8237_keep_dims_0, x = var_8232_cast_fp16)[name = string("op_8237_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8238_cast_fp16 = mul(x = s_297_cast_fp16, y = s_297_cast_fp16)[name = string("op_8238_cast_fp16")];
            fp32 var_8239_epsilon_0 = const()[name = string("op_8239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_8239_cast_fp16 = inverse(epsilon = var_8239_epsilon_0, x = var_8238_cast_fp16)[name = string("op_8239_cast_fp16")];
            fp16 var_8240_to_fp16 = const()[name = string("op_8240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_8241_cast_fp16 = mul(x = var_8239_cast_fp16, y = var_8240_to_fp16)[name = string("op_8241_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8243_cast_fp16 = add(x = var_8237_cast_fp16, y = var_8241_cast_fp16)[name = string("op_8243_cast_fp16")];
            fp32 var_8244_epsilon_0 = const()[name = string("op_8244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_8244_cast_fp16 = rsqrt(epsilon = var_8244_epsilon_0, x = var_8243_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_73_cast_fp16 = mul(x = y_297_cast_fp16, y = var_8244_cast_fp16)[name = string("v_73_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8246_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8246_cast_fp16")];
            tensor<int32, [4]> var_8251_begin_0 = const()[name = string("op_8251_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8251_end_0 = const()[name = string("op_8251_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_8251_end_mask_0 = const()[name = string("op_8251_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_8251_cast_fp16 = slice_by_index(begin = var_8251_begin_0, end = var_8251_end_0, end_mask = var_8251_end_mask_0, x = q_73_cast_fp16)[name = string("op_8251_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_8252_cast_fp16 = mul(x = var_8251_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8257_begin_0 = const()[name = string("op_8257_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8257_end_0 = const()[name = string("op_8257_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_8257_end_mask_0 = const()[name = string("op_8257_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_8257_cast_fp16 = slice_by_index(begin = var_8257_begin_0, end = var_8257_end_0, end_mask = var_8257_end_mask_0, x = q_73_cast_fp16)[name = string("op_8257_cast_fp16")];
            int32 var_8259 = const()[name = string("op_8259"), val = int32(-1)];
            bool var_8260_interleave_0 = const()[name = string("op_8260_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_8260_cast_fp16 = concat(axis = var_8259, interleave = var_8260_interleave_0, values = (var_8252_cast_fp16, var_8257_cast_fp16))[name = string("op_8260_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8261_cast_fp16 = mul(x = var_8260_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8261_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_75_cast_fp16 = add(x = var_8246_cast_fp16, y = var_8261_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8264_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8264_cast_fp16")];
            tensor<int32, [4]> var_8269_begin_0 = const()[name = string("op_8269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8269_end_0 = const()[name = string("op_8269_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_8269_end_mask_0 = const()[name = string("op_8269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_8269_cast_fp16 = slice_by_index(begin = var_8269_begin_0, end = var_8269_end_0, end_mask = var_8269_end_mask_0, x = k_109_cast_fp16)[name = string("op_8269_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_8270_cast_fp16 = mul(x = var_8269_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_8270_cast_fp16")];
            tensor<int32, [4]> var_8275_begin_0 = const()[name = string("op_8275_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8275_end_0 = const()[name = string("op_8275_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_8275_end_mask_0 = const()[name = string("op_8275_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_8275_cast_fp16 = slice_by_index(begin = var_8275_begin_0, end = var_8275_end_0, end_mask = var_8275_end_mask_0, x = k_109_cast_fp16)[name = string("op_8275_cast_fp16")];
            int32 var_8277 = const()[name = string("op_8277"), val = int32(-1)];
            bool var_8278_interleave_0 = const()[name = string("op_8278_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_8278_cast_fp16 = concat(axis = var_8277, interleave = var_8278_interleave_0, values = (var_8270_cast_fp16, var_8275_cast_fp16))[name = string("op_8278_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8279_cast_fp16 = mul(x = var_8278_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8279_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_111_cast_fp16 = add(x = var_8264_cast_fp16, y = var_8279_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_36_reps_0 = const()[name = string("tile_36_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_72_cast_fp16 = transpose(perm = transpose_72_perm_0, x = k_111_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 1, 64, 256]> tile_36_cast_fp16 = tile(reps = tile_36_reps_0, x = transpose_72_cast_fp16)[name = string("tile_36_cast_fp16")];
            tensor<int32, [5]> concat_72 = const()[name = string("concat_72"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_72_cast_fp16 = reshape(shape = concat_72, x = tile_36_cast_fp16)[name = string("reshape_72_cast_fp16")];
            tensor<int32, [5]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_73 = const()[name = string("concat_73"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = reshape_72_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 1, 64, 256]> reshape_73_cast_fp16 = reshape(shape = concat_73, x = transpose_73_cast_fp16)[name = string("reshape_73_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_37_reps_0 = const()[name = string("tile_37_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = v_73_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 1, 64, 256]> tile_37_cast_fp16 = tile(reps = tile_37_reps_0, x = transpose_74_cast_fp16)[name = string("tile_37_cast_fp16")];
            tensor<int32, [5]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_74_cast_fp16 = reshape(shape = concat_74, x = tile_37_cast_fp16)[name = string("reshape_74_cast_fp16")];
            tensor<int32, [5]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = reshape_74_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 1, 64, 256]> reshape_75_cast_fp16 = reshape(shape = concat_75, x = transpose_75_cast_fp16)[name = string("reshape_75_cast_fp16")];
            tensor<int32, [4]> v_75_perm_0 = const()[name = string("v_75_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8293_transpose_x_0 = const()[name = string("op_8293_transpose_x_0"), val = bool(false)];
            bool var_8293_transpose_y_0 = const()[name = string("op_8293_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = reshape_73_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 4, 64, 64]> var_8293_cast_fp16 = matmul(transpose_x = var_8293_transpose_x_0, transpose_y = var_8293_transpose_y_0, x = q_75_cast_fp16, y = transpose_114_cast_fp16)[name = string("op_8293_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_8295_cast_fp16 = add(x = var_8293_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8295_cast_fp16")];
            int32 var_8296 = const()[name = string("op_8296"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_297_cast_fp16 = softmax(axis = var_8296, x = var_8295_cast_fp16)[name = string("w_297_cast_fp16")];
            bool var_8299_transpose_x_0 = const()[name = string("op_8299_transpose_x_0"), val = bool(false)];
            bool var_8299_transpose_y_0 = const()[name = string("op_8299_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_75_cast_fp16 = transpose(perm = v_75_perm_0, x = reshape_75_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 4, 64, 256]> var_8299_cast_fp16 = matmul(transpose_x = var_8299_transpose_x_0, transpose_y = var_8299_transpose_y_0, x = w_297_cast_fp16, y = v_75_cast_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [4]> var_8302_perm_0 = const()[name = string("op_8302_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8306 = const()[name = string("op_8306"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_8302_cast_fp16 = transpose(perm = var_8302_perm_0, x = var_8299_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 64, 1024]> input_219_cast_fp16 = reshape(shape = var_8306, x = var_8302_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105361024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105885376))))[name = string("tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8311_cast_fp16 = abs(x = linear_166_cast_fp16)[name = string("op_8311_cast_fp16")];
            bool var_8314 = const()[name = string("op_8314"), val = bool(true)];
            tensor<int32, [1]> var_8315_axes_0 = const()[name = string("op_8315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8315_cast_fp16 = reduce_max(axes = var_8315_axes_0, keep_dims = var_8314, x = var_8311_cast_fp16)[name = string("op_8315_cast_fp16")];
            fp16 var_8316_to_fp16 = const()[name = string("op_8316_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_299_cast_fp16 = maximum(x = var_8315_cast_fp16, y = var_8316_to_fp16)[name = string("s_299_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_299_cast_fp16 = real_div(x = linear_166_cast_fp16, y = s_299_cast_fp16)[name = string("y_299_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8319_cast_fp16 = mul(x = y_299_cast_fp16, y = y_299_cast_fp16)[name = string("op_8319_cast_fp16")];
            tensor<int32, [1]> var_8324_axes_0 = const()[name = string("op_8324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8324_keep_dims_0 = const()[name = string("op_8324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8324_cast_fp16 = reduce_mean(axes = var_8324_axes_0, keep_dims = var_8324_keep_dims_0, x = var_8319_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8325_cast_fp16 = mul(x = s_299_cast_fp16, y = s_299_cast_fp16)[name = string("op_8325_cast_fp16")];
            fp32 var_8326_epsilon_0 = const()[name = string("op_8326_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8326_cast_fp16 = inverse(epsilon = var_8326_epsilon_0, x = var_8325_cast_fp16)[name = string("op_8326_cast_fp16")];
            fp16 var_8327_to_fp16 = const()[name = string("op_8327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8328_cast_fp16 = mul(x = var_8326_cast_fp16, y = var_8327_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8330_cast_fp16 = add(x = var_8324_cast_fp16, y = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_261_cast_fp16 = mul(x = y_299_cast_fp16, y = var_8331_cast_fp16)[name = string("out_261_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105886464)))];
            tensor<fp16, [1, 64, 512]> var_8333_cast_fp16 = mul(x = out_261_cast_fp16, y = tm_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("op_8333_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = var_8333_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8336_cast_fp16 = abs(x = x_335_cast_fp16)[name = string("op_8336_cast_fp16")];
            bool var_8339 = const()[name = string("op_8339"), val = bool(true)];
            tensor<int32, [1]> var_8340_axes_0 = const()[name = string("op_8340_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8340_cast_fp16 = reduce_max(axes = var_8340_axes_0, keep_dims = var_8339, x = var_8336_cast_fp16)[name = string("op_8340_cast_fp16")];
            fp16 var_8341_to_fp16 = const()[name = string("op_8341_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_301_cast_fp16 = maximum(x = var_8340_cast_fp16, y = var_8341_to_fp16)[name = string("s_301_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_301_cast_fp16 = real_div(x = x_335_cast_fp16, y = s_301_cast_fp16)[name = string("y_301_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8344_cast_fp16 = mul(x = y_301_cast_fp16, y = y_301_cast_fp16)[name = string("op_8344_cast_fp16")];
            tensor<int32, [1]> var_8349_axes_0 = const()[name = string("op_8349_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8349_keep_dims_0 = const()[name = string("op_8349_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8349_cast_fp16 = reduce_mean(axes = var_8349_axes_0, keep_dims = var_8349_keep_dims_0, x = var_8344_cast_fp16)[name = string("op_8349_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8350_cast_fp16 = mul(x = s_301_cast_fp16, y = s_301_cast_fp16)[name = string("op_8350_cast_fp16")];
            fp32 var_8351_epsilon_0 = const()[name = string("op_8351_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8351_cast_fp16 = inverse(epsilon = var_8351_epsilon_0, x = var_8350_cast_fp16)[name = string("op_8351_cast_fp16")];
            fp16 var_8352_to_fp16 = const()[name = string("op_8352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8353_cast_fp16 = mul(x = var_8351_cast_fp16, y = var_8352_to_fp16)[name = string("op_8353_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8355_cast_fp16 = add(x = var_8349_cast_fp16, y = var_8353_cast_fp16)[name = string("op_8355_cast_fp16")];
            fp32 var_8356_epsilon_0 = const()[name = string("op_8356_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8356_cast_fp16 = rsqrt(epsilon = var_8356_epsilon_0, x = var_8355_cast_fp16)[name = string("op_8356_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_263_cast_fp16 = mul(x = y_301_cast_fp16, y = var_8356_cast_fp16)[name = string("out_263_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105887552)))];
            tensor<fp16, [1, 64, 512]> input_221_cast_fp16 = mul(x = out_263_cast_fp16, y = tm_layers_18_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105888640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106937280))))[name = string("tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_167_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_167_cast_fp16")];
            string var_8363_mode_0 = const()[name = string("op_8363_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_8363_cast_fp16 = gelu(mode = var_8363_mode_0, x = linear_167_cast_fp16)[name = string("op_8363_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106941440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107990080))))[name = string("tm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_168_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_223_cast_fp16 = mul(x = var_8363_cast_fp16, y = linear_168_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107994240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109042880))))[name = string("tm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8371_cast_fp16 = abs(x = linear_169_cast_fp16)[name = string("op_8371_cast_fp16")];
            bool var_8374 = const()[name = string("op_8374"), val = bool(true)];
            tensor<int32, [1]> var_8375_axes_0 = const()[name = string("op_8375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8375_cast_fp16 = reduce_max(axes = var_8375_axes_0, keep_dims = var_8374, x = var_8371_cast_fp16)[name = string("op_8375_cast_fp16")];
            fp16 var_8376_to_fp16 = const()[name = string("op_8376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_303_cast_fp16 = maximum(x = var_8375_cast_fp16, y = var_8376_to_fp16)[name = string("s_303_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_303_cast_fp16 = real_div(x = linear_169_cast_fp16, y = s_303_cast_fp16)[name = string("y_303_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8379_cast_fp16 = mul(x = y_303_cast_fp16, y = y_303_cast_fp16)[name = string("op_8379_cast_fp16")];
            tensor<int32, [1]> var_8384_axes_0 = const()[name = string("op_8384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8384_keep_dims_0 = const()[name = string("op_8384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8384_cast_fp16 = reduce_mean(axes = var_8384_axes_0, keep_dims = var_8384_keep_dims_0, x = var_8379_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8385_cast_fp16 = mul(x = s_303_cast_fp16, y = s_303_cast_fp16)[name = string("op_8385_cast_fp16")];
            fp32 var_8386_epsilon_0 = const()[name = string("op_8386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8386_cast_fp16 = inverse(epsilon = var_8386_epsilon_0, x = var_8385_cast_fp16)[name = string("op_8386_cast_fp16")];
            fp16 var_8387_to_fp16 = const()[name = string("op_8387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8388_cast_fp16 = mul(x = var_8386_cast_fp16, y = var_8387_to_fp16)[name = string("op_8388_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8390_cast_fp16 = add(x = var_8384_cast_fp16, y = var_8388_cast_fp16)[name = string("op_8390_cast_fp16")];
            fp32 var_8391_epsilon_0 = const()[name = string("op_8391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8391_cast_fp16 = rsqrt(epsilon = var_8391_epsilon_0, x = var_8390_cast_fp16)[name = string("op_8391_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_265_cast_fp16 = mul(x = y_303_cast_fp16, y = var_8391_cast_fp16)[name = string("out_265_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109043968)))];
            tensor<fp16, [1, 64, 512]> var_8393_cast_fp16 = mul(x = out_265_cast_fp16, y = tm_layers_18_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8393_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_225_cast_fp16 = add(x = x_335_cast_fp16, y = var_8393_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109045056))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109307264))))[name = string("tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_170_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_170_cast_fp16")];
            string x_339_mode_0 = const()[name = string("x_339_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_339_cast_fp16 = gelu(mode = x_339_mode_0, x = linear_170_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [4]> var_8413_begin_0 = const()[name = string("op_8413_begin_0"), val = tensor<int32, [4]>([0, 0, 18, 0])];
            tensor<int32, [4]> var_8413_end_0 = const()[name = string("op_8413_end_0"), val = tensor<int32, [4]>([1, 64, 19, 512])];
            tensor<bool, [4]> var_8413_end_mask_0 = const()[name = string("op_8413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8413_squeeze_mask_0 = const()[name = string("op_8413_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_8413_cast_fp16 = slice_by_index(begin = var_8413_begin_0, end = var_8413_end_0, end_mask = var_8413_end_mask_0, squeeze_mask = var_8413_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_227_cast_fp16 = mul(x = x_339_cast_fp16, y = var_8413_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109308352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109570560))))[name = string("tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8423_cast_fp16 = abs(x = linear_171_cast_fp16)[name = string("op_8423_cast_fp16")];
            bool var_8426 = const()[name = string("op_8426"), val = bool(true)];
            tensor<int32, [1]> var_8427_axes_0 = const()[name = string("op_8427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8427_cast_fp16 = reduce_max(axes = var_8427_axes_0, keep_dims = var_8426, x = var_8423_cast_fp16)[name = string("op_8427_cast_fp16")];
            fp16 var_8428_to_fp16 = const()[name = string("op_8428_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_305_cast_fp16 = maximum(x = var_8427_cast_fp16, y = var_8428_to_fp16)[name = string("s_305_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_305_cast_fp16 = real_div(x = linear_171_cast_fp16, y = s_305_cast_fp16)[name = string("y_305_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8431_cast_fp16 = mul(x = y_305_cast_fp16, y = y_305_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<int32, [1]> var_8436_axes_0 = const()[name = string("op_8436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8436_keep_dims_0 = const()[name = string("op_8436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8436_cast_fp16 = reduce_mean(axes = var_8436_axes_0, keep_dims = var_8436_keep_dims_0, x = var_8431_cast_fp16)[name = string("op_8436_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8437_cast_fp16 = mul(x = s_305_cast_fp16, y = s_305_cast_fp16)[name = string("op_8437_cast_fp16")];
            fp32 var_8438_epsilon_0 = const()[name = string("op_8438_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8438_cast_fp16 = inverse(epsilon = var_8438_epsilon_0, x = var_8437_cast_fp16)[name = string("op_8438_cast_fp16")];
            fp16 var_8439_to_fp16 = const()[name = string("op_8439_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8440_cast_fp16 = mul(x = var_8438_cast_fp16, y = var_8439_to_fp16)[name = string("op_8440_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8442_cast_fp16 = add(x = var_8436_cast_fp16, y = var_8440_cast_fp16)[name = string("op_8442_cast_fp16")];
            fp32 var_8443_epsilon_0 = const()[name = string("op_8443_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8443_cast_fp16 = rsqrt(epsilon = var_8443_epsilon_0, x = var_8442_cast_fp16)[name = string("op_8443_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_267_cast_fp16 = mul(x = y_305_cast_fp16, y = var_8443_cast_fp16)[name = string("out_267_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109571648)))];
            tensor<fp16, [1, 64, 512]> var_8445_cast_fp16 = mul(x = out_267_cast_fp16, y = tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8445_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_37_cast_fp16 = add(x = input_225_cast_fp16, y = var_8445_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<fp16, [1]> tm_layers_18_layer_scalar_to_fp16 = const()[name = string("tm_layers_18_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.68p-1])];
            tensor<fp16, [1, 64, 512]> x_343_cast_fp16 = mul(x = h_37_cast_fp16, y = tm_layers_18_layer_scalar_to_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8449_cast_fp16 = abs(x = x_343_cast_fp16)[name = string("op_8449_cast_fp16")];
            bool var_8452 = const()[name = string("op_8452"), val = bool(true)];
            tensor<int32, [1]> var_8453_axes_0 = const()[name = string("op_8453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8453_cast_fp16 = reduce_max(axes = var_8453_axes_0, keep_dims = var_8452, x = var_8449_cast_fp16)[name = string("op_8453_cast_fp16")];
            fp16 var_8454_to_fp16 = const()[name = string("op_8454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_307_cast_fp16 = maximum(x = var_8453_cast_fp16, y = var_8454_to_fp16)[name = string("s_307_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_307_cast_fp16 = real_div(x = x_343_cast_fp16, y = s_307_cast_fp16)[name = string("y_307_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8457_cast_fp16 = mul(x = y_307_cast_fp16, y = y_307_cast_fp16)[name = string("op_8457_cast_fp16")];
            tensor<int32, [1]> var_8462_axes_0 = const()[name = string("op_8462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8462_keep_dims_0 = const()[name = string("op_8462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8462_cast_fp16 = reduce_mean(axes = var_8462_axes_0, keep_dims = var_8462_keep_dims_0, x = var_8457_cast_fp16)[name = string("op_8462_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8463_cast_fp16 = mul(x = s_307_cast_fp16, y = s_307_cast_fp16)[name = string("op_8463_cast_fp16")];
            fp32 var_8464_epsilon_0 = const()[name = string("op_8464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8464_cast_fp16 = inverse(epsilon = var_8464_epsilon_0, x = var_8463_cast_fp16)[name = string("op_8464_cast_fp16")];
            fp16 var_8465_to_fp16 = const()[name = string("op_8465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8466_cast_fp16 = mul(x = var_8464_cast_fp16, y = var_8465_to_fp16)[name = string("op_8466_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8468_cast_fp16 = add(x = var_8462_cast_fp16, y = var_8466_cast_fp16)[name = string("op_8468_cast_fp16")];
            fp32 var_8469_epsilon_0 = const()[name = string("op_8469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8469_cast_fp16 = rsqrt(epsilon = var_8469_epsilon_0, x = var_8468_cast_fp16)[name = string("op_8469_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_269_cast_fp16 = mul(x = y_307_cast_fp16, y = var_8469_cast_fp16)[name = string("out_269_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109572736)))];
            tensor<fp16, [1, 64, 512]> input_229_cast_fp16 = mul(x = out_269_cast_fp16, y = tm_layers_19_input_layernorm_weight_to_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109573824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110098176))))[name = string("tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [4]> var_8479 = const()[name = string("op_8479"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_8480_cast_fp16 = reshape(shape = var_8479, x = linear_172_cast_fp16)[name = string("op_8480_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110100288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110362496))))[name = string("tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<int32, [4]> var_8491 = const()[name = string("op_8491"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_8492_cast_fp16 = reshape(shape = var_8491, x = linear_173_cast_fp16)[name = string("op_8492_cast_fp16")];
            tensor<int32, [4]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110363584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110625792))))[name = string("tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<int32, [4]> var_8503 = const()[name = string("op_8503"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_8504_cast_fp16 = reshape(shape = var_8503, x = linear_174_cast_fp16)[name = string("op_8504_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_8480_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 4, 64, 256]> var_8508_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_8508_cast_fp16")];
            bool var_8511 = const()[name = string("op_8511"), val = bool(true)];
            tensor<int32, [1]> var_8512_axes_0 = const()[name = string("op_8512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_8512_cast_fp16 = reduce_max(axes = var_8512_axes_0, keep_dims = var_8511, x = var_8508_cast_fp16)[name = string("op_8512_cast_fp16")];
            fp16 var_8513_to_fp16 = const()[name = string("op_8513_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_309_cast_fp16 = maximum(x = var_8512_cast_fp16, y = var_8513_to_fp16)[name = string("s_309_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_309_cast_fp16 = real_div(x = x_345_cast_fp16, y = s_309_cast_fp16)[name = string("y_309_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8516_cast_fp16 = mul(x = y_309_cast_fp16, y = y_309_cast_fp16)[name = string("op_8516_cast_fp16")];
            tensor<int32, [1]> var_8521_axes_0 = const()[name = string("op_8521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8521_keep_dims_0 = const()[name = string("op_8521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_8521_cast_fp16 = reduce_mean(axes = var_8521_axes_0, keep_dims = var_8521_keep_dims_0, x = var_8516_cast_fp16)[name = string("op_8521_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_8522_cast_fp16 = mul(x = s_309_cast_fp16, y = s_309_cast_fp16)[name = string("op_8522_cast_fp16")];
            fp32 var_8523_epsilon_0 = const()[name = string("op_8523_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_8523_cast_fp16 = inverse(epsilon = var_8523_epsilon_0, x = var_8522_cast_fp16)[name = string("op_8523_cast_fp16")];
            fp16 var_8524_to_fp16 = const()[name = string("op_8524_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_8525_cast_fp16 = mul(x = var_8523_cast_fp16, y = var_8524_to_fp16)[name = string("op_8525_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_8527_cast_fp16 = add(x = var_8521_cast_fp16, y = var_8525_cast_fp16)[name = string("op_8527_cast_fp16")];
            fp32 var_8528_epsilon_0 = const()[name = string("op_8528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_8528_cast_fp16 = rsqrt(epsilon = var_8528_epsilon_0, x = var_8527_cast_fp16)[name = string("op_8528_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_271_cast_fp16 = mul(x = y_309_cast_fp16, y = var_8528_cast_fp16)[name = string("out_271_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_77_cast_fp16 = mul(x = out_271_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_77_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = var_8492_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 64, 256]> var_8531_cast_fp16 = abs(x = x_347_cast_fp16)[name = string("op_8531_cast_fp16")];
            bool var_8534 = const()[name = string("op_8534"), val = bool(true)];
            tensor<int32, [1]> var_8535_axes_0 = const()[name = string("op_8535_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_8535_cast_fp16 = reduce_max(axes = var_8535_axes_0, keep_dims = var_8534, x = var_8531_cast_fp16)[name = string("op_8535_cast_fp16")];
            fp16 var_8536_to_fp16 = const()[name = string("op_8536_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_311_cast_fp16 = maximum(x = var_8535_cast_fp16, y = var_8536_to_fp16)[name = string("s_311_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_311_cast_fp16 = real_div(x = x_347_cast_fp16, y = s_311_cast_fp16)[name = string("y_311_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8539_cast_fp16 = mul(x = y_311_cast_fp16, y = y_311_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<int32, [1]> var_8544_axes_0 = const()[name = string("op_8544_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8544_keep_dims_0 = const()[name = string("op_8544_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_8544_cast_fp16 = reduce_mean(axes = var_8544_axes_0, keep_dims = var_8544_keep_dims_0, x = var_8539_cast_fp16)[name = string("op_8544_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8545_cast_fp16 = mul(x = s_311_cast_fp16, y = s_311_cast_fp16)[name = string("op_8545_cast_fp16")];
            fp32 var_8546_epsilon_0 = const()[name = string("op_8546_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_8546_cast_fp16 = inverse(epsilon = var_8546_epsilon_0, x = var_8545_cast_fp16)[name = string("op_8546_cast_fp16")];
            fp16 var_8547_to_fp16 = const()[name = string("op_8547_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_8548_cast_fp16 = mul(x = var_8546_cast_fp16, y = var_8547_to_fp16)[name = string("op_8548_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8550_cast_fp16 = add(x = var_8544_cast_fp16, y = var_8548_cast_fp16)[name = string("op_8550_cast_fp16")];
            fp32 var_8551_epsilon_0 = const()[name = string("op_8551_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_8551_cast_fp16 = rsqrt(epsilon = var_8551_epsilon_0, x = var_8550_cast_fp16)[name = string("op_8551_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_273_cast_fp16 = mul(x = y_311_cast_fp16, y = var_8551_cast_fp16)[name = string("out_273_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_115_cast_fp16 = mul(x = out_273_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_115_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_8504_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 2, 64, 256]> var_8554_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_8554_cast_fp16")];
            bool var_8557 = const()[name = string("op_8557"), val = bool(true)];
            tensor<int32, [1]> var_8558_axes_0 = const()[name = string("op_8558_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_8558_cast_fp16 = reduce_max(axes = var_8558_axes_0, keep_dims = var_8557, x = var_8554_cast_fp16)[name = string("op_8558_cast_fp16")];
            fp16 var_8559_to_fp16 = const()[name = string("op_8559_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_313_cast_fp16 = maximum(x = var_8558_cast_fp16, y = var_8559_to_fp16)[name = string("s_313_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_313_cast_fp16 = real_div(x = x_349_cast_fp16, y = s_313_cast_fp16)[name = string("y_313_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8562_cast_fp16 = mul(x = y_313_cast_fp16, y = y_313_cast_fp16)[name = string("op_8562_cast_fp16")];
            tensor<int32, [1]> var_8567_axes_0 = const()[name = string("op_8567_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8567_keep_dims_0 = const()[name = string("op_8567_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_8567_cast_fp16 = reduce_mean(axes = var_8567_axes_0, keep_dims = var_8567_keep_dims_0, x = var_8562_cast_fp16)[name = string("op_8567_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8568_cast_fp16 = mul(x = s_313_cast_fp16, y = s_313_cast_fp16)[name = string("op_8568_cast_fp16")];
            fp32 var_8569_epsilon_0 = const()[name = string("op_8569_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_8569_cast_fp16 = inverse(epsilon = var_8569_epsilon_0, x = var_8568_cast_fp16)[name = string("op_8569_cast_fp16")];
            fp16 var_8570_to_fp16 = const()[name = string("op_8570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_8571_cast_fp16 = mul(x = var_8569_cast_fp16, y = var_8570_to_fp16)[name = string("op_8571_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8573_cast_fp16 = add(x = var_8567_cast_fp16, y = var_8571_cast_fp16)[name = string("op_8573_cast_fp16")];
            fp32 var_8574_epsilon_0 = const()[name = string("op_8574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_8574_cast_fp16 = rsqrt(epsilon = var_8574_epsilon_0, x = var_8573_cast_fp16)[name = string("op_8574_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_77_cast_fp16 = mul(x = y_313_cast_fp16, y = var_8574_cast_fp16)[name = string("v_77_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8576_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8576_cast_fp16")];
            tensor<int32, [4]> var_8581_begin_0 = const()[name = string("op_8581_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8581_end_0 = const()[name = string("op_8581_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_8581_end_mask_0 = const()[name = string("op_8581_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_8581_cast_fp16 = slice_by_index(begin = var_8581_begin_0, end = var_8581_end_0, end_mask = var_8581_end_mask_0, x = q_77_cast_fp16)[name = string("op_8581_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_8582_cast_fp16 = mul(x = var_8581_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_8582_cast_fp16")];
            tensor<int32, [4]> var_8587_begin_0 = const()[name = string("op_8587_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8587_end_0 = const()[name = string("op_8587_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_8587_end_mask_0 = const()[name = string("op_8587_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_8587_cast_fp16 = slice_by_index(begin = var_8587_begin_0, end = var_8587_end_0, end_mask = var_8587_end_mask_0, x = q_77_cast_fp16)[name = string("op_8587_cast_fp16")];
            int32 var_8589 = const()[name = string("op_8589"), val = int32(-1)];
            bool var_8590_interleave_0 = const()[name = string("op_8590_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_8590_cast_fp16 = concat(axis = var_8589, interleave = var_8590_interleave_0, values = (var_8582_cast_fp16, var_8587_cast_fp16))[name = string("op_8590_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8591_cast_fp16 = mul(x = var_8590_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8591_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_79_cast_fp16 = add(x = var_8576_cast_fp16, y = var_8591_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8594_cast_fp16 = mul(x = k_115_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8594_cast_fp16")];
            tensor<int32, [4]> var_8599_begin_0 = const()[name = string("op_8599_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8599_end_0 = const()[name = string("op_8599_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_8599_end_mask_0 = const()[name = string("op_8599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_8599_cast_fp16 = slice_by_index(begin = var_8599_begin_0, end = var_8599_end_0, end_mask = var_8599_end_mask_0, x = k_115_cast_fp16)[name = string("op_8599_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_8600_cast_fp16 = mul(x = var_8599_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_8600_cast_fp16")];
            tensor<int32, [4]> var_8605_begin_0 = const()[name = string("op_8605_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8605_end_0 = const()[name = string("op_8605_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_8605_end_mask_0 = const()[name = string("op_8605_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_8605_cast_fp16 = slice_by_index(begin = var_8605_begin_0, end = var_8605_end_0, end_mask = var_8605_end_mask_0, x = k_115_cast_fp16)[name = string("op_8605_cast_fp16")];
            int32 var_8607 = const()[name = string("op_8607"), val = int32(-1)];
            bool var_8608_interleave_0 = const()[name = string("op_8608_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_8608_cast_fp16 = concat(axis = var_8607, interleave = var_8608_interleave_0, values = (var_8600_cast_fp16, var_8605_cast_fp16))[name = string("op_8608_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8609_cast_fp16 = mul(x = var_8608_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8609_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_117_cast_fp16 = add(x = var_8594_cast_fp16, y = var_8609_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_38_reps_0 = const()[name = string("tile_38_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = k_117_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 1, 64, 256]> tile_38_cast_fp16 = tile(reps = tile_38_reps_0, x = transpose_76_cast_fp16)[name = string("tile_38_cast_fp16")];
            tensor<int32, [5]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_76_cast_fp16 = reshape(shape = concat_76, x = tile_38_cast_fp16)[name = string("reshape_76_cast_fp16")];
            tensor<int32, [5]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_77 = const()[name = string("concat_77"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_77_cast_fp16 = transpose(perm = transpose_77_perm_0, x = reshape_76_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 1, 64, 256]> reshape_77_cast_fp16 = reshape(shape = concat_77, x = transpose_77_cast_fp16)[name = string("reshape_77_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_39_reps_0 = const()[name = string("tile_39_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_78_cast_fp16 = transpose(perm = transpose_78_perm_0, x = v_77_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 1, 64, 256]> tile_39_cast_fp16 = tile(reps = tile_39_reps_0, x = transpose_78_cast_fp16)[name = string("tile_39_cast_fp16")];
            tensor<int32, [5]> concat_78 = const()[name = string("concat_78"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_78_cast_fp16 = reshape(shape = concat_78, x = tile_39_cast_fp16)[name = string("reshape_78_cast_fp16")];
            tensor<int32, [5]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = reshape_78_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 1, 64, 256]> reshape_79_cast_fp16 = reshape(shape = concat_79, x = transpose_79_cast_fp16)[name = string("reshape_79_cast_fp16")];
            tensor<int32, [4]> v_79_perm_0 = const()[name = string("v_79_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8623_transpose_x_0 = const()[name = string("op_8623_transpose_x_0"), val = bool(false)];
            bool var_8623_transpose_y_0 = const()[name = string("op_8623_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = reshape_77_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4, 64, 64]> var_8623_cast_fp16 = matmul(transpose_x = var_8623_transpose_x_0, transpose_y = var_8623_transpose_y_0, x = q_79_cast_fp16, y = transpose_115_cast_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8625_cast_fp16")];
            int32 var_8626 = const()[name = string("op_8626"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_313_cast_fp16 = softmax(axis = var_8626, x = var_8625_cast_fp16)[name = string("w_313_cast_fp16")];
            bool var_8629_transpose_x_0 = const()[name = string("op_8629_transpose_x_0"), val = bool(false)];
            bool var_8629_transpose_y_0 = const()[name = string("op_8629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = reshape_79_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 4, 64, 256]> var_8629_cast_fp16 = matmul(transpose_x = var_8629_transpose_x_0, transpose_y = var_8629_transpose_y_0, x = w_313_cast_fp16, y = v_79_cast_fp16)[name = string("op_8629_cast_fp16")];
            tensor<int32, [4]> var_8632_perm_0 = const()[name = string("op_8632_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_8632_cast_fp16 = transpose(perm = var_8632_perm_0, x = var_8629_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 64, 1024]> input_231_cast_fp16 = reshape(shape = var_8636, x = var_8632_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110626880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111151232))))[name = string("tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8641_cast_fp16 = abs(x = linear_175_cast_fp16)[name = string("op_8641_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<int32, [1]> var_8645_axes_0 = const()[name = string("op_8645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8645_cast_fp16 = reduce_max(axes = var_8645_axes_0, keep_dims = var_8644, x = var_8641_cast_fp16)[name = string("op_8645_cast_fp16")];
            fp16 var_8646_to_fp16 = const()[name = string("op_8646_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_315_cast_fp16 = maximum(x = var_8645_cast_fp16, y = var_8646_to_fp16)[name = string("s_315_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_315_cast_fp16 = real_div(x = linear_175_cast_fp16, y = s_315_cast_fp16)[name = string("y_315_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8649_cast_fp16 = mul(x = y_315_cast_fp16, y = y_315_cast_fp16)[name = string("op_8649_cast_fp16")];
            tensor<int32, [1]> var_8654_axes_0 = const()[name = string("op_8654_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8654_keep_dims_0 = const()[name = string("op_8654_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8654_cast_fp16 = reduce_mean(axes = var_8654_axes_0, keep_dims = var_8654_keep_dims_0, x = var_8649_cast_fp16)[name = string("op_8654_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8655_cast_fp16 = mul(x = s_315_cast_fp16, y = s_315_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp32 var_8656_epsilon_0 = const()[name = string("op_8656_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8656_cast_fp16 = inverse(epsilon = var_8656_epsilon_0, x = var_8655_cast_fp16)[name = string("op_8656_cast_fp16")];
            fp16 var_8657_to_fp16 = const()[name = string("op_8657_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8658_cast_fp16 = mul(x = var_8656_cast_fp16, y = var_8657_to_fp16)[name = string("op_8658_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8660_cast_fp16 = add(x = var_8654_cast_fp16, y = var_8658_cast_fp16)[name = string("op_8660_cast_fp16")];
            fp32 var_8661_epsilon_0 = const()[name = string("op_8661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8661_cast_fp16 = rsqrt(epsilon = var_8661_epsilon_0, x = var_8660_cast_fp16)[name = string("op_8661_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_275_cast_fp16 = mul(x = y_315_cast_fp16, y = var_8661_cast_fp16)[name = string("out_275_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111152320)))];
            tensor<fp16, [1, 64, 512]> var_8663_cast_fp16 = mul(x = out_275_cast_fp16, y = tm_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("op_8663_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_353_cast_fp16 = add(x = x_343_cast_fp16, y = var_8663_cast_fp16)[name = string("x_353_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8666_cast_fp16 = abs(x = x_353_cast_fp16)[name = string("op_8666_cast_fp16")];
            bool var_8669 = const()[name = string("op_8669"), val = bool(true)];
            tensor<int32, [1]> var_8670_axes_0 = const()[name = string("op_8670_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8670_cast_fp16 = reduce_max(axes = var_8670_axes_0, keep_dims = var_8669, x = var_8666_cast_fp16)[name = string("op_8670_cast_fp16")];
            fp16 var_8671_to_fp16 = const()[name = string("op_8671_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_317_cast_fp16 = maximum(x = var_8670_cast_fp16, y = var_8671_to_fp16)[name = string("s_317_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_317_cast_fp16 = real_div(x = x_353_cast_fp16, y = s_317_cast_fp16)[name = string("y_317_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8674_cast_fp16 = mul(x = y_317_cast_fp16, y = y_317_cast_fp16)[name = string("op_8674_cast_fp16")];
            tensor<int32, [1]> var_8679_axes_0 = const()[name = string("op_8679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8679_keep_dims_0 = const()[name = string("op_8679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8679_cast_fp16 = reduce_mean(axes = var_8679_axes_0, keep_dims = var_8679_keep_dims_0, x = var_8674_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8680_cast_fp16 = mul(x = s_317_cast_fp16, y = s_317_cast_fp16)[name = string("op_8680_cast_fp16")];
            fp32 var_8681_epsilon_0 = const()[name = string("op_8681_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8681_cast_fp16 = inverse(epsilon = var_8681_epsilon_0, x = var_8680_cast_fp16)[name = string("op_8681_cast_fp16")];
            fp16 var_8682_to_fp16 = const()[name = string("op_8682_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8683_cast_fp16 = mul(x = var_8681_cast_fp16, y = var_8682_to_fp16)[name = string("op_8683_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8685_cast_fp16 = add(x = var_8679_cast_fp16, y = var_8683_cast_fp16)[name = string("op_8685_cast_fp16")];
            fp32 var_8686_epsilon_0 = const()[name = string("op_8686_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8686_cast_fp16 = rsqrt(epsilon = var_8686_epsilon_0, x = var_8685_cast_fp16)[name = string("op_8686_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_277_cast_fp16 = mul(x = y_317_cast_fp16, y = var_8686_cast_fp16)[name = string("out_277_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111153408)))];
            tensor<fp16, [1, 64, 512]> input_233_cast_fp16 = mul(x = out_277_cast_fp16, y = tm_layers_19_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111154496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112203136))))[name = string("tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_176_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_176_cast_fp16")];
            string var_8693_mode_0 = const()[name = string("op_8693_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_8693_cast_fp16 = gelu(mode = var_8693_mode_0, x = linear_176_cast_fp16)[name = string("op_8693_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112207296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113255936))))[name = string("tm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_177_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_235_cast_fp16 = mul(x = var_8693_cast_fp16, y = linear_177_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113260096))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114308736))))[name = string("tm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8701_cast_fp16 = abs(x = linear_178_cast_fp16)[name = string("op_8701_cast_fp16")];
            bool var_8704 = const()[name = string("op_8704"), val = bool(true)];
            tensor<int32, [1]> var_8705_axes_0 = const()[name = string("op_8705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8705_cast_fp16 = reduce_max(axes = var_8705_axes_0, keep_dims = var_8704, x = var_8701_cast_fp16)[name = string("op_8705_cast_fp16")];
            fp16 var_8706_to_fp16 = const()[name = string("op_8706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_319_cast_fp16 = maximum(x = var_8705_cast_fp16, y = var_8706_to_fp16)[name = string("s_319_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_319_cast_fp16 = real_div(x = linear_178_cast_fp16, y = s_319_cast_fp16)[name = string("y_319_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8709_cast_fp16 = mul(x = y_319_cast_fp16, y = y_319_cast_fp16)[name = string("op_8709_cast_fp16")];
            tensor<int32, [1]> var_8714_axes_0 = const()[name = string("op_8714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8714_keep_dims_0 = const()[name = string("op_8714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8714_cast_fp16 = reduce_mean(axes = var_8714_axes_0, keep_dims = var_8714_keep_dims_0, x = var_8709_cast_fp16)[name = string("op_8714_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8715_cast_fp16 = mul(x = s_319_cast_fp16, y = s_319_cast_fp16)[name = string("op_8715_cast_fp16")];
            fp32 var_8716_epsilon_0 = const()[name = string("op_8716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8716_cast_fp16 = inverse(epsilon = var_8716_epsilon_0, x = var_8715_cast_fp16)[name = string("op_8716_cast_fp16")];
            fp16 var_8717_to_fp16 = const()[name = string("op_8717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8718_cast_fp16 = mul(x = var_8716_cast_fp16, y = var_8717_to_fp16)[name = string("op_8718_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8720_cast_fp16 = add(x = var_8714_cast_fp16, y = var_8718_cast_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8721_cast_fp16 = rsqrt(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_279_cast_fp16 = mul(x = y_319_cast_fp16, y = var_8721_cast_fp16)[name = string("out_279_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114309824)))];
            tensor<fp16, [1, 64, 512]> var_8723_cast_fp16 = mul(x = out_279_cast_fp16, y = tm_layers_19_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_237_cast_fp16 = add(x = x_353_cast_fp16, y = var_8723_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114310912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114573120))))[name = string("tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_179_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_237_cast_fp16)[name = string("linear_179_cast_fp16")];
            string x_357_mode_0 = const()[name = string("x_357_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_357_cast_fp16 = gelu(mode = x_357_mode_0, x = linear_179_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [4]> var_8743_begin_0 = const()[name = string("op_8743_begin_0"), val = tensor<int32, [4]>([0, 0, 19, 0])];
            tensor<int32, [4]> var_8743_end_0 = const()[name = string("op_8743_end_0"), val = tensor<int32, [4]>([1, 64, 20, 512])];
            tensor<bool, [4]> var_8743_end_mask_0 = const()[name = string("op_8743_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8743_squeeze_mask_0 = const()[name = string("op_8743_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_8743_cast_fp16 = slice_by_index(begin = var_8743_begin_0, end = var_8743_end_0, end_mask = var_8743_end_mask_0, squeeze_mask = var_8743_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_8743_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_239_cast_fp16 = mul(x = x_357_cast_fp16, y = var_8743_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114574208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114836416))))[name = string("tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8753_cast_fp16 = abs(x = linear_180_cast_fp16)[name = string("op_8753_cast_fp16")];
            bool var_8756 = const()[name = string("op_8756"), val = bool(true)];
            tensor<int32, [1]> var_8757_axes_0 = const()[name = string("op_8757_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8757_cast_fp16 = reduce_max(axes = var_8757_axes_0, keep_dims = var_8756, x = var_8753_cast_fp16)[name = string("op_8757_cast_fp16")];
            fp16 var_8758_to_fp16 = const()[name = string("op_8758_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_321_cast_fp16 = maximum(x = var_8757_cast_fp16, y = var_8758_to_fp16)[name = string("s_321_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_321_cast_fp16 = real_div(x = linear_180_cast_fp16, y = s_321_cast_fp16)[name = string("y_321_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8761_cast_fp16 = mul(x = y_321_cast_fp16, y = y_321_cast_fp16)[name = string("op_8761_cast_fp16")];
            tensor<int32, [1]> var_8766_axes_0 = const()[name = string("op_8766_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8766_keep_dims_0 = const()[name = string("op_8766_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8766_cast_fp16 = reduce_mean(axes = var_8766_axes_0, keep_dims = var_8766_keep_dims_0, x = var_8761_cast_fp16)[name = string("op_8766_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8767_cast_fp16 = mul(x = s_321_cast_fp16, y = s_321_cast_fp16)[name = string("op_8767_cast_fp16")];
            fp32 var_8768_epsilon_0 = const()[name = string("op_8768_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8768_cast_fp16 = inverse(epsilon = var_8768_epsilon_0, x = var_8767_cast_fp16)[name = string("op_8768_cast_fp16")];
            fp16 var_8769_to_fp16 = const()[name = string("op_8769_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8770_cast_fp16 = mul(x = var_8768_cast_fp16, y = var_8769_to_fp16)[name = string("op_8770_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8772_cast_fp16 = add(x = var_8766_cast_fp16, y = var_8770_cast_fp16)[name = string("op_8772_cast_fp16")];
            fp32 var_8773_epsilon_0 = const()[name = string("op_8773_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8773_cast_fp16 = rsqrt(epsilon = var_8773_epsilon_0, x = var_8772_cast_fp16)[name = string("op_8773_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_281_cast_fp16 = mul(x = y_321_cast_fp16, y = var_8773_cast_fp16)[name = string("out_281_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114837504)))];
            tensor<fp16, [1, 64, 512]> var_8775_cast_fp16 = mul(x = out_281_cast_fp16, y = tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_39_cast_fp16 = add(x = input_237_cast_fp16, y = var_8775_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<fp16, [1]> tm_layers_19_layer_scalar_to_fp16 = const()[name = string("tm_layers_19_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fp-1])];
            tensor<fp16, [1, 64, 512]> x_361_cast_fp16 = mul(x = h_39_cast_fp16, y = tm_layers_19_layer_scalar_to_fp16)[name = string("x_361_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8779_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_8779_cast_fp16")];
            bool var_8782 = const()[name = string("op_8782"), val = bool(true)];
            tensor<int32, [1]> var_8783_axes_0 = const()[name = string("op_8783_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8783_cast_fp16 = reduce_max(axes = var_8783_axes_0, keep_dims = var_8782, x = var_8779_cast_fp16)[name = string("op_8783_cast_fp16")];
            fp16 var_8784_to_fp16 = const()[name = string("op_8784_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_323_cast_fp16 = maximum(x = var_8783_cast_fp16, y = var_8784_to_fp16)[name = string("s_323_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_323_cast_fp16 = real_div(x = x_361_cast_fp16, y = s_323_cast_fp16)[name = string("y_323_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8787_cast_fp16 = mul(x = y_323_cast_fp16, y = y_323_cast_fp16)[name = string("op_8787_cast_fp16")];
            tensor<int32, [1]> var_8792_axes_0 = const()[name = string("op_8792_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8792_keep_dims_0 = const()[name = string("op_8792_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8792_cast_fp16 = reduce_mean(axes = var_8792_axes_0, keep_dims = var_8792_keep_dims_0, x = var_8787_cast_fp16)[name = string("op_8792_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8793_cast_fp16 = mul(x = s_323_cast_fp16, y = s_323_cast_fp16)[name = string("op_8793_cast_fp16")];
            fp32 var_8794_epsilon_0 = const()[name = string("op_8794_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8794_cast_fp16 = inverse(epsilon = var_8794_epsilon_0, x = var_8793_cast_fp16)[name = string("op_8794_cast_fp16")];
            fp16 var_8795_to_fp16 = const()[name = string("op_8795_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8796_cast_fp16 = mul(x = var_8794_cast_fp16, y = var_8795_to_fp16)[name = string("op_8796_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8798_cast_fp16 = add(x = var_8792_cast_fp16, y = var_8796_cast_fp16)[name = string("op_8798_cast_fp16")];
            fp32 var_8799_epsilon_0 = const()[name = string("op_8799_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8799_cast_fp16 = rsqrt(epsilon = var_8799_epsilon_0, x = var_8798_cast_fp16)[name = string("op_8799_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_283_cast_fp16 = mul(x = y_323_cast_fp16, y = var_8799_cast_fp16)[name = string("out_283_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114838592)))];
            tensor<fp16, [1, 64, 512]> input_241_cast_fp16 = mul(x = out_283_cast_fp16, y = tm_layers_20_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114839680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115364032))))[name = string("tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<int32, [4]> var_8809 = const()[name = string("op_8809"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_8810_cast_fp16 = reshape(shape = var_8809, x = linear_181_cast_fp16)[name = string("op_8810_cast_fp16")];
            tensor<int32, [4]> x_363_perm_0 = const()[name = string("x_363_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115366144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115628352))))[name = string("tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_8821 = const()[name = string("op_8821"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_8822_cast_fp16 = reshape(shape = var_8821, x = linear_182_cast_fp16)[name = string("op_8822_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115629440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115891648))))[name = string("tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_8833 = const()[name = string("op_8833"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_8834_cast_fp16 = reshape(shape = var_8833, x = linear_183_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<int32, [4]> x_367_perm_0 = const()[name = string("x_367_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_363_cast_fp16 = transpose(perm = x_363_perm_0, x = var_8810_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 4, 64, 256]> var_8838_cast_fp16 = abs(x = x_363_cast_fp16)[name = string("op_8838_cast_fp16")];
            bool var_8841 = const()[name = string("op_8841"), val = bool(true)];
            tensor<int32, [1]> var_8842_axes_0 = const()[name = string("op_8842_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_8842_cast_fp16 = reduce_max(axes = var_8842_axes_0, keep_dims = var_8841, x = var_8838_cast_fp16)[name = string("op_8842_cast_fp16")];
            fp16 var_8843_to_fp16 = const()[name = string("op_8843_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_325_cast_fp16 = maximum(x = var_8842_cast_fp16, y = var_8843_to_fp16)[name = string("s_325_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_325_cast_fp16 = real_div(x = x_363_cast_fp16, y = s_325_cast_fp16)[name = string("y_325_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8846_cast_fp16 = mul(x = y_325_cast_fp16, y = y_325_cast_fp16)[name = string("op_8846_cast_fp16")];
            tensor<int32, [1]> var_8851_axes_0 = const()[name = string("op_8851_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8851_keep_dims_0 = const()[name = string("op_8851_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_8851_cast_fp16 = reduce_mean(axes = var_8851_axes_0, keep_dims = var_8851_keep_dims_0, x = var_8846_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_8852_cast_fp16 = mul(x = s_325_cast_fp16, y = s_325_cast_fp16)[name = string("op_8852_cast_fp16")];
            fp32 var_8853_epsilon_0 = const()[name = string("op_8853_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_8853_cast_fp16 = inverse(epsilon = var_8853_epsilon_0, x = var_8852_cast_fp16)[name = string("op_8853_cast_fp16")];
            fp16 var_8854_to_fp16 = const()[name = string("op_8854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_8855_cast_fp16 = mul(x = var_8853_cast_fp16, y = var_8854_to_fp16)[name = string("op_8855_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_8857_cast_fp16 = add(x = var_8851_cast_fp16, y = var_8855_cast_fp16)[name = string("op_8857_cast_fp16")];
            fp32 var_8858_epsilon_0 = const()[name = string("op_8858_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_8858_cast_fp16 = rsqrt(epsilon = var_8858_epsilon_0, x = var_8857_cast_fp16)[name = string("op_8858_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_285_cast_fp16 = mul(x = y_325_cast_fp16, y = var_8858_cast_fp16)[name = string("out_285_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_81_cast_fp16 = mul(x = out_285_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_81_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_8822_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 64, 256]> var_8861_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_8861_cast_fp16")];
            bool var_8864 = const()[name = string("op_8864"), val = bool(true)];
            tensor<int32, [1]> var_8865_axes_0 = const()[name = string("op_8865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_8865_cast_fp16 = reduce_max(axes = var_8865_axes_0, keep_dims = var_8864, x = var_8861_cast_fp16)[name = string("op_8865_cast_fp16")];
            fp16 var_8866_to_fp16 = const()[name = string("op_8866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_327_cast_fp16 = maximum(x = var_8865_cast_fp16, y = var_8866_to_fp16)[name = string("s_327_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_327_cast_fp16 = real_div(x = x_365_cast_fp16, y = s_327_cast_fp16)[name = string("y_327_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8869_cast_fp16 = mul(x = y_327_cast_fp16, y = y_327_cast_fp16)[name = string("op_8869_cast_fp16")];
            tensor<int32, [1]> var_8874_axes_0 = const()[name = string("op_8874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8874_keep_dims_0 = const()[name = string("op_8874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_8874_cast_fp16 = reduce_mean(axes = var_8874_axes_0, keep_dims = var_8874_keep_dims_0, x = var_8869_cast_fp16)[name = string("op_8874_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8875_cast_fp16 = mul(x = s_327_cast_fp16, y = s_327_cast_fp16)[name = string("op_8875_cast_fp16")];
            fp32 var_8876_epsilon_0 = const()[name = string("op_8876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_8876_cast_fp16 = inverse(epsilon = var_8876_epsilon_0, x = var_8875_cast_fp16)[name = string("op_8876_cast_fp16")];
            fp16 var_8877_to_fp16 = const()[name = string("op_8877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_8878_cast_fp16 = mul(x = var_8876_cast_fp16, y = var_8877_to_fp16)[name = string("op_8878_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8880_cast_fp16 = add(x = var_8874_cast_fp16, y = var_8878_cast_fp16)[name = string("op_8880_cast_fp16")];
            fp32 var_8881_epsilon_0 = const()[name = string("op_8881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_8881_cast_fp16 = rsqrt(epsilon = var_8881_epsilon_0, x = var_8880_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_287_cast_fp16 = mul(x = y_327_cast_fp16, y = var_8881_cast_fp16)[name = string("out_287_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_121_cast_fp16 = mul(x = out_287_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_121_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_367_cast_fp16 = transpose(perm = x_367_perm_0, x = var_8834_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 64, 256]> var_8884_cast_fp16 = abs(x = x_367_cast_fp16)[name = string("op_8884_cast_fp16")];
            bool var_8887 = const()[name = string("op_8887"), val = bool(true)];
            tensor<int32, [1]> var_8888_axes_0 = const()[name = string("op_8888_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_8888_cast_fp16 = reduce_max(axes = var_8888_axes_0, keep_dims = var_8887, x = var_8884_cast_fp16)[name = string("op_8888_cast_fp16")];
            fp16 var_8889_to_fp16 = const()[name = string("op_8889_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_329_cast_fp16 = maximum(x = var_8888_cast_fp16, y = var_8889_to_fp16)[name = string("s_329_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_329_cast_fp16 = real_div(x = x_367_cast_fp16, y = s_329_cast_fp16)[name = string("y_329_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8892_cast_fp16 = mul(x = y_329_cast_fp16, y = y_329_cast_fp16)[name = string("op_8892_cast_fp16")];
            tensor<int32, [1]> var_8897_axes_0 = const()[name = string("op_8897_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8897_keep_dims_0 = const()[name = string("op_8897_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_8897_cast_fp16 = reduce_mean(axes = var_8897_axes_0, keep_dims = var_8897_keep_dims_0, x = var_8892_cast_fp16)[name = string("op_8897_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8898_cast_fp16 = mul(x = s_329_cast_fp16, y = s_329_cast_fp16)[name = string("op_8898_cast_fp16")];
            fp32 var_8899_epsilon_0 = const()[name = string("op_8899_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_8899_cast_fp16 = inverse(epsilon = var_8899_epsilon_0, x = var_8898_cast_fp16)[name = string("op_8899_cast_fp16")];
            fp16 var_8900_to_fp16 = const()[name = string("op_8900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_8901_cast_fp16 = mul(x = var_8899_cast_fp16, y = var_8900_to_fp16)[name = string("op_8901_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_8903_cast_fp16 = add(x = var_8897_cast_fp16, y = var_8901_cast_fp16)[name = string("op_8903_cast_fp16")];
            fp32 var_8904_epsilon_0 = const()[name = string("op_8904_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_8904_cast_fp16 = rsqrt(epsilon = var_8904_epsilon_0, x = var_8903_cast_fp16)[name = string("op_8904_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_81_cast_fp16 = mul(x = y_329_cast_fp16, y = var_8904_cast_fp16)[name = string("v_81_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8906_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8906_cast_fp16")];
            tensor<int32, [4]> var_8911_begin_0 = const()[name = string("op_8911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8911_end_0 = const()[name = string("op_8911_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_8911_end_mask_0 = const()[name = string("op_8911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_8911_cast_fp16 = slice_by_index(begin = var_8911_begin_0, end = var_8911_end_0, end_mask = var_8911_end_mask_0, x = q_81_cast_fp16)[name = string("op_8911_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_8912_cast_fp16 = mul(x = var_8911_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_8912_cast_fp16")];
            tensor<int32, [4]> var_8917_begin_0 = const()[name = string("op_8917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8917_end_0 = const()[name = string("op_8917_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_8917_end_mask_0 = const()[name = string("op_8917_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_8917_cast_fp16 = slice_by_index(begin = var_8917_begin_0, end = var_8917_end_0, end_mask = var_8917_end_mask_0, x = q_81_cast_fp16)[name = string("op_8917_cast_fp16")];
            int32 var_8919 = const()[name = string("op_8919"), val = int32(-1)];
            bool var_8920_interleave_0 = const()[name = string("op_8920_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_8920_cast_fp16 = concat(axis = var_8919, interleave = var_8920_interleave_0, values = (var_8912_cast_fp16, var_8917_cast_fp16))[name = string("op_8920_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_8921_cast_fp16 = mul(x = var_8920_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_83_cast_fp16 = add(x = var_8906_cast_fp16, y = var_8921_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8924_cast_fp16 = mul(x = k_121_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_8924_cast_fp16")];
            tensor<int32, [4]> var_8929_begin_0 = const()[name = string("op_8929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8929_end_0 = const()[name = string("op_8929_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_8929_end_mask_0 = const()[name = string("op_8929_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, x = k_121_cast_fp16)[name = string("op_8929_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_8930_cast_fp16 = mul(x = var_8929_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_8930_cast_fp16")];
            tensor<int32, [4]> var_8935_begin_0 = const()[name = string("op_8935_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8935_end_0 = const()[name = string("op_8935_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_8935_end_mask_0 = const()[name = string("op_8935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, x = k_121_cast_fp16)[name = string("op_8935_cast_fp16")];
            int32 var_8937 = const()[name = string("op_8937"), val = int32(-1)];
            bool var_8938_interleave_0 = const()[name = string("op_8938_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_8938_cast_fp16 = concat(axis = var_8937, interleave = var_8938_interleave_0, values = (var_8930_cast_fp16, var_8935_cast_fp16))[name = string("op_8938_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_8939_cast_fp16 = mul(x = var_8938_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_8939_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_123_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8939_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_40_reps_0 = const()[name = string("tile_40_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = k_123_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 1, 64, 256]> tile_40_cast_fp16 = tile(reps = tile_40_reps_0, x = transpose_80_cast_fp16)[name = string("tile_40_cast_fp16")];
            tensor<int32, [5]> concat_80 = const()[name = string("concat_80"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_80_cast_fp16 = reshape(shape = concat_80, x = tile_40_cast_fp16)[name = string("reshape_80_cast_fp16")];
            tensor<int32, [5]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_81 = const()[name = string("concat_81"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = reshape_80_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 1, 64, 256]> reshape_81_cast_fp16 = reshape(shape = concat_81, x = transpose_81_cast_fp16)[name = string("reshape_81_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_41_reps_0 = const()[name = string("tile_41_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_82_cast_fp16 = transpose(perm = transpose_82_perm_0, x = v_81_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 1, 64, 256]> tile_41_cast_fp16 = tile(reps = tile_41_reps_0, x = transpose_82_cast_fp16)[name = string("tile_41_cast_fp16")];
            tensor<int32, [5]> concat_82 = const()[name = string("concat_82"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_82_cast_fp16 = reshape(shape = concat_82, x = tile_41_cast_fp16)[name = string("reshape_82_cast_fp16")];
            tensor<int32, [5]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_83 = const()[name = string("concat_83"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = reshape_82_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 1, 64, 256]> reshape_83_cast_fp16 = reshape(shape = concat_83, x = transpose_83_cast_fp16)[name = string("reshape_83_cast_fp16")];
            tensor<int32, [4]> v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8953_transpose_x_0 = const()[name = string("op_8953_transpose_x_0"), val = bool(false)];
            bool var_8953_transpose_y_0 = const()[name = string("op_8953_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = reshape_81_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 4, 64, 64]> var_8953_cast_fp16 = matmul(transpose_x = var_8953_transpose_x_0, transpose_y = var_8953_transpose_y_0, x = q_83_cast_fp16, y = transpose_116_cast_fp16)[name = string("op_8953_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_8955_cast_fp16 = add(x = var_8953_cast_fp16, y = var_2173_cast_fp16)[name = string("op_8955_cast_fp16")];
            int32 var_8956 = const()[name = string("op_8956"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_329_cast_fp16 = softmax(axis = var_8956, x = var_8955_cast_fp16)[name = string("w_329_cast_fp16")];
            bool var_8959_transpose_x_0 = const()[name = string("op_8959_transpose_x_0"), val = bool(false)];
            bool var_8959_transpose_y_0 = const()[name = string("op_8959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_83_cast_fp16 = transpose(perm = v_83_perm_0, x = reshape_83_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 4, 64, 256]> var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = w_329_cast_fp16, y = v_83_cast_fp16)[name = string("op_8959_cast_fp16")];
            tensor<int32, [4]> var_8962_perm_0 = const()[name = string("op_8962_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8966 = const()[name = string("op_8966"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_8962_cast_fp16 = transpose(perm = var_8962_perm_0, x = var_8959_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 64, 1024]> input_243_cast_fp16 = reshape(shape = var_8966, x = var_8962_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115892736))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116417088))))[name = string("tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8971_cast_fp16 = abs(x = linear_184_cast_fp16)[name = string("op_8971_cast_fp16")];
            bool var_8974 = const()[name = string("op_8974"), val = bool(true)];
            tensor<int32, [1]> var_8975_axes_0 = const()[name = string("op_8975_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_8975_cast_fp16 = reduce_max(axes = var_8975_axes_0, keep_dims = var_8974, x = var_8971_cast_fp16)[name = string("op_8975_cast_fp16")];
            fp16 var_8976_to_fp16 = const()[name = string("op_8976_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_331_cast_fp16 = maximum(x = var_8975_cast_fp16, y = var_8976_to_fp16)[name = string("s_331_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_331_cast_fp16 = real_div(x = linear_184_cast_fp16, y = s_331_cast_fp16)[name = string("y_331_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8979_cast_fp16 = mul(x = y_331_cast_fp16, y = y_331_cast_fp16)[name = string("op_8979_cast_fp16")];
            tensor<int32, [1]> var_8984_axes_0 = const()[name = string("op_8984_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8984_keep_dims_0 = const()[name = string("op_8984_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_8984_cast_fp16 = reduce_mean(axes = var_8984_axes_0, keep_dims = var_8984_keep_dims_0, x = var_8979_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8985_cast_fp16 = mul(x = s_331_cast_fp16, y = s_331_cast_fp16)[name = string("op_8985_cast_fp16")];
            fp32 var_8986_epsilon_0 = const()[name = string("op_8986_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_8986_cast_fp16 = inverse(epsilon = var_8986_epsilon_0, x = var_8985_cast_fp16)[name = string("op_8986_cast_fp16")];
            fp16 var_8987_to_fp16 = const()[name = string("op_8987_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_8988_cast_fp16 = mul(x = var_8986_cast_fp16, y = var_8987_to_fp16)[name = string("op_8988_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_8990_cast_fp16 = add(x = var_8984_cast_fp16, y = var_8988_cast_fp16)[name = string("op_8990_cast_fp16")];
            fp32 var_8991_epsilon_0 = const()[name = string("op_8991_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_8991_cast_fp16 = rsqrt(epsilon = var_8991_epsilon_0, x = var_8990_cast_fp16)[name = string("op_8991_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_289_cast_fp16 = mul(x = y_331_cast_fp16, y = var_8991_cast_fp16)[name = string("out_289_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116418176)))];
            tensor<fp16, [1, 64, 512]> var_8993_cast_fp16 = mul(x = out_289_cast_fp16, y = tm_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("op_8993_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_371_cast_fp16 = add(x = x_361_cast_fp16, y = var_8993_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_8996_cast_fp16 = abs(x = x_371_cast_fp16)[name = string("op_8996_cast_fp16")];
            bool var_8999 = const()[name = string("op_8999"), val = bool(true)];
            tensor<int32, [1]> var_9000_axes_0 = const()[name = string("op_9000_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9000_cast_fp16 = reduce_max(axes = var_9000_axes_0, keep_dims = var_8999, x = var_8996_cast_fp16)[name = string("op_9000_cast_fp16")];
            fp16 var_9001_to_fp16 = const()[name = string("op_9001_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_333_cast_fp16 = maximum(x = var_9000_cast_fp16, y = var_9001_to_fp16)[name = string("s_333_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_333_cast_fp16 = real_div(x = x_371_cast_fp16, y = s_333_cast_fp16)[name = string("y_333_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9004_cast_fp16 = mul(x = y_333_cast_fp16, y = y_333_cast_fp16)[name = string("op_9004_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9004_cast_fp16)[name = string("op_9009_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9010_cast_fp16 = mul(x = s_333_cast_fp16, y = s_333_cast_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9015_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9015_cast_fp16")];
            fp32 var_9016_epsilon_0 = const()[name = string("op_9016_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9016_cast_fp16 = rsqrt(epsilon = var_9016_epsilon_0, x = var_9015_cast_fp16)[name = string("op_9016_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_291_cast_fp16 = mul(x = y_333_cast_fp16, y = var_9016_cast_fp16)[name = string("out_291_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116419264)))];
            tensor<fp16, [1, 64, 512]> input_245_cast_fp16 = mul(x = out_291_cast_fp16, y = tm_layers_20_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116420352))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117468992))))[name = string("tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_185_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_185_cast_fp16")];
            string var_9023_mode_0 = const()[name = string("op_9023_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_9023_cast_fp16 = gelu(mode = var_9023_mode_0, x = linear_185_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117473152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118521792))))[name = string("tm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_186_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_247_cast_fp16 = mul(x = var_9023_cast_fp16, y = linear_186_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118525952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119574592))))[name = string("tm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9031_cast_fp16 = abs(x = linear_187_cast_fp16)[name = string("op_9031_cast_fp16")];
            bool var_9034 = const()[name = string("op_9034"), val = bool(true)];
            tensor<int32, [1]> var_9035_axes_0 = const()[name = string("op_9035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9035_cast_fp16 = reduce_max(axes = var_9035_axes_0, keep_dims = var_9034, x = var_9031_cast_fp16)[name = string("op_9035_cast_fp16")];
            fp16 var_9036_to_fp16 = const()[name = string("op_9036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_335_cast_fp16 = maximum(x = var_9035_cast_fp16, y = var_9036_to_fp16)[name = string("s_335_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_335_cast_fp16 = real_div(x = linear_187_cast_fp16, y = s_335_cast_fp16)[name = string("y_335_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9039_cast_fp16 = mul(x = y_335_cast_fp16, y = y_335_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [1]> var_9044_axes_0 = const()[name = string("op_9044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9044_keep_dims_0 = const()[name = string("op_9044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9044_cast_fp16 = reduce_mean(axes = var_9044_axes_0, keep_dims = var_9044_keep_dims_0, x = var_9039_cast_fp16)[name = string("op_9044_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9045_cast_fp16 = mul(x = s_335_cast_fp16, y = s_335_cast_fp16)[name = string("op_9045_cast_fp16")];
            fp32 var_9046_epsilon_0 = const()[name = string("op_9046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9046_cast_fp16 = inverse(epsilon = var_9046_epsilon_0, x = var_9045_cast_fp16)[name = string("op_9046_cast_fp16")];
            fp16 var_9047_to_fp16 = const()[name = string("op_9047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9048_cast_fp16 = mul(x = var_9046_cast_fp16, y = var_9047_to_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9050_cast_fp16 = add(x = var_9044_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9050_cast_fp16")];
            fp32 var_9051_epsilon_0 = const()[name = string("op_9051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9051_cast_fp16 = rsqrt(epsilon = var_9051_epsilon_0, x = var_9050_cast_fp16)[name = string("op_9051_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_293_cast_fp16 = mul(x = y_335_cast_fp16, y = var_9051_cast_fp16)[name = string("out_293_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119575680)))];
            tensor<fp16, [1, 64, 512]> var_9053_cast_fp16 = mul(x = out_293_cast_fp16, y = tm_layers_20_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9053_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_249_cast_fp16 = add(x = x_371_cast_fp16, y = var_9053_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119576768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119838976))))[name = string("tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_188_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_188_cast_fp16")];
            string x_375_mode_0 = const()[name = string("x_375_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_375_cast_fp16 = gelu(mode = x_375_mode_0, x = linear_188_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [4]> var_9073_begin_0 = const()[name = string("op_9073_begin_0"), val = tensor<int32, [4]>([0, 0, 20, 0])];
            tensor<int32, [4]> var_9073_end_0 = const()[name = string("op_9073_end_0"), val = tensor<int32, [4]>([1, 64, 21, 512])];
            tensor<bool, [4]> var_9073_end_mask_0 = const()[name = string("op_9073_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9073_squeeze_mask_0 = const()[name = string("op_9073_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_9073_cast_fp16 = slice_by_index(begin = var_9073_begin_0, end = var_9073_end_0, end_mask = var_9073_end_mask_0, squeeze_mask = var_9073_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_251_cast_fp16 = mul(x = x_375_cast_fp16, y = var_9073_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119840064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120102272))))[name = string("tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9083_cast_fp16 = abs(x = linear_189_cast_fp16)[name = string("op_9083_cast_fp16")];
            bool var_9086 = const()[name = string("op_9086"), val = bool(true)];
            tensor<int32, [1]> var_9087_axes_0 = const()[name = string("op_9087_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9087_cast_fp16 = reduce_max(axes = var_9087_axes_0, keep_dims = var_9086, x = var_9083_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_337_cast_fp16 = maximum(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("s_337_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_337_cast_fp16 = real_div(x = linear_189_cast_fp16, y = s_337_cast_fp16)[name = string("y_337_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9091_cast_fp16 = mul(x = y_337_cast_fp16, y = y_337_cast_fp16)[name = string("op_9091_cast_fp16")];
            tensor<int32, [1]> var_9096_axes_0 = const()[name = string("op_9096_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9096_keep_dims_0 = const()[name = string("op_9096_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9096_cast_fp16 = reduce_mean(axes = var_9096_axes_0, keep_dims = var_9096_keep_dims_0, x = var_9091_cast_fp16)[name = string("op_9096_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9097_cast_fp16 = mul(x = s_337_cast_fp16, y = s_337_cast_fp16)[name = string("op_9097_cast_fp16")];
            fp32 var_9098_epsilon_0 = const()[name = string("op_9098_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9098_cast_fp16 = inverse(epsilon = var_9098_epsilon_0, x = var_9097_cast_fp16)[name = string("op_9098_cast_fp16")];
            fp16 var_9099_to_fp16 = const()[name = string("op_9099_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9100_cast_fp16 = mul(x = var_9098_cast_fp16, y = var_9099_to_fp16)[name = string("op_9100_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9102_cast_fp16 = add(x = var_9096_cast_fp16, y = var_9100_cast_fp16)[name = string("op_9102_cast_fp16")];
            fp32 var_9103_epsilon_0 = const()[name = string("op_9103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9103_cast_fp16 = rsqrt(epsilon = var_9103_epsilon_0, x = var_9102_cast_fp16)[name = string("op_9103_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_295_cast_fp16 = mul(x = y_337_cast_fp16, y = var_9103_cast_fp16)[name = string("out_295_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120103360)))];
            tensor<fp16, [1, 64, 512]> var_9105_cast_fp16 = mul(x = out_295_cast_fp16, y = tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9105_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_41_cast_fp16 = add(x = input_249_cast_fp16, y = var_9105_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<fp16, [1]> tm_layers_20_layer_scalar_to_fp16 = const()[name = string("tm_layers_20_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ecp-1])];
            tensor<fp16, [1, 64, 512]> x_379_cast_fp16 = mul(x = h_41_cast_fp16, y = tm_layers_20_layer_scalar_to_fp16)[name = string("x_379_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9109_cast_fp16 = abs(x = x_379_cast_fp16)[name = string("op_9109_cast_fp16")];
            bool var_9112 = const()[name = string("op_9112"), val = bool(true)];
            tensor<int32, [1]> var_9113_axes_0 = const()[name = string("op_9113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9113_cast_fp16 = reduce_max(axes = var_9113_axes_0, keep_dims = var_9112, x = var_9109_cast_fp16)[name = string("op_9113_cast_fp16")];
            fp16 var_9114_to_fp16 = const()[name = string("op_9114_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_339_cast_fp16 = maximum(x = var_9113_cast_fp16, y = var_9114_to_fp16)[name = string("s_339_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_339_cast_fp16 = real_div(x = x_379_cast_fp16, y = s_339_cast_fp16)[name = string("y_339_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9117_cast_fp16 = mul(x = y_339_cast_fp16, y = y_339_cast_fp16)[name = string("op_9117_cast_fp16")];
            tensor<int32, [1]> var_9122_axes_0 = const()[name = string("op_9122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9122_keep_dims_0 = const()[name = string("op_9122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9122_cast_fp16 = reduce_mean(axes = var_9122_axes_0, keep_dims = var_9122_keep_dims_0, x = var_9117_cast_fp16)[name = string("op_9122_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9123_cast_fp16 = mul(x = s_339_cast_fp16, y = s_339_cast_fp16)[name = string("op_9123_cast_fp16")];
            fp32 var_9124_epsilon_0 = const()[name = string("op_9124_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9124_cast_fp16 = inverse(epsilon = var_9124_epsilon_0, x = var_9123_cast_fp16)[name = string("op_9124_cast_fp16")];
            fp16 var_9125_to_fp16 = const()[name = string("op_9125_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9126_cast_fp16 = mul(x = var_9124_cast_fp16, y = var_9125_to_fp16)[name = string("op_9126_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9128_cast_fp16 = add(x = var_9122_cast_fp16, y = var_9126_cast_fp16)[name = string("op_9128_cast_fp16")];
            fp32 var_9129_epsilon_0 = const()[name = string("op_9129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9129_cast_fp16 = rsqrt(epsilon = var_9129_epsilon_0, x = var_9128_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_297_cast_fp16 = mul(x = y_339_cast_fp16, y = var_9129_cast_fp16)[name = string("out_297_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120104448)))];
            tensor<fp16, [1, 64, 512]> input_253_cast_fp16 = mul(x = out_297_cast_fp16, y = tm_layers_21_input_layernorm_weight_to_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120105536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120629888))))[name = string("tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9139 = const()[name = string("op_9139"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_9140_cast_fp16 = reshape(shape = var_9139, x = linear_190_cast_fp16)[name = string("op_9140_cast_fp16")];
            tensor<int32, [4]> x_381_perm_0 = const()[name = string("x_381_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120632000))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120894208))))[name = string("tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9151 = const()[name = string("op_9151"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_9152_cast_fp16 = reshape(shape = var_9151, x = linear_191_cast_fp16)[name = string("op_9152_cast_fp16")];
            tensor<int32, [4]> x_383_perm_0 = const()[name = string("x_383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120895296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121157504))))[name = string("tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_9163 = const()[name = string("op_9163"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_9164_cast_fp16 = reshape(shape = var_9163, x = linear_192_cast_fp16)[name = string("op_9164_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_381_cast_fp16 = transpose(perm = x_381_perm_0, x = var_9140_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 4, 64, 256]> var_9168_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_9168_cast_fp16")];
            bool var_9171 = const()[name = string("op_9171"), val = bool(true)];
            tensor<int32, [1]> var_9172_axes_0 = const()[name = string("op_9172_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_9172_cast_fp16 = reduce_max(axes = var_9172_axes_0, keep_dims = var_9171, x = var_9168_cast_fp16)[name = string("op_9172_cast_fp16")];
            fp16 var_9173_to_fp16 = const()[name = string("op_9173_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_341_cast_fp16 = maximum(x = var_9172_cast_fp16, y = var_9173_to_fp16)[name = string("s_341_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_341_cast_fp16 = real_div(x = x_381_cast_fp16, y = s_341_cast_fp16)[name = string("y_341_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_9176_cast_fp16 = mul(x = y_341_cast_fp16, y = y_341_cast_fp16)[name = string("op_9176_cast_fp16")];
            tensor<int32, [1]> var_9181_axes_0 = const()[name = string("op_9181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9181_keep_dims_0 = const()[name = string("op_9181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_9181_cast_fp16 = reduce_mean(axes = var_9181_axes_0, keep_dims = var_9181_keep_dims_0, x = var_9176_cast_fp16)[name = string("op_9181_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_9182_cast_fp16 = mul(x = s_341_cast_fp16, y = s_341_cast_fp16)[name = string("op_9182_cast_fp16")];
            fp32 var_9183_epsilon_0 = const()[name = string("op_9183_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_9183_cast_fp16 = inverse(epsilon = var_9183_epsilon_0, x = var_9182_cast_fp16)[name = string("op_9183_cast_fp16")];
            fp16 var_9184_to_fp16 = const()[name = string("op_9184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_9185_cast_fp16 = mul(x = var_9183_cast_fp16, y = var_9184_to_fp16)[name = string("op_9185_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_9187_cast_fp16 = add(x = var_9181_cast_fp16, y = var_9185_cast_fp16)[name = string("op_9187_cast_fp16")];
            fp32 var_9188_epsilon_0 = const()[name = string("op_9188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_9188_cast_fp16 = rsqrt(epsilon = var_9188_epsilon_0, x = var_9187_cast_fp16)[name = string("op_9188_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_299_cast_fp16 = mul(x = y_341_cast_fp16, y = var_9188_cast_fp16)[name = string("out_299_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_85_cast_fp16 = mul(x = out_299_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_85_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_383_cast_fp16 = transpose(perm = x_383_perm_0, x = var_9152_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2, 64, 256]> var_9191_cast_fp16 = abs(x = x_383_cast_fp16)[name = string("op_9191_cast_fp16")];
            bool var_9194 = const()[name = string("op_9194"), val = bool(true)];
            tensor<int32, [1]> var_9195_axes_0 = const()[name = string("op_9195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_9195_cast_fp16 = reduce_max(axes = var_9195_axes_0, keep_dims = var_9194, x = var_9191_cast_fp16)[name = string("op_9195_cast_fp16")];
            fp16 var_9196_to_fp16 = const()[name = string("op_9196_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_343_cast_fp16 = maximum(x = var_9195_cast_fp16, y = var_9196_to_fp16)[name = string("s_343_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_343_cast_fp16 = real_div(x = x_383_cast_fp16, y = s_343_cast_fp16)[name = string("y_343_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_9199_cast_fp16 = mul(x = y_343_cast_fp16, y = y_343_cast_fp16)[name = string("op_9199_cast_fp16")];
            tensor<int32, [1]> var_9204_axes_0 = const()[name = string("op_9204_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9204_keep_dims_0 = const()[name = string("op_9204_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_9204_cast_fp16 = reduce_mean(axes = var_9204_axes_0, keep_dims = var_9204_keep_dims_0, x = var_9199_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_9205_cast_fp16 = mul(x = s_343_cast_fp16, y = s_343_cast_fp16)[name = string("op_9205_cast_fp16")];
            fp32 var_9206_epsilon_0 = const()[name = string("op_9206_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_9206_cast_fp16 = inverse(epsilon = var_9206_epsilon_0, x = var_9205_cast_fp16)[name = string("op_9206_cast_fp16")];
            fp16 var_9207_to_fp16 = const()[name = string("op_9207_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_9208_cast_fp16 = mul(x = var_9206_cast_fp16, y = var_9207_to_fp16)[name = string("op_9208_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_9210_cast_fp16 = add(x = var_9204_cast_fp16, y = var_9208_cast_fp16)[name = string("op_9210_cast_fp16")];
            fp32 var_9211_epsilon_0 = const()[name = string("op_9211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_9211_cast_fp16 = rsqrt(epsilon = var_9211_epsilon_0, x = var_9210_cast_fp16)[name = string("op_9211_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_301_cast_fp16 = mul(x = y_343_cast_fp16, y = var_9211_cast_fp16)[name = string("out_301_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_127_cast_fp16 = mul(x = out_301_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_127_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_9164_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2, 64, 256]> var_9214_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_9214_cast_fp16")];
            bool var_9217 = const()[name = string("op_9217"), val = bool(true)];
            tensor<int32, [1]> var_9218_axes_0 = const()[name = string("op_9218_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_9218_cast_fp16 = reduce_max(axes = var_9218_axes_0, keep_dims = var_9217, x = var_9214_cast_fp16)[name = string("op_9218_cast_fp16")];
            fp16 var_9219_to_fp16 = const()[name = string("op_9219_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_345_cast_fp16 = maximum(x = var_9218_cast_fp16, y = var_9219_to_fp16)[name = string("s_345_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_345_cast_fp16 = real_div(x = x_385_cast_fp16, y = s_345_cast_fp16)[name = string("y_345_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_9222_cast_fp16 = mul(x = y_345_cast_fp16, y = y_345_cast_fp16)[name = string("op_9222_cast_fp16")];
            tensor<int32, [1]> var_9227_axes_0 = const()[name = string("op_9227_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9227_keep_dims_0 = const()[name = string("op_9227_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_9227_cast_fp16 = reduce_mean(axes = var_9227_axes_0, keep_dims = var_9227_keep_dims_0, x = var_9222_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_9228_cast_fp16 = mul(x = s_345_cast_fp16, y = s_345_cast_fp16)[name = string("op_9228_cast_fp16")];
            fp32 var_9229_epsilon_0 = const()[name = string("op_9229_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_9229_cast_fp16 = inverse(epsilon = var_9229_epsilon_0, x = var_9228_cast_fp16)[name = string("op_9229_cast_fp16")];
            fp16 var_9230_to_fp16 = const()[name = string("op_9230_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_9231_cast_fp16 = mul(x = var_9229_cast_fp16, y = var_9230_to_fp16)[name = string("op_9231_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_9233_cast_fp16 = add(x = var_9227_cast_fp16, y = var_9231_cast_fp16)[name = string("op_9233_cast_fp16")];
            fp32 var_9234_epsilon_0 = const()[name = string("op_9234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_9234_cast_fp16 = rsqrt(epsilon = var_9234_epsilon_0, x = var_9233_cast_fp16)[name = string("op_9234_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_85_cast_fp16 = mul(x = y_345_cast_fp16, y = var_9234_cast_fp16)[name = string("v_85_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_9236_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9236_cast_fp16")];
            tensor<int32, [4]> var_9241_begin_0 = const()[name = string("op_9241_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9241_end_0 = const()[name = string("op_9241_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_9241_end_mask_0 = const()[name = string("op_9241_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_9241_cast_fp16 = slice_by_index(begin = var_9241_begin_0, end = var_9241_end_0, end_mask = var_9241_end_mask_0, x = q_85_cast_fp16)[name = string("op_9241_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_9242_cast_fp16 = mul(x = var_9241_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_9242_cast_fp16")];
            tensor<int32, [4]> var_9247_begin_0 = const()[name = string("op_9247_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9247_end_0 = const()[name = string("op_9247_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_9247_end_mask_0 = const()[name = string("op_9247_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_9247_cast_fp16 = slice_by_index(begin = var_9247_begin_0, end = var_9247_end_0, end_mask = var_9247_end_mask_0, x = q_85_cast_fp16)[name = string("op_9247_cast_fp16")];
            int32 var_9249 = const()[name = string("op_9249"), val = int32(-1)];
            bool var_9250_interleave_0 = const()[name = string("op_9250_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_9250_cast_fp16 = concat(axis = var_9249, interleave = var_9250_interleave_0, values = (var_9242_cast_fp16, var_9247_cast_fp16))[name = string("op_9250_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_9251_cast_fp16 = mul(x = var_9250_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9251_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_87_cast_fp16 = add(x = var_9236_cast_fp16, y = var_9251_cast_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_9254_cast_fp16 = mul(x = k_127_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9254_cast_fp16")];
            tensor<int32, [4]> var_9259_begin_0 = const()[name = string("op_9259_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9259_end_0 = const()[name = string("op_9259_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_9259_end_mask_0 = const()[name = string("op_9259_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_9259_cast_fp16 = slice_by_index(begin = var_9259_begin_0, end = var_9259_end_0, end_mask = var_9259_end_mask_0, x = k_127_cast_fp16)[name = string("op_9259_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_9260_cast_fp16 = mul(x = var_9259_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_9260_cast_fp16")];
            tensor<int32, [4]> var_9265_begin_0 = const()[name = string("op_9265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9265_end_0 = const()[name = string("op_9265_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_9265_end_mask_0 = const()[name = string("op_9265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_9265_cast_fp16 = slice_by_index(begin = var_9265_begin_0, end = var_9265_end_0, end_mask = var_9265_end_mask_0, x = k_127_cast_fp16)[name = string("op_9265_cast_fp16")];
            int32 var_9267 = const()[name = string("op_9267"), val = int32(-1)];
            bool var_9268_interleave_0 = const()[name = string("op_9268_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_9268_cast_fp16 = concat(axis = var_9267, interleave = var_9268_interleave_0, values = (var_9260_cast_fp16, var_9265_cast_fp16))[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_9269_cast_fp16 = mul(x = var_9268_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_129_cast_fp16 = add(x = var_9254_cast_fp16, y = var_9269_cast_fp16)[name = string("k_129_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_42_reps_0 = const()[name = string("tile_42_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = k_129_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 1, 64, 256]> tile_42_cast_fp16 = tile(reps = tile_42_reps_0, x = transpose_84_cast_fp16)[name = string("tile_42_cast_fp16")];
            tensor<int32, [5]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_84_cast_fp16 = reshape(shape = concat_84, x = tile_42_cast_fp16)[name = string("reshape_84_cast_fp16")];
            tensor<int32, [5]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = reshape_84_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 1, 64, 256]> reshape_85_cast_fp16 = reshape(shape = concat_85, x = transpose_85_cast_fp16)[name = string("reshape_85_cast_fp16")];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_43_reps_0 = const()[name = string("tile_43_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_86_cast_fp16 = transpose(perm = transpose_86_perm_0, x = v_85_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 1, 64, 256]> tile_43_cast_fp16 = tile(reps = tile_43_reps_0, x = transpose_86_cast_fp16)[name = string("tile_43_cast_fp16")];
            tensor<int32, [5]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_86_cast_fp16 = reshape(shape = concat_86, x = tile_43_cast_fp16)[name = string("reshape_86_cast_fp16")];
            tensor<int32, [5]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_87_cast_fp16 = transpose(perm = transpose_87_perm_0, x = reshape_86_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 1, 64, 256]> reshape_87_cast_fp16 = reshape(shape = concat_87, x = transpose_87_cast_fp16)[name = string("reshape_87_cast_fp16")];
            tensor<int32, [4]> v_87_perm_0 = const()[name = string("v_87_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9283_transpose_x_0 = const()[name = string("op_9283_transpose_x_0"), val = bool(false)];
            bool var_9283_transpose_y_0 = const()[name = string("op_9283_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_117_cast_fp16 = transpose(perm = transpose_117_perm_0, x = reshape_85_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 4, 64, 64]> var_9283_cast_fp16 = matmul(transpose_x = var_9283_transpose_x_0, transpose_y = var_9283_transpose_y_0, x = q_87_cast_fp16, y = transpose_117_cast_fp16)[name = string("op_9283_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_9285_cast_fp16 = add(x = var_9283_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9285_cast_fp16")];
            int32 var_9286 = const()[name = string("op_9286"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_345_cast_fp16 = softmax(axis = var_9286, x = var_9285_cast_fp16)[name = string("w_345_cast_fp16")];
            bool var_9289_transpose_x_0 = const()[name = string("op_9289_transpose_x_0"), val = bool(false)];
            bool var_9289_transpose_y_0 = const()[name = string("op_9289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_87_cast_fp16 = transpose(perm = v_87_perm_0, x = reshape_87_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 4, 64, 256]> var_9289_cast_fp16 = matmul(transpose_x = var_9289_transpose_x_0, transpose_y = var_9289_transpose_y_0, x = w_345_cast_fp16, y = v_87_cast_fp16)[name = string("op_9289_cast_fp16")];
            tensor<int32, [4]> var_9292_perm_0 = const()[name = string("op_9292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9296 = const()[name = string("op_9296"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_9292_cast_fp16 = transpose(perm = var_9292_perm_0, x = var_9289_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 64, 1024]> input_255_cast_fp16 = reshape(shape = var_9296, x = var_9292_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121158592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121682944))))[name = string("tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9301_cast_fp16 = abs(x = linear_193_cast_fp16)[name = string("op_9301_cast_fp16")];
            bool var_9304 = const()[name = string("op_9304"), val = bool(true)];
            tensor<int32, [1]> var_9305_axes_0 = const()[name = string("op_9305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9305_cast_fp16 = reduce_max(axes = var_9305_axes_0, keep_dims = var_9304, x = var_9301_cast_fp16)[name = string("op_9305_cast_fp16")];
            fp16 var_9306_to_fp16 = const()[name = string("op_9306_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_347_cast_fp16 = maximum(x = var_9305_cast_fp16, y = var_9306_to_fp16)[name = string("s_347_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_347_cast_fp16 = real_div(x = linear_193_cast_fp16, y = s_347_cast_fp16)[name = string("y_347_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9309_cast_fp16 = mul(x = y_347_cast_fp16, y = y_347_cast_fp16)[name = string("op_9309_cast_fp16")];
            tensor<int32, [1]> var_9314_axes_0 = const()[name = string("op_9314_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9314_keep_dims_0 = const()[name = string("op_9314_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9314_cast_fp16 = reduce_mean(axes = var_9314_axes_0, keep_dims = var_9314_keep_dims_0, x = var_9309_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9315_cast_fp16 = mul(x = s_347_cast_fp16, y = s_347_cast_fp16)[name = string("op_9315_cast_fp16")];
            fp32 var_9316_epsilon_0 = const()[name = string("op_9316_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9316_cast_fp16 = inverse(epsilon = var_9316_epsilon_0, x = var_9315_cast_fp16)[name = string("op_9316_cast_fp16")];
            fp16 var_9317_to_fp16 = const()[name = string("op_9317_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9318_cast_fp16 = mul(x = var_9316_cast_fp16, y = var_9317_to_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9320_cast_fp16 = add(x = var_9314_cast_fp16, y = var_9318_cast_fp16)[name = string("op_9320_cast_fp16")];
            fp32 var_9321_epsilon_0 = const()[name = string("op_9321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9321_cast_fp16 = rsqrt(epsilon = var_9321_epsilon_0, x = var_9320_cast_fp16)[name = string("op_9321_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_303_cast_fp16 = mul(x = y_347_cast_fp16, y = var_9321_cast_fp16)[name = string("out_303_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121684032)))];
            tensor<fp16, [1, 64, 512]> var_9323_cast_fp16 = mul(x = out_303_cast_fp16, y = tm_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("op_9323_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_389_cast_fp16 = add(x = x_379_cast_fp16, y = var_9323_cast_fp16)[name = string("x_389_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9326_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_9326_cast_fp16")];
            bool var_9329 = const()[name = string("op_9329"), val = bool(true)];
            tensor<int32, [1]> var_9330_axes_0 = const()[name = string("op_9330_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9330_cast_fp16 = reduce_max(axes = var_9330_axes_0, keep_dims = var_9329, x = var_9326_cast_fp16)[name = string("op_9330_cast_fp16")];
            fp16 var_9331_to_fp16 = const()[name = string("op_9331_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_349_cast_fp16 = maximum(x = var_9330_cast_fp16, y = var_9331_to_fp16)[name = string("s_349_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_349_cast_fp16 = real_div(x = x_389_cast_fp16, y = s_349_cast_fp16)[name = string("y_349_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9334_cast_fp16 = mul(x = y_349_cast_fp16, y = y_349_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [1]> var_9339_axes_0 = const()[name = string("op_9339_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9339_keep_dims_0 = const()[name = string("op_9339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9339_cast_fp16 = reduce_mean(axes = var_9339_axes_0, keep_dims = var_9339_keep_dims_0, x = var_9334_cast_fp16)[name = string("op_9339_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9340_cast_fp16 = mul(x = s_349_cast_fp16, y = s_349_cast_fp16)[name = string("op_9340_cast_fp16")];
            fp32 var_9341_epsilon_0 = const()[name = string("op_9341_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9341_cast_fp16 = inverse(epsilon = var_9341_epsilon_0, x = var_9340_cast_fp16)[name = string("op_9341_cast_fp16")];
            fp16 var_9342_to_fp16 = const()[name = string("op_9342_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9343_cast_fp16 = mul(x = var_9341_cast_fp16, y = var_9342_to_fp16)[name = string("op_9343_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9345_cast_fp16 = add(x = var_9339_cast_fp16, y = var_9343_cast_fp16)[name = string("op_9345_cast_fp16")];
            fp32 var_9346_epsilon_0 = const()[name = string("op_9346_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9346_cast_fp16 = rsqrt(epsilon = var_9346_epsilon_0, x = var_9345_cast_fp16)[name = string("op_9346_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_305_cast_fp16 = mul(x = y_349_cast_fp16, y = var_9346_cast_fp16)[name = string("out_305_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121685120)))];
            tensor<fp16, [1, 64, 512]> input_257_cast_fp16 = mul(x = out_305_cast_fp16, y = tm_layers_21_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121686208))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122734848))))[name = string("tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_194_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_194_cast_fp16")];
            string var_9353_mode_0 = const()[name = string("op_9353_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_9353_cast_fp16 = gelu(mode = var_9353_mode_0, x = linear_194_cast_fp16)[name = string("op_9353_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122739008))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123787648))))[name = string("tm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_195_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_259_cast_fp16 = mul(x = var_9353_cast_fp16, y = linear_195_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123791808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124840448))))[name = string("tm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9361_cast_fp16 = abs(x = linear_196_cast_fp16)[name = string("op_9361_cast_fp16")];
            bool var_9364 = const()[name = string("op_9364"), val = bool(true)];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9365_cast_fp16 = reduce_max(axes = var_9365_axes_0, keep_dims = var_9364, x = var_9361_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9366_to_fp16 = const()[name = string("op_9366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_351_cast_fp16 = maximum(x = var_9365_cast_fp16, y = var_9366_to_fp16)[name = string("s_351_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_351_cast_fp16 = real_div(x = linear_196_cast_fp16, y = s_351_cast_fp16)[name = string("y_351_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9369_cast_fp16 = mul(x = y_351_cast_fp16, y = y_351_cast_fp16)[name = string("op_9369_cast_fp16")];
            tensor<int32, [1]> var_9374_axes_0 = const()[name = string("op_9374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9374_keep_dims_0 = const()[name = string("op_9374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9374_cast_fp16 = reduce_mean(axes = var_9374_axes_0, keep_dims = var_9374_keep_dims_0, x = var_9369_cast_fp16)[name = string("op_9374_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9375_cast_fp16 = mul(x = s_351_cast_fp16, y = s_351_cast_fp16)[name = string("op_9375_cast_fp16")];
            fp32 var_9376_epsilon_0 = const()[name = string("op_9376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9376_cast_fp16 = inverse(epsilon = var_9376_epsilon_0, x = var_9375_cast_fp16)[name = string("op_9376_cast_fp16")];
            fp16 var_9377_to_fp16 = const()[name = string("op_9377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9378_cast_fp16 = mul(x = var_9376_cast_fp16, y = var_9377_to_fp16)[name = string("op_9378_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9380_cast_fp16 = add(x = var_9374_cast_fp16, y = var_9378_cast_fp16)[name = string("op_9380_cast_fp16")];
            fp32 var_9381_epsilon_0 = const()[name = string("op_9381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9381_cast_fp16 = rsqrt(epsilon = var_9381_epsilon_0, x = var_9380_cast_fp16)[name = string("op_9381_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_307_cast_fp16 = mul(x = y_351_cast_fp16, y = var_9381_cast_fp16)[name = string("out_307_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124841536)))];
            tensor<fp16, [1, 64, 512]> var_9383_cast_fp16 = mul(x = out_307_cast_fp16, y = tm_layers_21_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9383_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_261_cast_fp16 = add(x = x_389_cast_fp16, y = var_9383_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124842624))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125104832))))[name = string("tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_197_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_197_cast_fp16")];
            string x_393_mode_0 = const()[name = string("x_393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_393_cast_fp16 = gelu(mode = x_393_mode_0, x = linear_197_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<int32, [4]> var_9403_begin_0 = const()[name = string("op_9403_begin_0"), val = tensor<int32, [4]>([0, 0, 21, 0])];
            tensor<int32, [4]> var_9403_end_0 = const()[name = string("op_9403_end_0"), val = tensor<int32, [4]>([1, 64, 22, 512])];
            tensor<bool, [4]> var_9403_end_mask_0 = const()[name = string("op_9403_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9403_squeeze_mask_0 = const()[name = string("op_9403_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_9403_cast_fp16 = slice_by_index(begin = var_9403_begin_0, end = var_9403_end_0, end_mask = var_9403_end_mask_0, squeeze_mask = var_9403_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9403_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_263_cast_fp16 = mul(x = x_393_cast_fp16, y = var_9403_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125105920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125368128))))[name = string("tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9413_cast_fp16 = abs(x = linear_198_cast_fp16)[name = string("op_9413_cast_fp16")];
            bool var_9416 = const()[name = string("op_9416"), val = bool(true)];
            tensor<int32, [1]> var_9417_axes_0 = const()[name = string("op_9417_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9417_cast_fp16 = reduce_max(axes = var_9417_axes_0, keep_dims = var_9416, x = var_9413_cast_fp16)[name = string("op_9417_cast_fp16")];
            fp16 var_9418_to_fp16 = const()[name = string("op_9418_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_353_cast_fp16 = maximum(x = var_9417_cast_fp16, y = var_9418_to_fp16)[name = string("s_353_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_353_cast_fp16 = real_div(x = linear_198_cast_fp16, y = s_353_cast_fp16)[name = string("y_353_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9421_cast_fp16 = mul(x = y_353_cast_fp16, y = y_353_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<int32, [1]> var_9426_axes_0 = const()[name = string("op_9426_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9426_keep_dims_0 = const()[name = string("op_9426_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9426_cast_fp16 = reduce_mean(axes = var_9426_axes_0, keep_dims = var_9426_keep_dims_0, x = var_9421_cast_fp16)[name = string("op_9426_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9427_cast_fp16 = mul(x = s_353_cast_fp16, y = s_353_cast_fp16)[name = string("op_9427_cast_fp16")];
            fp32 var_9428_epsilon_0 = const()[name = string("op_9428_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9428_cast_fp16 = inverse(epsilon = var_9428_epsilon_0, x = var_9427_cast_fp16)[name = string("op_9428_cast_fp16")];
            fp16 var_9429_to_fp16 = const()[name = string("op_9429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9430_cast_fp16 = mul(x = var_9428_cast_fp16, y = var_9429_to_fp16)[name = string("op_9430_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9432_cast_fp16 = add(x = var_9426_cast_fp16, y = var_9430_cast_fp16)[name = string("op_9432_cast_fp16")];
            fp32 var_9433_epsilon_0 = const()[name = string("op_9433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9433_cast_fp16 = rsqrt(epsilon = var_9433_epsilon_0, x = var_9432_cast_fp16)[name = string("op_9433_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_309_cast_fp16 = mul(x = y_353_cast_fp16, y = var_9433_cast_fp16)[name = string("out_309_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125369216)))];
            tensor<fp16, [1, 64, 512]> var_9435_cast_fp16 = mul(x = out_309_cast_fp16, y = tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9435_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_43_cast_fp16 = add(x = input_261_cast_fp16, y = var_9435_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<fp16, [1]> tm_layers_21_layer_scalar_to_fp16 = const()[name = string("tm_layers_21_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e8p-1])];
            tensor<fp16, [1, 64, 512]> x_397_cast_fp16 = mul(x = h_43_cast_fp16, y = tm_layers_21_layer_scalar_to_fp16)[name = string("x_397_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9439_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_9439_cast_fp16")];
            bool var_9442 = const()[name = string("op_9442"), val = bool(true)];
            tensor<int32, [1]> var_9443_axes_0 = const()[name = string("op_9443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9443_cast_fp16 = reduce_max(axes = var_9443_axes_0, keep_dims = var_9442, x = var_9439_cast_fp16)[name = string("op_9443_cast_fp16")];
            fp16 var_9444_to_fp16 = const()[name = string("op_9444_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_355_cast_fp16 = maximum(x = var_9443_cast_fp16, y = var_9444_to_fp16)[name = string("s_355_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_355_cast_fp16 = real_div(x = x_397_cast_fp16, y = s_355_cast_fp16)[name = string("y_355_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9447_cast_fp16 = mul(x = y_355_cast_fp16, y = y_355_cast_fp16)[name = string("op_9447_cast_fp16")];
            tensor<int32, [1]> var_9452_axes_0 = const()[name = string("op_9452_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9452_keep_dims_0 = const()[name = string("op_9452_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9452_cast_fp16 = reduce_mean(axes = var_9452_axes_0, keep_dims = var_9452_keep_dims_0, x = var_9447_cast_fp16)[name = string("op_9452_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9453_cast_fp16 = mul(x = s_355_cast_fp16, y = s_355_cast_fp16)[name = string("op_9453_cast_fp16")];
            fp32 var_9454_epsilon_0 = const()[name = string("op_9454_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9454_cast_fp16 = inverse(epsilon = var_9454_epsilon_0, x = var_9453_cast_fp16)[name = string("op_9454_cast_fp16")];
            fp16 var_9455_to_fp16 = const()[name = string("op_9455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9456_cast_fp16 = mul(x = var_9454_cast_fp16, y = var_9455_to_fp16)[name = string("op_9456_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9458_cast_fp16 = add(x = var_9452_cast_fp16, y = var_9456_cast_fp16)[name = string("op_9458_cast_fp16")];
            fp32 var_9459_epsilon_0 = const()[name = string("op_9459_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9459_cast_fp16 = rsqrt(epsilon = var_9459_epsilon_0, x = var_9458_cast_fp16)[name = string("op_9459_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_311_cast_fp16 = mul(x = y_355_cast_fp16, y = var_9459_cast_fp16)[name = string("out_311_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125370304)))];
            tensor<fp16, [1, 64, 512]> input_265_cast_fp16 = mul(x = out_311_cast_fp16, y = tm_layers_22_input_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125371392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125895744))))[name = string("tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 1024]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<int32, [4]> var_9469 = const()[name = string("op_9469"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 64, 4, 256]> var_9470_cast_fp16 = reshape(shape = var_9469, x = linear_199_cast_fp16)[name = string("op_9470_cast_fp16")];
            tensor<int32, [4]> x_399_perm_0 = const()[name = string("x_399_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125897856))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126160064))))[name = string("tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<int32, [4]> var_9481 = const()[name = string("op_9481"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_9482_cast_fp16 = reshape(shape = var_9481, x = linear_200_cast_fp16)[name = string("op_9482_cast_fp16")];
            tensor<int32, [4]> x_401_perm_0 = const()[name = string("x_401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126161152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126423360))))[name = string("tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<int32, [4]> var_9493 = const()[name = string("op_9493"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 64, 2, 256]> var_9494_cast_fp16 = reshape(shape = var_9493, x = linear_201_cast_fp16)[name = string("op_9494_cast_fp16")];
            tensor<int32, [4]> x_403_perm_0 = const()[name = string("x_403_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 256]> x_399_cast_fp16 = transpose(perm = x_399_perm_0, x = var_9470_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 4, 64, 256]> var_9498_cast_fp16 = abs(x = x_399_cast_fp16)[name = string("op_9498_cast_fp16")];
            bool var_9501 = const()[name = string("op_9501"), val = bool(true)];
            tensor<int32, [1]> var_9502_axes_0 = const()[name = string("op_9502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_9502_cast_fp16 = reduce_max(axes = var_9502_axes_0, keep_dims = var_9501, x = var_9498_cast_fp16)[name = string("op_9502_cast_fp16")];
            fp16 var_9503_to_fp16 = const()[name = string("op_9503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_357_cast_fp16 = maximum(x = var_9502_cast_fp16, y = var_9503_to_fp16)[name = string("s_357_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> y_357_cast_fp16 = real_div(x = x_399_cast_fp16, y = s_357_cast_fp16)[name = string("y_357_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_9506_cast_fp16 = mul(x = y_357_cast_fp16, y = y_357_cast_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9511_axes_0 = const()[name = string("op_9511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9511_keep_dims_0 = const()[name = string("op_9511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_9511_cast_fp16 = reduce_mean(axes = var_9511_axes_0, keep_dims = var_9511_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_9512_cast_fp16 = mul(x = s_357_cast_fp16, y = s_357_cast_fp16)[name = string("op_9512_cast_fp16")];
            fp32 var_9513_epsilon_0 = const()[name = string("op_9513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_9513_cast_fp16 = inverse(epsilon = var_9513_epsilon_0, x = var_9512_cast_fp16)[name = string("op_9513_cast_fp16")];
            fp16 var_9514_to_fp16 = const()[name = string("op_9514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_9515_cast_fp16 = mul(x = var_9513_cast_fp16, y = var_9514_to_fp16)[name = string("op_9515_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_9517_cast_fp16 = add(x = var_9511_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9517_cast_fp16")];
            fp32 var_9518_epsilon_0 = const()[name = string("op_9518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_9518_cast_fp16 = rsqrt(epsilon = var_9518_epsilon_0, x = var_9517_cast_fp16)[name = string("op_9518_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> out_313_cast_fp16 = mul(x = y_357_cast_fp16, y = var_9518_cast_fp16)[name = string("out_313_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_89_cast_fp16 = mul(x = out_313_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_89_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_401_cast_fp16 = transpose(perm = x_401_perm_0, x = var_9482_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 64, 256]> var_9521_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_9521_cast_fp16")];
            bool var_9524 = const()[name = string("op_9524"), val = bool(true)];
            tensor<int32, [1]> var_9525_axes_0 = const()[name = string("op_9525_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_9525_cast_fp16 = reduce_max(axes = var_9525_axes_0, keep_dims = var_9524, x = var_9521_cast_fp16)[name = string("op_9525_cast_fp16")];
            fp16 var_9526_to_fp16 = const()[name = string("op_9526_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_359_cast_fp16 = maximum(x = var_9525_cast_fp16, y = var_9526_to_fp16)[name = string("s_359_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_359_cast_fp16 = real_div(x = x_401_cast_fp16, y = s_359_cast_fp16)[name = string("y_359_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_9529_cast_fp16 = mul(x = y_359_cast_fp16, y = y_359_cast_fp16)[name = string("op_9529_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9529_cast_fp16)[name = string("op_9534_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_9535_cast_fp16 = mul(x = s_359_cast_fp16, y = s_359_cast_fp16)[name = string("op_9535_cast_fp16")];
            fp32 var_9536_epsilon_0 = const()[name = string("op_9536_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_9536_cast_fp16 = inverse(epsilon = var_9536_epsilon_0, x = var_9535_cast_fp16)[name = string("op_9536_cast_fp16")];
            fp16 var_9537_to_fp16 = const()[name = string("op_9537_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_9538_cast_fp16 = mul(x = var_9536_cast_fp16, y = var_9537_to_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_9540_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9538_cast_fp16)[name = string("op_9540_cast_fp16")];
            fp32 var_9541_epsilon_0 = const()[name = string("op_9541_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_9541_cast_fp16 = rsqrt(epsilon = var_9541_epsilon_0, x = var_9540_cast_fp16)[name = string("op_9541_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> out_315_cast_fp16 = mul(x = y_359_cast_fp16, y = var_9541_cast_fp16)[name = string("out_315_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_133_cast_fp16 = mul(x = out_315_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_133_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> x_403_cast_fp16 = transpose(perm = x_403_perm_0, x = var_9494_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 64, 256]> var_9544_cast_fp16 = abs(x = x_403_cast_fp16)[name = string("op_9544_cast_fp16")];
            bool var_9547 = const()[name = string("op_9547"), val = bool(true)];
            tensor<int32, [1]> var_9548_axes_0 = const()[name = string("op_9548_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 64, 1]> var_9548_cast_fp16 = reduce_max(axes = var_9548_axes_0, keep_dims = var_9547, x = var_9544_cast_fp16)[name = string("op_9548_cast_fp16")];
            fp16 var_9549_to_fp16 = const()[name = string("op_9549_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 64, 1]> s_361_cast_fp16 = maximum(x = var_9548_cast_fp16, y = var_9549_to_fp16)[name = string("s_361_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> y_361_cast_fp16 = real_div(x = x_403_cast_fp16, y = s_361_cast_fp16)[name = string("y_361_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_9552_cast_fp16 = mul(x = y_361_cast_fp16, y = y_361_cast_fp16)[name = string("op_9552_cast_fp16")];
            tensor<int32, [1]> var_9557_axes_0 = const()[name = string("op_9557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9557_keep_dims_0 = const()[name = string("op_9557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 64, 1]> var_9557_cast_fp16 = reduce_mean(axes = var_9557_axes_0, keep_dims = var_9557_keep_dims_0, x = var_9552_cast_fp16)[name = string("op_9557_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_9558_cast_fp16 = mul(x = s_361_cast_fp16, y = s_361_cast_fp16)[name = string("op_9558_cast_fp16")];
            fp32 var_9559_epsilon_0 = const()[name = string("op_9559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 64, 1]> var_9559_cast_fp16 = inverse(epsilon = var_9559_epsilon_0, x = var_9558_cast_fp16)[name = string("op_9559_cast_fp16")];
            fp16 var_9560_to_fp16 = const()[name = string("op_9560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 64, 1]> var_9561_cast_fp16 = mul(x = var_9559_cast_fp16, y = var_9560_to_fp16)[name = string("op_9561_cast_fp16")];
            tensor<fp16, [1, 2, 64, 1]> var_9563_cast_fp16 = add(x = var_9557_cast_fp16, y = var_9561_cast_fp16)[name = string("op_9563_cast_fp16")];
            fp32 var_9564_epsilon_0 = const()[name = string("op_9564_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 64, 1]> var_9564_cast_fp16 = rsqrt(epsilon = var_9564_epsilon_0, x = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> v_89_cast_fp16 = mul(x = y_361_cast_fp16, y = var_9564_cast_fp16)[name = string("v_89_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_9566_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9566_cast_fp16")];
            tensor<int32, [4]> var_9571_begin_0 = const()[name = string("op_9571_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9571_end_0 = const()[name = string("op_9571_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_9571_end_mask_0 = const()[name = string("op_9571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 128]> var_9571_cast_fp16 = slice_by_index(begin = var_9571_begin_0, end = var_9571_end_0, end_mask = var_9571_end_mask_0, x = q_89_cast_fp16)[name = string("op_9571_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 128]> var_9572_cast_fp16 = mul(x = var_9571_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_9572_cast_fp16")];
            tensor<int32, [4]> var_9577_begin_0 = const()[name = string("op_9577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9577_end_0 = const()[name = string("op_9577_end_0"), val = tensor<int32, [4]>([1, 4, 64, 128])];
            tensor<bool, [4]> var_9577_end_mask_0 = const()[name = string("op_9577_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 128]> var_9577_cast_fp16 = slice_by_index(begin = var_9577_begin_0, end = var_9577_end_0, end_mask = var_9577_end_mask_0, x = q_89_cast_fp16)[name = string("op_9577_cast_fp16")];
            int32 var_9579 = const()[name = string("op_9579"), val = int32(-1)];
            bool var_9580_interleave_0 = const()[name = string("op_9580_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> var_9580_cast_fp16 = concat(axis = var_9579, interleave = var_9580_interleave_0, values = (var_9572_cast_fp16, var_9577_cast_fp16))[name = string("op_9580_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> var_9581_cast_fp16 = mul(x = var_9580_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9581_cast_fp16")];
            tensor<fp16, [1, 4, 64, 256]> q_91_cast_fp16 = add(x = var_9566_cast_fp16, y = var_9581_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_9584_cast_fp16 = mul(x = k_133_cast_fp16, y = cos_1_to_fp16_quantized)[name = string("op_9584_cast_fp16")];
            tensor<int32, [4]> var_9589_begin_0 = const()[name = string("op_9589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9589_end_0 = const()[name = string("op_9589_end_0"), val = tensor<int32, [4]>([1, 2, 64, 256])];
            tensor<bool, [4]> var_9589_end_mask_0 = const()[name = string("op_9589_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 128]> var_9589_cast_fp16 = slice_by_index(begin = var_9589_begin_0, end = var_9589_end_0, end_mask = var_9589_end_mask_0, x = k_133_cast_fp16)[name = string("op_9589_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 128]> var_9590_cast_fp16 = mul(x = var_9589_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_9590_cast_fp16")];
            tensor<int32, [4]> var_9595_begin_0 = const()[name = string("op_9595_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9595_end_0 = const()[name = string("op_9595_end_0"), val = tensor<int32, [4]>([1, 2, 64, 128])];
            tensor<bool, [4]> var_9595_end_mask_0 = const()[name = string("op_9595_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 128]> var_9595_cast_fp16 = slice_by_index(begin = var_9595_begin_0, end = var_9595_end_0, end_mask = var_9595_end_mask_0, x = k_133_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool var_9598_interleave_0 = const()[name = string("op_9598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 256]> var_9598_cast_fp16 = concat(axis = var_9597, interleave = var_9598_interleave_0, values = (var_9590_cast_fp16, var_9595_cast_fp16))[name = string("op_9598_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> var_9599_cast_fp16 = mul(x = var_9598_cast_fp16, y = sin_1_to_fp16_quantized)[name = string("op_9599_cast_fp16")];
            tensor<fp16, [1, 2, 64, 256]> k_135_cast_fp16 = add(x = var_9584_cast_fp16, y = var_9599_cast_fp16)[name = string("k_135_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_44_reps_0 = const()[name = string("tile_44_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = k_135_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 1, 64, 256]> tile_44_cast_fp16 = tile(reps = tile_44_reps_0, x = transpose_88_cast_fp16)[name = string("tile_44_cast_fp16")];
            tensor<int32, [5]> concat_88 = const()[name = string("concat_88"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_88_cast_fp16 = reshape(shape = concat_88, x = tile_44_cast_fp16)[name = string("reshape_88_cast_fp16")];
            tensor<int32, [5]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = reshape_88_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 1, 64, 256]> reshape_89_cast_fp16 = reshape(shape = concat_89, x = transpose_89_cast_fp16)[name = string("reshape_89_cast_fp16")];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_45_reps_0 = const()[name = string("tile_45_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 64, 256]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = v_89_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 1, 64, 256]> tile_45_cast_fp16 = tile(reps = tile_45_reps_0, x = transpose_90_cast_fp16)[name = string("tile_45_cast_fp16")];
            tensor<int32, [5]> concat_90 = const()[name = string("concat_90"), val = tensor<int32, [5]>([2, 2, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> reshape_90_cast_fp16 = reshape(shape = concat_90, x = tile_45_cast_fp16)[name = string("reshape_90_cast_fp16")];
            tensor<int32, [5]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_91 = const()[name = string("concat_91"), val = tensor<int32, [4]>([-1, 1, 64, 256])];
            tensor<fp16, [2, 2, 1, 64, 256]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = reshape_90_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 1, 64, 256]> reshape_91_cast_fp16 = reshape(shape = concat_91, x = transpose_91_cast_fp16)[name = string("reshape_91_cast_fp16")];
            tensor<int32, [4]> v_91_perm_0 = const()[name = string("v_91_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9613_transpose_x_0 = const()[name = string("op_9613_transpose_x_0"), val = bool(false)];
            bool var_9613_transpose_y_0 = const()[name = string("op_9613_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 64]> transpose_118_cast_fp16 = transpose(perm = transpose_118_perm_0, x = reshape_89_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 4, 64, 64]> var_9613_cast_fp16 = matmul(transpose_x = var_9613_transpose_x_0, transpose_y = var_9613_transpose_y_0, x = q_91_cast_fp16, y = transpose_118_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_9615_cast_fp16 = add(x = var_9613_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9615_cast_fp16")];
            int32 var_9616 = const()[name = string("op_9616"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_361_cast_fp16 = softmax(axis = var_9616, x = var_9615_cast_fp16)[name = string("w_361_cast_fp16")];
            bool var_9619_transpose_x_0 = const()[name = string("op_9619_transpose_x_0"), val = bool(false)];
            bool var_9619_transpose_y_0 = const()[name = string("op_9619_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 256]> v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = reshape_91_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 4, 64, 256]> var_9619_cast_fp16 = matmul(transpose_x = var_9619_transpose_x_0, transpose_y = var_9619_transpose_y_0, x = w_361_cast_fp16, y = v_91_cast_fp16)[name = string("op_9619_cast_fp16")];
            tensor<int32, [4]> var_9622_perm_0 = const()[name = string("op_9622_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9626 = const()[name = string("op_9626"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 64, 4, 256]> var_9622_cast_fp16 = transpose(perm = var_9622_perm_0, x = var_9619_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 64, 1024]> input_267_cast_fp16 = reshape(shape = var_9626, x = var_9622_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126424448))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126948800))))[name = string("tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9631_cast_fp16 = abs(x = linear_202_cast_fp16)[name = string("op_9631_cast_fp16")];
            bool var_9634 = const()[name = string("op_9634"), val = bool(true)];
            tensor<int32, [1]> var_9635_axes_0 = const()[name = string("op_9635_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9635_cast_fp16 = reduce_max(axes = var_9635_axes_0, keep_dims = var_9634, x = var_9631_cast_fp16)[name = string("op_9635_cast_fp16")];
            fp16 var_9636_to_fp16 = const()[name = string("op_9636_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_363_cast_fp16 = maximum(x = var_9635_cast_fp16, y = var_9636_to_fp16)[name = string("s_363_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_363_cast_fp16 = real_div(x = linear_202_cast_fp16, y = s_363_cast_fp16)[name = string("y_363_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9639_cast_fp16 = mul(x = y_363_cast_fp16, y = y_363_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [1]> var_9644_axes_0 = const()[name = string("op_9644_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9644_keep_dims_0 = const()[name = string("op_9644_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9644_cast_fp16 = reduce_mean(axes = var_9644_axes_0, keep_dims = var_9644_keep_dims_0, x = var_9639_cast_fp16)[name = string("op_9644_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9645_cast_fp16 = mul(x = s_363_cast_fp16, y = s_363_cast_fp16)[name = string("op_9645_cast_fp16")];
            fp32 var_9646_epsilon_0 = const()[name = string("op_9646_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9646_cast_fp16 = inverse(epsilon = var_9646_epsilon_0, x = var_9645_cast_fp16)[name = string("op_9646_cast_fp16")];
            fp16 var_9647_to_fp16 = const()[name = string("op_9647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9648_cast_fp16 = mul(x = var_9646_cast_fp16, y = var_9647_to_fp16)[name = string("op_9648_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9650_cast_fp16 = add(x = var_9644_cast_fp16, y = var_9648_cast_fp16)[name = string("op_9650_cast_fp16")];
            fp32 var_9651_epsilon_0 = const()[name = string("op_9651_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9651_cast_fp16 = rsqrt(epsilon = var_9651_epsilon_0, x = var_9650_cast_fp16)[name = string("op_9651_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_317_cast_fp16 = mul(x = y_363_cast_fp16, y = var_9651_cast_fp16)[name = string("out_317_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126949888)))];
            tensor<fp16, [1, 64, 512]> var_9653_cast_fp16 = mul(x = out_317_cast_fp16, y = tm_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_407_cast_fp16 = add(x = x_397_cast_fp16, y = var_9653_cast_fp16)[name = string("x_407_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9656_cast_fp16 = abs(x = x_407_cast_fp16)[name = string("op_9656_cast_fp16")];
            bool var_9659 = const()[name = string("op_9659"), val = bool(true)];
            tensor<int32, [1]> var_9660_axes_0 = const()[name = string("op_9660_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9660_cast_fp16 = reduce_max(axes = var_9660_axes_0, keep_dims = var_9659, x = var_9656_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp16 var_9661_to_fp16 = const()[name = string("op_9661_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_365_cast_fp16 = maximum(x = var_9660_cast_fp16, y = var_9661_to_fp16)[name = string("s_365_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_365_cast_fp16 = real_div(x = x_407_cast_fp16, y = s_365_cast_fp16)[name = string("y_365_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9664_cast_fp16 = mul(x = y_365_cast_fp16, y = y_365_cast_fp16)[name = string("op_9664_cast_fp16")];
            tensor<int32, [1]> var_9669_axes_0 = const()[name = string("op_9669_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9669_keep_dims_0 = const()[name = string("op_9669_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9669_cast_fp16 = reduce_mean(axes = var_9669_axes_0, keep_dims = var_9669_keep_dims_0, x = var_9664_cast_fp16)[name = string("op_9669_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9670_cast_fp16 = mul(x = s_365_cast_fp16, y = s_365_cast_fp16)[name = string("op_9670_cast_fp16")];
            fp32 var_9671_epsilon_0 = const()[name = string("op_9671_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9671_cast_fp16 = inverse(epsilon = var_9671_epsilon_0, x = var_9670_cast_fp16)[name = string("op_9671_cast_fp16")];
            fp16 var_9672_to_fp16 = const()[name = string("op_9672_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9673_cast_fp16 = mul(x = var_9671_cast_fp16, y = var_9672_to_fp16)[name = string("op_9673_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9675_cast_fp16 = add(x = var_9669_cast_fp16, y = var_9673_cast_fp16)[name = string("op_9675_cast_fp16")];
            fp32 var_9676_epsilon_0 = const()[name = string("op_9676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9676_cast_fp16 = rsqrt(epsilon = var_9676_epsilon_0, x = var_9675_cast_fp16)[name = string("op_9676_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_319_cast_fp16 = mul(x = y_365_cast_fp16, y = var_9676_cast_fp16)[name = string("out_319_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126950976)))];
            tensor<fp16, [1, 64, 512]> input_269_cast_fp16 = mul(x = out_319_cast_fp16, y = tm_layers_22_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126952064))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128000704))))[name = string("tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_203_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_203_cast_fp16")];
            string var_9683_mode_0 = const()[name = string("op_9683_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_9683_cast_fp16 = gelu(mode = var_9683_mode_0, x = linear_203_cast_fp16)[name = string("op_9683_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128004864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129053504))))[name = string("tm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_204_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_271_cast_fp16 = mul(x = var_9683_cast_fp16, y = linear_204_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129057664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130106304))))[name = string("tm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9691_cast_fp16 = abs(x = linear_205_cast_fp16)[name = string("op_9691_cast_fp16")];
            bool var_9694 = const()[name = string("op_9694"), val = bool(true)];
            tensor<int32, [1]> var_9695_axes_0 = const()[name = string("op_9695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9695_cast_fp16 = reduce_max(axes = var_9695_axes_0, keep_dims = var_9694, x = var_9691_cast_fp16)[name = string("op_9695_cast_fp16")];
            fp16 var_9696_to_fp16 = const()[name = string("op_9696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_367_cast_fp16 = maximum(x = var_9695_cast_fp16, y = var_9696_to_fp16)[name = string("s_367_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_367_cast_fp16 = real_div(x = linear_205_cast_fp16, y = s_367_cast_fp16)[name = string("y_367_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9699_cast_fp16 = mul(x = y_367_cast_fp16, y = y_367_cast_fp16)[name = string("op_9699_cast_fp16")];
            tensor<int32, [1]> var_9704_axes_0 = const()[name = string("op_9704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9704_keep_dims_0 = const()[name = string("op_9704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9704_cast_fp16 = reduce_mean(axes = var_9704_axes_0, keep_dims = var_9704_keep_dims_0, x = var_9699_cast_fp16)[name = string("op_9704_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9705_cast_fp16 = mul(x = s_367_cast_fp16, y = s_367_cast_fp16)[name = string("op_9705_cast_fp16")];
            fp32 var_9706_epsilon_0 = const()[name = string("op_9706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9706_cast_fp16 = inverse(epsilon = var_9706_epsilon_0, x = var_9705_cast_fp16)[name = string("op_9706_cast_fp16")];
            fp16 var_9707_to_fp16 = const()[name = string("op_9707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9708_cast_fp16 = mul(x = var_9706_cast_fp16, y = var_9707_to_fp16)[name = string("op_9708_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9710_cast_fp16 = add(x = var_9704_cast_fp16, y = var_9708_cast_fp16)[name = string("op_9710_cast_fp16")];
            fp32 var_9711_epsilon_0 = const()[name = string("op_9711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9711_cast_fp16 = rsqrt(epsilon = var_9711_epsilon_0, x = var_9710_cast_fp16)[name = string("op_9711_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_321_cast_fp16 = mul(x = y_367_cast_fp16, y = var_9711_cast_fp16)[name = string("out_321_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130107392)))];
            tensor<fp16, [1, 64, 512]> var_9713_cast_fp16 = mul(x = out_321_cast_fp16, y = tm_layers_22_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9713_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_273_cast_fp16 = add(x = x_407_cast_fp16, y = var_9713_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130108480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130370688))))[name = string("tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_206_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_206_cast_fp16")];
            string x_411_mode_0 = const()[name = string("x_411_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_411_cast_fp16 = gelu(mode = x_411_mode_0, x = linear_206_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [4]> var_9733_begin_0 = const()[name = string("op_9733_begin_0"), val = tensor<int32, [4]>([0, 0, 22, 0])];
            tensor<int32, [4]> var_9733_end_0 = const()[name = string("op_9733_end_0"), val = tensor<int32, [4]>([1, 64, 23, 512])];
            tensor<bool, [4]> var_9733_end_mask_0 = const()[name = string("op_9733_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9733_squeeze_mask_0 = const()[name = string("op_9733_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_9733_cast_fp16 = slice_by_index(begin = var_9733_begin_0, end = var_9733_end_0, end_mask = var_9733_end_mask_0, squeeze_mask = var_9733_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_9733_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_275_cast_fp16 = mul(x = x_411_cast_fp16, y = var_9733_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130371776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130633984))))[name = string("tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9743_cast_fp16 = abs(x = linear_207_cast_fp16)[name = string("op_9743_cast_fp16")];
            bool var_9746 = const()[name = string("op_9746"), val = bool(true)];
            tensor<int32, [1]> var_9747_axes_0 = const()[name = string("op_9747_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9747_cast_fp16 = reduce_max(axes = var_9747_axes_0, keep_dims = var_9746, x = var_9743_cast_fp16)[name = string("op_9747_cast_fp16")];
            fp16 var_9748_to_fp16 = const()[name = string("op_9748_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_369_cast_fp16 = maximum(x = var_9747_cast_fp16, y = var_9748_to_fp16)[name = string("s_369_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_369_cast_fp16 = real_div(x = linear_207_cast_fp16, y = s_369_cast_fp16)[name = string("y_369_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9751_cast_fp16 = mul(x = y_369_cast_fp16, y = y_369_cast_fp16)[name = string("op_9751_cast_fp16")];
            tensor<int32, [1]> var_9756_axes_0 = const()[name = string("op_9756_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9756_keep_dims_0 = const()[name = string("op_9756_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9756_cast_fp16 = reduce_mean(axes = var_9756_axes_0, keep_dims = var_9756_keep_dims_0, x = var_9751_cast_fp16)[name = string("op_9756_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9757_cast_fp16 = mul(x = s_369_cast_fp16, y = s_369_cast_fp16)[name = string("op_9757_cast_fp16")];
            fp32 var_9758_epsilon_0 = const()[name = string("op_9758_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9758_cast_fp16 = inverse(epsilon = var_9758_epsilon_0, x = var_9757_cast_fp16)[name = string("op_9758_cast_fp16")];
            fp16 var_9759_to_fp16 = const()[name = string("op_9759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9760_cast_fp16 = mul(x = var_9758_cast_fp16, y = var_9759_to_fp16)[name = string("op_9760_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9762_cast_fp16 = add(x = var_9756_cast_fp16, y = var_9760_cast_fp16)[name = string("op_9762_cast_fp16")];
            fp32 var_9763_epsilon_0 = const()[name = string("op_9763_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9763_cast_fp16 = rsqrt(epsilon = var_9763_epsilon_0, x = var_9762_cast_fp16)[name = string("op_9763_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_323_cast_fp16 = mul(x = y_369_cast_fp16, y = var_9763_cast_fp16)[name = string("out_323_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130635072)))];
            tensor<fp16, [1, 64, 512]> var_9765_cast_fp16 = mul(x = out_323_cast_fp16, y = tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9765_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_45_cast_fp16 = add(x = input_273_cast_fp16, y = var_9765_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<fp16, [1]> tm_layers_22_layer_scalar_to_fp16 = const()[name = string("tm_layers_22_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 64, 512]> x_415_cast_fp16 = mul(x = h_45_cast_fp16, y = tm_layers_22_layer_scalar_to_fp16)[name = string("x_415_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9769_cast_fp16 = abs(x = x_415_cast_fp16)[name = string("op_9769_cast_fp16")];
            bool var_9772 = const()[name = string("op_9772"), val = bool(true)];
            tensor<int32, [1]> var_9773_axes_0 = const()[name = string("op_9773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9773_cast_fp16 = reduce_max(axes = var_9773_axes_0, keep_dims = var_9772, x = var_9769_cast_fp16)[name = string("op_9773_cast_fp16")];
            fp16 var_9774_to_fp16 = const()[name = string("op_9774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_371_cast_fp16 = maximum(x = var_9773_cast_fp16, y = var_9774_to_fp16)[name = string("s_371_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_371_cast_fp16 = real_div(x = x_415_cast_fp16, y = s_371_cast_fp16)[name = string("y_371_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9777_cast_fp16 = mul(x = y_371_cast_fp16, y = y_371_cast_fp16)[name = string("op_9777_cast_fp16")];
            tensor<int32, [1]> var_9782_axes_0 = const()[name = string("op_9782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9782_keep_dims_0 = const()[name = string("op_9782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9782_cast_fp16 = reduce_mean(axes = var_9782_axes_0, keep_dims = var_9782_keep_dims_0, x = var_9777_cast_fp16)[name = string("op_9782_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9783_cast_fp16 = mul(x = s_371_cast_fp16, y = s_371_cast_fp16)[name = string("op_9783_cast_fp16")];
            fp32 var_9784_epsilon_0 = const()[name = string("op_9784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9784_cast_fp16 = inverse(epsilon = var_9784_epsilon_0, x = var_9783_cast_fp16)[name = string("op_9784_cast_fp16")];
            fp16 var_9785_to_fp16 = const()[name = string("op_9785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9786_cast_fp16 = mul(x = var_9784_cast_fp16, y = var_9785_to_fp16)[name = string("op_9786_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9788_cast_fp16 = add(x = var_9782_cast_fp16, y = var_9786_cast_fp16)[name = string("op_9788_cast_fp16")];
            fp32 var_9789_epsilon_0 = const()[name = string("op_9789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9789_cast_fp16 = rsqrt(epsilon = var_9789_epsilon_0, x = var_9788_cast_fp16)[name = string("op_9789_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_325_cast_fp16 = mul(x = y_371_cast_fp16, y = var_9789_cast_fp16)[name = string("out_325_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130636160)))];
            tensor<fp16, [1, 64, 512]> input_277_cast_fp16 = mul(x = out_325_cast_fp16, y = tm_layers_23_input_layernorm_weight_to_fp16)[name = string("input_277_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130637248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131685888))))[name = string("tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_208_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<int32, [4]> var_9799 = const()[name = string("op_9799"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 64, 4, 512]> var_9800_cast_fp16 = reshape(shape = var_9799, x = linear_208_cast_fp16)[name = string("op_9800_cast_fp16")];
            tensor<int32, [4]> x_417_perm_0 = const()[name = string("x_417_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131690048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131952256))))[name = string("tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<int32, [4]> var_9811 = const()[name = string("op_9811"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 64, 1, 512]> var_9812_cast_fp16 = reshape(shape = var_9811, x = linear_209_cast_fp16)[name = string("op_9812_cast_fp16")];
            tensor<int32, [4]> x_419_perm_0 = const()[name = string("x_419_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131953344))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132215552))))[name = string("tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<int32, [4]> var_9823 = const()[name = string("op_9823"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 64, 1, 512]> var_9824_cast_fp16 = reshape(shape = var_9823, x = linear_210_cast_fp16)[name = string("op_9824_cast_fp16")];
            tensor<int32, [4]> x_421_perm_0 = const()[name = string("x_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 64, 512]> x_417_cast_fp16 = transpose(perm = x_417_perm_0, x = var_9800_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 4, 64, 512]> var_9828_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_9828_cast_fp16")];
            bool var_9831 = const()[name = string("op_9831"), val = bool(true)];
            tensor<int32, [1]> var_9832_axes_0 = const()[name = string("op_9832_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 64, 1]> var_9832_cast_fp16 = reduce_max(axes = var_9832_axes_0, keep_dims = var_9831, x = var_9828_cast_fp16)[name = string("op_9832_cast_fp16")];
            fp16 var_9833_to_fp16 = const()[name = string("op_9833_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 64, 1]> s_373_cast_fp16 = maximum(x = var_9832_cast_fp16, y = var_9833_to_fp16)[name = string("s_373_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> y_373_cast_fp16 = real_div(x = x_417_cast_fp16, y = s_373_cast_fp16)[name = string("y_373_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_9836_cast_fp16 = mul(x = y_373_cast_fp16, y = y_373_cast_fp16)[name = string("op_9836_cast_fp16")];
            tensor<int32, [1]> var_9841_axes_0 = const()[name = string("op_9841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9841_keep_dims_0 = const()[name = string("op_9841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 64, 1]> var_9841_cast_fp16 = reduce_mean(axes = var_9841_axes_0, keep_dims = var_9841_keep_dims_0, x = var_9836_cast_fp16)[name = string("op_9841_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_9842_cast_fp16 = mul(x = s_373_cast_fp16, y = s_373_cast_fp16)[name = string("op_9842_cast_fp16")];
            fp32 var_9843_epsilon_0 = const()[name = string("op_9843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 64, 1]> var_9843_cast_fp16 = inverse(epsilon = var_9843_epsilon_0, x = var_9842_cast_fp16)[name = string("op_9843_cast_fp16")];
            fp16 var_9844_to_fp16 = const()[name = string("op_9844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 64, 1]> var_9845_cast_fp16 = mul(x = var_9843_cast_fp16, y = var_9844_to_fp16)[name = string("op_9845_cast_fp16")];
            tensor<fp16, [1, 4, 64, 1]> var_9847_cast_fp16 = add(x = var_9841_cast_fp16, y = var_9845_cast_fp16)[name = string("op_9847_cast_fp16")];
            fp32 var_9848_epsilon_0 = const()[name = string("op_9848_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 64, 1]> var_9848_cast_fp16 = rsqrt(epsilon = var_9848_epsilon_0, x = var_9847_cast_fp16)[name = string("op_9848_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> out_327_cast_fp16 = mul(x = y_373_cast_fp16, y = var_9848_cast_fp16)[name = string("out_327_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132216640)))];
            tensor<fp16, [1, 4, 64, 512]> q_93_cast_fp16 = mul(x = out_327_cast_fp16, y = tm_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("q_93_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> x_419_cast_fp16 = transpose(perm = x_419_perm_0, x = var_9812_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1, 64, 512]> var_9851_cast_fp16 = abs(x = x_419_cast_fp16)[name = string("op_9851_cast_fp16")];
            bool var_9854 = const()[name = string("op_9854"), val = bool(true)];
            tensor<int32, [1]> var_9855_axes_0 = const()[name = string("op_9855_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 64, 1]> var_9855_cast_fp16 = reduce_max(axes = var_9855_axes_0, keep_dims = var_9854, x = var_9851_cast_fp16)[name = string("op_9855_cast_fp16")];
            fp16 var_9856_to_fp16 = const()[name = string("op_9856_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 64, 1]> s_375_cast_fp16 = maximum(x = var_9855_cast_fp16, y = var_9856_to_fp16)[name = string("s_375_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> y_375_cast_fp16 = real_div(x = x_419_cast_fp16, y = s_375_cast_fp16)[name = string("y_375_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_9859_cast_fp16 = mul(x = y_375_cast_fp16, y = y_375_cast_fp16)[name = string("op_9859_cast_fp16")];
            tensor<int32, [1]> var_9864_axes_0 = const()[name = string("op_9864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9864_keep_dims_0 = const()[name = string("op_9864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 64, 1]> var_9864_cast_fp16 = reduce_mean(axes = var_9864_axes_0, keep_dims = var_9864_keep_dims_0, x = var_9859_cast_fp16)[name = string("op_9864_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_9865_cast_fp16 = mul(x = s_375_cast_fp16, y = s_375_cast_fp16)[name = string("op_9865_cast_fp16")];
            fp32 var_9866_epsilon_0 = const()[name = string("op_9866_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 64, 1]> var_9866_cast_fp16 = inverse(epsilon = var_9866_epsilon_0, x = var_9865_cast_fp16)[name = string("op_9866_cast_fp16")];
            fp16 var_9867_to_fp16 = const()[name = string("op_9867_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 64, 1]> var_9868_cast_fp16 = mul(x = var_9866_cast_fp16, y = var_9867_to_fp16)[name = string("op_9868_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_9870_cast_fp16 = add(x = var_9864_cast_fp16, y = var_9868_cast_fp16)[name = string("op_9870_cast_fp16")];
            fp32 var_9871_epsilon_0 = const()[name = string("op_9871_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 64, 1]> var_9871_cast_fp16 = rsqrt(epsilon = var_9871_epsilon_0, x = var_9870_cast_fp16)[name = string("op_9871_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> out_329_cast_fp16 = mul(x = y_375_cast_fp16, y = var_9871_cast_fp16)[name = string("out_329_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132217728)))];
            tensor<fp16, [1, 1, 64, 512]> k_139_cast_fp16 = mul(x = out_329_cast_fp16, y = tm_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("k_139_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> x_421_cast_fp16 = transpose(perm = x_421_perm_0, x = var_9824_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 1, 64, 512]> var_9874_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_9874_cast_fp16")];
            bool var_9877 = const()[name = string("op_9877"), val = bool(true)];
            tensor<int32, [1]> var_9878_axes_0 = const()[name = string("op_9878_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 64, 1]> var_9878_cast_fp16 = reduce_max(axes = var_9878_axes_0, keep_dims = var_9877, x = var_9874_cast_fp16)[name = string("op_9878_cast_fp16")];
            fp16 var_9879_to_fp16 = const()[name = string("op_9879_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 64, 1]> s_377_cast_fp16 = maximum(x = var_9878_cast_fp16, y = var_9879_to_fp16)[name = string("s_377_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> y_377_cast_fp16 = real_div(x = x_421_cast_fp16, y = s_377_cast_fp16)[name = string("y_377_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_9882_cast_fp16 = mul(x = y_377_cast_fp16, y = y_377_cast_fp16)[name = string("op_9882_cast_fp16")];
            tensor<int32, [1]> var_9887_axes_0 = const()[name = string("op_9887_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9887_keep_dims_0 = const()[name = string("op_9887_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 64, 1]> var_9887_cast_fp16 = reduce_mean(axes = var_9887_axes_0, keep_dims = var_9887_keep_dims_0, x = var_9882_cast_fp16)[name = string("op_9887_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_9888_cast_fp16 = mul(x = s_377_cast_fp16, y = s_377_cast_fp16)[name = string("op_9888_cast_fp16")];
            fp32 var_9889_epsilon_0 = const()[name = string("op_9889_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 64, 1]> var_9889_cast_fp16 = inverse(epsilon = var_9889_epsilon_0, x = var_9888_cast_fp16)[name = string("op_9889_cast_fp16")];
            fp16 var_9890_to_fp16 = const()[name = string("op_9890_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 64, 1]> var_9891_cast_fp16 = mul(x = var_9889_cast_fp16, y = var_9890_to_fp16)[name = string("op_9891_cast_fp16")];
            tensor<fp16, [1, 1, 64, 1]> var_9893_cast_fp16 = add(x = var_9887_cast_fp16, y = var_9891_cast_fp16)[name = string("op_9893_cast_fp16")];
            fp32 var_9894_epsilon_0 = const()[name = string("op_9894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 64, 1]> var_9894_cast_fp16 = rsqrt(epsilon = var_9894_epsilon_0, x = var_9893_cast_fp16)[name = string("op_9894_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> v_93_cast_fp16 = mul(x = y_377_cast_fp16, y = var_9894_cast_fp16)[name = string("v_93_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_9896_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9896_cast_fp16")];
            tensor<int32, [4]> var_9901_begin_0 = const()[name = string("op_9901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9901_end_0 = const()[name = string("op_9901_end_0"), val = tensor<int32, [4]>([1, 4, 64, 512])];
            tensor<bool, [4]> var_9901_end_mask_0 = const()[name = string("op_9901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 64, 256]> var_9901_cast_fp16 = slice_by_index(begin = var_9901_begin_0, end = var_9901_end_0, end_mask = var_9901_end_mask_0, x = q_93_cast_fp16)[name = string("op_9901_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 64, 256]> var_9902_cast_fp16 = mul(x = var_9901_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_9902_cast_fp16")];
            tensor<int32, [4]> var_9907_begin_0 = const()[name = string("op_9907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9907_end_0 = const()[name = string("op_9907_end_0"), val = tensor<int32, [4]>([1, 4, 64, 256])];
            tensor<bool, [4]> var_9907_end_mask_0 = const()[name = string("op_9907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 64, 256]> var_9907_cast_fp16 = slice_by_index(begin = var_9907_begin_0, end = var_9907_end_0, end_mask = var_9907_end_mask_0, x = q_93_cast_fp16)[name = string("op_9907_cast_fp16")];
            int32 var_9909 = const()[name = string("op_9909"), val = int32(-1)];
            bool var_9910_interleave_0 = const()[name = string("op_9910_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 512]> var_9910_cast_fp16 = concat(axis = var_9909, interleave = var_9910_interleave_0, values = (var_9902_cast_fp16, var_9907_cast_fp16))[name = string("op_9910_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> var_9911_cast_fp16 = mul(x = var_9910_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9911_cast_fp16")];
            tensor<fp16, [1, 4, 64, 512]> q_cast_fp16 = add(x = var_9896_cast_fp16, y = var_9911_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_9914_cast_fp16 = mul(x = k_139_cast_fp16, y = cos_to_fp16_quantized)[name = string("op_9914_cast_fp16")];
            tensor<int32, [4]> var_9919_begin_0 = const()[name = string("op_9919_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9919_end_0 = const()[name = string("op_9919_end_0"), val = tensor<int32, [4]>([1, 1, 64, 512])];
            tensor<bool, [4]> var_9919_end_mask_0 = const()[name = string("op_9919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 64, 256]> var_9919_cast_fp16 = slice_by_index(begin = var_9919_begin_0, end = var_9919_end_0, end_mask = var_9919_end_mask_0, x = k_139_cast_fp16)[name = string("op_9919_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 64, 256]> var_9920_cast_fp16 = mul(x = var_9919_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_9920_cast_fp16")];
            tensor<int32, [4]> var_9925_begin_0 = const()[name = string("op_9925_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9925_end_0 = const()[name = string("op_9925_end_0"), val = tensor<int32, [4]>([1, 1, 64, 256])];
            tensor<bool, [4]> var_9925_end_mask_0 = const()[name = string("op_9925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 64, 256]> var_9925_cast_fp16 = slice_by_index(begin = var_9925_begin_0, end = var_9925_end_0, end_mask = var_9925_end_mask_0, x = k_139_cast_fp16)[name = string("op_9925_cast_fp16")];
            int32 var_9927 = const()[name = string("op_9927"), val = int32(-1)];
            bool var_9928_interleave_0 = const()[name = string("op_9928_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 64, 512]> var_9928_cast_fp16 = concat(axis = var_9927, interleave = var_9928_interleave_0, values = (var_9920_cast_fp16, var_9925_cast_fp16))[name = string("op_9928_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> var_9929_cast_fp16 = mul(x = var_9928_cast_fp16, y = sin_to_fp16_quantized)[name = string("op_9929_cast_fp16")];
            tensor<fp16, [1, 1, 64, 512]> k_141_cast_fp16 = add(x = var_9914_cast_fp16, y = var_9929_cast_fp16)[name = string("k_141_cast_fp16")];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_46_reps_0 = const()[name = string("tile_46_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 64, 512]> transpose_92_cast_fp16 = transpose(perm = transpose_92_perm_0, x = k_141_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 1, 64, 512]> tile_46_cast_fp16 = tile(reps = tile_46_reps_0, x = transpose_92_cast_fp16)[name = string("tile_46_cast_fp16")];
            tensor<int32, [5]> concat_92 = const()[name = string("concat_92"), val = tensor<int32, [5]>([4, 1, 1, 64, 512])];
            tensor<fp16, [4, 1, 1, 64, 512]> reshape_92_cast_fp16 = reshape(shape = concat_92, x = tile_46_cast_fp16)[name = string("reshape_92_cast_fp16")];
            tensor<int32, [5]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_93 = const()[name = string("concat_93"), val = tensor<int32, [4]>([-1, 1, 64, 512])];
            tensor<fp16, [1, 4, 1, 64, 512]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = reshape_92_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 1, 64, 512]> reshape_93_cast_fp16 = reshape(shape = concat_93, x = transpose_93_cast_fp16)[name = string("reshape_93_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_47_reps_0 = const()[name = string("tile_47_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 64, 512]> transpose_94_cast_fp16 = transpose(perm = transpose_94_perm_0, x = v_93_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 1, 64, 512]> tile_47_cast_fp16 = tile(reps = tile_47_reps_0, x = transpose_94_cast_fp16)[name = string("tile_47_cast_fp16")];
            tensor<int32, [5]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [5]>([4, 1, 1, 64, 512])];
            tensor<fp16, [4, 1, 1, 64, 512]> reshape_94_cast_fp16 = reshape(shape = concat_94, x = tile_47_cast_fp16)[name = string("reshape_94_cast_fp16")];
            tensor<int32, [5]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [4]>([-1, 1, 64, 512])];
            tensor<fp16, [1, 4, 1, 64, 512]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = reshape_94_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 1, 64, 512]> reshape_95_cast_fp16 = reshape(shape = concat_95, x = transpose_95_cast_fp16)[name = string("reshape_95_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9943_transpose_x_0 = const()[name = string("op_9943_transpose_x_0"), val = bool(false)];
            bool var_9943_transpose_y_0 = const()[name = string("op_9943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 64]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = reshape_93_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 4, 64, 64]> var_9943_cast_fp16 = matmul(transpose_x = var_9943_transpose_x_0, transpose_y = var_9943_transpose_y_0, x = q_cast_fp16, y = transpose_119_cast_fp16)[name = string("op_9943_cast_fp16")];
            tensor<fp16, [1, 4, 64, 64]> var_9945_cast_fp16 = add(x = var_9943_cast_fp16, y = var_2173_cast_fp16)[name = string("op_9945_cast_fp16")];
            int32 var_9946 = const()[name = string("op_9946"), val = int32(-1)];
            tensor<fp16, [1, 4, 64, 64]> w_377_cast_fp16 = softmax(axis = var_9946, x = var_9945_cast_fp16)[name = string("w_377_cast_fp16")];
            bool var_9949_transpose_x_0 = const()[name = string("op_9949_transpose_x_0"), val = bool(false)];
            bool var_9949_transpose_y_0 = const()[name = string("op_9949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 64, 512]> v_cast_fp16 = transpose(perm = v_perm_0, x = reshape_95_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 4, 64, 512]> var_9949_cast_fp16 = matmul(transpose_x = var_9949_transpose_x_0, transpose_y = var_9949_transpose_y_0, x = w_377_cast_fp16, y = v_cast_fp16)[name = string("op_9949_cast_fp16")];
            tensor<int32, [4]> var_9952_perm_0 = const()[name = string("op_9952_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9956 = const()[name = string("op_9956"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 64, 4, 512]> var_9952_cast_fp16 = transpose(perm = var_9952_perm_0, x = var_9949_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 64, 2048]> input_279_cast_fp16 = reshape(shape = var_9956, x = var_9952_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132218816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133267456))))[name = string("tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9961_cast_fp16 = abs(x = linear_211_cast_fp16)[name = string("op_9961_cast_fp16")];
            bool var_9964 = const()[name = string("op_9964"), val = bool(true)];
            tensor<int32, [1]> var_9965_axes_0 = const()[name = string("op_9965_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9965_cast_fp16 = reduce_max(axes = var_9965_axes_0, keep_dims = var_9964, x = var_9961_cast_fp16)[name = string("op_9965_cast_fp16")];
            fp16 var_9966_to_fp16 = const()[name = string("op_9966_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_379_cast_fp16 = maximum(x = var_9965_cast_fp16, y = var_9966_to_fp16)[name = string("s_379_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_379_cast_fp16 = real_div(x = linear_211_cast_fp16, y = s_379_cast_fp16)[name = string("y_379_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9969_cast_fp16 = mul(x = y_379_cast_fp16, y = y_379_cast_fp16)[name = string("op_9969_cast_fp16")];
            tensor<int32, [1]> var_9974_axes_0 = const()[name = string("op_9974_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9974_keep_dims_0 = const()[name = string("op_9974_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9974_cast_fp16 = reduce_mean(axes = var_9974_axes_0, keep_dims = var_9974_keep_dims_0, x = var_9969_cast_fp16)[name = string("op_9974_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9975_cast_fp16 = mul(x = s_379_cast_fp16, y = s_379_cast_fp16)[name = string("op_9975_cast_fp16")];
            fp32 var_9976_epsilon_0 = const()[name = string("op_9976_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_9976_cast_fp16 = inverse(epsilon = var_9976_epsilon_0, x = var_9975_cast_fp16)[name = string("op_9976_cast_fp16")];
            fp16 var_9977_to_fp16 = const()[name = string("op_9977_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_9978_cast_fp16 = mul(x = var_9976_cast_fp16, y = var_9977_to_fp16)[name = string("op_9978_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_9980_cast_fp16 = add(x = var_9974_cast_fp16, y = var_9978_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp32 var_9981_epsilon_0 = const()[name = string("op_9981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_9981_cast_fp16 = rsqrt(epsilon = var_9981_epsilon_0, x = var_9980_cast_fp16)[name = string("op_9981_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_331_cast_fp16 = mul(x = y_379_cast_fp16, y = var_9981_cast_fp16)[name = string("out_331_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133268544)))];
            tensor<fp16, [1, 64, 512]> var_9983_cast_fp16 = mul(x = out_331_cast_fp16, y = tm_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("op_9983_cast_fp16")];
            tensor<fp16, [1, 64, 512]> x_425_cast_fp16 = add(x = x_415_cast_fp16, y = var_9983_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9986_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_9986_cast_fp16")];
            bool var_9989 = const()[name = string("op_9989"), val = bool(true)];
            tensor<int32, [1]> var_9990_axes_0 = const()[name = string("op_9990_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_9990_cast_fp16 = reduce_max(axes = var_9990_axes_0, keep_dims = var_9989, x = var_9986_cast_fp16)[name = string("op_9990_cast_fp16")];
            fp16 var_9991_to_fp16 = const()[name = string("op_9991_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_381_cast_fp16 = maximum(x = var_9990_cast_fp16, y = var_9991_to_fp16)[name = string("s_381_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_381_cast_fp16 = real_div(x = x_425_cast_fp16, y = s_381_cast_fp16)[name = string("y_381_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_9994_cast_fp16 = mul(x = y_381_cast_fp16, y = y_381_cast_fp16)[name = string("op_9994_cast_fp16")];
            tensor<int32, [1]> var_9999_axes_0 = const()[name = string("op_9999_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9999_keep_dims_0 = const()[name = string("op_9999_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9999_cast_fp16 = reduce_mean(axes = var_9999_axes_0, keep_dims = var_9999_keep_dims_0, x = var_9994_cast_fp16)[name = string("op_9999_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_10000_cast_fp16 = mul(x = s_381_cast_fp16, y = s_381_cast_fp16)[name = string("op_10000_cast_fp16")];
            fp32 var_10001_epsilon_0 = const()[name = string("op_10001_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_10001_cast_fp16 = inverse(epsilon = var_10001_epsilon_0, x = var_10000_cast_fp16)[name = string("op_10001_cast_fp16")];
            fp16 var_10002_to_fp16 = const()[name = string("op_10002_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_10003_cast_fp16 = mul(x = var_10001_cast_fp16, y = var_10002_to_fp16)[name = string("op_10003_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_10005_cast_fp16 = add(x = var_9999_cast_fp16, y = var_10003_cast_fp16)[name = string("op_10005_cast_fp16")];
            fp32 var_10006_epsilon_0 = const()[name = string("op_10006_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_10006_cast_fp16 = rsqrt(epsilon = var_10006_epsilon_0, x = var_10005_cast_fp16)[name = string("op_10006_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_333_cast_fp16 = mul(x = y_381_cast_fp16, y = var_10006_cast_fp16)[name = string("out_333_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133269632)))];
            tensor<fp16, [1, 64, 512]> input_281_cast_fp16 = mul(x = out_333_cast_fp16, y = tm_layers_23_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133270720))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134319360))))[name = string("tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_212_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_212_cast_fp16")];
            string var_10013_mode_0 = const()[name = string("op_10013_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 2048]> var_10013_cast_fp16 = gelu(mode = var_10013_mode_0, x = linear_212_cast_fp16)[name = string("op_10013_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134323520))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135372160))))[name = string("tm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 2048]> linear_213_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_213_cast_fp16")];
            tensor<fp16, [1, 64, 2048]> input_283_cast_fp16 = mul(x = var_10013_cast_fp16, y = linear_213_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135376320))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136424960))))[name = string("tm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_283_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_10021_cast_fp16 = abs(x = linear_214_cast_fp16)[name = string("op_10021_cast_fp16")];
            bool var_10024 = const()[name = string("op_10024"), val = bool(true)];
            tensor<int32, [1]> var_10025_axes_0 = const()[name = string("op_10025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_10025_cast_fp16 = reduce_max(axes = var_10025_axes_0, keep_dims = var_10024, x = var_10021_cast_fp16)[name = string("op_10025_cast_fp16")];
            fp16 var_10026_to_fp16 = const()[name = string("op_10026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_383_cast_fp16 = maximum(x = var_10025_cast_fp16, y = var_10026_to_fp16)[name = string("s_383_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_383_cast_fp16 = real_div(x = linear_214_cast_fp16, y = s_383_cast_fp16)[name = string("y_383_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_10029_cast_fp16 = mul(x = y_383_cast_fp16, y = y_383_cast_fp16)[name = string("op_10029_cast_fp16")];
            tensor<int32, [1]> var_10034_axes_0 = const()[name = string("op_10034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10034_keep_dims_0 = const()[name = string("op_10034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_10034_cast_fp16 = reduce_mean(axes = var_10034_axes_0, keep_dims = var_10034_keep_dims_0, x = var_10029_cast_fp16)[name = string("op_10034_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_10035_cast_fp16 = mul(x = s_383_cast_fp16, y = s_383_cast_fp16)[name = string("op_10035_cast_fp16")];
            fp32 var_10036_epsilon_0 = const()[name = string("op_10036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_10036_cast_fp16 = inverse(epsilon = var_10036_epsilon_0, x = var_10035_cast_fp16)[name = string("op_10036_cast_fp16")];
            fp16 var_10037_to_fp16 = const()[name = string("op_10037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_10038_cast_fp16 = mul(x = var_10036_cast_fp16, y = var_10037_to_fp16)[name = string("op_10038_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_10040_cast_fp16 = add(x = var_10034_cast_fp16, y = var_10038_cast_fp16)[name = string("op_10040_cast_fp16")];
            fp32 var_10041_epsilon_0 = const()[name = string("op_10041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_10041_cast_fp16 = rsqrt(epsilon = var_10041_epsilon_0, x = var_10040_cast_fp16)[name = string("op_10041_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_335_cast_fp16 = mul(x = y_383_cast_fp16, y = var_10041_cast_fp16)[name = string("out_335_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136426048)))];
            tensor<fp16, [1, 64, 512]> var_10043_cast_fp16 = mul(x = out_335_cast_fp16, y = tm_layers_23_post_feedforward_layernorm_weight_to_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_285_cast_fp16 = add(x = x_425_cast_fp16, y = var_10043_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136427136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136689344))))[name = string("tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_215_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = string("linear_215_cast_fp16")];
            string x_429_mode_0 = const()[name = string("x_429_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 64, 512]> x_429_cast_fp16 = gelu(mode = x_429_mode_0, x = linear_215_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [4]> var_10063_begin_0 = const()[name = string("op_10063_begin_0"), val = tensor<int32, [4]>([0, 0, 23, 0])];
            tensor<int32, [4]> var_10063_end_0 = const()[name = string("op_10063_end_0"), val = tensor<int32, [4]>([1, 64, 24, 512])];
            tensor<bool, [4]> var_10063_end_mask_0 = const()[name = string("op_10063_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_10063_squeeze_mask_0 = const()[name = string("op_10063_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 64, 512]> var_10063_cast_fp16 = slice_by_index(begin = var_10063_begin_0, end = var_10063_end_0, end_mask = var_10063_end_mask_0, squeeze_mask = var_10063_squeeze_mask_0, x = ple_3_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 64, 512]> input_287_cast_fp16 = mul(x = x_429_cast_fp16, y = var_10063_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136690432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136952640))))[name = string("tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 512]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_10073_cast_fp16 = abs(x = linear_216_cast_fp16)[name = string("op_10073_cast_fp16")];
            bool var_10076 = const()[name = string("op_10076"), val = bool(true)];
            tensor<int32, [1]> var_10077_axes_0 = const()[name = string("op_10077_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_10077_cast_fp16 = reduce_max(axes = var_10077_axes_0, keep_dims = var_10076, x = var_10073_cast_fp16)[name = string("op_10077_cast_fp16")];
            fp16 var_10078_to_fp16 = const()[name = string("op_10078_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_385_cast_fp16 = maximum(x = var_10077_cast_fp16, y = var_10078_to_fp16)[name = string("s_385_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_385_cast_fp16 = real_div(x = linear_216_cast_fp16, y = s_385_cast_fp16)[name = string("y_385_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_10081_cast_fp16 = mul(x = y_385_cast_fp16, y = y_385_cast_fp16)[name = string("op_10081_cast_fp16")];
            tensor<int32, [1]> var_10086_axes_0 = const()[name = string("op_10086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10086_keep_dims_0 = const()[name = string("op_10086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_10086_cast_fp16 = reduce_mean(axes = var_10086_axes_0, keep_dims = var_10086_keep_dims_0, x = var_10081_cast_fp16)[name = string("op_10086_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_10087_cast_fp16 = mul(x = s_385_cast_fp16, y = s_385_cast_fp16)[name = string("op_10087_cast_fp16")];
            fp32 var_10088_epsilon_0 = const()[name = string("op_10088_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_10088_cast_fp16 = inverse(epsilon = var_10088_epsilon_0, x = var_10087_cast_fp16)[name = string("op_10088_cast_fp16")];
            fp16 var_10089_to_fp16 = const()[name = string("op_10089_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_10090_cast_fp16 = mul(x = var_10088_cast_fp16, y = var_10089_to_fp16)[name = string("op_10090_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_10092_cast_fp16 = add(x = var_10086_cast_fp16, y = var_10090_cast_fp16)[name = string("op_10092_cast_fp16")];
            fp32 var_10093_epsilon_0 = const()[name = string("op_10093_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_10093_cast_fp16 = rsqrt(epsilon = var_10093_epsilon_0, x = var_10092_cast_fp16)[name = string("op_10093_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_337_cast_fp16 = mul(x = y_385_cast_fp16, y = var_10093_cast_fp16)[name = string("out_337_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136953728)))];
            tensor<fp16, [1, 64, 512]> var_10095_cast_fp16 = mul(x = out_337_cast_fp16, y = tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_10095_cast_fp16")];
            tensor<fp16, [1, 64, 512]> h_47_cast_fp16 = add(x = input_285_cast_fp16, y = var_10095_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<fp16, [1]> tm_layers_23_layer_scalar_to_fp16 = const()[name = string("tm_layers_23_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d6p-4])];
            tensor<fp16, [1, 64, 512]> x_cast_fp16 = mul(x = h_47_cast_fp16, y = tm_layers_23_layer_scalar_to_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_10099_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_10099_cast_fp16")];
            bool var_10102 = const()[name = string("op_10102"), val = bool(true)];
            tensor<int32, [1]> var_10103_axes_0 = const()[name = string("op_10103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_10103_cast_fp16 = reduce_max(axes = var_10103_axes_0, keep_dims = var_10102, x = var_10099_cast_fp16)[name = string("op_10103_cast_fp16")];
            fp16 var_10104_to_fp16 = const()[name = string("op_10104_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 64, 1]> s_387_cast_fp16 = maximum(x = var_10103_cast_fp16, y = var_10104_to_fp16)[name = string("s_387_cast_fp16")];
            tensor<fp16, [1, 64, 512]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_387_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 64, 512]> var_10107_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_10107_cast_fp16")];
            tensor<int32, [1]> var_10112_axes_0 = const()[name = string("op_10112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10112_keep_dims_0 = const()[name = string("op_10112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_10112_cast_fp16 = reduce_mean(axes = var_10112_axes_0, keep_dims = var_10112_keep_dims_0, x = var_10107_cast_fp16)[name = string("op_10112_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_10113_cast_fp16 = mul(x = s_387_cast_fp16, y = s_387_cast_fp16)[name = string("op_10113_cast_fp16")];
            fp32 var_10114_epsilon_0 = const()[name = string("op_10114_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_10114_cast_fp16 = inverse(epsilon = var_10114_epsilon_0, x = var_10113_cast_fp16)[name = string("op_10114_cast_fp16")];
            fp16 var_10115_to_fp16 = const()[name = string("op_10115_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_10116_cast_fp16 = mul(x = var_10114_cast_fp16, y = var_10115_to_fp16)[name = string("op_10116_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_10118_cast_fp16 = add(x = var_10112_cast_fp16, y = var_10116_cast_fp16)[name = string("op_10118_cast_fp16")];
            fp32 var_10119_epsilon_0 = const()[name = string("op_10119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_10119_cast_fp16 = rsqrt(epsilon = var_10119_epsilon_0, x = var_10118_cast_fp16)[name = string("op_10119_cast_fp16")];
            tensor<fp16, [1, 64, 512]> out_cast_fp16 = mul(x = y_cast_fp16, y = var_10119_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [512]> tm_norm_weight_to_fp16 = const()[name = string("tm_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136954816)))];
            tensor<fp16, [1, 64, 512]> h_cast_fp16 = mul(x = out_cast_fp16, y = tm_norm_weight_to_fp16)[name = string("h_cast_fp16")];
            tensor<int32, [1]> var_10126_axes_0 = const()[name = string("op_10126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10126_keep_dims_0 = const()[name = string("op_10126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10126_cast_fp16 = reduce_sum(axes = var_10126_axes_0, keep_dims = var_10126_keep_dims_0, x = attention_mask)[name = string("op_10126_cast_fp16")];
            tensor<fp16, [1, 64]> var_10127_cast_fp16 = real_div(x = attention_mask, y = var_10126_cast_fp16)[name = string("op_10127_cast_fp16")];
            tensor<int32, [1]> var_10134_axes_0 = const()[name = string("op_10134_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 64]> var_10134_cast_fp16 = expand_dims(axes = var_10134_axes_0, x = var_10127_cast_fp16)[name = string("op_10134_cast_fp16")];
            bool input_transpose_x_0 = const()[name = string("input_transpose_x_0"), val = bool(false)];
            bool input_transpose_y_0 = const()[name = string("input_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1, 512]> input_cast_fp16 = matmul(transpose_x = input_transpose_x_0, transpose_y = input_transpose_y_0, x = var_10134_cast_fp16, y = h_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [768, 512]> tm_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136955904))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137349184))))[name = string("tm_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_217_bias_0_to_fp16 = const()[name = string("linear_217_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137350784)))];
            tensor<fp16, [1, 1, 768]> linear_217_cast_fp16 = linear(bias = linear_217_bias_0_to_fp16, weight = tm_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<int32, [3]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [3]>([1, 1, 768])];
            tensor<bool, [3]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_10151_squeeze_mask_0 = const()[name = string("op_10151_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [1, 768]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, squeeze_mask = var_10151_squeeze_mask_0, x = linear_217_cast_fp16)[name = string("op_10151_cast_fp16")];
            tensor<fp16, [1, 768]> var_10157_cast_fp16 = abs(x = var_10151_cast_fp16)[name = string("op_10157_cast_fp16")];
            bool var_10160 = const()[name = string("op_10160"), val = bool(true)];
            tensor<int32, [1]> s_axes_0 = const()[name = string("s_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1]> s_cast_fp16 = reduce_max(axes = s_axes_0, keep_dims = var_10160, x = var_10157_cast_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [1, 768]> pooled_cast_fp16 = real_div(x = var_10151_cast_fp16, y = s_cast_fp16)[name = string("pooled_cast_fp16")];
            tensor<fp16, [1, 768]> var_10163_cast_fp16 = mul(x = pooled_cast_fp16, y = pooled_cast_fp16)[name = string("op_10163_cast_fp16")];
            tensor<int32, [1]> var_10168_axes_0 = const()[name = string("op_10168_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10168_keep_dims_0 = const()[name = string("op_10168_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1]> var_10168_cast_fp16 = reduce_sum(axes = var_10168_axes_0, keep_dims = var_10168_keep_dims_0, x = var_10163_cast_fp16)[name = string("op_10168_cast_fp16")];
            fp32 var_10169_epsilon_0 = const()[name = string("op_10169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1]> var_10169_cast_fp16 = rsqrt(epsilon = var_10169_epsilon_0, x = var_10168_cast_fp16)[name = string("op_10169_cast_fp16")];
            tensor<fp16, [1, 768]> var_10170_cast_fp16 = mul(x = pooled_cast_fp16, y = var_10169_cast_fp16)[name = string("op_10170_cast_fp16")];
            string var_10170_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_10170_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 768]> embedding = cast(dtype = var_10170_cast_fp16_to_fp32_dtype_0, x = var_10170_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func pack_256<ios18>(tensor<fp16, [1, 1, 256, 256]> attention_bias, tensor<fp16, [1, 256, 512]> inputs_embeds, tensor<fp16, [8, 256]> pool, tensor<fp16, [256, 1]> positions) {
            tensor<fp16, [1, 512]> inv_full_attention_to_fp16 = const()[name = string("inv_full_attention_to_fp16"), val = tensor<fp16, [1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [256, 512]> var_2107_cast_fp16 = mul(x = positions, y = inv_full_attention_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [256, 512]> cos_cast_fp16 = cos(x = var_2107_cast_fp16)[name = string("cos_cast_fp16")];
            tensor<fp16, [256, 512]> sin_cast_fp16 = sin(x = var_2107_cast_fp16)[name = string("sin_cast_fp16")];
            tensor<fp16, [1, 256]> inv_sliding_attention_to_fp16 = const()[name = string("inv_sliding_attention_to_fp16"), val = tensor<fp16, [1, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1152)))];
            tensor<fp16, [256, 256]> var_2111_cast_fp16 = mul(x = positions, y = inv_sliding_attention_to_fp16)[name = string("op_2111_cast_fp16")];
            tensor<fp16, [256, 256]> cos_1_cast_fp16 = cos(x = var_2111_cast_fp16)[name = string("cos_1_cast_fp16")];
            tensor<fp16, [256, 256]> sin_1_cast_fp16 = sin(x = var_2111_cast_fp16)[name = string("sin_1_cast_fp16")];
            tensor<fp16, [12288, 512]> tm_ple_per_layer_model_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [12288, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1728))), scale = tensor<fp16, [12288, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6293248))))[name = string("tm_ple_per_layer_model_projection_weight_to_fp16_quantized")];
            tensor<fp16, [12288]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [12288]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6317888)))];
            tensor<fp16, [1, 256, 12288]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = tm_ple_per_layer_model_projection_weight_to_fp16_quantized, x = inputs_embeds)[name = string("linear_0_cast_fp16")];
            fp16 var_2118_to_fp16 = const()[name = string("op_2118_to_fp16"), val = fp16(0x1.6ap-5)];
            tensor<fp16, [1, 256, 12288]> var_2119_cast_fp16 = mul(x = linear_0_cast_fp16, y = var_2118_to_fp16)[name = string("op_2119_cast_fp16")];
            tensor<int32, [4]> var_2124 = const()[name = string("op_2124"), val = tensor<int32, [4]>([1, -1, 24, 512])];
            tensor<fp16, [1, 256, 24, 512]> x_1_cast_fp16 = reshape(shape = var_2124, x = var_2119_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 256, 24, 512]> var_2126_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_2126_cast_fp16")];
            bool var_2129 = const()[name = string("op_2129"), val = bool(true)];
            tensor<int32, [1]> var_2130_axes_0 = const()[name = string("op_2130_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 24, 1]> var_2130_cast_fp16 = reduce_max(axes = var_2130_axes_0, keep_dims = var_2129, x = var_2126_cast_fp16)[name = string("op_2130_cast_fp16")];
            fp16 var_2131_to_fp16 = const()[name = string("op_2131_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 24, 1]> s_1_cast_fp16 = maximum(x = var_2130_cast_fp16, y = var_2131_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 256, 24, 512]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 256, 24, 512]> var_2134_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [1]> var_2139_axes_0 = const()[name = string("op_2139_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2139_keep_dims_0 = const()[name = string("op_2139_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 24, 1]> var_2139_cast_fp16 = reduce_mean(axes = var_2139_axes_0, keep_dims = var_2139_keep_dims_0, x = var_2134_cast_fp16)[name = string("op_2139_cast_fp16")];
            tensor<fp16, [1, 256, 24, 1]> var_2140_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_2140_cast_fp16")];
            fp32 var_2141_epsilon_0 = const()[name = string("op_2141_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 24, 1]> var_2141_cast_fp16 = inverse(epsilon = var_2141_epsilon_0, x = var_2140_cast_fp16)[name = string("op_2141_cast_fp16")];
            fp16 var_2142_to_fp16 = const()[name = string("op_2142_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 24, 1]> var_2143_cast_fp16 = mul(x = var_2141_cast_fp16, y = var_2142_to_fp16)[name = string("op_2143_cast_fp16")];
            tensor<fp16, [1, 256, 24, 1]> var_2145_cast_fp16 = add(x = var_2139_cast_fp16, y = var_2143_cast_fp16)[name = string("op_2145_cast_fp16")];
            fp32 var_2146_epsilon_0 = const()[name = string("op_2146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 24, 1]> var_2146_cast_fp16 = rsqrt(epsilon = var_2146_epsilon_0, x = var_2145_cast_fp16)[name = string("op_2146_cast_fp16")];
            tensor<fp16, [1, 256, 24, 512]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_2146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [512]> tm_ple_per_layer_projection_norm_weight_to_fp16 = const()[name = string("tm_ple_per_layer_projection_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6342528)))];
            tensor<fp16, [1, 256, 24, 512]> ple_1_cast_fp16 = mul(x = out_1_cast_fp16, y = tm_ple_per_layer_projection_norm_weight_to_fp16)[name = string("ple_1_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2149_cast_fp16 = abs(x = inputs_embeds)[name = string("op_2149_cast_fp16")];
            bool var_2152 = const()[name = string("op_2152"), val = bool(true)];
            tensor<int32, [1]> var_2153_axes_0 = const()[name = string("op_2153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2153_cast_fp16 = reduce_max(axes = var_2153_axes_0, keep_dims = var_2152, x = var_2149_cast_fp16)[name = string("op_2153_cast_fp16")];
            fp16 var_2154_to_fp16 = const()[name = string("op_2154_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_3_cast_fp16 = maximum(x = var_2153_cast_fp16, y = var_2154_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_3_cast_fp16 = real_div(x = inputs_embeds, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2157_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<int32, [1]> var_2162_axes_0 = const()[name = string("op_2162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2162_keep_dims_0 = const()[name = string("op_2162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2162_cast_fp16 = reduce_mean(axes = var_2162_axes_0, keep_dims = var_2162_keep_dims_0, x = var_2157_cast_fp16)[name = string("op_2162_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2163_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_2163_cast_fp16")];
            fp32 var_2164_epsilon_0 = const()[name = string("op_2164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2164_cast_fp16 = inverse(epsilon = var_2164_epsilon_0, x = var_2163_cast_fp16)[name = string("op_2164_cast_fp16")];
            fp16 var_2165_to_fp16 = const()[name = string("op_2165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2166_cast_fp16 = mul(x = var_2164_cast_fp16, y = var_2165_to_fp16)[name = string("op_2166_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2168_cast_fp16 = add(x = var_2162_cast_fp16, y = var_2166_cast_fp16)[name = string("op_2168_cast_fp16")];
            fp32 var_2169_epsilon_0 = const()[name = string("op_2169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2169_cast_fp16 = rsqrt(epsilon = var_2169_epsilon_0, x = var_2168_cast_fp16)[name = string("op_2169_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_2169_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6343616)))];
            tensor<fp16, [1, 256, 512]> input_1_cast_fp16 = mul(x = out_3_cast_fp16, y = tm_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6344704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6869056))))[name = string("tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6871168)))];
            tensor<fp16, [1, 256, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_2179 = const()[name = string("op_2179"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_2180_cast_fp16 = reshape(shape = var_2179, x = linear_1_cast_fp16)[name = string("op_2180_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6873280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7135488))))[name = string("tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_2_bias_0_to_fp16 = const()[name = string("linear_2_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7136576)))];
            tensor<fp16, [1, 256, 512]> linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_2191 = const()[name = string("op_2191"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2192_cast_fp16 = reshape(shape = var_2191, x = linear_2_cast_fp16)[name = string("op_2192_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7137664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7399872))))[name = string("tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_3_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_2203 = const()[name = string("op_2203"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2204_cast_fp16 = reshape(shape = var_2203, x = linear_3_cast_fp16)[name = string("op_2204_cast_fp16")];
            tensor<int32, [4]> x_7_perm_0 = const()[name = string("x_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_2180_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 4, 256, 256]> var_2208_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_2208_cast_fp16")];
            bool var_2211 = const()[name = string("op_2211"), val = bool(true)];
            tensor<int32, [1]> var_2212_axes_0 = const()[name = string("op_2212_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_2212_cast_fp16 = reduce_max(axes = var_2212_axes_0, keep_dims = var_2211, x = var_2208_cast_fp16)[name = string("op_2212_cast_fp16")];
            fp16 var_2213_to_fp16 = const()[name = string("op_2213_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_5_cast_fp16 = maximum(x = var_2212_cast_fp16, y = var_2213_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2216_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<int32, [1]> var_2221_axes_0 = const()[name = string("op_2221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2221_keep_dims_0 = const()[name = string("op_2221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_2221_cast_fp16 = reduce_mean(axes = var_2221_axes_0, keep_dims = var_2221_keep_dims_0, x = var_2216_cast_fp16)[name = string("op_2221_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2222_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2227_cast_fp16 = add(x = var_2221_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2227_cast_fp16")];
            fp32 var_2228_epsilon_0 = const()[name = string("op_2228_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_2228_cast_fp16 = rsqrt(epsilon = var_2228_epsilon_0, x = var_2227_cast_fp16)[name = string("op_2228_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_2228_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7400960)))];
            tensor<fp16, [1, 4, 256, 256]> q_1_cast_fp16 = mul(x = out_5_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_2192_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 2, 256, 256]> var_2231_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_2231_cast_fp16")];
            bool var_2234 = const()[name = string("op_2234"), val = bool(true)];
            tensor<int32, [1]> var_2235_axes_0 = const()[name = string("op_2235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2235_cast_fp16 = reduce_max(axes = var_2235_axes_0, keep_dims = var_2234, x = var_2231_cast_fp16)[name = string("op_2235_cast_fp16")];
            fp16 var_2236_to_fp16 = const()[name = string("op_2236_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_7_cast_fp16 = maximum(x = var_2235_cast_fp16, y = var_2236_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2239_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_2239_cast_fp16")];
            tensor<int32, [1]> var_2244_axes_0 = const()[name = string("op_2244_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2244_keep_dims_0 = const()[name = string("op_2244_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2244_cast_fp16 = reduce_mean(axes = var_2244_axes_0, keep_dims = var_2244_keep_dims_0, x = var_2239_cast_fp16)[name = string("op_2244_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2245_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_2245_cast_fp16")];
            fp32 var_2246_epsilon_0 = const()[name = string("op_2246_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2246_cast_fp16 = inverse(epsilon = var_2246_epsilon_0, x = var_2245_cast_fp16)[name = string("op_2246_cast_fp16")];
            fp16 var_2247_to_fp16 = const()[name = string("op_2247_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2248_cast_fp16 = mul(x = var_2246_cast_fp16, y = var_2247_to_fp16)[name = string("op_2248_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2250_cast_fp16 = add(x = var_2244_cast_fp16, y = var_2248_cast_fp16)[name = string("op_2250_cast_fp16")];
            fp32 var_2251_epsilon_0 = const()[name = string("op_2251_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2251_cast_fp16 = rsqrt(epsilon = var_2251_epsilon_0, x = var_2250_cast_fp16)[name = string("op_2251_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_7_cast_fp16 = mul(x = y_7_cast_fp16, y = var_2251_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [256]> tm_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7401536)))];
            tensor<fp16, [1, 2, 256, 256]> k_1_cast_fp16 = mul(x = out_7_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_7_cast_fp16 = transpose(perm = x_7_perm_0, x = var_2204_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 2, 256, 256]> var_2254_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_2254_cast_fp16")];
            bool var_2257 = const()[name = string("op_2257"), val = bool(true)];
            tensor<int32, [1]> var_2258_axes_0 = const()[name = string("op_2258_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2258_cast_fp16 = reduce_max(axes = var_2258_axes_0, keep_dims = var_2257, x = var_2254_cast_fp16)[name = string("op_2258_cast_fp16")];
            fp16 var_2259_to_fp16 = const()[name = string("op_2259_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_9_cast_fp16 = maximum(x = var_2258_cast_fp16, y = var_2259_to_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_9_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2262_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_2262_cast_fp16")];
            tensor<int32, [1]> var_2267_axes_0 = const()[name = string("op_2267_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2267_keep_dims_0 = const()[name = string("op_2267_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2267_cast_fp16 = reduce_mean(axes = var_2267_axes_0, keep_dims = var_2267_keep_dims_0, x = var_2262_cast_fp16)[name = string("op_2267_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2268_cast_fp16 = mul(x = s_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_2268_cast_fp16")];
            fp32 var_2269_epsilon_0 = const()[name = string("op_2269_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2269_cast_fp16 = inverse(epsilon = var_2269_epsilon_0, x = var_2268_cast_fp16)[name = string("op_2269_cast_fp16")];
            fp16 var_2270_to_fp16 = const()[name = string("op_2270_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2271_cast_fp16 = mul(x = var_2269_cast_fp16, y = var_2270_to_fp16)[name = string("op_2271_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2273_cast_fp16 = add(x = var_2267_cast_fp16, y = var_2271_cast_fp16)[name = string("op_2273_cast_fp16")];
            fp32 var_2274_epsilon_0 = const()[name = string("op_2274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2274_cast_fp16 = rsqrt(epsilon = var_2274_epsilon_0, x = var_2273_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_1_cast_fp16 = mul(x = y_9_cast_fp16, y = var_2274_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2276_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2276_cast_fp16")];
            tensor<int32, [4]> var_2281_begin_0 = const()[name = string("op_2281_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2281_end_0 = const()[name = string("op_2281_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_2281_end_mask_0 = const()[name = string("op_2281_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_2281_cast_fp16 = slice_by_index(begin = var_2281_begin_0, end = var_2281_end_0, end_mask = var_2281_end_mask_0, x = q_1_cast_fp16)[name = string("op_2281_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_2282_cast_fp16 = mul(x = var_2281_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_2282_cast_fp16")];
            tensor<int32, [4]> var_2287_begin_0 = const()[name = string("op_2287_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2287_end_0 = const()[name = string("op_2287_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_2287_end_mask_0 = const()[name = string("op_2287_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_2287_cast_fp16 = slice_by_index(begin = var_2287_begin_0, end = var_2287_end_0, end_mask = var_2287_end_mask_0, x = q_1_cast_fp16)[name = string("op_2287_cast_fp16")];
            int32 var_2289 = const()[name = string("op_2289"), val = int32(-1)];
            bool var_2290_interleave_0 = const()[name = string("op_2290_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_2290_cast_fp16 = concat(axis = var_2289, interleave = var_2290_interleave_0, values = (var_2282_cast_fp16, var_2287_cast_fp16))[name = string("op_2290_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2291_cast_fp16 = mul(x = var_2290_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_3_cast_fp16 = add(x = var_2276_cast_fp16, y = var_2291_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2294_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2294_cast_fp16")];
            tensor<int32, [4]> var_2299_begin_0 = const()[name = string("op_2299_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2299_end_0 = const()[name = string("op_2299_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_2299_end_mask_0 = const()[name = string("op_2299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_2299_cast_fp16 = slice_by_index(begin = var_2299_begin_0, end = var_2299_end_0, end_mask = var_2299_end_mask_0, x = k_1_cast_fp16)[name = string("op_2299_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_2300_cast_fp16 = mul(x = var_2299_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [4]> var_2305_begin_0 = const()[name = string("op_2305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2305_end_0 = const()[name = string("op_2305_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_2305_end_mask_0 = const()[name = string("op_2305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_2305_cast_fp16 = slice_by_index(begin = var_2305_begin_0, end = var_2305_end_0, end_mask = var_2305_end_mask_0, x = k_1_cast_fp16)[name = string("op_2305_cast_fp16")];
            int32 var_2307 = const()[name = string("op_2307"), val = int32(-1)];
            bool var_2308_interleave_0 = const()[name = string("op_2308_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_2308_cast_fp16 = concat(axis = var_2307, interleave = var_2308_interleave_0, values = (var_2300_cast_fp16, var_2305_cast_fp16))[name = string("op_2308_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2309_cast_fp16 = mul(x = var_2308_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2309_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_3_cast_fp16 = add(x = var_2294_cast_fp16, y = var_2309_cast_fp16)[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_0_reps_0 = const()[name = string("tile_0_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = k_3_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [4, 1, 256, 256]> tile_0_cast_fp16 = tile(reps = tile_0_reps_0, x = transpose_0_cast_fp16)[name = string("tile_0_cast_fp16")];
            tensor<int32, [5]> concat_0 = const()[name = string("concat_0"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_0_cast_fp16 = reshape(shape = concat_0, x = tile_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [5]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_1 = const()[name = string("concat_1"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = reshape_0_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [4, 1, 256, 256]> reshape_1_cast_fp16 = reshape(shape = concat_1, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_1_reps_0 = const()[name = string("tile_1_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = v_1_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [4, 1, 256, 256]> tile_1_cast_fp16 = tile(reps = tile_1_reps_0, x = transpose_2_cast_fp16)[name = string("tile_1_cast_fp16")];
            tensor<int32, [5]> concat_2 = const()[name = string("concat_2"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_2_cast_fp16 = reshape(shape = concat_2, x = tile_1_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [5]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_3 = const()[name = string("concat_3"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [4, 1, 256, 256]> reshape_3_cast_fp16 = reshape(shape = concat_3, x = transpose_3_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2323_transpose_x_0 = const()[name = string("op_2323_transpose_x_0"), val = bool(false)];
            bool var_2323_transpose_y_0 = const()[name = string("op_2323_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = reshape_1_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 4, 256, 256]> var_2323_cast_fp16 = matmul(transpose_x = var_2323_transpose_x_0, transpose_y = var_2323_transpose_y_0, x = q_3_cast_fp16, y = transpose_96_cast_fp16)[name = string("op_2323_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2325_cast_fp16 = add(x = var_2323_cast_fp16, y = attention_bias)[name = string("op_2325_cast_fp16")];
            int32 var_2326 = const()[name = string("op_2326"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_2328_cast_fp16 = softmax(axis = var_2326, x = var_2325_cast_fp16)[name = string("op_2328_cast_fp16")];
            bool var_2329_transpose_x_0 = const()[name = string("op_2329_transpose_x_0"), val = bool(false)];
            bool var_2329_transpose_y_0 = const()[name = string("op_2329_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = reshape_3_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 4, 256, 256]> var_2329_cast_fp16 = matmul(transpose_x = var_2329_transpose_x_0, transpose_y = var_2329_transpose_y_0, x = var_2328_cast_fp16, y = v_3_cast_fp16)[name = string("op_2329_cast_fp16")];
            tensor<int32, [4]> var_2332_perm_0 = const()[name = string("op_2332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_2332_cast_fp16 = transpose(perm = var_2332_perm_0, x = var_2329_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 256, 1024]> input_3_cast_fp16 = reshape(shape = var_2336, x = var_2332_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7402112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7926464))))[name = string("tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_4_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2341_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_2341_cast_fp16")];
            bool var_2344 = const()[name = string("op_2344"), val = bool(true)];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2345_cast_fp16 = reduce_max(axes = var_2345_axes_0, keep_dims = var_2344, x = var_2341_cast_fp16)[name = string("op_2345_cast_fp16")];
            fp16 var_2346_to_fp16 = const()[name = string("op_2346_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_11_cast_fp16 = maximum(x = var_2345_cast_fp16, y = var_2346_to_fp16)[name = string("s_11_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_11_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_11_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2349_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_2349_cast_fp16")];
            tensor<int32, [1]> var_2354_axes_0 = const()[name = string("op_2354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2354_keep_dims_0 = const()[name = string("op_2354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2354_cast_fp16 = reduce_mean(axes = var_2354_axes_0, keep_dims = var_2354_keep_dims_0, x = var_2349_cast_fp16)[name = string("op_2354_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2355_cast_fp16 = mul(x = s_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2355_cast_fp16")];
            fp32 var_2356_epsilon_0 = const()[name = string("op_2356_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2356_cast_fp16 = inverse(epsilon = var_2356_epsilon_0, x = var_2355_cast_fp16)[name = string("op_2356_cast_fp16")];
            fp16 var_2357_to_fp16 = const()[name = string("op_2357_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2358_cast_fp16 = mul(x = var_2356_cast_fp16, y = var_2357_to_fp16)[name = string("op_2358_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2360_cast_fp16 = add(x = var_2354_cast_fp16, y = var_2358_cast_fp16)[name = string("op_2360_cast_fp16")];
            fp32 var_2361_epsilon_0 = const()[name = string("op_2361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2361_cast_fp16 = rsqrt(epsilon = var_2361_epsilon_0, x = var_2360_cast_fp16)[name = string("op_2361_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_2361_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7927552)))];
            tensor<fp16, [1, 256, 512]> var_2363_cast_fp16 = mul(x = out_9_cast_fp16, y = tm_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_2363_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_11_cast_fp16 = add(x = inputs_embeds, y = var_2363_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2366_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_2366_cast_fp16")];
            bool var_2369 = const()[name = string("op_2369"), val = bool(true)];
            tensor<int32, [1]> var_2370_axes_0 = const()[name = string("op_2370_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2370_cast_fp16 = reduce_max(axes = var_2370_axes_0, keep_dims = var_2369, x = var_2366_cast_fp16)[name = string("op_2370_cast_fp16")];
            fp16 var_2371_to_fp16 = const()[name = string("op_2371_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_13_cast_fp16 = maximum(x = var_2370_cast_fp16, y = var_2371_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = s_13_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2374_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_2374_cast_fp16")];
            tensor<int32, [1]> var_2379_axes_0 = const()[name = string("op_2379_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2379_keep_dims_0 = const()[name = string("op_2379_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2379_cast_fp16 = reduce_mean(axes = var_2379_axes_0, keep_dims = var_2379_keep_dims_0, x = var_2374_cast_fp16)[name = string("op_2379_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2380_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_2380_cast_fp16")];
            fp32 var_2381_epsilon_0 = const()[name = string("op_2381_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2381_cast_fp16 = inverse(epsilon = var_2381_epsilon_0, x = var_2380_cast_fp16)[name = string("op_2381_cast_fp16")];
            fp16 var_2382_to_fp16 = const()[name = string("op_2382_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2383_cast_fp16 = mul(x = var_2381_cast_fp16, y = var_2382_to_fp16)[name = string("op_2383_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2385_cast_fp16 = add(x = var_2379_cast_fp16, y = var_2383_cast_fp16)[name = string("op_2385_cast_fp16")];
            fp32 var_2386_epsilon_0 = const()[name = string("op_2386_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2386_cast_fp16 = rsqrt(epsilon = var_2386_epsilon_0, x = var_2385_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_2386_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7928640)))];
            tensor<fp16, [1, 256, 512]> input_5_cast_fp16 = mul(x = out_11_cast_fp16, y = tm_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7929728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8978368))))[name = string("tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8982528)))];
            tensor<fp16, [1, 256, 2048]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_2393_mode_0 = const()[name = string("op_2393_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_2393_cast_fp16 = gelu(mode = var_2393_mode_0, x = linear_5_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8986688))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10035328))))[name = string("tm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_7_cast_fp16 = mul(x = var_2393_cast_fp16, y = linear_6_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10039488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11088128))))[name = string("tm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2401_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_2401_cast_fp16")];
            bool var_2404 = const()[name = string("op_2404"), val = bool(true)];
            tensor<int32, [1]> var_2405_axes_0 = const()[name = string("op_2405_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2405_cast_fp16 = reduce_max(axes = var_2405_axes_0, keep_dims = var_2404, x = var_2401_cast_fp16)[name = string("op_2405_cast_fp16")];
            fp16 var_2406_to_fp16 = const()[name = string("op_2406_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_15_cast_fp16 = maximum(x = var_2405_cast_fp16, y = var_2406_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_15_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_15_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2409_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_2409_cast_fp16")];
            tensor<int32, [1]> var_2414_axes_0 = const()[name = string("op_2414_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2414_keep_dims_0 = const()[name = string("op_2414_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2414_cast_fp16 = reduce_mean(axes = var_2414_axes_0, keep_dims = var_2414_keep_dims_0, x = var_2409_cast_fp16)[name = string("op_2414_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2415_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_2415_cast_fp16")];
            fp32 var_2416_epsilon_0 = const()[name = string("op_2416_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2416_cast_fp16 = inverse(epsilon = var_2416_epsilon_0, x = var_2415_cast_fp16)[name = string("op_2416_cast_fp16")];
            fp16 var_2417_to_fp16 = const()[name = string("op_2417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2418_cast_fp16 = mul(x = var_2416_cast_fp16, y = var_2417_to_fp16)[name = string("op_2418_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2420_cast_fp16 = add(x = var_2414_cast_fp16, y = var_2418_cast_fp16)[name = string("op_2420_cast_fp16")];
            fp32 var_2421_epsilon_0 = const()[name = string("op_2421_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2421_cast_fp16 = rsqrt(epsilon = var_2421_epsilon_0, x = var_2420_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_2421_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11089216)))];
            tensor<fp16, [1, 256, 512]> var_2423_cast_fp16 = mul(x = out_13_cast_fp16, y = tm_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_9_cast_fp16 = add(x = x_11_cast_fp16, y = var_2423_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11090304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11352512))))[name = string("tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_8_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_8_cast_fp16")];
            string var_2430_mode_0 = const()[name = string("op_2430_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_2430_cast_fp16 = gelu(mode = var_2430_mode_0, x = linear_8_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> var_2443_begin_0 = const()[name = string("op_2443_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2443_end_0 = const()[name = string("op_2443_end_0"), val = tensor<int32, [4]>([1, 256, 1, 512])];
            tensor<bool, [4]> var_2443_end_mask_0 = const()[name = string("op_2443_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2443_squeeze_mask_0 = const()[name = string("op_2443_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_2443_cast_fp16 = slice_by_index(begin = var_2443_begin_0, end = var_2443_end_0, end_mask = var_2443_end_mask_0, squeeze_mask = var_2443_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_2443_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_11_cast_fp16 = mul(x = var_2430_cast_fp16, y = var_2443_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11353600))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11615808))))[name = string("tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_0_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2453_cast_fp16 = abs(x = linear_9_cast_fp16)[name = string("op_2453_cast_fp16")];
            bool var_2456 = const()[name = string("op_2456"), val = bool(true)];
            tensor<int32, [1]> var_2457_axes_0 = const()[name = string("op_2457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2457_cast_fp16 = reduce_max(axes = var_2457_axes_0, keep_dims = var_2456, x = var_2453_cast_fp16)[name = string("op_2457_cast_fp16")];
            fp16 var_2458_to_fp16 = const()[name = string("op_2458_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_17_cast_fp16 = maximum(x = var_2457_cast_fp16, y = var_2458_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_17_cast_fp16 = real_div(x = linear_9_cast_fp16, y = s_17_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2461_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_2461_cast_fp16")];
            tensor<int32, [1]> var_2466_axes_0 = const()[name = string("op_2466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2466_keep_dims_0 = const()[name = string("op_2466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2466_cast_fp16 = reduce_mean(axes = var_2466_axes_0, keep_dims = var_2466_keep_dims_0, x = var_2461_cast_fp16)[name = string("op_2466_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2467_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_2467_cast_fp16")];
            fp32 var_2468_epsilon_0 = const()[name = string("op_2468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2468_cast_fp16 = inverse(epsilon = var_2468_epsilon_0, x = var_2467_cast_fp16)[name = string("op_2468_cast_fp16")];
            fp16 var_2469_to_fp16 = const()[name = string("op_2469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2470_cast_fp16 = mul(x = var_2468_cast_fp16, y = var_2469_to_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2472_cast_fp16 = add(x = var_2466_cast_fp16, y = var_2470_cast_fp16)[name = string("op_2472_cast_fp16")];
            fp32 var_2473_epsilon_0 = const()[name = string("op_2473_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2473_cast_fp16 = rsqrt(epsilon = var_2473_epsilon_0, x = var_2472_cast_fp16)[name = string("op_2473_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_2473_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [512]> tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11616896)))];
            tensor<fp16, [1, 256, 512]> var_2475_cast_fp16 = mul(x = out_15_cast_fp16, y = tm_layers_0_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2475_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_1_cast_fp16 = add(x = input_9_cast_fp16, y = var_2475_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<fp16, [1]> tm_layers_0_layer_scalar_to_fp16 = const()[name = string("tm_layers_0_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9p-2])];
            tensor<fp16, [1, 256, 512]> x_17_cast_fp16 = mul(x = h_1_cast_fp16, y = tm_layers_0_layer_scalar_to_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2479_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_2479_cast_fp16")];
            bool var_2482 = const()[name = string("op_2482"), val = bool(true)];
            tensor<int32, [1]> var_2483_axes_0 = const()[name = string("op_2483_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2483_cast_fp16 = reduce_max(axes = var_2483_axes_0, keep_dims = var_2482, x = var_2479_cast_fp16)[name = string("op_2483_cast_fp16")];
            fp16 var_2484_to_fp16 = const()[name = string("op_2484_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_19_cast_fp16 = maximum(x = var_2483_cast_fp16, y = var_2484_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = s_19_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2487_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_2487_cast_fp16")];
            tensor<int32, [1]> var_2492_axes_0 = const()[name = string("op_2492_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2492_keep_dims_0 = const()[name = string("op_2492_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2492_cast_fp16 = reduce_mean(axes = var_2492_axes_0, keep_dims = var_2492_keep_dims_0, x = var_2487_cast_fp16)[name = string("op_2492_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2493_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_2493_cast_fp16")];
            fp32 var_2494_epsilon_0 = const()[name = string("op_2494_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2494_cast_fp16 = inverse(epsilon = var_2494_epsilon_0, x = var_2493_cast_fp16)[name = string("op_2494_cast_fp16")];
            fp16 var_2495_to_fp16 = const()[name = string("op_2495_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2496_cast_fp16 = mul(x = var_2494_cast_fp16, y = var_2495_to_fp16)[name = string("op_2496_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2498_cast_fp16 = add(x = var_2492_cast_fp16, y = var_2496_cast_fp16)[name = string("op_2498_cast_fp16")];
            fp32 var_2499_epsilon_0 = const()[name = string("op_2499_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2499_cast_fp16 = rsqrt(epsilon = var_2499_epsilon_0, x = var_2498_cast_fp16)[name = string("op_2499_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_2499_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11617984)))];
            tensor<fp16, [1, 256, 512]> input_13_cast_fp16 = mul(x = out_17_cast_fp16, y = tm_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11619072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12143424))))[name = string("tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_2510_cast_fp16 = reshape(shape = var_2509, x = linear_10_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<int32, [4]> x_19_perm_0 = const()[name = string("x_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12145536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12407744))))[name = string("tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<int32, [4]> var_2521 = const()[name = string("op_2521"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2522_cast_fp16 = reshape(shape = var_2521, x = linear_11_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<int32, [4]> x_21_perm_0 = const()[name = string("x_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12408832))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12671040))))[name = string("tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<int32, [4]> var_2533 = const()[name = string("op_2533"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2534_cast_fp16 = reshape(shape = var_2533, x = linear_12_cast_fp16)[name = string("op_2534_cast_fp16")];
            tensor<int32, [4]> x_23_perm_0 = const()[name = string("x_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_19_cast_fp16 = transpose(perm = x_19_perm_0, x = var_2510_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 4, 256, 256]> var_2538_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_2538_cast_fp16")];
            bool var_2541 = const()[name = string("op_2541"), val = bool(true)];
            tensor<int32, [1]> var_2542_axes_0 = const()[name = string("op_2542_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_2542_cast_fp16 = reduce_max(axes = var_2542_axes_0, keep_dims = var_2541, x = var_2538_cast_fp16)[name = string("op_2542_cast_fp16")];
            fp16 var_2543_to_fp16 = const()[name = string("op_2543_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_21_cast_fp16 = maximum(x = var_2542_cast_fp16, y = var_2543_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_21_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2546_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_2546_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2551_keep_dims_0 = const()[name = string("op_2551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_2551_cast_fp16 = reduce_mean(axes = var_2551_axes_0, keep_dims = var_2551_keep_dims_0, x = var_2546_cast_fp16)[name = string("op_2551_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2552_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_2552_cast_fp16")];
            fp32 var_2553_epsilon_0 = const()[name = string("op_2553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_2553_cast_fp16 = inverse(epsilon = var_2553_epsilon_0, x = var_2552_cast_fp16)[name = string("op_2553_cast_fp16")];
            fp16 var_2554_to_fp16 = const()[name = string("op_2554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_2555_cast_fp16 = mul(x = var_2553_cast_fp16, y = var_2554_to_fp16)[name = string("op_2555_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2557_cast_fp16 = add(x = var_2551_cast_fp16, y = var_2555_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_2558_cast_fp16 = rsqrt(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_19_cast_fp16 = mul(x = y_21_cast_fp16, y = var_2558_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672128)))];
            tensor<fp16, [1, 4, 256, 256]> q_5_cast_fp16 = mul(x = out_19_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_21_cast_fp16 = transpose(perm = x_21_perm_0, x = var_2522_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 2, 256, 256]> var_2561_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_2561_cast_fp16")];
            bool var_2564 = const()[name = string("op_2564"), val = bool(true)];
            tensor<int32, [1]> var_2565_axes_0 = const()[name = string("op_2565_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2565_cast_fp16 = reduce_max(axes = var_2565_axes_0, keep_dims = var_2564, x = var_2561_cast_fp16)[name = string("op_2565_cast_fp16")];
            fp16 var_2566_to_fp16 = const()[name = string("op_2566_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_23_cast_fp16 = maximum(x = var_2565_cast_fp16, y = var_2566_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_23_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2569_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2569_cast_fp16")];
            tensor<int32, [1]> var_2574_axes_0 = const()[name = string("op_2574_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2574_keep_dims_0 = const()[name = string("op_2574_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2574_cast_fp16 = reduce_mean(axes = var_2574_axes_0, keep_dims = var_2574_keep_dims_0, x = var_2569_cast_fp16)[name = string("op_2574_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2575_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_2575_cast_fp16")];
            fp32 var_2576_epsilon_0 = const()[name = string("op_2576_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2576_cast_fp16 = inverse(epsilon = var_2576_epsilon_0, x = var_2575_cast_fp16)[name = string("op_2576_cast_fp16")];
            fp16 var_2577_to_fp16 = const()[name = string("op_2577_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2578_cast_fp16 = mul(x = var_2576_cast_fp16, y = var_2577_to_fp16)[name = string("op_2578_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2580_cast_fp16 = add(x = var_2574_cast_fp16, y = var_2578_cast_fp16)[name = string("op_2580_cast_fp16")];
            fp32 var_2581_epsilon_0 = const()[name = string("op_2581_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2581_cast_fp16 = rsqrt(epsilon = var_2581_epsilon_0, x = var_2580_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_21_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2581_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [256]> tm_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12672704)))];
            tensor<fp16, [1, 2, 256, 256]> k_7_cast_fp16 = mul(x = out_21_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_23_cast_fp16 = transpose(perm = x_23_perm_0, x = var_2534_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 2, 256, 256]> var_2584_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_2584_cast_fp16")];
            bool var_2587 = const()[name = string("op_2587"), val = bool(true)];
            tensor<int32, [1]> var_2588_axes_0 = const()[name = string("op_2588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2588_cast_fp16 = reduce_max(axes = var_2588_axes_0, keep_dims = var_2587, x = var_2584_cast_fp16)[name = string("op_2588_cast_fp16")];
            fp16 var_2589_to_fp16 = const()[name = string("op_2589_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_25_cast_fp16 = maximum(x = var_2588_cast_fp16, y = var_2589_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_25_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2592_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2592_cast_fp16")];
            tensor<int32, [1]> var_2597_axes_0 = const()[name = string("op_2597_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2597_keep_dims_0 = const()[name = string("op_2597_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2597_cast_fp16 = reduce_mean(axes = var_2597_axes_0, keep_dims = var_2597_keep_dims_0, x = var_2592_cast_fp16)[name = string("op_2597_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2598_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_2598_cast_fp16")];
            fp32 var_2599_epsilon_0 = const()[name = string("op_2599_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2599_cast_fp16 = inverse(epsilon = var_2599_epsilon_0, x = var_2598_cast_fp16)[name = string("op_2599_cast_fp16")];
            fp16 var_2600_to_fp16 = const()[name = string("op_2600_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2601_cast_fp16 = mul(x = var_2599_cast_fp16, y = var_2600_to_fp16)[name = string("op_2601_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2603_cast_fp16 = add(x = var_2597_cast_fp16, y = var_2601_cast_fp16)[name = string("op_2603_cast_fp16")];
            fp32 var_2604_epsilon_0 = const()[name = string("op_2604_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2604_cast_fp16 = rsqrt(epsilon = var_2604_epsilon_0, x = var_2603_cast_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_5_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2604_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2606_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2606_cast_fp16")];
            tensor<int32, [4]> var_2611_begin_0 = const()[name = string("op_2611_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2611_end_0 = const()[name = string("op_2611_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_2611_end_mask_0 = const()[name = string("op_2611_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_2611_cast_fp16 = slice_by_index(begin = var_2611_begin_0, end = var_2611_end_0, end_mask = var_2611_end_mask_0, x = q_5_cast_fp16)[name = string("op_2611_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_2612_cast_fp16 = mul(x = var_2611_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_2612_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = q_5_cast_fp16)[name = string("op_2617_cast_fp16")];
            int32 var_2619 = const()[name = string("op_2619"), val = int32(-1)];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_2620_cast_fp16 = concat(axis = var_2619, interleave = var_2620_interleave_0, values = (var_2612_cast_fp16, var_2617_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_7_cast_fp16 = add(x = var_2606_cast_fp16, y = var_2621_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2624_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2624_cast_fp16")];
            tensor<int32, [4]> var_2629_begin_0 = const()[name = string("op_2629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2629_end_0 = const()[name = string("op_2629_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_2629_end_mask_0 = const()[name = string("op_2629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_2629_cast_fp16 = slice_by_index(begin = var_2629_begin_0, end = var_2629_end_0, end_mask = var_2629_end_mask_0, x = k_7_cast_fp16)[name = string("op_2629_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_2630_cast_fp16 = mul(x = var_2629_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [4]> var_2635_begin_0 = const()[name = string("op_2635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2635_end_0 = const()[name = string("op_2635_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_2635_end_mask_0 = const()[name = string("op_2635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_2635_cast_fp16 = slice_by_index(begin = var_2635_begin_0, end = var_2635_end_0, end_mask = var_2635_end_mask_0, x = k_7_cast_fp16)[name = string("op_2635_cast_fp16")];
            int32 var_2637 = const()[name = string("op_2637"), val = int32(-1)];
            bool var_2638_interleave_0 = const()[name = string("op_2638_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_2638_cast_fp16 = concat(axis = var_2637, interleave = var_2638_interleave_0, values = (var_2630_cast_fp16, var_2635_cast_fp16))[name = string("op_2638_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2639_cast_fp16 = mul(x = var_2638_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_9_cast_fp16 = add(x = var_2624_cast_fp16, y = var_2639_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_2_reps_0 = const()[name = string("tile_2_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = k_9_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [4, 1, 256, 256]> tile_2_cast_fp16 = tile(reps = tile_2_reps_0, x = transpose_4_cast_fp16)[name = string("tile_2_cast_fp16")];
            tensor<int32, [5]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_4_cast_fp16 = reshape(shape = concat_4, x = tile_2_cast_fp16)[name = string("reshape_4_cast_fp16")];
            tensor<int32, [5]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = reshape_4_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [4, 1, 256, 256]> reshape_5_cast_fp16 = reshape(shape = concat_5, x = transpose_5_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_3_reps_0 = const()[name = string("tile_3_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = v_5_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [4, 1, 256, 256]> tile_3_cast_fp16 = tile(reps = tile_3_reps_0, x = transpose_6_cast_fp16)[name = string("tile_3_cast_fp16")];
            tensor<int32, [5]> concat_6 = const()[name = string("concat_6"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_6_cast_fp16 = reshape(shape = concat_6, x = tile_3_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [5]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = reshape_6_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [4, 1, 256, 256]> reshape_7_cast_fp16 = reshape(shape = concat_7, x = transpose_7_cast_fp16)[name = string("reshape_7_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2653_transpose_x_0 = const()[name = string("op_2653_transpose_x_0"), val = bool(false)];
            bool var_2653_transpose_y_0 = const()[name = string("op_2653_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_97_cast_fp16 = transpose(perm = transpose_97_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 4, 256, 256]> var_2653_cast_fp16 = matmul(transpose_x = var_2653_transpose_x_0, transpose_y = var_2653_transpose_y_0, x = q_7_cast_fp16, y = transpose_97_cast_fp16)[name = string("op_2653_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2655_cast_fp16 = add(x = var_2653_cast_fp16, y = attention_bias)[name = string("op_2655_cast_fp16")];
            int32 var_2656 = const()[name = string("op_2656"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_2658_cast_fp16 = softmax(axis = var_2656, x = var_2655_cast_fp16)[name = string("op_2658_cast_fp16")];
            bool var_2659_transpose_x_0 = const()[name = string("op_2659_transpose_x_0"), val = bool(false)];
            bool var_2659_transpose_y_0 = const()[name = string("op_2659_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = reshape_7_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 4, 256, 256]> var_2659_cast_fp16 = matmul(transpose_x = var_2659_transpose_x_0, transpose_y = var_2659_transpose_y_0, x = var_2658_cast_fp16, y = v_7_cast_fp16)[name = string("op_2659_cast_fp16")];
            tensor<int32, [4]> var_2662_perm_0 = const()[name = string("op_2662_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2666 = const()[name = string("op_2666"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_2662_cast_fp16 = transpose(perm = var_2662_perm_0, x = var_2659_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 256, 1024]> input_15_cast_fp16 = reshape(shape = var_2666, x = var_2662_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12673280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13197632))))[name = string("tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2671_cast_fp16 = abs(x = linear_13_cast_fp16)[name = string("op_2671_cast_fp16")];
            bool var_2674 = const()[name = string("op_2674"), val = bool(true)];
            tensor<int32, [1]> var_2675_axes_0 = const()[name = string("op_2675_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2675_cast_fp16 = reduce_max(axes = var_2675_axes_0, keep_dims = var_2674, x = var_2671_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp16 var_2676_to_fp16 = const()[name = string("op_2676_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_27_cast_fp16 = maximum(x = var_2675_cast_fp16, y = var_2676_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_27_cast_fp16 = real_div(x = linear_13_cast_fp16, y = s_27_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2679_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2679_cast_fp16")];
            tensor<int32, [1]> var_2684_axes_0 = const()[name = string("op_2684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2684_keep_dims_0 = const()[name = string("op_2684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2684_cast_fp16 = reduce_mean(axes = var_2684_axes_0, keep_dims = var_2684_keep_dims_0, x = var_2679_cast_fp16)[name = string("op_2684_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2685_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2685_cast_fp16")];
            fp32 var_2686_epsilon_0 = const()[name = string("op_2686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2686_cast_fp16 = inverse(epsilon = var_2686_epsilon_0, x = var_2685_cast_fp16)[name = string("op_2686_cast_fp16")];
            fp16 var_2687_to_fp16 = const()[name = string("op_2687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2688_cast_fp16 = mul(x = var_2686_cast_fp16, y = var_2687_to_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2690_cast_fp16 = add(x = var_2684_cast_fp16, y = var_2688_cast_fp16)[name = string("op_2690_cast_fp16")];
            fp32 var_2691_epsilon_0 = const()[name = string("op_2691_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2691_cast_fp16 = rsqrt(epsilon = var_2691_epsilon_0, x = var_2690_cast_fp16)[name = string("op_2691_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2691_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13198720)))];
            tensor<fp16, [1, 256, 512]> var_2693_cast_fp16 = mul(x = out_23_cast_fp16, y = tm_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2693_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_27_cast_fp16 = add(x = x_17_cast_fp16, y = var_2693_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2696_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_2696_cast_fp16")];
            bool var_2699 = const()[name = string("op_2699"), val = bool(true)];
            tensor<int32, [1]> var_2700_axes_0 = const()[name = string("op_2700_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2700_cast_fp16 = reduce_max(axes = var_2700_axes_0, keep_dims = var_2699, x = var_2696_cast_fp16)[name = string("op_2700_cast_fp16")];
            fp16 var_2701_to_fp16 = const()[name = string("op_2701_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_29_cast_fp16 = maximum(x = var_2700_cast_fp16, y = var_2701_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = s_29_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2704_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2704_cast_fp16")];
            tensor<int32, [1]> var_2709_axes_0 = const()[name = string("op_2709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2709_keep_dims_0 = const()[name = string("op_2709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2709_cast_fp16 = reduce_mean(axes = var_2709_axes_0, keep_dims = var_2709_keep_dims_0, x = var_2704_cast_fp16)[name = string("op_2709_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2710_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2710_cast_fp16")];
            fp32 var_2711_epsilon_0 = const()[name = string("op_2711_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2711_cast_fp16 = inverse(epsilon = var_2711_epsilon_0, x = var_2710_cast_fp16)[name = string("op_2711_cast_fp16")];
            fp16 var_2712_to_fp16 = const()[name = string("op_2712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2713_cast_fp16 = mul(x = var_2711_cast_fp16, y = var_2712_to_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2715_cast_fp16 = add(x = var_2709_cast_fp16, y = var_2713_cast_fp16)[name = string("op_2715_cast_fp16")];
            fp32 var_2716_epsilon_0 = const()[name = string("op_2716_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2716_cast_fp16 = rsqrt(epsilon = var_2716_epsilon_0, x = var_2715_cast_fp16)[name = string("op_2716_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2716_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13199808)))];
            tensor<fp16, [1, 256, 512]> input_17_cast_fp16 = mul(x = out_25_cast_fp16, y = tm_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13200896))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14249536))))[name = string("tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_14_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_14_cast_fp16")];
            string var_2723_mode_0 = const()[name = string("op_2723_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_2723_cast_fp16 = gelu(mode = var_2723_mode_0, x = linear_14_cast_fp16)[name = string("op_2723_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14253696))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15302336))))[name = string("tm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_15_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_19_cast_fp16 = mul(x = var_2723_cast_fp16, y = linear_15_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15306496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16355136))))[name = string("tm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2731_cast_fp16 = abs(x = linear_16_cast_fp16)[name = string("op_2731_cast_fp16")];
            bool var_2734 = const()[name = string("op_2734"), val = bool(true)];
            tensor<int32, [1]> var_2735_axes_0 = const()[name = string("op_2735_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2735_cast_fp16 = reduce_max(axes = var_2735_axes_0, keep_dims = var_2734, x = var_2731_cast_fp16)[name = string("op_2735_cast_fp16")];
            fp16 var_2736_to_fp16 = const()[name = string("op_2736_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_31_cast_fp16 = maximum(x = var_2735_cast_fp16, y = var_2736_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_31_cast_fp16 = real_div(x = linear_16_cast_fp16, y = s_31_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2739_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2739_cast_fp16")];
            tensor<int32, [1]> var_2744_axes_0 = const()[name = string("op_2744_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2744_keep_dims_0 = const()[name = string("op_2744_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2744_cast_fp16 = reduce_mean(axes = var_2744_axes_0, keep_dims = var_2744_keep_dims_0, x = var_2739_cast_fp16)[name = string("op_2744_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2745_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2745_cast_fp16")];
            fp32 var_2746_epsilon_0 = const()[name = string("op_2746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2746_cast_fp16 = inverse(epsilon = var_2746_epsilon_0, x = var_2745_cast_fp16)[name = string("op_2746_cast_fp16")];
            fp16 var_2747_to_fp16 = const()[name = string("op_2747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2748_cast_fp16 = mul(x = var_2746_cast_fp16, y = var_2747_to_fp16)[name = string("op_2748_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2750_cast_fp16 = add(x = var_2744_cast_fp16, y = var_2748_cast_fp16)[name = string("op_2750_cast_fp16")];
            fp32 var_2751_epsilon_0 = const()[name = string("op_2751_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2751_cast_fp16 = rsqrt(epsilon = var_2751_epsilon_0, x = var_2750_cast_fp16)[name = string("op_2751_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2751_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16356224)))];
            tensor<fp16, [1, 256, 512]> var_2753_cast_fp16 = mul(x = out_27_cast_fp16, y = tm_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_21_cast_fp16 = add(x = x_27_cast_fp16, y = var_2753_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16357312))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16619520))))[name = string("tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_17_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            string var_2760_mode_0 = const()[name = string("op_2760_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_2760_cast_fp16 = gelu(mode = var_2760_mode_0, x = linear_17_cast_fp16)[name = string("op_2760_cast_fp16")];
            tensor<int32, [4]> var_2773_begin_0 = const()[name = string("op_2773_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2773_end_0 = const()[name = string("op_2773_end_0"), val = tensor<int32, [4]>([1, 256, 2, 512])];
            tensor<bool, [4]> var_2773_end_mask_0 = const()[name = string("op_2773_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_2773_squeeze_mask_0 = const()[name = string("op_2773_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_2773_cast_fp16 = slice_by_index(begin = var_2773_begin_0, end = var_2773_end_0, end_mask = var_2773_end_mask_0, squeeze_mask = var_2773_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_2773_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_23_cast_fp16 = mul(x = var_2760_cast_fp16, y = var_2773_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16620608))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16882816))))[name = string("tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_1_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2783_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2783_cast_fp16")];
            bool var_2786 = const()[name = string("op_2786"), val = bool(true)];
            tensor<int32, [1]> var_2787_axes_0 = const()[name = string("op_2787_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2787_cast_fp16 = reduce_max(axes = var_2787_axes_0, keep_dims = var_2786, x = var_2783_cast_fp16)[name = string("op_2787_cast_fp16")];
            fp16 var_2788_to_fp16 = const()[name = string("op_2788_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_33_cast_fp16 = maximum(x = var_2787_cast_fp16, y = var_2788_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_33_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_33_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2791_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2791_cast_fp16")];
            tensor<int32, [1]> var_2796_axes_0 = const()[name = string("op_2796_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2796_keep_dims_0 = const()[name = string("op_2796_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2796_cast_fp16 = reduce_mean(axes = var_2796_axes_0, keep_dims = var_2796_keep_dims_0, x = var_2791_cast_fp16)[name = string("op_2796_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2797_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2797_cast_fp16")];
            fp32 var_2798_epsilon_0 = const()[name = string("op_2798_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2798_cast_fp16 = inverse(epsilon = var_2798_epsilon_0, x = var_2797_cast_fp16)[name = string("op_2798_cast_fp16")];
            fp16 var_2799_to_fp16 = const()[name = string("op_2799_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2800_cast_fp16 = mul(x = var_2798_cast_fp16, y = var_2799_to_fp16)[name = string("op_2800_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2802_cast_fp16 = add(x = var_2796_cast_fp16, y = var_2800_cast_fp16)[name = string("op_2802_cast_fp16")];
            fp32 var_2803_epsilon_0 = const()[name = string("op_2803_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2803_cast_fp16 = rsqrt(epsilon = var_2803_epsilon_0, x = var_2802_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2803_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [512]> tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16883904)))];
            tensor<fp16, [1, 256, 512]> var_2805_cast_fp16 = mul(x = out_29_cast_fp16, y = tm_layers_1_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_2805_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_3_cast_fp16 = add(x = input_21_cast_fp16, y = var_2805_cast_fp16)[name = string("h_3_cast_fp16")];
            tensor<fp16, [1]> tm_layers_1_layer_scalar_to_fp16 = const()[name = string("tm_layers_1_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.5ap-3])];
            tensor<fp16, [1, 256, 512]> x_33_cast_fp16 = mul(x = h_3_cast_fp16, y = tm_layers_1_layer_scalar_to_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2809_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_2809_cast_fp16")];
            bool var_2812 = const()[name = string("op_2812"), val = bool(true)];
            tensor<int32, [1]> var_2813_axes_0 = const()[name = string("op_2813_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2813_cast_fp16 = reduce_max(axes = var_2813_axes_0, keep_dims = var_2812, x = var_2809_cast_fp16)[name = string("op_2813_cast_fp16")];
            fp16 var_2814_to_fp16 = const()[name = string("op_2814_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_35_cast_fp16 = maximum(x = var_2813_cast_fp16, y = var_2814_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = s_35_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_2817_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2817_cast_fp16")];
            tensor<int32, [1]> var_2822_axes_0 = const()[name = string("op_2822_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2822_keep_dims_0 = const()[name = string("op_2822_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_2822_cast_fp16 = reduce_mean(axes = var_2822_axes_0, keep_dims = var_2822_keep_dims_0, x = var_2817_cast_fp16)[name = string("op_2822_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2823_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2823_cast_fp16")];
            fp32 var_2824_epsilon_0 = const()[name = string("op_2824_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2824_cast_fp16 = inverse(epsilon = var_2824_epsilon_0, x = var_2823_cast_fp16)[name = string("op_2824_cast_fp16")];
            fp16 var_2825_to_fp16 = const()[name = string("op_2825_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2826_cast_fp16 = mul(x = var_2824_cast_fp16, y = var_2825_to_fp16)[name = string("op_2826_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2828_cast_fp16 = add(x = var_2822_cast_fp16, y = var_2826_cast_fp16)[name = string("op_2828_cast_fp16")];
            fp32 var_2829_epsilon_0 = const()[name = string("op_2829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2829_cast_fp16 = rsqrt(epsilon = var_2829_epsilon_0, x = var_2828_cast_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_31_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2829_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16884992)))];
            tensor<fp16, [1, 256, 512]> input_25_cast_fp16 = mul(x = out_31_cast_fp16, y = tm_layers_2_input_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16886080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17410432))))[name = string("tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [4]> var_2839 = const()[name = string("op_2839"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_2840_cast_fp16 = reshape(shape = var_2839, x = linear_19_cast_fp16)[name = string("op_2840_cast_fp16")];
            tensor<int32, [4]> x_35_perm_0 = const()[name = string("x_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17412544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17674752))))[name = string("tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<int32, [4]> var_2851 = const()[name = string("op_2851"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2852_cast_fp16 = reshape(shape = var_2851, x = linear_20_cast_fp16)[name = string("op_2852_cast_fp16")];
            tensor<int32, [4]> x_37_perm_0 = const()[name = string("x_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17675840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17938048))))[name = string("tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2863 = const()[name = string("op_2863"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_2864_cast_fp16 = reshape(shape = var_2863, x = linear_21_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_35_cast_fp16 = transpose(perm = x_35_perm_0, x = var_2840_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 4, 256, 256]> var_2868_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_2868_cast_fp16")];
            bool var_2871 = const()[name = string("op_2871"), val = bool(true)];
            tensor<int32, [1]> var_2872_axes_0 = const()[name = string("op_2872_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_2872_cast_fp16 = reduce_max(axes = var_2872_axes_0, keep_dims = var_2871, x = var_2868_cast_fp16)[name = string("op_2872_cast_fp16")];
            fp16 var_2873_to_fp16 = const()[name = string("op_2873_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_37_cast_fp16 = maximum(x = var_2872_cast_fp16, y = var_2873_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = s_37_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2876_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2876_cast_fp16")];
            tensor<int32, [1]> var_2881_axes_0 = const()[name = string("op_2881_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2881_keep_dims_0 = const()[name = string("op_2881_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_2881_cast_fp16 = reduce_mean(axes = var_2881_axes_0, keep_dims = var_2881_keep_dims_0, x = var_2876_cast_fp16)[name = string("op_2881_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2882_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2882_cast_fp16")];
            fp32 var_2883_epsilon_0 = const()[name = string("op_2883_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_2883_cast_fp16 = inverse(epsilon = var_2883_epsilon_0, x = var_2882_cast_fp16)[name = string("op_2883_cast_fp16")];
            fp16 var_2884_to_fp16 = const()[name = string("op_2884_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_2885_cast_fp16 = mul(x = var_2883_cast_fp16, y = var_2884_to_fp16)[name = string("op_2885_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_2887_cast_fp16 = add(x = var_2881_cast_fp16, y = var_2885_cast_fp16)[name = string("op_2887_cast_fp16")];
            fp32 var_2888_epsilon_0 = const()[name = string("op_2888_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_2888_cast_fp16 = rsqrt(epsilon = var_2888_epsilon_0, x = var_2887_cast_fp16)[name = string("op_2888_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_33_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2888_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_9_cast_fp16 = mul(x = out_33_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_9_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_37_cast_fp16 = transpose(perm = x_37_perm_0, x = var_2852_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 2, 256, 256]> var_2891_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2891_cast_fp16")];
            bool var_2894 = const()[name = string("op_2894"), val = bool(true)];
            tensor<int32, [1]> var_2895_axes_0 = const()[name = string("op_2895_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2895_cast_fp16 = reduce_max(axes = var_2895_axes_0, keep_dims = var_2894, x = var_2891_cast_fp16)[name = string("op_2895_cast_fp16")];
            fp16 var_2896_to_fp16 = const()[name = string("op_2896_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_39_cast_fp16 = maximum(x = var_2895_cast_fp16, y = var_2896_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_39_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2899_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<int32, [1]> var_2904_axes_0 = const()[name = string("op_2904_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2904_keep_dims_0 = const()[name = string("op_2904_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2904_cast_fp16 = reduce_mean(axes = var_2904_axes_0, keep_dims = var_2904_keep_dims_0, x = var_2899_cast_fp16)[name = string("op_2904_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2905_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2905_cast_fp16")];
            fp32 var_2906_epsilon_0 = const()[name = string("op_2906_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2906_cast_fp16 = inverse(epsilon = var_2906_epsilon_0, x = var_2905_cast_fp16)[name = string("op_2906_cast_fp16")];
            fp16 var_2907_to_fp16 = const()[name = string("op_2907_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2908_cast_fp16 = mul(x = var_2906_cast_fp16, y = var_2907_to_fp16)[name = string("op_2908_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2910_cast_fp16 = add(x = var_2904_cast_fp16, y = var_2908_cast_fp16)[name = string("op_2910_cast_fp16")];
            fp32 var_2911_epsilon_0 = const()[name = string("op_2911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2911_cast_fp16 = rsqrt(epsilon = var_2911_epsilon_0, x = var_2910_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_35_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2911_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_13_cast_fp16 = mul(x = out_35_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2864_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 2, 256, 256]> var_2914_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2914_cast_fp16")];
            bool var_2917 = const()[name = string("op_2917"), val = bool(true)];
            tensor<int32, [1]> var_2918_axes_0 = const()[name = string("op_2918_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_2918_cast_fp16 = reduce_max(axes = var_2918_axes_0, keep_dims = var_2917, x = var_2914_cast_fp16)[name = string("op_2918_cast_fp16")];
            fp16 var_2919_to_fp16 = const()[name = string("op_2919_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_41_cast_fp16 = maximum(x = var_2918_cast_fp16, y = var_2919_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_41_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2922_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2922_cast_fp16")];
            tensor<int32, [1]> var_2927_axes_0 = const()[name = string("op_2927_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2927_keep_dims_0 = const()[name = string("op_2927_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_2927_cast_fp16 = reduce_mean(axes = var_2927_axes_0, keep_dims = var_2927_keep_dims_0, x = var_2922_cast_fp16)[name = string("op_2927_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2928_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2928_cast_fp16")];
            fp32 var_2929_epsilon_0 = const()[name = string("op_2929_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_2929_cast_fp16 = inverse(epsilon = var_2929_epsilon_0, x = var_2928_cast_fp16)[name = string("op_2929_cast_fp16")];
            fp16 var_2930_to_fp16 = const()[name = string("op_2930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_2931_cast_fp16 = mul(x = var_2929_cast_fp16, y = var_2930_to_fp16)[name = string("op_2931_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_2933_cast_fp16 = add(x = var_2927_cast_fp16, y = var_2931_cast_fp16)[name = string("op_2933_cast_fp16")];
            fp32 var_2934_epsilon_0 = const()[name = string("op_2934_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_2934_cast_fp16 = rsqrt(epsilon = var_2934_epsilon_0, x = var_2933_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_9_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2934_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2936_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2936_cast_fp16")];
            tensor<int32, [4]> var_2941_begin_0 = const()[name = string("op_2941_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2941_end_0 = const()[name = string("op_2941_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_2941_end_mask_0 = const()[name = string("op_2941_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_2941_cast_fp16 = slice_by_index(begin = var_2941_begin_0, end = var_2941_end_0, end_mask = var_2941_end_mask_0, x = q_9_cast_fp16)[name = string("op_2941_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_2942_cast_fp16 = mul(x = var_2941_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_2942_cast_fp16")];
            tensor<int32, [4]> var_2947_begin_0 = const()[name = string("op_2947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2947_end_0 = const()[name = string("op_2947_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_2947_end_mask_0 = const()[name = string("op_2947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_2947_cast_fp16 = slice_by_index(begin = var_2947_begin_0, end = var_2947_end_0, end_mask = var_2947_end_mask_0, x = q_9_cast_fp16)[name = string("op_2947_cast_fp16")];
            int32 var_2949 = const()[name = string("op_2949"), val = int32(-1)];
            bool var_2950_interleave_0 = const()[name = string("op_2950_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_2950_cast_fp16 = concat(axis = var_2949, interleave = var_2950_interleave_0, values = (var_2942_cast_fp16, var_2947_cast_fp16))[name = string("op_2950_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2951_cast_fp16 = mul(x = var_2950_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2951_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_11_cast_fp16 = add(x = var_2936_cast_fp16, y = var_2951_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2954_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2954_cast_fp16")];
            tensor<int32, [4]> var_2959_begin_0 = const()[name = string("op_2959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_2959_end_0 = const()[name = string("op_2959_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_2959_end_mask_0 = const()[name = string("op_2959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_2959_cast_fp16 = slice_by_index(begin = var_2959_begin_0, end = var_2959_end_0, end_mask = var_2959_end_mask_0, x = k_13_cast_fp16)[name = string("op_2959_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_2960_cast_fp16 = mul(x = var_2959_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2960_cast_fp16")];
            tensor<int32, [4]> var_2965_begin_0 = const()[name = string("op_2965_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2965_end_0 = const()[name = string("op_2965_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_2965_end_mask_0 = const()[name = string("op_2965_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_2965_cast_fp16 = slice_by_index(begin = var_2965_begin_0, end = var_2965_end_0, end_mask = var_2965_end_mask_0, x = k_13_cast_fp16)[name = string("op_2965_cast_fp16")];
            int32 var_2967 = const()[name = string("op_2967"), val = int32(-1)];
            bool var_2968_interleave_0 = const()[name = string("op_2968_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_2968_cast_fp16 = concat(axis = var_2967, interleave = var_2968_interleave_0, values = (var_2960_cast_fp16, var_2965_cast_fp16))[name = string("op_2968_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_2969_cast_fp16 = mul(x = var_2968_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_15_cast_fp16 = add(x = var_2954_cast_fp16, y = var_2969_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_4_reps_0 = const()[name = string("tile_4_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = k_15_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [4, 1, 256, 256]> tile_4_cast_fp16 = tile(reps = tile_4_reps_0, x = transpose_8_cast_fp16)[name = string("tile_4_cast_fp16")];
            tensor<int32, [5]> concat_8 = const()[name = string("concat_8"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_8_cast_fp16 = reshape(shape = concat_8, x = tile_4_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [5]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [4, 1, 256, 256]> reshape_9_cast_fp16 = reshape(shape = concat_9, x = transpose_9_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_5_reps_0 = const()[name = string("tile_5_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = v_9_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [4, 1, 256, 256]> tile_5_cast_fp16 = tile(reps = tile_5_reps_0, x = transpose_10_cast_fp16)[name = string("tile_5_cast_fp16")];
            tensor<int32, [5]> concat_10 = const()[name = string("concat_10"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_10_cast_fp16 = reshape(shape = concat_10, x = tile_5_cast_fp16)[name = string("reshape_10_cast_fp16")];
            tensor<int32, [5]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_11 = const()[name = string("concat_11"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = reshape_10_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [4, 1, 256, 256]> reshape_11_cast_fp16 = reshape(shape = concat_11, x = transpose_11_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_2983_transpose_x_0 = const()[name = string("op_2983_transpose_x_0"), val = bool(false)];
            bool var_2983_transpose_y_0 = const()[name = string("op_2983_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = reshape_9_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 4, 256, 256]> var_2983_cast_fp16 = matmul(transpose_x = var_2983_transpose_x_0, transpose_y = var_2983_transpose_y_0, x = q_11_cast_fp16, y = transpose_98_cast_fp16)[name = string("op_2983_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_2985_cast_fp16 = add(x = var_2983_cast_fp16, y = attention_bias)[name = string("op_2985_cast_fp16")];
            int32 var_2986 = const()[name = string("op_2986"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_2988_cast_fp16 = softmax(axis = var_2986, x = var_2985_cast_fp16)[name = string("op_2988_cast_fp16")];
            bool var_2989_transpose_x_0 = const()[name = string("op_2989_transpose_x_0"), val = bool(false)];
            bool var_2989_transpose_y_0 = const()[name = string("op_2989_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 4, 256, 256]> var_2989_cast_fp16 = matmul(transpose_x = var_2989_transpose_x_0, transpose_y = var_2989_transpose_y_0, x = var_2988_cast_fp16, y = v_11_cast_fp16)[name = string("op_2989_cast_fp16")];
            tensor<int32, [4]> var_2992_perm_0 = const()[name = string("op_2992_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2996 = const()[name = string("op_2996"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_2992_cast_fp16 = transpose(perm = var_2992_perm_0, x = var_2989_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 256, 1024]> input_27_cast_fp16 = reshape(shape = var_2996, x = var_2992_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17939136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18463488))))[name = string("tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3001_cast_fp16 = abs(x = linear_22_cast_fp16)[name = string("op_3001_cast_fp16")];
            bool var_3004 = const()[name = string("op_3004"), val = bool(true)];
            tensor<int32, [1]> var_3005_axes_0 = const()[name = string("op_3005_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3005_cast_fp16 = reduce_max(axes = var_3005_axes_0, keep_dims = var_3004, x = var_3001_cast_fp16)[name = string("op_3005_cast_fp16")];
            fp16 var_3006_to_fp16 = const()[name = string("op_3006_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_43_cast_fp16 = maximum(x = var_3005_cast_fp16, y = var_3006_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_43_cast_fp16 = real_div(x = linear_22_cast_fp16, y = s_43_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3009_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_3009_cast_fp16")];
            tensor<int32, [1]> var_3014_axes_0 = const()[name = string("op_3014_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3014_keep_dims_0 = const()[name = string("op_3014_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3014_cast_fp16 = reduce_mean(axes = var_3014_axes_0, keep_dims = var_3014_keep_dims_0, x = var_3009_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3015_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_3015_cast_fp16")];
            fp32 var_3016_epsilon_0 = const()[name = string("op_3016_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3016_cast_fp16 = inverse(epsilon = var_3016_epsilon_0, x = var_3015_cast_fp16)[name = string("op_3016_cast_fp16")];
            fp16 var_3017_to_fp16 = const()[name = string("op_3017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3018_cast_fp16 = mul(x = var_3016_cast_fp16, y = var_3017_to_fp16)[name = string("op_3018_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3020_cast_fp16 = add(x = var_3014_cast_fp16, y = var_3018_cast_fp16)[name = string("op_3020_cast_fp16")];
            fp32 var_3021_epsilon_0 = const()[name = string("op_3021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3021_cast_fp16 = rsqrt(epsilon = var_3021_epsilon_0, x = var_3020_cast_fp16)[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_3021_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18464576)))];
            tensor<fp16, [1, 256, 512]> var_3023_cast_fp16 = mul(x = out_37_cast_fp16, y = tm_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_3023_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_43_cast_fp16 = add(x = x_33_cast_fp16, y = var_3023_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3026_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_3026_cast_fp16")];
            bool var_3029 = const()[name = string("op_3029"), val = bool(true)];
            tensor<int32, [1]> var_3030_axes_0 = const()[name = string("op_3030_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3030_cast_fp16 = reduce_max(axes = var_3030_axes_0, keep_dims = var_3029, x = var_3026_cast_fp16)[name = string("op_3030_cast_fp16")];
            fp16 var_3031_to_fp16 = const()[name = string("op_3031_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_45_cast_fp16 = maximum(x = var_3030_cast_fp16, y = var_3031_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_45_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3034_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [1]> var_3039_axes_0 = const()[name = string("op_3039_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3039_keep_dims_0 = const()[name = string("op_3039_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3039_cast_fp16 = reduce_mean(axes = var_3039_axes_0, keep_dims = var_3039_keep_dims_0, x = var_3034_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3040_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_3040_cast_fp16")];
            fp32 var_3041_epsilon_0 = const()[name = string("op_3041_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3041_cast_fp16 = inverse(epsilon = var_3041_epsilon_0, x = var_3040_cast_fp16)[name = string("op_3041_cast_fp16")];
            fp16 var_3042_to_fp16 = const()[name = string("op_3042_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3043_cast_fp16 = mul(x = var_3041_cast_fp16, y = var_3042_to_fp16)[name = string("op_3043_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3045_cast_fp16 = add(x = var_3039_cast_fp16, y = var_3043_cast_fp16)[name = string("op_3045_cast_fp16")];
            fp32 var_3046_epsilon_0 = const()[name = string("op_3046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3046_cast_fp16 = rsqrt(epsilon = var_3046_epsilon_0, x = var_3045_cast_fp16)[name = string("op_3046_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_3046_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18465664)))];
            tensor<fp16, [1, 256, 512]> input_29_cast_fp16 = mul(x = out_39_cast_fp16, y = tm_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18466752))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19515392))))[name = string("tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_23_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_23_cast_fp16")];
            string var_3053_mode_0 = const()[name = string("op_3053_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_3053_cast_fp16 = gelu(mode = var_3053_mode_0, x = linear_23_cast_fp16)[name = string("op_3053_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19519552))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20568192))))[name = string("tm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_24_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_31_cast_fp16 = mul(x = var_3053_cast_fp16, y = linear_24_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20572352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21620992))))[name = string("tm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3061_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_3061_cast_fp16")];
            bool var_3064 = const()[name = string("op_3064"), val = bool(true)];
            tensor<int32, [1]> var_3065_axes_0 = const()[name = string("op_3065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3065_cast_fp16 = reduce_max(axes = var_3065_axes_0, keep_dims = var_3064, x = var_3061_cast_fp16)[name = string("op_3065_cast_fp16")];
            fp16 var_3066_to_fp16 = const()[name = string("op_3066_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_47_cast_fp16 = maximum(x = var_3065_cast_fp16, y = var_3066_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_47_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_47_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3069_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<int32, [1]> var_3074_axes_0 = const()[name = string("op_3074_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3074_keep_dims_0 = const()[name = string("op_3074_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3074_cast_fp16 = reduce_mean(axes = var_3074_axes_0, keep_dims = var_3074_keep_dims_0, x = var_3069_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3075_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_3075_cast_fp16")];
            fp32 var_3076_epsilon_0 = const()[name = string("op_3076_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3076_cast_fp16 = inverse(epsilon = var_3076_epsilon_0, x = var_3075_cast_fp16)[name = string("op_3076_cast_fp16")];
            fp16 var_3077_to_fp16 = const()[name = string("op_3077_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3078_cast_fp16 = mul(x = var_3076_cast_fp16, y = var_3077_to_fp16)[name = string("op_3078_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3080_cast_fp16 = add(x = var_3074_cast_fp16, y = var_3078_cast_fp16)[name = string("op_3080_cast_fp16")];
            fp32 var_3081_epsilon_0 = const()[name = string("op_3081_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3081_cast_fp16 = rsqrt(epsilon = var_3081_epsilon_0, x = var_3080_cast_fp16)[name = string("op_3081_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_3081_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21622080)))];
            tensor<fp16, [1, 256, 512]> var_3083_cast_fp16 = mul(x = out_41_cast_fp16, y = tm_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3083_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_33_cast_fp16 = add(x = x_43_cast_fp16, y = var_3083_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21623168))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21885376))))[name = string("tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_26_cast_fp16")];
            string var_3090_mode_0 = const()[name = string("op_3090_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_3090_cast_fp16 = gelu(mode = var_3090_mode_0, x = linear_26_cast_fp16)[name = string("op_3090_cast_fp16")];
            tensor<int32, [4]> var_3103_begin_0 = const()[name = string("op_3103_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_3103_end_0 = const()[name = string("op_3103_end_0"), val = tensor<int32, [4]>([1, 256, 3, 512])];
            tensor<bool, [4]> var_3103_end_mask_0 = const()[name = string("op_3103_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3103_squeeze_mask_0 = const()[name = string("op_3103_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_3103_cast_fp16 = slice_by_index(begin = var_3103_begin_0, end = var_3103_end_0, end_mask = var_3103_end_mask_0, squeeze_mask = var_3103_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_3103_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_35_cast_fp16 = mul(x = var_3090_cast_fp16, y = var_3103_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21886464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22148672))))[name = string("tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_2_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3113_cast_fp16 = abs(x = linear_27_cast_fp16)[name = string("op_3113_cast_fp16")];
            bool var_3116 = const()[name = string("op_3116"), val = bool(true)];
            tensor<int32, [1]> var_3117_axes_0 = const()[name = string("op_3117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3117_cast_fp16 = reduce_max(axes = var_3117_axes_0, keep_dims = var_3116, x = var_3113_cast_fp16)[name = string("op_3117_cast_fp16")];
            fp16 var_3118_to_fp16 = const()[name = string("op_3118_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_49_cast_fp16 = maximum(x = var_3117_cast_fp16, y = var_3118_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_49_cast_fp16 = real_div(x = linear_27_cast_fp16, y = s_49_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3121_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_3121_cast_fp16")];
            tensor<int32, [1]> var_3126_axes_0 = const()[name = string("op_3126_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3126_keep_dims_0 = const()[name = string("op_3126_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3126_cast_fp16 = reduce_mean(axes = var_3126_axes_0, keep_dims = var_3126_keep_dims_0, x = var_3121_cast_fp16)[name = string("op_3126_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3127_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_3127_cast_fp16")];
            fp32 var_3128_epsilon_0 = const()[name = string("op_3128_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3128_cast_fp16 = inverse(epsilon = var_3128_epsilon_0, x = var_3127_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp16 var_3129_to_fp16 = const()[name = string("op_3129_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3130_cast_fp16 = mul(x = var_3128_cast_fp16, y = var_3129_to_fp16)[name = string("op_3130_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3132_cast_fp16 = add(x = var_3126_cast_fp16, y = var_3130_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_43_cast_fp16 = mul(x = y_49_cast_fp16, y = var_3133_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [512]> tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22149760)))];
            tensor<fp16, [1, 256, 512]> var_3135_cast_fp16 = mul(x = out_43_cast_fp16, y = tm_layers_2_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3135_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_5_cast_fp16 = add(x = input_33_cast_fp16, y = var_3135_cast_fp16)[name = string("h_5_cast_fp16")];
            tensor<fp16, [1]> tm_layers_2_layer_scalar_to_fp16 = const()[name = string("tm_layers_2_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ep-1])];
            tensor<fp16, [1, 256, 512]> x_49_cast_fp16 = mul(x = h_5_cast_fp16, y = tm_layers_2_layer_scalar_to_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3139_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_3139_cast_fp16")];
            bool var_3142 = const()[name = string("op_3142"), val = bool(true)];
            tensor<int32, [1]> var_3143_axes_0 = const()[name = string("op_3143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3143_cast_fp16 = reduce_max(axes = var_3143_axes_0, keep_dims = var_3142, x = var_3139_cast_fp16)[name = string("op_3143_cast_fp16")];
            fp16 var_3144_to_fp16 = const()[name = string("op_3144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_51_cast_fp16 = maximum(x = var_3143_cast_fp16, y = var_3144_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = s_51_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3147_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_3147_cast_fp16")];
            tensor<int32, [1]> var_3152_axes_0 = const()[name = string("op_3152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3152_keep_dims_0 = const()[name = string("op_3152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3152_cast_fp16 = reduce_mean(axes = var_3152_axes_0, keep_dims = var_3152_keep_dims_0, x = var_3147_cast_fp16)[name = string("op_3152_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3153_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_3153_cast_fp16")];
            fp32 var_3154_epsilon_0 = const()[name = string("op_3154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3154_cast_fp16 = inverse(epsilon = var_3154_epsilon_0, x = var_3153_cast_fp16)[name = string("op_3154_cast_fp16")];
            fp16 var_3155_to_fp16 = const()[name = string("op_3155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3156_cast_fp16 = mul(x = var_3154_cast_fp16, y = var_3155_to_fp16)[name = string("op_3156_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3158_cast_fp16 = add(x = var_3152_cast_fp16, y = var_3156_cast_fp16)[name = string("op_3158_cast_fp16")];
            fp32 var_3159_epsilon_0 = const()[name = string("op_3159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3159_cast_fp16 = rsqrt(epsilon = var_3159_epsilon_0, x = var_3158_cast_fp16)[name = string("op_3159_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_45_cast_fp16 = mul(x = y_51_cast_fp16, y = var_3159_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22150848)))];
            tensor<fp16, [1, 256, 512]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = tm_layers_3_input_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22151936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22676288))))[name = string("tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_3169 = const()[name = string("op_3169"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_3170_cast_fp16 = reshape(shape = var_3169, x = linear_28_cast_fp16)[name = string("op_3170_cast_fp16")];
            tensor<int32, [4]> x_51_perm_0 = const()[name = string("x_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22678400))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22940608))))[name = string("tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_3181 = const()[name = string("op_3181"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_3182_cast_fp16 = reshape(shape = var_3181, x = linear_29_cast_fp16)[name = string("op_3182_cast_fp16")];
            tensor<int32, [4]> x_53_perm_0 = const()[name = string("x_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22941696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23203904))))[name = string("tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_3193 = const()[name = string("op_3193"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_3194_cast_fp16 = reshape(shape = var_3193, x = linear_30_cast_fp16)[name = string("op_3194_cast_fp16")];
            tensor<int32, [4]> x_55_perm_0 = const()[name = string("x_55_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_51_cast_fp16 = transpose(perm = x_51_perm_0, x = var_3170_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 4, 256, 256]> var_3198_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_3198_cast_fp16")];
            bool var_3201 = const()[name = string("op_3201"), val = bool(true)];
            tensor<int32, [1]> var_3202_axes_0 = const()[name = string("op_3202_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_3202_cast_fp16 = reduce_max(axes = var_3202_axes_0, keep_dims = var_3201, x = var_3198_cast_fp16)[name = string("op_3202_cast_fp16")];
            fp16 var_3203_to_fp16 = const()[name = string("op_3203_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_53_cast_fp16 = maximum(x = var_3202_cast_fp16, y = var_3203_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = s_53_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3206_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<int32, [1]> var_3211_axes_0 = const()[name = string("op_3211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3211_keep_dims_0 = const()[name = string("op_3211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_3211_cast_fp16 = reduce_mean(axes = var_3211_axes_0, keep_dims = var_3211_keep_dims_0, x = var_3206_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3212_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_3212_cast_fp16")];
            fp32 var_3213_epsilon_0 = const()[name = string("op_3213_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_3213_cast_fp16 = inverse(epsilon = var_3213_epsilon_0, x = var_3212_cast_fp16)[name = string("op_3213_cast_fp16")];
            fp16 var_3214_to_fp16 = const()[name = string("op_3214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_3215_cast_fp16 = mul(x = var_3213_cast_fp16, y = var_3214_to_fp16)[name = string("op_3215_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3217_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3215_cast_fp16)[name = string("op_3217_cast_fp16")];
            fp32 var_3218_epsilon_0 = const()[name = string("op_3218_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_3218_cast_fp16 = rsqrt(epsilon = var_3218_epsilon_0, x = var_3217_cast_fp16)[name = string("op_3218_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_47_cast_fp16 = mul(x = y_53_cast_fp16, y = var_3218_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23204992)))];
            tensor<fp16, [1, 4, 256, 256]> q_13_cast_fp16 = mul(x = out_47_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_53_cast_fp16 = transpose(perm = x_53_perm_0, x = var_3182_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 2, 256, 256]> var_3221_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_3221_cast_fp16")];
            bool var_3224 = const()[name = string("op_3224"), val = bool(true)];
            tensor<int32, [1]> var_3225_axes_0 = const()[name = string("op_3225_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_3225_cast_fp16 = reduce_max(axes = var_3225_axes_0, keep_dims = var_3224, x = var_3221_cast_fp16)[name = string("op_3225_cast_fp16")];
            fp16 var_3226_to_fp16 = const()[name = string("op_3226_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_55_cast_fp16 = maximum(x = var_3225_cast_fp16, y = var_3226_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = s_55_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3229_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_3229_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3234_keep_dims_0 = const()[name = string("op_3234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_3234_cast_fp16 = reduce_mean(axes = var_3234_axes_0, keep_dims = var_3234_keep_dims_0, x = var_3229_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3235_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_3235_cast_fp16")];
            fp32 var_3236_epsilon_0 = const()[name = string("op_3236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_3236_cast_fp16 = inverse(epsilon = var_3236_epsilon_0, x = var_3235_cast_fp16)[name = string("op_3236_cast_fp16")];
            fp16 var_3237_to_fp16 = const()[name = string("op_3237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_3238_cast_fp16 = mul(x = var_3236_cast_fp16, y = var_3237_to_fp16)[name = string("op_3238_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3240_cast_fp16 = add(x = var_3234_cast_fp16, y = var_3238_cast_fp16)[name = string("op_3240_cast_fp16")];
            fp32 var_3241_epsilon_0 = const()[name = string("op_3241_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_3241_cast_fp16 = rsqrt(epsilon = var_3241_epsilon_0, x = var_3240_cast_fp16)[name = string("op_3241_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_49_cast_fp16 = mul(x = y_55_cast_fp16, y = var_3241_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [256]> tm_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23205568)))];
            tensor<fp16, [1, 2, 256, 256]> k_19_cast_fp16 = mul(x = out_49_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_19_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_55_cast_fp16 = transpose(perm = x_55_perm_0, x = var_3194_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 2, 256, 256]> var_3244_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_3244_cast_fp16")];
            bool var_3247 = const()[name = string("op_3247"), val = bool(true)];
            tensor<int32, [1]> var_3248_axes_0 = const()[name = string("op_3248_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_3248_cast_fp16 = reduce_max(axes = var_3248_axes_0, keep_dims = var_3247, x = var_3244_cast_fp16)[name = string("op_3248_cast_fp16")];
            fp16 var_3249_to_fp16 = const()[name = string("op_3249_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_57_cast_fp16 = maximum(x = var_3248_cast_fp16, y = var_3249_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_57_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3252_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_3252_cast_fp16")];
            tensor<int32, [1]> var_3257_axes_0 = const()[name = string("op_3257_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3257_keep_dims_0 = const()[name = string("op_3257_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_3257_cast_fp16 = reduce_mean(axes = var_3257_axes_0, keep_dims = var_3257_keep_dims_0, x = var_3252_cast_fp16)[name = string("op_3257_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3258_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_3258_cast_fp16")];
            fp32 var_3259_epsilon_0 = const()[name = string("op_3259_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_3259_cast_fp16 = inverse(epsilon = var_3259_epsilon_0, x = var_3258_cast_fp16)[name = string("op_3259_cast_fp16")];
            fp16 var_3260_to_fp16 = const()[name = string("op_3260_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_3261_cast_fp16 = mul(x = var_3259_cast_fp16, y = var_3260_to_fp16)[name = string("op_3261_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3263_cast_fp16 = add(x = var_3257_cast_fp16, y = var_3261_cast_fp16)[name = string("op_3263_cast_fp16")];
            fp32 var_3264_epsilon_0 = const()[name = string("op_3264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_3264_cast_fp16 = rsqrt(epsilon = var_3264_epsilon_0, x = var_3263_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_13_cast_fp16 = mul(x = y_57_cast_fp16, y = var_3264_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3266_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3266_cast_fp16")];
            tensor<int32, [4]> var_3271_begin_0 = const()[name = string("op_3271_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3271_end_0 = const()[name = string("op_3271_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_3271_end_mask_0 = const()[name = string("op_3271_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_3271_cast_fp16 = slice_by_index(begin = var_3271_begin_0, end = var_3271_end_0, end_mask = var_3271_end_mask_0, x = q_13_cast_fp16)[name = string("op_3271_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_3272_cast_fp16 = mul(x = var_3271_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_3272_cast_fp16")];
            tensor<int32, [4]> var_3277_begin_0 = const()[name = string("op_3277_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3277_end_0 = const()[name = string("op_3277_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_3277_end_mask_0 = const()[name = string("op_3277_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_3277_cast_fp16 = slice_by_index(begin = var_3277_begin_0, end = var_3277_end_0, end_mask = var_3277_end_mask_0, x = q_13_cast_fp16)[name = string("op_3277_cast_fp16")];
            int32 var_3279 = const()[name = string("op_3279"), val = int32(-1)];
            bool var_3280_interleave_0 = const()[name = string("op_3280_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_3280_cast_fp16 = concat(axis = var_3279, interleave = var_3280_interleave_0, values = (var_3272_cast_fp16, var_3277_cast_fp16))[name = string("op_3280_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3281_cast_fp16 = mul(x = var_3280_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3281_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_15_cast_fp16 = add(x = var_3266_cast_fp16, y = var_3281_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3284_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<int32, [4]> var_3289_begin_0 = const()[name = string("op_3289_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3289_end_0 = const()[name = string("op_3289_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_3289_end_mask_0 = const()[name = string("op_3289_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_3289_cast_fp16 = slice_by_index(begin = var_3289_begin_0, end = var_3289_end_0, end_mask = var_3289_end_mask_0, x = k_19_cast_fp16)[name = string("op_3289_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_3290_cast_fp16 = mul(x = var_3289_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_3290_cast_fp16")];
            tensor<int32, [4]> var_3295_begin_0 = const()[name = string("op_3295_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3295_end_0 = const()[name = string("op_3295_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_3295_end_mask_0 = const()[name = string("op_3295_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_3295_cast_fp16 = slice_by_index(begin = var_3295_begin_0, end = var_3295_end_0, end_mask = var_3295_end_mask_0, x = k_19_cast_fp16)[name = string("op_3295_cast_fp16")];
            int32 var_3297 = const()[name = string("op_3297"), val = int32(-1)];
            bool var_3298_interleave_0 = const()[name = string("op_3298_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_3298_cast_fp16 = concat(axis = var_3297, interleave = var_3298_interleave_0, values = (var_3290_cast_fp16, var_3295_cast_fp16))[name = string("op_3298_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3299_cast_fp16 = mul(x = var_3298_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3299_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_21_cast_fp16 = add(x = var_3284_cast_fp16, y = var_3299_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_6_reps_0 = const()[name = string("tile_6_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = k_21_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [4, 1, 256, 256]> tile_6_cast_fp16 = tile(reps = tile_6_reps_0, x = transpose_12_cast_fp16)[name = string("tile_6_cast_fp16")];
            tensor<int32, [5]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_12_cast_fp16 = reshape(shape = concat_12, x = tile_6_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [5]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_13 = const()[name = string("concat_13"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = reshape_12_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [4, 1, 256, 256]> reshape_13_cast_fp16 = reshape(shape = concat_13, x = transpose_13_cast_fp16)[name = string("reshape_13_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_7_reps_0 = const()[name = string("tile_7_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = v_13_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [4, 1, 256, 256]> tile_7_cast_fp16 = tile(reps = tile_7_reps_0, x = transpose_14_cast_fp16)[name = string("tile_7_cast_fp16")];
            tensor<int32, [5]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_14_cast_fp16 = reshape(shape = concat_14, x = tile_7_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [5]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [4, 1, 256, 256]> reshape_15_cast_fp16 = reshape(shape = concat_15, x = transpose_15_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3313_transpose_x_0 = const()[name = string("op_3313_transpose_x_0"), val = bool(false)];
            bool var_3313_transpose_y_0 = const()[name = string("op_3313_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = reshape_13_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 4, 256, 256]> var_3313_cast_fp16 = matmul(transpose_x = var_3313_transpose_x_0, transpose_y = var_3313_transpose_y_0, x = q_15_cast_fp16, y = transpose_99_cast_fp16)[name = string("op_3313_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3315_cast_fp16 = add(x = var_3313_cast_fp16, y = attention_bias)[name = string("op_3315_cast_fp16")];
            int32 var_3316 = const()[name = string("op_3316"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_3318_cast_fp16 = softmax(axis = var_3316, x = var_3315_cast_fp16)[name = string("op_3318_cast_fp16")];
            bool var_3319_transpose_x_0 = const()[name = string("op_3319_transpose_x_0"), val = bool(false)];
            bool var_3319_transpose_y_0 = const()[name = string("op_3319_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = reshape_15_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 4, 256, 256]> var_3319_cast_fp16 = matmul(transpose_x = var_3319_transpose_x_0, transpose_y = var_3319_transpose_y_0, x = var_3318_cast_fp16, y = v_15_cast_fp16)[name = string("op_3319_cast_fp16")];
            tensor<int32, [4]> var_3322_perm_0 = const()[name = string("op_3322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3326 = const()[name = string("op_3326"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_3322_cast_fp16 = transpose(perm = var_3322_perm_0, x = var_3319_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 256, 1024]> input_39_cast_fp16 = reshape(shape = var_3326, x = var_3322_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23206144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23730496))))[name = string("tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3331_cast_fp16 = abs(x = linear_31_cast_fp16)[name = string("op_3331_cast_fp16")];
            bool var_3334 = const()[name = string("op_3334"), val = bool(true)];
            tensor<int32, [1]> var_3335_axes_0 = const()[name = string("op_3335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3335_cast_fp16 = reduce_max(axes = var_3335_axes_0, keep_dims = var_3334, x = var_3331_cast_fp16)[name = string("op_3335_cast_fp16")];
            fp16 var_3336_to_fp16 = const()[name = string("op_3336_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_59_cast_fp16 = maximum(x = var_3335_cast_fp16, y = var_3336_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_59_cast_fp16 = real_div(x = linear_31_cast_fp16, y = s_59_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3339_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_3339_cast_fp16")];
            tensor<int32, [1]> var_3344_axes_0 = const()[name = string("op_3344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3344_keep_dims_0 = const()[name = string("op_3344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3344_cast_fp16 = reduce_mean(axes = var_3344_axes_0, keep_dims = var_3344_keep_dims_0, x = var_3339_cast_fp16)[name = string("op_3344_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3345_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_3345_cast_fp16")];
            fp32 var_3346_epsilon_0 = const()[name = string("op_3346_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3346_cast_fp16 = inverse(epsilon = var_3346_epsilon_0, x = var_3345_cast_fp16)[name = string("op_3346_cast_fp16")];
            fp16 var_3347_to_fp16 = const()[name = string("op_3347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3348_cast_fp16 = mul(x = var_3346_cast_fp16, y = var_3347_to_fp16)[name = string("op_3348_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3350_cast_fp16 = add(x = var_3344_cast_fp16, y = var_3348_cast_fp16)[name = string("op_3350_cast_fp16")];
            fp32 var_3351_epsilon_0 = const()[name = string("op_3351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3351_cast_fp16 = rsqrt(epsilon = var_3351_epsilon_0, x = var_3350_cast_fp16)[name = string("op_3351_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_3351_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23731584)))];
            tensor<fp16, [1, 256, 512]> var_3353_cast_fp16 = mul(x = out_51_cast_fp16, y = tm_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_3353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_59_cast_fp16 = add(x = x_49_cast_fp16, y = var_3353_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3356_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_3356_cast_fp16")];
            bool var_3359 = const()[name = string("op_3359"), val = bool(true)];
            tensor<int32, [1]> var_3360_axes_0 = const()[name = string("op_3360_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3360_cast_fp16 = reduce_max(axes = var_3360_axes_0, keep_dims = var_3359, x = var_3356_cast_fp16)[name = string("op_3360_cast_fp16")];
            fp16 var_3361_to_fp16 = const()[name = string("op_3361_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_61_cast_fp16 = maximum(x = var_3360_cast_fp16, y = var_3361_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_61_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3364_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [1]> var_3369_axes_0 = const()[name = string("op_3369_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3369_keep_dims_0 = const()[name = string("op_3369_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3369_cast_fp16 = reduce_mean(axes = var_3369_axes_0, keep_dims = var_3369_keep_dims_0, x = var_3364_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3370_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_3370_cast_fp16")];
            fp32 var_3371_epsilon_0 = const()[name = string("op_3371_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3371_cast_fp16 = inverse(epsilon = var_3371_epsilon_0, x = var_3370_cast_fp16)[name = string("op_3371_cast_fp16")];
            fp16 var_3372_to_fp16 = const()[name = string("op_3372_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3373_cast_fp16 = mul(x = var_3371_cast_fp16, y = var_3372_to_fp16)[name = string("op_3373_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3375_cast_fp16 = add(x = var_3369_cast_fp16, y = var_3373_cast_fp16)[name = string("op_3375_cast_fp16")];
            fp32 var_3376_epsilon_0 = const()[name = string("op_3376_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3376_cast_fp16 = rsqrt(epsilon = var_3376_epsilon_0, x = var_3375_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_3376_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23732672)))];
            tensor<fp16, [1, 256, 512]> input_41_cast_fp16 = mul(x = out_53_cast_fp16, y = tm_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23733760))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24782400))))[name = string("tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_32_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_32_cast_fp16")];
            string var_3383_mode_0 = const()[name = string("op_3383_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_3383_cast_fp16 = gelu(mode = var_3383_mode_0, x = linear_32_cast_fp16)[name = string("op_3383_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24786560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25835200))))[name = string("tm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_43_cast_fp16 = mul(x = var_3383_cast_fp16, y = linear_33_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25839360))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26888000))))[name = string("tm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3391_cast_fp16 = abs(x = linear_34_cast_fp16)[name = string("op_3391_cast_fp16")];
            bool var_3394 = const()[name = string("op_3394"), val = bool(true)];
            tensor<int32, [1]> var_3395_axes_0 = const()[name = string("op_3395_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3395_cast_fp16 = reduce_max(axes = var_3395_axes_0, keep_dims = var_3394, x = var_3391_cast_fp16)[name = string("op_3395_cast_fp16")];
            fp16 var_3396_to_fp16 = const()[name = string("op_3396_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_63_cast_fp16 = maximum(x = var_3395_cast_fp16, y = var_3396_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_63_cast_fp16 = real_div(x = linear_34_cast_fp16, y = s_63_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3399_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_3399_cast_fp16")];
            tensor<int32, [1]> var_3404_axes_0 = const()[name = string("op_3404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3404_keep_dims_0 = const()[name = string("op_3404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3404_cast_fp16 = reduce_mean(axes = var_3404_axes_0, keep_dims = var_3404_keep_dims_0, x = var_3399_cast_fp16)[name = string("op_3404_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3405_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_3405_cast_fp16")];
            fp32 var_3406_epsilon_0 = const()[name = string("op_3406_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3406_cast_fp16 = inverse(epsilon = var_3406_epsilon_0, x = var_3405_cast_fp16)[name = string("op_3406_cast_fp16")];
            fp16 var_3407_to_fp16 = const()[name = string("op_3407_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3408_cast_fp16 = mul(x = var_3406_cast_fp16, y = var_3407_to_fp16)[name = string("op_3408_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3410_cast_fp16 = add(x = var_3404_cast_fp16, y = var_3408_cast_fp16)[name = string("op_3410_cast_fp16")];
            fp32 var_3411_epsilon_0 = const()[name = string("op_3411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3411_cast_fp16 = rsqrt(epsilon = var_3411_epsilon_0, x = var_3410_cast_fp16)[name = string("op_3411_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_55_cast_fp16 = mul(x = y_63_cast_fp16, y = var_3411_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26889088)))];
            tensor<fp16, [1, 256, 512]> var_3413_cast_fp16 = mul(x = out_55_cast_fp16, y = tm_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_45_cast_fp16 = add(x = x_59_cast_fp16, y = var_3413_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26890176))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27152384))))[name = string("tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_35_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_35_cast_fp16")];
            string var_3420_mode_0 = const()[name = string("op_3420_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_3420_cast_fp16 = gelu(mode = var_3420_mode_0, x = linear_35_cast_fp16)[name = string("op_3420_cast_fp16")];
            tensor<int32, [4]> var_3433_begin_0 = const()[name = string("op_3433_begin_0"), val = tensor<int32, [4]>([0, 0, 3, 0])];
            tensor<int32, [4]> var_3433_end_0 = const()[name = string("op_3433_end_0"), val = tensor<int32, [4]>([1, 256, 4, 512])];
            tensor<bool, [4]> var_3433_end_mask_0 = const()[name = string("op_3433_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3433_squeeze_mask_0 = const()[name = string("op_3433_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_3433_cast_fp16 = slice_by_index(begin = var_3433_begin_0, end = var_3433_end_0, end_mask = var_3433_end_mask_0, squeeze_mask = var_3433_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_47_cast_fp16 = mul(x = var_3420_cast_fp16, y = var_3433_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27153472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27415680))))[name = string("tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_3_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3443_cast_fp16 = abs(x = linear_36_cast_fp16)[name = string("op_3443_cast_fp16")];
            bool var_3446 = const()[name = string("op_3446"), val = bool(true)];
            tensor<int32, [1]> var_3447_axes_0 = const()[name = string("op_3447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3447_cast_fp16 = reduce_max(axes = var_3447_axes_0, keep_dims = var_3446, x = var_3443_cast_fp16)[name = string("op_3447_cast_fp16")];
            fp16 var_3448_to_fp16 = const()[name = string("op_3448_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_65_cast_fp16 = maximum(x = var_3447_cast_fp16, y = var_3448_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_65_cast_fp16 = real_div(x = linear_36_cast_fp16, y = s_65_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3451_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3451_cast_fp16")];
            tensor<int32, [1]> var_3456_axes_0 = const()[name = string("op_3456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3456_keep_dims_0 = const()[name = string("op_3456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3456_cast_fp16 = reduce_mean(axes = var_3456_axes_0, keep_dims = var_3456_keep_dims_0, x = var_3451_cast_fp16)[name = string("op_3456_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3457_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp32 var_3458_epsilon_0 = const()[name = string("op_3458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3458_cast_fp16 = inverse(epsilon = var_3458_epsilon_0, x = var_3457_cast_fp16)[name = string("op_3458_cast_fp16")];
            fp16 var_3459_to_fp16 = const()[name = string("op_3459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3460_cast_fp16 = mul(x = var_3458_cast_fp16, y = var_3459_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3462_cast_fp16 = add(x = var_3456_cast_fp16, y = var_3460_cast_fp16)[name = string("op_3462_cast_fp16")];
            fp32 var_3463_epsilon_0 = const()[name = string("op_3463_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3463_cast_fp16 = rsqrt(epsilon = var_3463_epsilon_0, x = var_3462_cast_fp16)[name = string("op_3463_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_57_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3463_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [512]> tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27416768)))];
            tensor<fp16, [1, 256, 512]> var_3465_cast_fp16 = mul(x = out_57_cast_fp16, y = tm_layers_3_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3465_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_7_cast_fp16 = add(x = input_45_cast_fp16, y = var_3465_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<fp16, [1]> tm_layers_3_layer_scalar_to_fp16 = const()[name = string("tm_layers_3_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 256, 512]> x_65_cast_fp16 = mul(x = h_7_cast_fp16, y = tm_layers_3_layer_scalar_to_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3469_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_3469_cast_fp16")];
            bool var_3472 = const()[name = string("op_3472"), val = bool(true)];
            tensor<int32, [1]> var_3473_axes_0 = const()[name = string("op_3473_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3473_cast_fp16 = reduce_max(axes = var_3473_axes_0, keep_dims = var_3472, x = var_3469_cast_fp16)[name = string("op_3473_cast_fp16")];
            fp16 var_3474_to_fp16 = const()[name = string("op_3474_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_67_cast_fp16 = maximum(x = var_3473_cast_fp16, y = var_3474_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = s_67_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3477_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3477_cast_fp16")];
            tensor<int32, [1]> var_3482_axes_0 = const()[name = string("op_3482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3482_keep_dims_0 = const()[name = string("op_3482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3482_cast_fp16 = reduce_mean(axes = var_3482_axes_0, keep_dims = var_3482_keep_dims_0, x = var_3477_cast_fp16)[name = string("op_3482_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3483_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_3483_cast_fp16")];
            fp32 var_3484_epsilon_0 = const()[name = string("op_3484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3484_cast_fp16 = inverse(epsilon = var_3484_epsilon_0, x = var_3483_cast_fp16)[name = string("op_3484_cast_fp16")];
            fp16 var_3485_to_fp16 = const()[name = string("op_3485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3486_cast_fp16 = mul(x = var_3484_cast_fp16, y = var_3485_to_fp16)[name = string("op_3486_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3488_cast_fp16 = add(x = var_3482_cast_fp16, y = var_3486_cast_fp16)[name = string("op_3488_cast_fp16")];
            fp32 var_3489_epsilon_0 = const()[name = string("op_3489_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3489_cast_fp16 = rsqrt(epsilon = var_3489_epsilon_0, x = var_3488_cast_fp16)[name = string("op_3489_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_59_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3489_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27417856)))];
            tensor<fp16, [1, 256, 512]> input_49_cast_fp16 = mul(x = out_59_cast_fp16, y = tm_layers_4_input_layernorm_weight_to_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27418944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27943296))))[name = string("tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3499 = const()[name = string("op_3499"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_3500_cast_fp16 = reshape(shape = var_3499, x = linear_37_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<int32, [4]> x_67_perm_0 = const()[name = string("x_67_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27945408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28207616))))[name = string("tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3511 = const()[name = string("op_3511"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_3512_cast_fp16 = reshape(shape = var_3511, x = linear_38_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<int32, [4]> x_69_perm_0 = const()[name = string("x_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28208704))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28470912))))[name = string("tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_3523 = const()[name = string("op_3523"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_3524_cast_fp16 = reshape(shape = var_3523, x = linear_39_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<int32, [4]> x_71_perm_0 = const()[name = string("x_71_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_67_cast_fp16 = transpose(perm = x_67_perm_0, x = var_3500_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 4, 256, 256]> var_3528_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_3528_cast_fp16")];
            bool var_3531 = const()[name = string("op_3531"), val = bool(true)];
            tensor<int32, [1]> var_3532_axes_0 = const()[name = string("op_3532_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_3532_cast_fp16 = reduce_max(axes = var_3532_axes_0, keep_dims = var_3531, x = var_3528_cast_fp16)[name = string("op_3532_cast_fp16")];
            fp16 var_3533_to_fp16 = const()[name = string("op_3533_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_69_cast_fp16 = maximum(x = var_3532_cast_fp16, y = var_3533_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = s_69_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3536_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [1]> var_3541_axes_0 = const()[name = string("op_3541_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3541_keep_dims_0 = const()[name = string("op_3541_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_3541_cast_fp16 = reduce_mean(axes = var_3541_axes_0, keep_dims = var_3541_keep_dims_0, x = var_3536_cast_fp16)[name = string("op_3541_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3542_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3542_cast_fp16")];
            fp32 var_3543_epsilon_0 = const()[name = string("op_3543_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_3543_cast_fp16 = inverse(epsilon = var_3543_epsilon_0, x = var_3542_cast_fp16)[name = string("op_3543_cast_fp16")];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_3545_cast_fp16 = mul(x = var_3543_cast_fp16, y = var_3544_to_fp16)[name = string("op_3545_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3547_cast_fp16 = add(x = var_3541_cast_fp16, y = var_3545_cast_fp16)[name = string("op_3547_cast_fp16")];
            fp32 var_3548_epsilon_0 = const()[name = string("op_3548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_3548_cast_fp16 = rsqrt(epsilon = var_3548_epsilon_0, x = var_3547_cast_fp16)[name = string("op_3548_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_61_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3548_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [256]> tm_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472000)))];
            tensor<fp16, [1, 4, 256, 256]> q_17_cast_fp16 = mul(x = out_61_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_69_cast_fp16 = transpose(perm = x_69_perm_0, x = var_3512_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 2, 256, 256]> var_3551_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_3551_cast_fp16")];
            bool var_3554 = const()[name = string("op_3554"), val = bool(true)];
            tensor<int32, [1]> var_3555_axes_0 = const()[name = string("op_3555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_3555_cast_fp16 = reduce_max(axes = var_3555_axes_0, keep_dims = var_3554, x = var_3551_cast_fp16)[name = string("op_3555_cast_fp16")];
            fp16 var_3556_to_fp16 = const()[name = string("op_3556_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_71_cast_fp16 = maximum(x = var_3555_cast_fp16, y = var_3556_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = s_71_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3559_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3559_cast_fp16")];
            tensor<int32, [1]> var_3564_axes_0 = const()[name = string("op_3564_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3564_keep_dims_0 = const()[name = string("op_3564_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_3564_cast_fp16 = reduce_mean(axes = var_3564_axes_0, keep_dims = var_3564_keep_dims_0, x = var_3559_cast_fp16)[name = string("op_3564_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3565_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3565_cast_fp16")];
            fp32 var_3566_epsilon_0 = const()[name = string("op_3566_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_3566_cast_fp16 = inverse(epsilon = var_3566_epsilon_0, x = var_3565_cast_fp16)[name = string("op_3566_cast_fp16")];
            fp16 var_3567_to_fp16 = const()[name = string("op_3567_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_3568_cast_fp16 = mul(x = var_3566_cast_fp16, y = var_3567_to_fp16)[name = string("op_3568_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3570_cast_fp16 = add(x = var_3564_cast_fp16, y = var_3568_cast_fp16)[name = string("op_3570_cast_fp16")];
            fp32 var_3571_epsilon_0 = const()[name = string("op_3571_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_3571_cast_fp16 = rsqrt(epsilon = var_3571_epsilon_0, x = var_3570_cast_fp16)[name = string("op_3571_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_63_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3571_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28472576)))];
            tensor<fp16, [1, 2, 256, 256]> k_25_cast_fp16 = mul(x = out_63_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_71_cast_fp16 = transpose(perm = x_71_perm_0, x = var_3524_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 2, 256, 256]> var_3574_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_3574_cast_fp16")];
            bool var_3577 = const()[name = string("op_3577"), val = bool(true)];
            tensor<int32, [1]> var_3578_axes_0 = const()[name = string("op_3578_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_3578_cast_fp16 = reduce_max(axes = var_3578_axes_0, keep_dims = var_3577, x = var_3574_cast_fp16)[name = string("op_3578_cast_fp16")];
            fp16 var_3579_to_fp16 = const()[name = string("op_3579_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_73_cast_fp16 = maximum(x = var_3578_cast_fp16, y = var_3579_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = s_73_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3582_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3582_cast_fp16")];
            tensor<int32, [1]> var_3587_axes_0 = const()[name = string("op_3587_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3587_keep_dims_0 = const()[name = string("op_3587_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_3587_cast_fp16 = reduce_mean(axes = var_3587_axes_0, keep_dims = var_3587_keep_dims_0, x = var_3582_cast_fp16)[name = string("op_3587_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3588_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3588_cast_fp16")];
            fp32 var_3589_epsilon_0 = const()[name = string("op_3589_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_3589_cast_fp16 = inverse(epsilon = var_3589_epsilon_0, x = var_3588_cast_fp16)[name = string("op_3589_cast_fp16")];
            fp16 var_3590_to_fp16 = const()[name = string("op_3590_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_3591_cast_fp16 = mul(x = var_3589_cast_fp16, y = var_3590_to_fp16)[name = string("op_3591_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_3593_cast_fp16 = add(x = var_3587_cast_fp16, y = var_3591_cast_fp16)[name = string("op_3593_cast_fp16")];
            fp32 var_3594_epsilon_0 = const()[name = string("op_3594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_3594_cast_fp16 = rsqrt(epsilon = var_3594_epsilon_0, x = var_3593_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_17_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3594_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3596_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3596_cast_fp16")];
            tensor<int32, [4]> var_3601_begin_0 = const()[name = string("op_3601_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3601_end_0 = const()[name = string("op_3601_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_3601_end_mask_0 = const()[name = string("op_3601_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_3601_cast_fp16 = slice_by_index(begin = var_3601_begin_0, end = var_3601_end_0, end_mask = var_3601_end_mask_0, x = q_17_cast_fp16)[name = string("op_3601_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_3602_cast_fp16 = mul(x = var_3601_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_3602_cast_fp16")];
            tensor<int32, [4]> var_3607_begin_0 = const()[name = string("op_3607_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3607_end_0 = const()[name = string("op_3607_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_3607_end_mask_0 = const()[name = string("op_3607_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_3607_cast_fp16 = slice_by_index(begin = var_3607_begin_0, end = var_3607_end_0, end_mask = var_3607_end_mask_0, x = q_17_cast_fp16)[name = string("op_3607_cast_fp16")];
            int32 var_3609 = const()[name = string("op_3609"), val = int32(-1)];
            bool var_3610_interleave_0 = const()[name = string("op_3610_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_3610_cast_fp16 = concat(axis = var_3609, interleave = var_3610_interleave_0, values = (var_3602_cast_fp16, var_3607_cast_fp16))[name = string("op_3610_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3611_cast_fp16 = mul(x = var_3610_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3611_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_19_cast_fp16 = add(x = var_3596_cast_fp16, y = var_3611_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3614_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3614_cast_fp16")];
            tensor<int32, [4]> var_3619_begin_0 = const()[name = string("op_3619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_3619_end_0 = const()[name = string("op_3619_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_3619_end_mask_0 = const()[name = string("op_3619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_3619_cast_fp16 = slice_by_index(begin = var_3619_begin_0, end = var_3619_end_0, end_mask = var_3619_end_mask_0, x = k_25_cast_fp16)[name = string("op_3619_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_3620_cast_fp16 = mul(x = var_3619_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_3620_cast_fp16")];
            tensor<int32, [4]> var_3625_begin_0 = const()[name = string("op_3625_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3625_end_0 = const()[name = string("op_3625_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_3625_end_mask_0 = const()[name = string("op_3625_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_3625_cast_fp16 = slice_by_index(begin = var_3625_begin_0, end = var_3625_end_0, end_mask = var_3625_end_mask_0, x = k_25_cast_fp16)[name = string("op_3625_cast_fp16")];
            int32 var_3627 = const()[name = string("op_3627"), val = int32(-1)];
            bool var_3628_interleave_0 = const()[name = string("op_3628_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_3628_cast_fp16 = concat(axis = var_3627, interleave = var_3628_interleave_0, values = (var_3620_cast_fp16, var_3625_cast_fp16))[name = string("op_3628_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_3629_cast_fp16 = mul(x = var_3628_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3629_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_27_cast_fp16 = add(x = var_3614_cast_fp16, y = var_3629_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_8_reps_0 = const()[name = string("tile_8_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = k_27_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [4, 1, 256, 256]> tile_8_cast_fp16 = tile(reps = tile_8_reps_0, x = transpose_16_cast_fp16)[name = string("tile_8_cast_fp16")];
            tensor<int32, [5]> concat_16 = const()[name = string("concat_16"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_16_cast_fp16 = reshape(shape = concat_16, x = tile_8_cast_fp16)[name = string("reshape_16_cast_fp16")];
            tensor<int32, [5]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_17 = const()[name = string("concat_17"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = reshape_16_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [4, 1, 256, 256]> reshape_17_cast_fp16 = reshape(shape = concat_17, x = transpose_17_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_9_reps_0 = const()[name = string("tile_9_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = v_17_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [4, 1, 256, 256]> tile_9_cast_fp16 = tile(reps = tile_9_reps_0, x = transpose_18_cast_fp16)[name = string("tile_9_cast_fp16")];
            tensor<int32, [5]> concat_18 = const()[name = string("concat_18"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_18_cast_fp16 = reshape(shape = concat_18, x = tile_9_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [5]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = reshape_18_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [4, 1, 256, 256]> reshape_19_cast_fp16 = reshape(shape = concat_19, x = transpose_19_cast_fp16)[name = string("reshape_19_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3643_transpose_x_0 = const()[name = string("op_3643_transpose_x_0"), val = bool(false)];
            bool var_3643_transpose_y_0 = const()[name = string("op_3643_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 4, 256, 256]> var_3643_cast_fp16 = matmul(transpose_x = var_3643_transpose_x_0, transpose_y = var_3643_transpose_y_0, x = q_19_cast_fp16, y = transpose_100_cast_fp16)[name = string("op_3643_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3645_cast_fp16 = add(x = var_3643_cast_fp16, y = attention_bias)[name = string("op_3645_cast_fp16")];
            int32 var_3646 = const()[name = string("op_3646"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_3648_cast_fp16 = softmax(axis = var_3646, x = var_3645_cast_fp16)[name = string("op_3648_cast_fp16")];
            bool var_3649_transpose_x_0 = const()[name = string("op_3649_transpose_x_0"), val = bool(false)];
            bool var_3649_transpose_y_0 = const()[name = string("op_3649_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = reshape_19_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 4, 256, 256]> var_3649_cast_fp16 = matmul(transpose_x = var_3649_transpose_x_0, transpose_y = var_3649_transpose_y_0, x = var_3648_cast_fp16, y = v_19_cast_fp16)[name = string("op_3649_cast_fp16")];
            tensor<int32, [4]> var_3652_perm_0 = const()[name = string("op_3652_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3656 = const()[name = string("op_3656"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_3652_cast_fp16 = transpose(perm = var_3652_perm_0, x = var_3649_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 256, 1024]> input_51_cast_fp16 = reshape(shape = var_3656, x = var_3652_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28473152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28997504))))[name = string("tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3661_cast_fp16 = abs(x = linear_40_cast_fp16)[name = string("op_3661_cast_fp16")];
            bool var_3664 = const()[name = string("op_3664"), val = bool(true)];
            tensor<int32, [1]> var_3665_axes_0 = const()[name = string("op_3665_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3665_cast_fp16 = reduce_max(axes = var_3665_axes_0, keep_dims = var_3664, x = var_3661_cast_fp16)[name = string("op_3665_cast_fp16")];
            fp16 var_3666_to_fp16 = const()[name = string("op_3666_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_75_cast_fp16 = maximum(x = var_3665_cast_fp16, y = var_3666_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_75_cast_fp16 = real_div(x = linear_40_cast_fp16, y = s_75_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3669_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3669_cast_fp16")];
            tensor<int32, [1]> var_3674_axes_0 = const()[name = string("op_3674_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3674_keep_dims_0 = const()[name = string("op_3674_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3674_cast_fp16 = reduce_mean(axes = var_3674_axes_0, keep_dims = var_3674_keep_dims_0, x = var_3669_cast_fp16)[name = string("op_3674_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3675_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3675_cast_fp16")];
            fp32 var_3676_epsilon_0 = const()[name = string("op_3676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3676_cast_fp16 = inverse(epsilon = var_3676_epsilon_0, x = var_3675_cast_fp16)[name = string("op_3676_cast_fp16")];
            fp16 var_3677_to_fp16 = const()[name = string("op_3677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3678_cast_fp16 = mul(x = var_3676_cast_fp16, y = var_3677_to_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3680_cast_fp16 = add(x = var_3674_cast_fp16, y = var_3678_cast_fp16)[name = string("op_3680_cast_fp16")];
            fp32 var_3681_epsilon_0 = const()[name = string("op_3681_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3681_cast_fp16 = rsqrt(epsilon = var_3681_epsilon_0, x = var_3680_cast_fp16)[name = string("op_3681_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3681_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28998592)))];
            tensor<fp16, [1, 256, 512]> var_3683_cast_fp16 = mul(x = out_65_cast_fp16, y = tm_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3683_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_75_cast_fp16 = add(x = x_65_cast_fp16, y = var_3683_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3686_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3686_cast_fp16")];
            bool var_3689 = const()[name = string("op_3689"), val = bool(true)];
            tensor<int32, [1]> var_3690_axes_0 = const()[name = string("op_3690_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3690_cast_fp16 = reduce_max(axes = var_3690_axes_0, keep_dims = var_3689, x = var_3686_cast_fp16)[name = string("op_3690_cast_fp16")];
            fp16 var_3691_to_fp16 = const()[name = string("op_3691_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_77_cast_fp16 = maximum(x = var_3690_cast_fp16, y = var_3691_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_77_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3694_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3694_cast_fp16")];
            tensor<int32, [1]> var_3699_axes_0 = const()[name = string("op_3699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3699_keep_dims_0 = const()[name = string("op_3699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3699_cast_fp16 = reduce_mean(axes = var_3699_axes_0, keep_dims = var_3699_keep_dims_0, x = var_3694_cast_fp16)[name = string("op_3699_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3700_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3700_cast_fp16")];
            fp32 var_3701_epsilon_0 = const()[name = string("op_3701_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3701_cast_fp16 = inverse(epsilon = var_3701_epsilon_0, x = var_3700_cast_fp16)[name = string("op_3701_cast_fp16")];
            fp16 var_3702_to_fp16 = const()[name = string("op_3702_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3703_cast_fp16 = mul(x = var_3701_cast_fp16, y = var_3702_to_fp16)[name = string("op_3703_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3705_cast_fp16 = add(x = var_3699_cast_fp16, y = var_3703_cast_fp16)[name = string("op_3705_cast_fp16")];
            fp32 var_3706_epsilon_0 = const()[name = string("op_3706_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3706_cast_fp16 = rsqrt(epsilon = var_3706_epsilon_0, x = var_3705_cast_fp16)[name = string("op_3706_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_67_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3706_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28999680)))];
            tensor<fp16, [1, 256, 512]> input_53_cast_fp16 = mul(x = out_67_cast_fp16, y = tm_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29000768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30049408))))[name = string("tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_41_cast_fp16")];
            string var_3713_mode_0 = const()[name = string("op_3713_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_3713_cast_fp16 = gelu(mode = var_3713_mode_0, x = linear_41_cast_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30053568))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31102208))))[name = string("tm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_42_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_55_cast_fp16 = mul(x = var_3713_cast_fp16, y = linear_42_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31106368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32155008))))[name = string("tm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3721_cast_fp16 = abs(x = linear_43_cast_fp16)[name = string("op_3721_cast_fp16")];
            bool var_3724 = const()[name = string("op_3724"), val = bool(true)];
            tensor<int32, [1]> var_3725_axes_0 = const()[name = string("op_3725_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3725_cast_fp16 = reduce_max(axes = var_3725_axes_0, keep_dims = var_3724, x = var_3721_cast_fp16)[name = string("op_3725_cast_fp16")];
            fp16 var_3726_to_fp16 = const()[name = string("op_3726_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_79_cast_fp16 = maximum(x = var_3725_cast_fp16, y = var_3726_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_79_cast_fp16 = real_div(x = linear_43_cast_fp16, y = s_79_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3729_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_3734_axes_0 = const()[name = string("op_3734_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3734_keep_dims_0 = const()[name = string("op_3734_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3734_cast_fp16 = reduce_mean(axes = var_3734_axes_0, keep_dims = var_3734_keep_dims_0, x = var_3729_cast_fp16)[name = string("op_3734_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3735_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3735_cast_fp16")];
            fp32 var_3736_epsilon_0 = const()[name = string("op_3736_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3736_cast_fp16 = inverse(epsilon = var_3736_epsilon_0, x = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp16 var_3737_to_fp16 = const()[name = string("op_3737_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3738_cast_fp16 = mul(x = var_3736_cast_fp16, y = var_3737_to_fp16)[name = string("op_3738_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3740_cast_fp16 = add(x = var_3734_cast_fp16, y = var_3738_cast_fp16)[name = string("op_3740_cast_fp16")];
            fp32 var_3741_epsilon_0 = const()[name = string("op_3741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3741_cast_fp16 = rsqrt(epsilon = var_3741_epsilon_0, x = var_3740_cast_fp16)[name = string("op_3741_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_69_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3741_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32156096)))];
            tensor<fp16, [1, 256, 512]> var_3743_cast_fp16 = mul(x = out_69_cast_fp16, y = tm_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3743_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_57_cast_fp16 = add(x = x_75_cast_fp16, y = var_3743_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32157184))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32419392))))[name = string("tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_44_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_44_cast_fp16")];
            string var_3750_mode_0 = const()[name = string("op_3750_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_3750_cast_fp16 = gelu(mode = var_3750_mode_0, x = linear_44_cast_fp16)[name = string("op_3750_cast_fp16")];
            tensor<int32, [4]> var_3763_begin_0 = const()[name = string("op_3763_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_3763_end_0 = const()[name = string("op_3763_end_0"), val = tensor<int32, [4]>([1, 256, 5, 512])];
            tensor<bool, [4]> var_3763_end_mask_0 = const()[name = string("op_3763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_3763_squeeze_mask_0 = const()[name = string("op_3763_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_3763_cast_fp16 = slice_by_index(begin = var_3763_begin_0, end = var_3763_end_0, end_mask = var_3763_end_mask_0, squeeze_mask = var_3763_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_3763_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_59_cast_fp16 = mul(x = var_3750_cast_fp16, y = var_3763_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32420480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32682688))))[name = string("tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_4_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3773_cast_fp16 = abs(x = linear_45_cast_fp16)[name = string("op_3773_cast_fp16")];
            bool var_3776 = const()[name = string("op_3776"), val = bool(true)];
            tensor<int32, [1]> var_3777_axes_0 = const()[name = string("op_3777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3777_cast_fp16 = reduce_max(axes = var_3777_axes_0, keep_dims = var_3776, x = var_3773_cast_fp16)[name = string("op_3777_cast_fp16")];
            fp16 var_3778_to_fp16 = const()[name = string("op_3778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_81_cast_fp16 = maximum(x = var_3777_cast_fp16, y = var_3778_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_81_cast_fp16 = real_div(x = linear_45_cast_fp16, y = s_81_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3781_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3781_cast_fp16")];
            tensor<int32, [1]> var_3786_axes_0 = const()[name = string("op_3786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3786_keep_dims_0 = const()[name = string("op_3786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3786_cast_fp16 = reduce_mean(axes = var_3786_axes_0, keep_dims = var_3786_keep_dims_0, x = var_3781_cast_fp16)[name = string("op_3786_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3787_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3787_cast_fp16")];
            fp32 var_3788_epsilon_0 = const()[name = string("op_3788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3788_cast_fp16 = inverse(epsilon = var_3788_epsilon_0, x = var_3787_cast_fp16)[name = string("op_3788_cast_fp16")];
            fp16 var_3789_to_fp16 = const()[name = string("op_3789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3790_cast_fp16 = mul(x = var_3788_cast_fp16, y = var_3789_to_fp16)[name = string("op_3790_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3792_cast_fp16 = add(x = var_3786_cast_fp16, y = var_3790_cast_fp16)[name = string("op_3792_cast_fp16")];
            fp32 var_3793_epsilon_0 = const()[name = string("op_3793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3793_cast_fp16 = rsqrt(epsilon = var_3793_epsilon_0, x = var_3792_cast_fp16)[name = string("op_3793_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_71_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3793_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [512]> tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32683776)))];
            tensor<fp16, [1, 256, 512]> var_3795_cast_fp16 = mul(x = out_71_cast_fp16, y = tm_layers_4_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_3795_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_9_cast_fp16 = add(x = input_57_cast_fp16, y = var_3795_cast_fp16)[name = string("h_9_cast_fp16")];
            tensor<fp16, [1]> tm_layers_4_layer_scalar_to_fp16 = const()[name = string("tm_layers_4_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.62p-1])];
            tensor<fp16, [1, 256, 512]> x_81_cast_fp16 = mul(x = h_9_cast_fp16, y = tm_layers_4_layer_scalar_to_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3799_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3799_cast_fp16")];
            bool var_3802 = const()[name = string("op_3802"), val = bool(true)];
            tensor<int32, [1]> var_3803_axes_0 = const()[name = string("op_3803_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3803_cast_fp16 = reduce_max(axes = var_3803_axes_0, keep_dims = var_3802, x = var_3799_cast_fp16)[name = string("op_3803_cast_fp16")];
            fp16 var_3804_to_fp16 = const()[name = string("op_3804_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_83_cast_fp16 = maximum(x = var_3803_cast_fp16, y = var_3804_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = s_83_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3807_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3807_cast_fp16")];
            tensor<int32, [1]> var_3812_axes_0 = const()[name = string("op_3812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3812_keep_dims_0 = const()[name = string("op_3812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3812_cast_fp16 = reduce_mean(axes = var_3812_axes_0, keep_dims = var_3812_keep_dims_0, x = var_3807_cast_fp16)[name = string("op_3812_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3813_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3813_cast_fp16")];
            fp32 var_3814_epsilon_0 = const()[name = string("op_3814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3814_cast_fp16 = inverse(epsilon = var_3814_epsilon_0, x = var_3813_cast_fp16)[name = string("op_3814_cast_fp16")];
            fp16 var_3815_to_fp16 = const()[name = string("op_3815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3816_cast_fp16 = mul(x = var_3814_cast_fp16, y = var_3815_to_fp16)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3818_cast_fp16 = add(x = var_3812_cast_fp16, y = var_3816_cast_fp16)[name = string("op_3818_cast_fp16")];
            fp32 var_3819_epsilon_0 = const()[name = string("op_3819_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3819_cast_fp16 = rsqrt(epsilon = var_3819_epsilon_0, x = var_3818_cast_fp16)[name = string("op_3819_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_73_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3819_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32684864)))];
            tensor<fp16, [1, 256, 512]> input_61_cast_fp16 = mul(x = out_73_cast_fp16, y = tm_layers_5_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32685952))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33734592))))[name = string("tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_46_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<int32, [4]> var_3829 = const()[name = string("op_3829"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 256, 4, 512]> var_3830_cast_fp16 = reshape(shape = var_3829, x = linear_46_cast_fp16)[name = string("op_3830_cast_fp16")];
            tensor<int32, [4]> x_83_perm_0 = const()[name = string("x_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33738752))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34000960))))[name = string("tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<int32, [4]> var_3841 = const()[name = string("op_3841"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_3842_cast_fp16 = reshape(shape = var_3841, x = linear_47_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<int32, [4]> x_85_perm_0 = const()[name = string("x_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34002048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34264256))))[name = string("tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<int32, [4]> var_3853 = const()[name = string("op_3853"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_3854_cast_fp16 = reshape(shape = var_3853, x = linear_48_cast_fp16)[name = string("op_3854_cast_fp16")];
            tensor<int32, [4]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 512]> x_83_cast_fp16 = transpose(perm = x_83_perm_0, x = var_3830_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 4, 256, 512]> var_3858_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3858_cast_fp16")];
            bool var_3861 = const()[name = string("op_3861"), val = bool(true)];
            tensor<int32, [1]> var_3862_axes_0 = const()[name = string("op_3862_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_3862_cast_fp16 = reduce_max(axes = var_3862_axes_0, keep_dims = var_3861, x = var_3858_cast_fp16)[name = string("op_3862_cast_fp16")];
            fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_85_cast_fp16 = maximum(x = var_3862_cast_fp16, y = var_3863_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> y_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = s_85_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_3866_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3866_cast_fp16")];
            tensor<int32, [1]> var_3871_axes_0 = const()[name = string("op_3871_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3871_keep_dims_0 = const()[name = string("op_3871_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_3871_cast_fp16 = reduce_mean(axes = var_3871_axes_0, keep_dims = var_3871_keep_dims_0, x = var_3866_cast_fp16)[name = string("op_3871_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3872_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3872_cast_fp16")];
            fp32 var_3873_epsilon_0 = const()[name = string("op_3873_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_3873_cast_fp16 = inverse(epsilon = var_3873_epsilon_0, x = var_3872_cast_fp16)[name = string("op_3873_cast_fp16")];
            fp16 var_3874_to_fp16 = const()[name = string("op_3874_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_3875_cast_fp16 = mul(x = var_3873_cast_fp16, y = var_3874_to_fp16)[name = string("op_3875_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_3877_cast_fp16 = add(x = var_3871_cast_fp16, y = var_3875_cast_fp16)[name = string("op_3877_cast_fp16")];
            fp32 var_3878_epsilon_0 = const()[name = string("op_3878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_3878_cast_fp16 = rsqrt(epsilon = var_3878_epsilon_0, x = var_3877_cast_fp16)[name = string("op_3878_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> out_75_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3878_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34265344)))];
            tensor<fp16, [1, 4, 256, 512]> q_21_cast_fp16 = mul(x = out_75_cast_fp16, y = tm_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("q_21_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_85_cast_fp16 = transpose(perm = x_85_perm_0, x = var_3842_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1, 256, 512]> var_3881_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3881_cast_fp16")];
            bool var_3884 = const()[name = string("op_3884"), val = bool(true)];
            tensor<int32, [1]> var_3885_axes_0 = const()[name = string("op_3885_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_3885_cast_fp16 = reduce_max(axes = var_3885_axes_0, keep_dims = var_3884, x = var_3881_cast_fp16)[name = string("op_3885_cast_fp16")];
            fp16 var_3886_to_fp16 = const()[name = string("op_3886_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_87_cast_fp16 = maximum(x = var_3885_cast_fp16, y = var_3886_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = s_87_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_3889_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3889_cast_fp16")];
            tensor<int32, [1]> var_3894_axes_0 = const()[name = string("op_3894_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3894_keep_dims_0 = const()[name = string("op_3894_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_3894_cast_fp16 = reduce_mean(axes = var_3894_axes_0, keep_dims = var_3894_keep_dims_0, x = var_3889_cast_fp16)[name = string("op_3894_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_3895_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3895_cast_fp16")];
            fp32 var_3896_epsilon_0 = const()[name = string("op_3896_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_3896_cast_fp16 = inverse(epsilon = var_3896_epsilon_0, x = var_3895_cast_fp16)[name = string("op_3896_cast_fp16")];
            fp16 var_3897_to_fp16 = const()[name = string("op_3897_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_3898_cast_fp16 = mul(x = var_3896_cast_fp16, y = var_3897_to_fp16)[name = string("op_3898_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_3900_cast_fp16 = add(x = var_3894_cast_fp16, y = var_3898_cast_fp16)[name = string("op_3900_cast_fp16")];
            fp32 var_3901_epsilon_0 = const()[name = string("op_3901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_3901_cast_fp16 = rsqrt(epsilon = var_3901_epsilon_0, x = var_3900_cast_fp16)[name = string("op_3901_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> out_77_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3901_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34266432)))];
            tensor<fp16, [1, 1, 256, 512]> k_31_cast_fp16 = mul(x = out_77_cast_fp16, y = tm_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("k_31_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_3854_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 1, 256, 512]> var_3904_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3904_cast_fp16")];
            bool var_3907 = const()[name = string("op_3907"), val = bool(true)];
            tensor<int32, [1]> var_3908_axes_0 = const()[name = string("op_3908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_3908_cast_fp16 = reduce_max(axes = var_3908_axes_0, keep_dims = var_3907, x = var_3904_cast_fp16)[name = string("op_3908_cast_fp16")];
            fp16 var_3909_to_fp16 = const()[name = string("op_3909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_89_cast_fp16 = maximum(x = var_3908_cast_fp16, y = var_3909_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = s_89_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_3912_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3912_cast_fp16")];
            tensor<int32, [1]> var_3917_axes_0 = const()[name = string("op_3917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3917_keep_dims_0 = const()[name = string("op_3917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_3917_cast_fp16 = reduce_mean(axes = var_3917_axes_0, keep_dims = var_3917_keep_dims_0, x = var_3912_cast_fp16)[name = string("op_3917_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_3918_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3918_cast_fp16")];
            fp32 var_3919_epsilon_0 = const()[name = string("op_3919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_3919_cast_fp16 = inverse(epsilon = var_3919_epsilon_0, x = var_3918_cast_fp16)[name = string("op_3919_cast_fp16")];
            fp16 var_3920_to_fp16 = const()[name = string("op_3920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_3921_cast_fp16 = mul(x = var_3919_cast_fp16, y = var_3920_to_fp16)[name = string("op_3921_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_3923_cast_fp16 = add(x = var_3917_cast_fp16, y = var_3921_cast_fp16)[name = string("op_3923_cast_fp16")];
            fp32 var_3924_epsilon_0 = const()[name = string("op_3924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_3924_cast_fp16 = rsqrt(epsilon = var_3924_epsilon_0, x = var_3923_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> v_21_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3924_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_3926_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_cast_fp16)[name = string("op_3926_cast_fp16")];
            tensor<int32, [4]> var_3931_begin_0 = const()[name = string("op_3931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3931_end_0 = const()[name = string("op_3931_end_0"), val = tensor<int32, [4]>([1, 4, 256, 512])];
            tensor<bool, [4]> var_3931_end_mask_0 = const()[name = string("op_3931_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 256]> var_3931_cast_fp16 = slice_by_index(begin = var_3931_begin_0, end = var_3931_end_0, end_mask = var_3931_end_mask_0, x = q_21_cast_fp16)[name = string("op_3931_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 256]> var_3932_cast_fp16 = mul(x = var_3931_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_3932_cast_fp16")];
            tensor<int32, [4]> var_3937_begin_0 = const()[name = string("op_3937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3937_end_0 = const()[name = string("op_3937_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_3937_end_mask_0 = const()[name = string("op_3937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 256]> var_3937_cast_fp16 = slice_by_index(begin = var_3937_begin_0, end = var_3937_end_0, end_mask = var_3937_end_mask_0, x = q_21_cast_fp16)[name = string("op_3937_cast_fp16")];
            int32 var_3939 = const()[name = string("op_3939"), val = int32(-1)];
            bool var_3940_interleave_0 = const()[name = string("op_3940_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> var_3940_cast_fp16 = concat(axis = var_3939, interleave = var_3940_interleave_0, values = (var_3932_cast_fp16, var_3937_cast_fp16))[name = string("op_3940_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_3941_cast_fp16 = mul(x = var_3940_cast_fp16, y = sin_cast_fp16)[name = string("op_3941_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> q_23_cast_fp16 = add(x = var_3926_cast_fp16, y = var_3941_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_3944_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_cast_fp16)[name = string("op_3944_cast_fp16")];
            tensor<int32, [4]> var_3949_begin_0 = const()[name = string("op_3949_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_3949_end_0 = const()[name = string("op_3949_end_0"), val = tensor<int32, [4]>([1, 1, 256, 512])];
            tensor<bool, [4]> var_3949_end_mask_0 = const()[name = string("op_3949_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 256, 256]> var_3949_cast_fp16 = slice_by_index(begin = var_3949_begin_0, end = var_3949_end_0, end_mask = var_3949_end_mask_0, x = k_31_cast_fp16)[name = string("op_3949_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 256, 256]> var_3950_cast_fp16 = mul(x = var_3949_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_3950_cast_fp16")];
            tensor<int32, [4]> var_3955_begin_0 = const()[name = string("op_3955_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3955_end_0 = const()[name = string("op_3955_end_0"), val = tensor<int32, [4]>([1, 1, 256, 256])];
            tensor<bool, [4]> var_3955_end_mask_0 = const()[name = string("op_3955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 256, 256]> var_3955_cast_fp16 = slice_by_index(begin = var_3955_begin_0, end = var_3955_end_0, end_mask = var_3955_end_mask_0, x = k_31_cast_fp16)[name = string("op_3955_cast_fp16")];
            int32 var_3957 = const()[name = string("op_3957"), val = int32(-1)];
            bool var_3958_interleave_0 = const()[name = string("op_3958_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 256, 512]> var_3958_cast_fp16 = concat(axis = var_3957, interleave = var_3958_interleave_0, values = (var_3950_cast_fp16, var_3955_cast_fp16))[name = string("op_3958_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_3959_cast_fp16 = mul(x = var_3958_cast_fp16, y = sin_cast_fp16)[name = string("op_3959_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> k_33_cast_fp16 = add(x = var_3944_cast_fp16, y = var_3959_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_10_reps_0 = const()[name = string("tile_10_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = k_33_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [4, 1, 256, 512]> tile_10_cast_fp16 = tile(reps = tile_10_reps_0, x = transpose_20_cast_fp16)[name = string("tile_10_cast_fp16")];
            tensor<int32, [5]> concat_20 = const()[name = string("concat_20"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_20_cast_fp16 = reshape(shape = concat_20, x = tile_10_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [5]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_21 = const()[name = string("concat_21"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [4, 1, 256, 512]> reshape_21_cast_fp16 = reshape(shape = concat_21, x = transpose_21_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_11_reps_0 = const()[name = string("tile_11_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = v_21_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [4, 1, 256, 512]> tile_11_cast_fp16 = tile(reps = tile_11_reps_0, x = transpose_22_cast_fp16)[name = string("tile_11_cast_fp16")];
            tensor<int32, [5]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_22_cast_fp16 = reshape(shape = concat_22, x = tile_11_cast_fp16)[name = string("reshape_22_cast_fp16")];
            tensor<int32, [5]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_23 = const()[name = string("concat_23"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = reshape_22_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [4, 1, 256, 512]> reshape_23_cast_fp16 = reshape(shape = concat_23, x = transpose_23_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_3973_transpose_x_0 = const()[name = string("op_3973_transpose_x_0"), val = bool(false)];
            bool var_3973_transpose_y_0 = const()[name = string("op_3973_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = reshape_21_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 4, 256, 256]> var_3973_cast_fp16 = matmul(transpose_x = var_3973_transpose_x_0, transpose_y = var_3973_transpose_y_0, x = q_23_cast_fp16, y = transpose_101_cast_fp16)[name = string("op_3973_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_3975_cast_fp16 = add(x = var_3973_cast_fp16, y = attention_bias)[name = string("op_3975_cast_fp16")];
            int32 var_3976 = const()[name = string("op_3976"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_3978_cast_fp16 = softmax(axis = var_3976, x = var_3975_cast_fp16)[name = string("op_3978_cast_fp16")];
            bool var_3979_transpose_x_0 = const()[name = string("op_3979_transpose_x_0"), val = bool(false)];
            bool var_3979_transpose_y_0 = const()[name = string("op_3979_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 4, 256, 512]> var_3979_cast_fp16 = matmul(transpose_x = var_3979_transpose_x_0, transpose_y = var_3979_transpose_y_0, x = var_3978_cast_fp16, y = v_23_cast_fp16)[name = string("op_3979_cast_fp16")];
            tensor<int32, [4]> var_3982_perm_0 = const()[name = string("op_3982_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3986 = const()[name = string("op_3986"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 256, 4, 512]> var_3982_cast_fp16 = transpose(perm = var_3982_perm_0, x = var_3979_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 256, 2048]> input_63_cast_fp16 = reshape(shape = var_3986, x = var_3982_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34267520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35316160))))[name = string("tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3991_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_3991_cast_fp16")];
            bool var_3994 = const()[name = string("op_3994"), val = bool(true)];
            tensor<int32, [1]> var_3995_axes_0 = const()[name = string("op_3995_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3995_cast_fp16 = reduce_max(axes = var_3995_axes_0, keep_dims = var_3994, x = var_3991_cast_fp16)[name = string("op_3995_cast_fp16")];
            fp16 var_3996_to_fp16 = const()[name = string("op_3996_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_91_cast_fp16 = maximum(x = var_3995_cast_fp16, y = var_3996_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_91_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_91_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_3999_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_3999_cast_fp16")];
            tensor<int32, [1]> var_4004_axes_0 = const()[name = string("op_4004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4004_keep_dims_0 = const()[name = string("op_4004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4004_cast_fp16 = reduce_mean(axes = var_4004_axes_0, keep_dims = var_4004_keep_dims_0, x = var_3999_cast_fp16)[name = string("op_4004_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4005_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_4005_cast_fp16")];
            fp32 var_4006_epsilon_0 = const()[name = string("op_4006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4006_cast_fp16 = inverse(epsilon = var_4006_epsilon_0, x = var_4005_cast_fp16)[name = string("op_4006_cast_fp16")];
            fp16 var_4007_to_fp16 = const()[name = string("op_4007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4008_cast_fp16 = mul(x = var_4006_cast_fp16, y = var_4007_to_fp16)[name = string("op_4008_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4010_cast_fp16 = add(x = var_4004_cast_fp16, y = var_4008_cast_fp16)[name = string("op_4010_cast_fp16")];
            fp32 var_4011_epsilon_0 = const()[name = string("op_4011_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4011_cast_fp16 = rsqrt(epsilon = var_4011_epsilon_0, x = var_4010_cast_fp16)[name = string("op_4011_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_79_cast_fp16 = mul(x = y_91_cast_fp16, y = var_4011_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35317248)))];
            tensor<fp16, [1, 256, 512]> var_4013_cast_fp16 = mul(x = out_79_cast_fp16, y = tm_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_4013_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_91_cast_fp16 = add(x = x_81_cast_fp16, y = var_4013_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4016_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_4016_cast_fp16")];
            bool var_4019 = const()[name = string("op_4019"), val = bool(true)];
            tensor<int32, [1]> var_4020_axes_0 = const()[name = string("op_4020_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4020_cast_fp16 = reduce_max(axes = var_4020_axes_0, keep_dims = var_4019, x = var_4016_cast_fp16)[name = string("op_4020_cast_fp16")];
            fp16 var_4021_to_fp16 = const()[name = string("op_4021_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_93_cast_fp16 = maximum(x = var_4020_cast_fp16, y = var_4021_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_93_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4024_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_4024_cast_fp16")];
            tensor<int32, [1]> var_4029_axes_0 = const()[name = string("op_4029_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4029_keep_dims_0 = const()[name = string("op_4029_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4029_cast_fp16 = reduce_mean(axes = var_4029_axes_0, keep_dims = var_4029_keep_dims_0, x = var_4024_cast_fp16)[name = string("op_4029_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4030_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_4030_cast_fp16")];
            fp32 var_4031_epsilon_0 = const()[name = string("op_4031_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4031_cast_fp16 = inverse(epsilon = var_4031_epsilon_0, x = var_4030_cast_fp16)[name = string("op_4031_cast_fp16")];
            fp16 var_4032_to_fp16 = const()[name = string("op_4032_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4033_cast_fp16 = mul(x = var_4031_cast_fp16, y = var_4032_to_fp16)[name = string("op_4033_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4035_cast_fp16 = add(x = var_4029_cast_fp16, y = var_4033_cast_fp16)[name = string("op_4035_cast_fp16")];
            fp32 var_4036_epsilon_0 = const()[name = string("op_4036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4036_cast_fp16 = rsqrt(epsilon = var_4036_epsilon_0, x = var_4035_cast_fp16)[name = string("op_4036_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_81_cast_fp16 = mul(x = y_93_cast_fp16, y = var_4036_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35318336)))];
            tensor<fp16, [1, 256, 512]> input_65_cast_fp16 = mul(x = out_81_cast_fp16, y = tm_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35319424))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36368064))))[name = string("tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_50_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_50_cast_fp16")];
            string var_4043_mode_0 = const()[name = string("op_4043_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_4043_cast_fp16 = gelu(mode = var_4043_mode_0, x = linear_50_cast_fp16)[name = string("op_4043_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36372224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37420864))))[name = string("tm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_51_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_67_cast_fp16 = mul(x = var_4043_cast_fp16, y = linear_51_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37425024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38473664))))[name = string("tm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4051_cast_fp16 = abs(x = linear_52_cast_fp16)[name = string("op_4051_cast_fp16")];
            bool var_4054 = const()[name = string("op_4054"), val = bool(true)];
            tensor<int32, [1]> var_4055_axes_0 = const()[name = string("op_4055_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4055_cast_fp16 = reduce_max(axes = var_4055_axes_0, keep_dims = var_4054, x = var_4051_cast_fp16)[name = string("op_4055_cast_fp16")];
            fp16 var_4056_to_fp16 = const()[name = string("op_4056_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_95_cast_fp16 = maximum(x = var_4055_cast_fp16, y = var_4056_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_95_cast_fp16 = real_div(x = linear_52_cast_fp16, y = s_95_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4059_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<int32, [1]> var_4064_axes_0 = const()[name = string("op_4064_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4064_keep_dims_0 = const()[name = string("op_4064_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4064_cast_fp16 = reduce_mean(axes = var_4064_axes_0, keep_dims = var_4064_keep_dims_0, x = var_4059_cast_fp16)[name = string("op_4064_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4065_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp32 var_4066_epsilon_0 = const()[name = string("op_4066_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4066_cast_fp16 = inverse(epsilon = var_4066_epsilon_0, x = var_4065_cast_fp16)[name = string("op_4066_cast_fp16")];
            fp16 var_4067_to_fp16 = const()[name = string("op_4067_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4068_cast_fp16 = mul(x = var_4066_cast_fp16, y = var_4067_to_fp16)[name = string("op_4068_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4070_cast_fp16 = add(x = var_4064_cast_fp16, y = var_4068_cast_fp16)[name = string("op_4070_cast_fp16")];
            fp32 var_4071_epsilon_0 = const()[name = string("op_4071_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4071_cast_fp16 = rsqrt(epsilon = var_4071_epsilon_0, x = var_4070_cast_fp16)[name = string("op_4071_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_83_cast_fp16 = mul(x = y_95_cast_fp16, y = var_4071_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38474752)))];
            tensor<fp16, [1, 256, 512]> var_4073_cast_fp16 = mul(x = out_83_cast_fp16, y = tm_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4073_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_69_cast_fp16 = add(x = x_91_cast_fp16, y = var_4073_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38475840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38738048))))[name = string("tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_53_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_53_cast_fp16")];
            string var_4080_mode_0 = const()[name = string("op_4080_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_4080_cast_fp16 = gelu(mode = var_4080_mode_0, x = linear_53_cast_fp16)[name = string("op_4080_cast_fp16")];
            tensor<int32, [4]> var_4093_begin_0 = const()[name = string("op_4093_begin_0"), val = tensor<int32, [4]>([0, 0, 5, 0])];
            tensor<int32, [4]> var_4093_end_0 = const()[name = string("op_4093_end_0"), val = tensor<int32, [4]>([1, 256, 6, 512])];
            tensor<bool, [4]> var_4093_end_mask_0 = const()[name = string("op_4093_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4093_squeeze_mask_0 = const()[name = string("op_4093_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_4093_cast_fp16 = slice_by_index(begin = var_4093_begin_0, end = var_4093_end_0, end_mask = var_4093_end_mask_0, squeeze_mask = var_4093_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_4093_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_71_cast_fp16 = mul(x = var_4080_cast_fp16, y = var_4093_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38739136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39001344))))[name = string("tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_5_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4103_cast_fp16 = abs(x = linear_54_cast_fp16)[name = string("op_4103_cast_fp16")];
            bool var_4106 = const()[name = string("op_4106"), val = bool(true)];
            tensor<int32, [1]> var_4107_axes_0 = const()[name = string("op_4107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4107_cast_fp16 = reduce_max(axes = var_4107_axes_0, keep_dims = var_4106, x = var_4103_cast_fp16)[name = string("op_4107_cast_fp16")];
            fp16 var_4108_to_fp16 = const()[name = string("op_4108_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_97_cast_fp16 = maximum(x = var_4107_cast_fp16, y = var_4108_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_97_cast_fp16 = real_div(x = linear_54_cast_fp16, y = s_97_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4111_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_4111_cast_fp16")];
            tensor<int32, [1]> var_4116_axes_0 = const()[name = string("op_4116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4116_keep_dims_0 = const()[name = string("op_4116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4116_cast_fp16 = reduce_mean(axes = var_4116_axes_0, keep_dims = var_4116_keep_dims_0, x = var_4111_cast_fp16)[name = string("op_4116_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4117_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_4117_cast_fp16")];
            fp32 var_4118_epsilon_0 = const()[name = string("op_4118_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4118_cast_fp16 = inverse(epsilon = var_4118_epsilon_0, x = var_4117_cast_fp16)[name = string("op_4118_cast_fp16")];
            fp16 var_4119_to_fp16 = const()[name = string("op_4119_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4120_cast_fp16 = mul(x = var_4118_cast_fp16, y = var_4119_to_fp16)[name = string("op_4120_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4122_cast_fp16 = add(x = var_4116_cast_fp16, y = var_4120_cast_fp16)[name = string("op_4122_cast_fp16")];
            fp32 var_4123_epsilon_0 = const()[name = string("op_4123_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4123_cast_fp16 = rsqrt(epsilon = var_4123_epsilon_0, x = var_4122_cast_fp16)[name = string("op_4123_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_85_cast_fp16 = mul(x = y_97_cast_fp16, y = var_4123_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [512]> tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39002432)))];
            tensor<fp16, [1, 256, 512]> var_4125_cast_fp16 = mul(x = out_85_cast_fp16, y = tm_layers_5_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4125_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_11_cast_fp16 = add(x = input_69_cast_fp16, y = var_4125_cast_fp16)[name = string("h_11_cast_fp16")];
            tensor<fp16, [1]> tm_layers_5_layer_scalar_to_fp16 = const()[name = string("tm_layers_5_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-2])];
            tensor<fp16, [1, 256, 512]> x_97_cast_fp16 = mul(x = h_11_cast_fp16, y = tm_layers_5_layer_scalar_to_fp16)[name = string("x_97_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4129_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_4129_cast_fp16")];
            bool var_4132 = const()[name = string("op_4132"), val = bool(true)];
            tensor<int32, [1]> var_4133_axes_0 = const()[name = string("op_4133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4133_cast_fp16 = reduce_max(axes = var_4133_axes_0, keep_dims = var_4132, x = var_4129_cast_fp16)[name = string("op_4133_cast_fp16")];
            fp16 var_4134_to_fp16 = const()[name = string("op_4134_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_99_cast_fp16 = maximum(x = var_4133_cast_fp16, y = var_4134_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_99_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_99_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4137_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_4137_cast_fp16")];
            tensor<int32, [1]> var_4142_axes_0 = const()[name = string("op_4142_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4142_keep_dims_0 = const()[name = string("op_4142_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4142_cast_fp16 = reduce_mean(axes = var_4142_axes_0, keep_dims = var_4142_keep_dims_0, x = var_4137_cast_fp16)[name = string("op_4142_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4143_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_4143_cast_fp16")];
            fp32 var_4144_epsilon_0 = const()[name = string("op_4144_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4144_cast_fp16 = inverse(epsilon = var_4144_epsilon_0, x = var_4143_cast_fp16)[name = string("op_4144_cast_fp16")];
            fp16 var_4145_to_fp16 = const()[name = string("op_4145_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4146_cast_fp16 = mul(x = var_4144_cast_fp16, y = var_4145_to_fp16)[name = string("op_4146_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4148_cast_fp16 = add(x = var_4142_cast_fp16, y = var_4146_cast_fp16)[name = string("op_4148_cast_fp16")];
            fp32 var_4149_epsilon_0 = const()[name = string("op_4149_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4149_cast_fp16 = rsqrt(epsilon = var_4149_epsilon_0, x = var_4148_cast_fp16)[name = string("op_4149_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_87_cast_fp16 = mul(x = y_99_cast_fp16, y = var_4149_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39003520)))];
            tensor<fp16, [1, 256, 512]> input_73_cast_fp16 = mul(x = out_87_cast_fp16, y = tm_layers_6_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39004608))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39528960))))[name = string("tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<int32, [4]> var_4159 = const()[name = string("op_4159"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_4160_cast_fp16 = reshape(shape = var_4159, x = linear_55_cast_fp16)[name = string("op_4160_cast_fp16")];
            tensor<int32, [4]> x_99_perm_0 = const()[name = string("x_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39531072))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39793280))))[name = string("tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_4171 = const()[name = string("op_4171"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4172_cast_fp16 = reshape(shape = var_4171, x = linear_56_cast_fp16)[name = string("op_4172_cast_fp16")];
            tensor<int32, [4]> x_101_perm_0 = const()[name = string("x_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39794368))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40056576))))[name = string("tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4183 = const()[name = string("op_4183"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4184_cast_fp16 = reshape(shape = var_4183, x = linear_57_cast_fp16)[name = string("op_4184_cast_fp16")];
            tensor<int32, [4]> x_103_perm_0 = const()[name = string("x_103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_99_cast_fp16 = transpose(perm = x_99_perm_0, x = var_4160_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 4, 256, 256]> var_4188_cast_fp16 = abs(x = x_99_cast_fp16)[name = string("op_4188_cast_fp16")];
            bool var_4191 = const()[name = string("op_4191"), val = bool(true)];
            tensor<int32, [1]> var_4192_axes_0 = const()[name = string("op_4192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_4192_cast_fp16 = reduce_max(axes = var_4192_axes_0, keep_dims = var_4191, x = var_4188_cast_fp16)[name = string("op_4192_cast_fp16")];
            fp16 var_4193_to_fp16 = const()[name = string("op_4193_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_101_cast_fp16 = maximum(x = var_4192_cast_fp16, y = var_4193_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_101_cast_fp16 = real_div(x = x_99_cast_fp16, y = s_101_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4196_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_4196_cast_fp16")];
            tensor<int32, [1]> var_4201_axes_0 = const()[name = string("op_4201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4201_keep_dims_0 = const()[name = string("op_4201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_4201_cast_fp16 = reduce_mean(axes = var_4201_axes_0, keep_dims = var_4201_keep_dims_0, x = var_4196_cast_fp16)[name = string("op_4201_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4202_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_4202_cast_fp16")];
            fp32 var_4203_epsilon_0 = const()[name = string("op_4203_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_4203_cast_fp16 = inverse(epsilon = var_4203_epsilon_0, x = var_4202_cast_fp16)[name = string("op_4203_cast_fp16")];
            fp16 var_4204_to_fp16 = const()[name = string("op_4204_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_4205_cast_fp16 = mul(x = var_4203_cast_fp16, y = var_4204_to_fp16)[name = string("op_4205_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4207_cast_fp16 = add(x = var_4201_cast_fp16, y = var_4205_cast_fp16)[name = string("op_4207_cast_fp16")];
            fp32 var_4208_epsilon_0 = const()[name = string("op_4208_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_4208_cast_fp16 = rsqrt(epsilon = var_4208_epsilon_0, x = var_4207_cast_fp16)[name = string("op_4208_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_89_cast_fp16 = mul(x = y_101_cast_fp16, y = var_4208_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40057664)))];
            tensor<fp16, [1, 4, 256, 256]> q_25_cast_fp16 = mul(x = out_89_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_101_cast_fp16 = transpose(perm = x_101_perm_0, x = var_4172_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 2, 256, 256]> var_4211_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_4211_cast_fp16")];
            bool var_4214 = const()[name = string("op_4214"), val = bool(true)];
            tensor<int32, [1]> var_4215_axes_0 = const()[name = string("op_4215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4215_cast_fp16 = reduce_max(axes = var_4215_axes_0, keep_dims = var_4214, x = var_4211_cast_fp16)[name = string("op_4215_cast_fp16")];
            fp16 var_4216_to_fp16 = const()[name = string("op_4216_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_103_cast_fp16 = maximum(x = var_4215_cast_fp16, y = var_4216_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_103_cast_fp16 = real_div(x = x_101_cast_fp16, y = s_103_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4219_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_4219_cast_fp16")];
            tensor<int32, [1]> var_4224_axes_0 = const()[name = string("op_4224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4224_keep_dims_0 = const()[name = string("op_4224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4224_cast_fp16 = reduce_mean(axes = var_4224_axes_0, keep_dims = var_4224_keep_dims_0, x = var_4219_cast_fp16)[name = string("op_4224_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4225_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_4225_cast_fp16")];
            fp32 var_4226_epsilon_0 = const()[name = string("op_4226_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4226_cast_fp16 = inverse(epsilon = var_4226_epsilon_0, x = var_4225_cast_fp16)[name = string("op_4226_cast_fp16")];
            fp16 var_4227_to_fp16 = const()[name = string("op_4227_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4228_cast_fp16 = mul(x = var_4226_cast_fp16, y = var_4227_to_fp16)[name = string("op_4228_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4230_cast_fp16 = add(x = var_4224_cast_fp16, y = var_4228_cast_fp16)[name = string("op_4230_cast_fp16")];
            fp32 var_4231_epsilon_0 = const()[name = string("op_4231_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4231_cast_fp16 = rsqrt(epsilon = var_4231_epsilon_0, x = var_4230_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_91_cast_fp16 = mul(x = y_103_cast_fp16, y = var_4231_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [256]> tm_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058240)))];
            tensor<fp16, [1, 2, 256, 256]> k_37_cast_fp16 = mul(x = out_91_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_37_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_103_cast_fp16 = transpose(perm = x_103_perm_0, x = var_4184_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 2, 256, 256]> var_4234_cast_fp16 = abs(x = x_103_cast_fp16)[name = string("op_4234_cast_fp16")];
            bool var_4237 = const()[name = string("op_4237"), val = bool(true)];
            tensor<int32, [1]> var_4238_axes_0 = const()[name = string("op_4238_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4238_cast_fp16 = reduce_max(axes = var_4238_axes_0, keep_dims = var_4237, x = var_4234_cast_fp16)[name = string("op_4238_cast_fp16")];
            fp16 var_4239_to_fp16 = const()[name = string("op_4239_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_105_cast_fp16 = maximum(x = var_4238_cast_fp16, y = var_4239_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_105_cast_fp16 = real_div(x = x_103_cast_fp16, y = s_105_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4242_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_4242_cast_fp16")];
            tensor<int32, [1]> var_4247_axes_0 = const()[name = string("op_4247_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4247_keep_dims_0 = const()[name = string("op_4247_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4247_cast_fp16 = reduce_mean(axes = var_4247_axes_0, keep_dims = var_4247_keep_dims_0, x = var_4242_cast_fp16)[name = string("op_4247_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4248_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_4248_cast_fp16")];
            fp32 var_4249_epsilon_0 = const()[name = string("op_4249_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4249_cast_fp16 = inverse(epsilon = var_4249_epsilon_0, x = var_4248_cast_fp16)[name = string("op_4249_cast_fp16")];
            fp16 var_4250_to_fp16 = const()[name = string("op_4250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4251_cast_fp16 = mul(x = var_4249_cast_fp16, y = var_4250_to_fp16)[name = string("op_4251_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4253_cast_fp16 = add(x = var_4247_cast_fp16, y = var_4251_cast_fp16)[name = string("op_4253_cast_fp16")];
            fp32 var_4254_epsilon_0 = const()[name = string("op_4254_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4254_cast_fp16 = rsqrt(epsilon = var_4254_epsilon_0, x = var_4253_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_25_cast_fp16 = mul(x = y_105_cast_fp16, y = var_4254_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4256_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4256_cast_fp16")];
            tensor<int32, [4]> var_4261_begin_0 = const()[name = string("op_4261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4261_end_0 = const()[name = string("op_4261_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_4261_end_mask_0 = const()[name = string("op_4261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_4261_cast_fp16 = slice_by_index(begin = var_4261_begin_0, end = var_4261_end_0, end_mask = var_4261_end_mask_0, x = q_25_cast_fp16)[name = string("op_4261_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_4262_cast_fp16 = mul(x = var_4261_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_4262_cast_fp16")];
            tensor<int32, [4]> var_4267_begin_0 = const()[name = string("op_4267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4267_end_0 = const()[name = string("op_4267_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_4267_end_mask_0 = const()[name = string("op_4267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_4267_cast_fp16 = slice_by_index(begin = var_4267_begin_0, end = var_4267_end_0, end_mask = var_4267_end_mask_0, x = q_25_cast_fp16)[name = string("op_4267_cast_fp16")];
            int32 var_4269 = const()[name = string("op_4269"), val = int32(-1)];
            bool var_4270_interleave_0 = const()[name = string("op_4270_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_4270_cast_fp16 = concat(axis = var_4269, interleave = var_4270_interleave_0, values = (var_4262_cast_fp16, var_4267_cast_fp16))[name = string("op_4270_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4271_cast_fp16 = mul(x = var_4270_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4271_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_27_cast_fp16 = add(x = var_4256_cast_fp16, y = var_4271_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4274_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4274_cast_fp16")];
            tensor<int32, [4]> var_4279_begin_0 = const()[name = string("op_4279_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4279_end_0 = const()[name = string("op_4279_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_4279_end_mask_0 = const()[name = string("op_4279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_4279_cast_fp16 = slice_by_index(begin = var_4279_begin_0, end = var_4279_end_0, end_mask = var_4279_end_mask_0, x = k_37_cast_fp16)[name = string("op_4279_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_4280_cast_fp16 = mul(x = var_4279_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_4280_cast_fp16")];
            tensor<int32, [4]> var_4285_begin_0 = const()[name = string("op_4285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4285_end_0 = const()[name = string("op_4285_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_4285_end_mask_0 = const()[name = string("op_4285_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_4285_cast_fp16 = slice_by_index(begin = var_4285_begin_0, end = var_4285_end_0, end_mask = var_4285_end_mask_0, x = k_37_cast_fp16)[name = string("op_4285_cast_fp16")];
            int32 var_4287 = const()[name = string("op_4287"), val = int32(-1)];
            bool var_4288_interleave_0 = const()[name = string("op_4288_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_4288_cast_fp16 = concat(axis = var_4287, interleave = var_4288_interleave_0, values = (var_4280_cast_fp16, var_4285_cast_fp16))[name = string("op_4288_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4289_cast_fp16 = mul(x = var_4288_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4289_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_39_cast_fp16 = add(x = var_4274_cast_fp16, y = var_4289_cast_fp16)[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_12_reps_0 = const()[name = string("tile_12_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = k_39_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [4, 1, 256, 256]> tile_12_cast_fp16 = tile(reps = tile_12_reps_0, x = transpose_24_cast_fp16)[name = string("tile_12_cast_fp16")];
            tensor<int32, [5]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_24_cast_fp16 = reshape(shape = concat_24, x = tile_12_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [5]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = reshape_24_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [4, 1, 256, 256]> reshape_25_cast_fp16 = reshape(shape = concat_25, x = transpose_25_cast_fp16)[name = string("reshape_25_cast_fp16")];
            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_13_reps_0 = const()[name = string("tile_13_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = v_25_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [4, 1, 256, 256]> tile_13_cast_fp16 = tile(reps = tile_13_reps_0, x = transpose_26_cast_fp16)[name = string("tile_13_cast_fp16")];
            tensor<int32, [5]> concat_26 = const()[name = string("concat_26"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_26_cast_fp16 = reshape(shape = concat_26, x = tile_13_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [5]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_27 = const()[name = string("concat_27"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [4, 1, 256, 256]> reshape_27_cast_fp16 = reshape(shape = concat_27, x = transpose_27_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4303_transpose_x_0 = const()[name = string("op_4303_transpose_x_0"), val = bool(false)];
            bool var_4303_transpose_y_0 = const()[name = string("op_4303_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_102_cast_fp16 = transpose(perm = transpose_102_perm_0, x = reshape_25_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 4, 256, 256]> var_4303_cast_fp16 = matmul(transpose_x = var_4303_transpose_x_0, transpose_y = var_4303_transpose_y_0, x = q_27_cast_fp16, y = transpose_102_cast_fp16)[name = string("op_4303_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4305_cast_fp16 = add(x = var_4303_cast_fp16, y = attention_bias)[name = string("op_4305_cast_fp16")];
            int32 var_4306 = const()[name = string("op_4306"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_4308_cast_fp16 = softmax(axis = var_4306, x = var_4305_cast_fp16)[name = string("op_4308_cast_fp16")];
            bool var_4309_transpose_x_0 = const()[name = string("op_4309_transpose_x_0"), val = bool(false)];
            bool var_4309_transpose_y_0 = const()[name = string("op_4309_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = reshape_27_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 4, 256, 256]> var_4309_cast_fp16 = matmul(transpose_x = var_4309_transpose_x_0, transpose_y = var_4309_transpose_y_0, x = var_4308_cast_fp16, y = v_27_cast_fp16)[name = string("op_4309_cast_fp16")];
            tensor<int32, [4]> var_4312_perm_0 = const()[name = string("op_4312_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4316 = const()[name = string("op_4316"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_4312_cast_fp16 = transpose(perm = var_4312_perm_0, x = var_4309_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 256, 1024]> input_75_cast_fp16 = reshape(shape = var_4316, x = var_4312_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40058816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40583168))))[name = string("tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4321_cast_fp16 = abs(x = linear_58_cast_fp16)[name = string("op_4321_cast_fp16")];
            bool var_4324 = const()[name = string("op_4324"), val = bool(true)];
            tensor<int32, [1]> var_4325_axes_0 = const()[name = string("op_4325_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4325_cast_fp16 = reduce_max(axes = var_4325_axes_0, keep_dims = var_4324, x = var_4321_cast_fp16)[name = string("op_4325_cast_fp16")];
            fp16 var_4326_to_fp16 = const()[name = string("op_4326_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_107_cast_fp16 = maximum(x = var_4325_cast_fp16, y = var_4326_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_107_cast_fp16 = real_div(x = linear_58_cast_fp16, y = s_107_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4329_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4329_cast_fp16")];
            tensor<int32, [1]> var_4334_axes_0 = const()[name = string("op_4334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4334_keep_dims_0 = const()[name = string("op_4334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4334_cast_fp16 = reduce_mean(axes = var_4334_axes_0, keep_dims = var_4334_keep_dims_0, x = var_4329_cast_fp16)[name = string("op_4334_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4335_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_4335_cast_fp16")];
            fp32 var_4336_epsilon_0 = const()[name = string("op_4336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4336_cast_fp16 = inverse(epsilon = var_4336_epsilon_0, x = var_4335_cast_fp16)[name = string("op_4336_cast_fp16")];
            fp16 var_4337_to_fp16 = const()[name = string("op_4337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4338_cast_fp16 = mul(x = var_4336_cast_fp16, y = var_4337_to_fp16)[name = string("op_4338_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4340_cast_fp16 = add(x = var_4334_cast_fp16, y = var_4338_cast_fp16)[name = string("op_4340_cast_fp16")];
            fp32 var_4341_epsilon_0 = const()[name = string("op_4341_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4341_cast_fp16 = rsqrt(epsilon = var_4341_epsilon_0, x = var_4340_cast_fp16)[name = string("op_4341_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_93_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4341_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40584256)))];
            tensor<fp16, [1, 256, 512]> var_4343_cast_fp16 = mul(x = out_93_cast_fp16, y = tm_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_4343_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_107_cast_fp16 = add(x = x_97_cast_fp16, y = var_4343_cast_fp16)[name = string("x_107_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4346_cast_fp16 = abs(x = x_107_cast_fp16)[name = string("op_4346_cast_fp16")];
            bool var_4349 = const()[name = string("op_4349"), val = bool(true)];
            tensor<int32, [1]> var_4350_axes_0 = const()[name = string("op_4350_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4350_cast_fp16 = reduce_max(axes = var_4350_axes_0, keep_dims = var_4349, x = var_4346_cast_fp16)[name = string("op_4350_cast_fp16")];
            fp16 var_4351_to_fp16 = const()[name = string("op_4351_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_109_cast_fp16 = maximum(x = var_4350_cast_fp16, y = var_4351_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_109_cast_fp16 = real_div(x = x_107_cast_fp16, y = s_109_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4354_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4354_cast_fp16")];
            tensor<int32, [1]> var_4359_axes_0 = const()[name = string("op_4359_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4359_keep_dims_0 = const()[name = string("op_4359_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4359_cast_fp16 = reduce_mean(axes = var_4359_axes_0, keep_dims = var_4359_keep_dims_0, x = var_4354_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4360_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_4360_cast_fp16")];
            fp32 var_4361_epsilon_0 = const()[name = string("op_4361_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4361_cast_fp16 = inverse(epsilon = var_4361_epsilon_0, x = var_4360_cast_fp16)[name = string("op_4361_cast_fp16")];
            fp16 var_4362_to_fp16 = const()[name = string("op_4362_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4363_cast_fp16 = mul(x = var_4361_cast_fp16, y = var_4362_to_fp16)[name = string("op_4363_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4365_cast_fp16 = add(x = var_4359_cast_fp16, y = var_4363_cast_fp16)[name = string("op_4365_cast_fp16")];
            fp32 var_4366_epsilon_0 = const()[name = string("op_4366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4366_cast_fp16 = rsqrt(epsilon = var_4366_epsilon_0, x = var_4365_cast_fp16)[name = string("op_4366_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_95_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4366_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40585344)))];
            tensor<fp16, [1, 256, 512]> input_77_cast_fp16 = mul(x = out_95_cast_fp16, y = tm_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40586432))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41635072))))[name = string("tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_59_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_59_cast_fp16")];
            string var_4373_mode_0 = const()[name = string("op_4373_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_4373_cast_fp16 = gelu(mode = var_4373_mode_0, x = linear_59_cast_fp16)[name = string("op_4373_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41639232))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42687872))))[name = string("tm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_60_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_79_cast_fp16 = mul(x = var_4373_cast_fp16, y = linear_60_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42692032))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43740672))))[name = string("tm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4381_cast_fp16 = abs(x = linear_61_cast_fp16)[name = string("op_4381_cast_fp16")];
            bool var_4384 = const()[name = string("op_4384"), val = bool(true)];
            tensor<int32, [1]> var_4385_axes_0 = const()[name = string("op_4385_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4385_cast_fp16 = reduce_max(axes = var_4385_axes_0, keep_dims = var_4384, x = var_4381_cast_fp16)[name = string("op_4385_cast_fp16")];
            fp16 var_4386_to_fp16 = const()[name = string("op_4386_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_111_cast_fp16 = maximum(x = var_4385_cast_fp16, y = var_4386_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_111_cast_fp16 = real_div(x = linear_61_cast_fp16, y = s_111_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4389_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4389_cast_fp16")];
            tensor<int32, [1]> var_4394_axes_0 = const()[name = string("op_4394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4394_keep_dims_0 = const()[name = string("op_4394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4394_cast_fp16 = reduce_mean(axes = var_4394_axes_0, keep_dims = var_4394_keep_dims_0, x = var_4389_cast_fp16)[name = string("op_4394_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4395_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4395_cast_fp16")];
            fp32 var_4396_epsilon_0 = const()[name = string("op_4396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4396_cast_fp16 = inverse(epsilon = var_4396_epsilon_0, x = var_4395_cast_fp16)[name = string("op_4396_cast_fp16")];
            fp16 var_4397_to_fp16 = const()[name = string("op_4397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4398_cast_fp16 = mul(x = var_4396_cast_fp16, y = var_4397_to_fp16)[name = string("op_4398_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4400_cast_fp16 = add(x = var_4394_cast_fp16, y = var_4398_cast_fp16)[name = string("op_4400_cast_fp16")];
            fp32 var_4401_epsilon_0 = const()[name = string("op_4401_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4401_cast_fp16 = rsqrt(epsilon = var_4401_epsilon_0, x = var_4400_cast_fp16)[name = string("op_4401_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_97_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4401_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43741760)))];
            tensor<fp16, [1, 256, 512]> var_4403_cast_fp16 = mul(x = out_97_cast_fp16, y = tm_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4403_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_81_cast_fp16 = add(x = x_107_cast_fp16, y = var_4403_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43742848))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44005056))))[name = string("tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_62_cast_fp16")];
            string var_4410_mode_0 = const()[name = string("op_4410_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_4410_cast_fp16 = gelu(mode = var_4410_mode_0, x = linear_62_cast_fp16)[name = string("op_4410_cast_fp16")];
            tensor<int32, [4]> var_4423_begin_0 = const()[name = string("op_4423_begin_0"), val = tensor<int32, [4]>([0, 0, 6, 0])];
            tensor<int32, [4]> var_4423_end_0 = const()[name = string("op_4423_end_0"), val = tensor<int32, [4]>([1, 256, 7, 512])];
            tensor<bool, [4]> var_4423_end_mask_0 = const()[name = string("op_4423_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4423_squeeze_mask_0 = const()[name = string("op_4423_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_4423_cast_fp16 = slice_by_index(begin = var_4423_begin_0, end = var_4423_end_0, end_mask = var_4423_end_mask_0, squeeze_mask = var_4423_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_4423_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_83_cast_fp16 = mul(x = var_4410_cast_fp16, y = var_4423_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44006144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44268352))))[name = string("tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_6_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4433_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4433_cast_fp16")];
            bool var_4436 = const()[name = string("op_4436"), val = bool(true)];
            tensor<int32, [1]> var_4437_axes_0 = const()[name = string("op_4437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4437_cast_fp16 = reduce_max(axes = var_4437_axes_0, keep_dims = var_4436, x = var_4433_cast_fp16)[name = string("op_4437_cast_fp16")];
            fp16 var_4438_to_fp16 = const()[name = string("op_4438_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_113_cast_fp16 = maximum(x = var_4437_cast_fp16, y = var_4438_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_113_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_113_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4441_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4441_cast_fp16")];
            tensor<int32, [1]> var_4446_axes_0 = const()[name = string("op_4446_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4446_keep_dims_0 = const()[name = string("op_4446_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4446_cast_fp16 = reduce_mean(axes = var_4446_axes_0, keep_dims = var_4446_keep_dims_0, x = var_4441_cast_fp16)[name = string("op_4446_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4447_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_fp16)[name = string("op_4447_cast_fp16")];
            fp32 var_4448_epsilon_0 = const()[name = string("op_4448_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4448_cast_fp16 = inverse(epsilon = var_4448_epsilon_0, x = var_4447_cast_fp16)[name = string("op_4448_cast_fp16")];
            fp16 var_4449_to_fp16 = const()[name = string("op_4449_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4450_cast_fp16 = mul(x = var_4448_cast_fp16, y = var_4449_to_fp16)[name = string("op_4450_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4452_cast_fp16 = add(x = var_4446_cast_fp16, y = var_4450_cast_fp16)[name = string("op_4452_cast_fp16")];
            fp32 var_4453_epsilon_0 = const()[name = string("op_4453_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4453_cast_fp16 = rsqrt(epsilon = var_4453_epsilon_0, x = var_4452_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_99_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4453_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [512]> tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44269440)))];
            tensor<fp16, [1, 256, 512]> var_4455_cast_fp16 = mul(x = out_99_cast_fp16, y = tm_layers_6_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4455_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_13_cast_fp16 = add(x = input_81_cast_fp16, y = var_4455_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<fp16, [1]> tm_layers_6_layer_scalar_to_fp16 = const()[name = string("tm_layers_6_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.cep-3])];
            tensor<fp16, [1, 256, 512]> x_113_cast_fp16 = mul(x = h_13_cast_fp16, y = tm_layers_6_layer_scalar_to_fp16)[name = string("x_113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4459_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_4459_cast_fp16")];
            bool var_4462 = const()[name = string("op_4462"), val = bool(true)];
            tensor<int32, [1]> var_4463_axes_0 = const()[name = string("op_4463_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4463_cast_fp16 = reduce_max(axes = var_4463_axes_0, keep_dims = var_4462, x = var_4459_cast_fp16)[name = string("op_4463_cast_fp16")];
            fp16 var_4464_to_fp16 = const()[name = string("op_4464_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_115_cast_fp16 = maximum(x = var_4463_cast_fp16, y = var_4464_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_115_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_115_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4467_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4467_cast_fp16")];
            tensor<int32, [1]> var_4472_axes_0 = const()[name = string("op_4472_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4472_keep_dims_0 = const()[name = string("op_4472_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4472_cast_fp16 = reduce_mean(axes = var_4472_axes_0, keep_dims = var_4472_keep_dims_0, x = var_4467_cast_fp16)[name = string("op_4472_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4473_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4473_cast_fp16")];
            fp32 var_4474_epsilon_0 = const()[name = string("op_4474_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4474_cast_fp16 = inverse(epsilon = var_4474_epsilon_0, x = var_4473_cast_fp16)[name = string("op_4474_cast_fp16")];
            fp16 var_4475_to_fp16 = const()[name = string("op_4475_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4476_cast_fp16 = mul(x = var_4474_cast_fp16, y = var_4475_to_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4478_cast_fp16 = add(x = var_4472_cast_fp16, y = var_4476_cast_fp16)[name = string("op_4478_cast_fp16")];
            fp32 var_4479_epsilon_0 = const()[name = string("op_4479_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4479_cast_fp16 = rsqrt(epsilon = var_4479_epsilon_0, x = var_4478_cast_fp16)[name = string("op_4479_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_101_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4479_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44270528)))];
            tensor<fp16, [1, 256, 512]> input_85_cast_fp16 = mul(x = out_101_cast_fp16, y = tm_layers_7_input_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44271616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44795968))))[name = string("tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4489 = const()[name = string("op_4489"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_4490_cast_fp16 = reshape(shape = var_4489, x = linear_64_cast_fp16)[name = string("op_4490_cast_fp16")];
            tensor<int32, [4]> x_115_perm_0 = const()[name = string("x_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44798080))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45060288))))[name = string("tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4501 = const()[name = string("op_4501"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4502_cast_fp16 = reshape(shape = var_4501, x = linear_65_cast_fp16)[name = string("op_4502_cast_fp16")];
            tensor<int32, [4]> x_117_perm_0 = const()[name = string("x_117_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45061376))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45323584))))[name = string("tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4513 = const()[name = string("op_4513"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4514_cast_fp16 = reshape(shape = var_4513, x = linear_66_cast_fp16)[name = string("op_4514_cast_fp16")];
            tensor<int32, [4]> x_119_perm_0 = const()[name = string("x_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_115_cast_fp16 = transpose(perm = x_115_perm_0, x = var_4490_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 4, 256, 256]> var_4518_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_4518_cast_fp16")];
            bool var_4521 = const()[name = string("op_4521"), val = bool(true)];
            tensor<int32, [1]> var_4522_axes_0 = const()[name = string("op_4522_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_4522_cast_fp16 = reduce_max(axes = var_4522_axes_0, keep_dims = var_4521, x = var_4518_cast_fp16)[name = string("op_4522_cast_fp16")];
            fp16 var_4523_to_fp16 = const()[name = string("op_4523_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_117_cast_fp16 = maximum(x = var_4522_cast_fp16, y = var_4523_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_117_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_117_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4526_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4526_cast_fp16")];
            tensor<int32, [1]> var_4531_axes_0 = const()[name = string("op_4531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4531_keep_dims_0 = const()[name = string("op_4531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_4531_cast_fp16 = reduce_mean(axes = var_4531_axes_0, keep_dims = var_4531_keep_dims_0, x = var_4526_cast_fp16)[name = string("op_4531_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4532_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4532_cast_fp16")];
            fp32 var_4533_epsilon_0 = const()[name = string("op_4533_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_4533_cast_fp16 = inverse(epsilon = var_4533_epsilon_0, x = var_4532_cast_fp16)[name = string("op_4533_cast_fp16")];
            fp16 var_4534_to_fp16 = const()[name = string("op_4534_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_4535_cast_fp16 = mul(x = var_4533_cast_fp16, y = var_4534_to_fp16)[name = string("op_4535_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4537_cast_fp16 = add(x = var_4531_cast_fp16, y = var_4535_cast_fp16)[name = string("op_4537_cast_fp16")];
            fp32 var_4538_epsilon_0 = const()[name = string("op_4538_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_4538_cast_fp16 = rsqrt(epsilon = var_4538_epsilon_0, x = var_4537_cast_fp16)[name = string("op_4538_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_103_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4538_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_29_cast_fp16 = mul(x = out_103_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_117_cast_fp16 = transpose(perm = x_117_perm_0, x = var_4502_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 2, 256, 256]> var_4541_cast_fp16 = abs(x = x_117_cast_fp16)[name = string("op_4541_cast_fp16")];
            bool var_4544 = const()[name = string("op_4544"), val = bool(true)];
            tensor<int32, [1]> var_4545_axes_0 = const()[name = string("op_4545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4545_cast_fp16 = reduce_max(axes = var_4545_axes_0, keep_dims = var_4544, x = var_4541_cast_fp16)[name = string("op_4545_cast_fp16")];
            fp16 var_4546_to_fp16 = const()[name = string("op_4546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_119_cast_fp16 = maximum(x = var_4545_cast_fp16, y = var_4546_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_119_cast_fp16 = real_div(x = x_117_cast_fp16, y = s_119_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4549_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4549_cast_fp16")];
            tensor<int32, [1]> var_4554_axes_0 = const()[name = string("op_4554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4554_keep_dims_0 = const()[name = string("op_4554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4554_cast_fp16 = reduce_mean(axes = var_4554_axes_0, keep_dims = var_4554_keep_dims_0, x = var_4549_cast_fp16)[name = string("op_4554_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4555_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4555_cast_fp16")];
            fp32 var_4556_epsilon_0 = const()[name = string("op_4556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4556_cast_fp16 = inverse(epsilon = var_4556_epsilon_0, x = var_4555_cast_fp16)[name = string("op_4556_cast_fp16")];
            fp16 var_4557_to_fp16 = const()[name = string("op_4557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4558_cast_fp16 = mul(x = var_4556_cast_fp16, y = var_4557_to_fp16)[name = string("op_4558_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4560_cast_fp16 = add(x = var_4554_cast_fp16, y = var_4558_cast_fp16)[name = string("op_4560_cast_fp16")];
            fp32 var_4561_epsilon_0 = const()[name = string("op_4561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4561_cast_fp16 = rsqrt(epsilon = var_4561_epsilon_0, x = var_4560_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_105_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4561_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [256]> tm_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45324672)))];
            tensor<fp16, [1, 2, 256, 256]> k_43_cast_fp16 = mul(x = out_105_cast_fp16, y = tm_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("k_43_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_119_cast_fp16 = transpose(perm = x_119_perm_0, x = var_4514_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 2, 256, 256]> var_4564_cast_fp16 = abs(x = x_119_cast_fp16)[name = string("op_4564_cast_fp16")];
            bool var_4567 = const()[name = string("op_4567"), val = bool(true)];
            tensor<int32, [1]> var_4568_axes_0 = const()[name = string("op_4568_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4568_cast_fp16 = reduce_max(axes = var_4568_axes_0, keep_dims = var_4567, x = var_4564_cast_fp16)[name = string("op_4568_cast_fp16")];
            fp16 var_4569_to_fp16 = const()[name = string("op_4569_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_121_cast_fp16 = maximum(x = var_4568_cast_fp16, y = var_4569_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_121_cast_fp16 = real_div(x = x_119_cast_fp16, y = s_121_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4572_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4572_cast_fp16")];
            tensor<int32, [1]> var_4577_axes_0 = const()[name = string("op_4577_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4577_keep_dims_0 = const()[name = string("op_4577_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4577_cast_fp16 = reduce_mean(axes = var_4577_axes_0, keep_dims = var_4577_keep_dims_0, x = var_4572_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4578_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4578_cast_fp16")];
            fp32 var_4579_epsilon_0 = const()[name = string("op_4579_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4579_cast_fp16 = inverse(epsilon = var_4579_epsilon_0, x = var_4578_cast_fp16)[name = string("op_4579_cast_fp16")];
            fp16 var_4580_to_fp16 = const()[name = string("op_4580_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4581_cast_fp16 = mul(x = var_4579_cast_fp16, y = var_4580_to_fp16)[name = string("op_4581_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4583_cast_fp16 = add(x = var_4577_cast_fp16, y = var_4581_cast_fp16)[name = string("op_4583_cast_fp16")];
            fp32 var_4584_epsilon_0 = const()[name = string("op_4584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4584_cast_fp16 = rsqrt(epsilon = var_4584_epsilon_0, x = var_4583_cast_fp16)[name = string("op_4584_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_29_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4584_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4586_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4586_cast_fp16")];
            tensor<int32, [4]> var_4591_begin_0 = const()[name = string("op_4591_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4591_end_0 = const()[name = string("op_4591_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_4591_end_mask_0 = const()[name = string("op_4591_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_4591_cast_fp16 = slice_by_index(begin = var_4591_begin_0, end = var_4591_end_0, end_mask = var_4591_end_mask_0, x = q_29_cast_fp16)[name = string("op_4591_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_4592_cast_fp16 = mul(x = var_4591_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_4592_cast_fp16")];
            tensor<int32, [4]> var_4597_begin_0 = const()[name = string("op_4597_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4597_end_0 = const()[name = string("op_4597_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_4597_end_mask_0 = const()[name = string("op_4597_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_4597_cast_fp16 = slice_by_index(begin = var_4597_begin_0, end = var_4597_end_0, end_mask = var_4597_end_mask_0, x = q_29_cast_fp16)[name = string("op_4597_cast_fp16")];
            int32 var_4599 = const()[name = string("op_4599"), val = int32(-1)];
            bool var_4600_interleave_0 = const()[name = string("op_4600_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_4600_cast_fp16 = concat(axis = var_4599, interleave = var_4600_interleave_0, values = (var_4592_cast_fp16, var_4597_cast_fp16))[name = string("op_4600_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4601_cast_fp16 = mul(x = var_4600_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4601_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_31_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4601_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4604_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4604_cast_fp16")];
            tensor<int32, [4]> var_4609_begin_0 = const()[name = string("op_4609_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4609_end_0 = const()[name = string("op_4609_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_4609_end_mask_0 = const()[name = string("op_4609_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_4609_cast_fp16 = slice_by_index(begin = var_4609_begin_0, end = var_4609_end_0, end_mask = var_4609_end_mask_0, x = k_43_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_4610_cast_fp16 = mul(x = var_4609_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_4610_cast_fp16")];
            tensor<int32, [4]> var_4615_begin_0 = const()[name = string("op_4615_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4615_end_0 = const()[name = string("op_4615_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_4615_end_mask_0 = const()[name = string("op_4615_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_4615_cast_fp16 = slice_by_index(begin = var_4615_begin_0, end = var_4615_end_0, end_mask = var_4615_end_mask_0, x = k_43_cast_fp16)[name = string("op_4615_cast_fp16")];
            int32 var_4617 = const()[name = string("op_4617"), val = int32(-1)];
            bool var_4618_interleave_0 = const()[name = string("op_4618_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_4618_cast_fp16 = concat(axis = var_4617, interleave = var_4618_interleave_0, values = (var_4610_cast_fp16, var_4615_cast_fp16))[name = string("op_4618_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4619_cast_fp16 = mul(x = var_4618_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4619_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_45_cast_fp16 = add(x = var_4604_cast_fp16, y = var_4619_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_14_reps_0 = const()[name = string("tile_14_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = k_45_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [4, 1, 256, 256]> tile_14_cast_fp16 = tile(reps = tile_14_reps_0, x = transpose_28_cast_fp16)[name = string("tile_14_cast_fp16")];
            tensor<int32, [5]> concat_28 = const()[name = string("concat_28"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_28_cast_fp16 = reshape(shape = concat_28, x = tile_14_cast_fp16)[name = string("reshape_28_cast_fp16")];
            tensor<int32, [5]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = reshape_28_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [4, 1, 256, 256]> reshape_29_cast_fp16 = reshape(shape = concat_29, x = transpose_29_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_15_reps_0 = const()[name = string("tile_15_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = v_29_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [4, 1, 256, 256]> tile_15_cast_fp16 = tile(reps = tile_15_reps_0, x = transpose_30_cast_fp16)[name = string("tile_15_cast_fp16")];
            tensor<int32, [5]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_30_cast_fp16 = reshape(shape = concat_30, x = tile_15_cast_fp16)[name = string("reshape_30_cast_fp16")];
            tensor<int32, [5]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_31 = const()[name = string("concat_31"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = reshape_30_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [4, 1, 256, 256]> reshape_31_cast_fp16 = reshape(shape = concat_31, x = transpose_31_cast_fp16)[name = string("reshape_31_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4633_transpose_x_0 = const()[name = string("op_4633_transpose_x_0"), val = bool(false)];
            bool var_4633_transpose_y_0 = const()[name = string("op_4633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 4, 256, 256]> var_4633_cast_fp16 = matmul(transpose_x = var_4633_transpose_x_0, transpose_y = var_4633_transpose_y_0, x = q_31_cast_fp16, y = transpose_103_cast_fp16)[name = string("op_4633_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4635_cast_fp16 = add(x = var_4633_cast_fp16, y = attention_bias)[name = string("op_4635_cast_fp16")];
            int32 var_4636 = const()[name = string("op_4636"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_4638_cast_fp16 = softmax(axis = var_4636, x = var_4635_cast_fp16)[name = string("op_4638_cast_fp16")];
            bool var_4639_transpose_x_0 = const()[name = string("op_4639_transpose_x_0"), val = bool(false)];
            bool var_4639_transpose_y_0 = const()[name = string("op_4639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = reshape_31_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 4, 256, 256]> var_4639_cast_fp16 = matmul(transpose_x = var_4639_transpose_x_0, transpose_y = var_4639_transpose_y_0, x = var_4638_cast_fp16, y = v_31_cast_fp16)[name = string("op_4639_cast_fp16")];
            tensor<int32, [4]> var_4642_perm_0 = const()[name = string("op_4642_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4646 = const()[name = string("op_4646"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_4642_cast_fp16 = transpose(perm = var_4642_perm_0, x = var_4639_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 256, 1024]> input_87_cast_fp16 = reshape(shape = var_4646, x = var_4642_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45325248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45849600))))[name = string("tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4651_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4651_cast_fp16")];
            bool var_4654 = const()[name = string("op_4654"), val = bool(true)];
            tensor<int32, [1]> var_4655_axes_0 = const()[name = string("op_4655_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4655_cast_fp16 = reduce_max(axes = var_4655_axes_0, keep_dims = var_4654, x = var_4651_cast_fp16)[name = string("op_4655_cast_fp16")];
            fp16 var_4656_to_fp16 = const()[name = string("op_4656_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_123_cast_fp16 = maximum(x = var_4655_cast_fp16, y = var_4656_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_123_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_123_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4659_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4659_cast_fp16")];
            tensor<int32, [1]> var_4664_axes_0 = const()[name = string("op_4664_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4664_keep_dims_0 = const()[name = string("op_4664_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4664_cast_fp16 = reduce_mean(axes = var_4664_axes_0, keep_dims = var_4664_keep_dims_0, x = var_4659_cast_fp16)[name = string("op_4664_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4665_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4665_cast_fp16")];
            fp32 var_4666_epsilon_0 = const()[name = string("op_4666_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4666_cast_fp16 = inverse(epsilon = var_4666_epsilon_0, x = var_4665_cast_fp16)[name = string("op_4666_cast_fp16")];
            fp16 var_4667_to_fp16 = const()[name = string("op_4667_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4668_cast_fp16 = mul(x = var_4666_cast_fp16, y = var_4667_to_fp16)[name = string("op_4668_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4670_cast_fp16 = add(x = var_4664_cast_fp16, y = var_4668_cast_fp16)[name = string("op_4670_cast_fp16")];
            fp32 var_4671_epsilon_0 = const()[name = string("op_4671_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4671_cast_fp16 = rsqrt(epsilon = var_4671_epsilon_0, x = var_4670_cast_fp16)[name = string("op_4671_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_107_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4671_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45850688)))];
            tensor<fp16, [1, 256, 512]> var_4673_cast_fp16 = mul(x = out_107_cast_fp16, y = tm_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4673_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_123_cast_fp16 = add(x = x_113_cast_fp16, y = var_4673_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4676_cast_fp16 = abs(x = x_123_cast_fp16)[name = string("op_4676_cast_fp16")];
            bool var_4679 = const()[name = string("op_4679"), val = bool(true)];
            tensor<int32, [1]> var_4680_axes_0 = const()[name = string("op_4680_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4680_cast_fp16 = reduce_max(axes = var_4680_axes_0, keep_dims = var_4679, x = var_4676_cast_fp16)[name = string("op_4680_cast_fp16")];
            fp16 var_4681_to_fp16 = const()[name = string("op_4681_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_125_cast_fp16 = maximum(x = var_4680_cast_fp16, y = var_4681_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_125_cast_fp16 = real_div(x = x_123_cast_fp16, y = s_125_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4684_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4684_cast_fp16")];
            tensor<int32, [1]> var_4689_axes_0 = const()[name = string("op_4689_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4689_keep_dims_0 = const()[name = string("op_4689_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4689_cast_fp16 = reduce_mean(axes = var_4689_axes_0, keep_dims = var_4689_keep_dims_0, x = var_4684_cast_fp16)[name = string("op_4689_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4690_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4690_cast_fp16")];
            fp32 var_4691_epsilon_0 = const()[name = string("op_4691_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4691_cast_fp16 = inverse(epsilon = var_4691_epsilon_0, x = var_4690_cast_fp16)[name = string("op_4691_cast_fp16")];
            fp16 var_4692_to_fp16 = const()[name = string("op_4692_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4693_cast_fp16 = mul(x = var_4691_cast_fp16, y = var_4692_to_fp16)[name = string("op_4693_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4695_cast_fp16 = add(x = var_4689_cast_fp16, y = var_4693_cast_fp16)[name = string("op_4695_cast_fp16")];
            fp32 var_4696_epsilon_0 = const()[name = string("op_4696_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4696_cast_fp16 = rsqrt(epsilon = var_4696_epsilon_0, x = var_4695_cast_fp16)[name = string("op_4696_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_109_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4696_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45851776)))];
            tensor<fp16, [1, 256, 512]> input_89_cast_fp16 = mul(x = out_109_cast_fp16, y = tm_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45852864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46901504))))[name = string("tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4703_mode_0 = const()[name = string("op_4703_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_4703_cast_fp16 = gelu(mode = var_4703_mode_0, x = linear_68_cast_fp16)[name = string("op_4703_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46905664))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47954304))))[name = string("tm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_91_cast_fp16 = mul(x = var_4703_cast_fp16, y = linear_69_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47958464))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49007104))))[name = string("tm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4711_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4711_cast_fp16")];
            bool var_4714 = const()[name = string("op_4714"), val = bool(true)];
            tensor<int32, [1]> var_4715_axes_0 = const()[name = string("op_4715_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4715_cast_fp16 = reduce_max(axes = var_4715_axes_0, keep_dims = var_4714, x = var_4711_cast_fp16)[name = string("op_4715_cast_fp16")];
            fp16 var_4716_to_fp16 = const()[name = string("op_4716_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_127_cast_fp16 = maximum(x = var_4715_cast_fp16, y = var_4716_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_127_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_127_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4719_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4719_cast_fp16")];
            tensor<int32, [1]> var_4724_axes_0 = const()[name = string("op_4724_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4724_keep_dims_0 = const()[name = string("op_4724_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4724_cast_fp16 = reduce_mean(axes = var_4724_axes_0, keep_dims = var_4724_keep_dims_0, x = var_4719_cast_fp16)[name = string("op_4724_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4725_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4725_cast_fp16")];
            fp32 var_4726_epsilon_0 = const()[name = string("op_4726_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4726_cast_fp16 = inverse(epsilon = var_4726_epsilon_0, x = var_4725_cast_fp16)[name = string("op_4726_cast_fp16")];
            fp16 var_4727_to_fp16 = const()[name = string("op_4727_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4728_cast_fp16 = mul(x = var_4726_cast_fp16, y = var_4727_to_fp16)[name = string("op_4728_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4730_cast_fp16 = add(x = var_4724_cast_fp16, y = var_4728_cast_fp16)[name = string("op_4730_cast_fp16")];
            fp32 var_4731_epsilon_0 = const()[name = string("op_4731_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4731_cast_fp16 = rsqrt(epsilon = var_4731_epsilon_0, x = var_4730_cast_fp16)[name = string("op_4731_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_111_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4731_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49008192)))];
            tensor<fp16, [1, 256, 512]> var_4733_cast_fp16 = mul(x = out_111_cast_fp16, y = tm_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4733_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_93_cast_fp16 = add(x = x_123_cast_fp16, y = var_4733_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49009280))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49271488))))[name = string("tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_71_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_71_cast_fp16")];
            string var_4740_mode_0 = const()[name = string("op_4740_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_4740_cast_fp16 = gelu(mode = var_4740_mode_0, x = linear_71_cast_fp16)[name = string("op_4740_cast_fp16")];
            tensor<int32, [4]> var_4753_begin_0 = const()[name = string("op_4753_begin_0"), val = tensor<int32, [4]>([0, 0, 7, 0])];
            tensor<int32, [4]> var_4753_end_0 = const()[name = string("op_4753_end_0"), val = tensor<int32, [4]>([1, 256, 8, 512])];
            tensor<bool, [4]> var_4753_end_mask_0 = const()[name = string("op_4753_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_4753_squeeze_mask_0 = const()[name = string("op_4753_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_4753_cast_fp16 = slice_by_index(begin = var_4753_begin_0, end = var_4753_end_0, end_mask = var_4753_end_mask_0, squeeze_mask = var_4753_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_4753_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_95_cast_fp16 = mul(x = var_4740_cast_fp16, y = var_4753_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49272576))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49534784))))[name = string("tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_7_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4763_cast_fp16 = abs(x = linear_72_cast_fp16)[name = string("op_4763_cast_fp16")];
            bool var_4766 = const()[name = string("op_4766"), val = bool(true)];
            tensor<int32, [1]> var_4767_axes_0 = const()[name = string("op_4767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4767_cast_fp16 = reduce_max(axes = var_4767_axes_0, keep_dims = var_4766, x = var_4763_cast_fp16)[name = string("op_4767_cast_fp16")];
            fp16 var_4768_to_fp16 = const()[name = string("op_4768_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_129_cast_fp16 = maximum(x = var_4767_cast_fp16, y = var_4768_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_129_cast_fp16 = real_div(x = linear_72_cast_fp16, y = s_129_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4771_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<int32, [1]> var_4776_axes_0 = const()[name = string("op_4776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4776_keep_dims_0 = const()[name = string("op_4776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4776_cast_fp16 = reduce_mean(axes = var_4776_axes_0, keep_dims = var_4776_keep_dims_0, x = var_4771_cast_fp16)[name = string("op_4776_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4777_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4777_cast_fp16")];
            fp32 var_4778_epsilon_0 = const()[name = string("op_4778_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4778_cast_fp16 = inverse(epsilon = var_4778_epsilon_0, x = var_4777_cast_fp16)[name = string("op_4778_cast_fp16")];
            fp16 var_4779_to_fp16 = const()[name = string("op_4779_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4780_cast_fp16 = mul(x = var_4778_cast_fp16, y = var_4779_to_fp16)[name = string("op_4780_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4782_cast_fp16 = add(x = var_4776_cast_fp16, y = var_4780_cast_fp16)[name = string("op_4782_cast_fp16")];
            fp32 var_4783_epsilon_0 = const()[name = string("op_4783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4783_cast_fp16 = rsqrt(epsilon = var_4783_epsilon_0, x = var_4782_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_113_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4783_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [512]> tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49535872)))];
            tensor<fp16, [1, 256, 512]> var_4785_cast_fp16 = mul(x = out_113_cast_fp16, y = tm_layers_7_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_4785_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_15_cast_fp16 = add(x = input_93_cast_fp16, y = var_4785_cast_fp16)[name = string("h_15_cast_fp16")];
            tensor<fp16, [1]> tm_layers_7_layer_scalar_to_fp16 = const()[name = string("tm_layers_7_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d4p-1])];
            tensor<fp16, [1, 256, 512]> x_129_cast_fp16 = mul(x = h_15_cast_fp16, y = tm_layers_7_layer_scalar_to_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4789_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_4789_cast_fp16")];
            bool var_4792 = const()[name = string("op_4792"), val = bool(true)];
            tensor<int32, [1]> var_4793_axes_0 = const()[name = string("op_4793_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4793_cast_fp16 = reduce_max(axes = var_4793_axes_0, keep_dims = var_4792, x = var_4789_cast_fp16)[name = string("op_4793_cast_fp16")];
            fp16 var_4794_to_fp16 = const()[name = string("op_4794_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_131_cast_fp16 = maximum(x = var_4793_cast_fp16, y = var_4794_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_131_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_131_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4797_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4797_cast_fp16")];
            tensor<int32, [1]> var_4802_axes_0 = const()[name = string("op_4802_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4802_keep_dims_0 = const()[name = string("op_4802_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4802_cast_fp16 = reduce_mean(axes = var_4802_axes_0, keep_dims = var_4802_keep_dims_0, x = var_4797_cast_fp16)[name = string("op_4802_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4803_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4803_cast_fp16")];
            fp32 var_4804_epsilon_0 = const()[name = string("op_4804_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4804_cast_fp16 = inverse(epsilon = var_4804_epsilon_0, x = var_4803_cast_fp16)[name = string("op_4804_cast_fp16")];
            fp16 var_4805_to_fp16 = const()[name = string("op_4805_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4806_cast_fp16 = mul(x = var_4804_cast_fp16, y = var_4805_to_fp16)[name = string("op_4806_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4808_cast_fp16 = add(x = var_4802_cast_fp16, y = var_4806_cast_fp16)[name = string("op_4808_cast_fp16")];
            fp32 var_4809_epsilon_0 = const()[name = string("op_4809_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_4809_cast_fp16 = rsqrt(epsilon = var_4809_epsilon_0, x = var_4808_cast_fp16)[name = string("op_4809_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_115_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4809_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49536960)))];
            tensor<fp16, [1, 256, 512]> input_97_cast_fp16 = mul(x = out_115_cast_fp16, y = tm_layers_8_input_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49538048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50062400))))[name = string("tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4819 = const()[name = string("op_4819"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_4820_cast_fp16 = reshape(shape = var_4819, x = linear_73_cast_fp16)[name = string("op_4820_cast_fp16")];
            tensor<int32, [4]> x_131_perm_0 = const()[name = string("x_131_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50064512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50326720))))[name = string("tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [4]> var_4831 = const()[name = string("op_4831"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4832_cast_fp16 = reshape(shape = var_4831, x = linear_74_cast_fp16)[name = string("op_4832_cast_fp16")];
            tensor<int32, [4]> x_133_perm_0 = const()[name = string("x_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50327808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50590016))))[name = string("tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<int32, [4]> var_4843 = const()[name = string("op_4843"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_4844_cast_fp16 = reshape(shape = var_4843, x = linear_75_cast_fp16)[name = string("op_4844_cast_fp16")];
            tensor<int32, [4]> x_135_perm_0 = const()[name = string("x_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_131_cast_fp16 = transpose(perm = x_131_perm_0, x = var_4820_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 4, 256, 256]> var_4848_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_4848_cast_fp16")];
            bool var_4851 = const()[name = string("op_4851"), val = bool(true)];
            tensor<int32, [1]> var_4852_axes_0 = const()[name = string("op_4852_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_4852_cast_fp16 = reduce_max(axes = var_4852_axes_0, keep_dims = var_4851, x = var_4848_cast_fp16)[name = string("op_4852_cast_fp16")];
            fp16 var_4853_to_fp16 = const()[name = string("op_4853_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_133_cast_fp16 = maximum(x = var_4852_cast_fp16, y = var_4853_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_133_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_133_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4856_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4856_cast_fp16")];
            tensor<int32, [1]> var_4861_axes_0 = const()[name = string("op_4861_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4861_keep_dims_0 = const()[name = string("op_4861_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_4861_cast_fp16 = reduce_mean(axes = var_4861_axes_0, keep_dims = var_4861_keep_dims_0, x = var_4856_cast_fp16)[name = string("op_4861_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4862_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4862_cast_fp16")];
            fp32 var_4863_epsilon_0 = const()[name = string("op_4863_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_4863_cast_fp16 = inverse(epsilon = var_4863_epsilon_0, x = var_4862_cast_fp16)[name = string("op_4863_cast_fp16")];
            fp16 var_4864_to_fp16 = const()[name = string("op_4864_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_4865_cast_fp16 = mul(x = var_4863_cast_fp16, y = var_4864_to_fp16)[name = string("op_4865_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_4867_cast_fp16 = add(x = var_4861_cast_fp16, y = var_4865_cast_fp16)[name = string("op_4867_cast_fp16")];
            fp32 var_4868_epsilon_0 = const()[name = string("op_4868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_4868_cast_fp16 = rsqrt(epsilon = var_4868_epsilon_0, x = var_4867_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_117_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4868_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591104)))];
            tensor<fp16, [1, 4, 256, 256]> q_33_cast_fp16 = mul(x = out_117_cast_fp16, y = tm_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("q_33_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_133_cast_fp16 = transpose(perm = x_133_perm_0, x = var_4832_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 2, 256, 256]> var_4871_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_4871_cast_fp16")];
            bool var_4874 = const()[name = string("op_4874"), val = bool(true)];
            tensor<int32, [1]> var_4875_axes_0 = const()[name = string("op_4875_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4875_cast_fp16 = reduce_max(axes = var_4875_axes_0, keep_dims = var_4874, x = var_4871_cast_fp16)[name = string("op_4875_cast_fp16")];
            fp16 var_4876_to_fp16 = const()[name = string("op_4876_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_135_cast_fp16 = maximum(x = var_4875_cast_fp16, y = var_4876_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_135_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_135_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4879_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4879_cast_fp16")];
            tensor<int32, [1]> var_4884_axes_0 = const()[name = string("op_4884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4884_keep_dims_0 = const()[name = string("op_4884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4884_cast_fp16 = reduce_mean(axes = var_4884_axes_0, keep_dims = var_4884_keep_dims_0, x = var_4879_cast_fp16)[name = string("op_4884_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4885_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4885_cast_fp16")];
            fp32 var_4886_epsilon_0 = const()[name = string("op_4886_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4886_cast_fp16 = inverse(epsilon = var_4886_epsilon_0, x = var_4885_cast_fp16)[name = string("op_4886_cast_fp16")];
            fp16 var_4887_to_fp16 = const()[name = string("op_4887_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4888_cast_fp16 = mul(x = var_4886_cast_fp16, y = var_4887_to_fp16)[name = string("op_4888_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4890_cast_fp16 = add(x = var_4884_cast_fp16, y = var_4888_cast_fp16)[name = string("op_4890_cast_fp16")];
            fp32 var_4891_epsilon_0 = const()[name = string("op_4891_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4891_cast_fp16 = rsqrt(epsilon = var_4891_epsilon_0, x = var_4890_cast_fp16)[name = string("op_4891_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_119_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4891_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [256]> tm_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50591680)))];
            tensor<fp16, [1, 2, 256, 256]> k_49_cast_fp16 = mul(x = out_119_cast_fp16, y = tm_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_135_cast_fp16 = transpose(perm = x_135_perm_0, x = var_4844_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 2, 256, 256]> var_4894_cast_fp16 = abs(x = x_135_cast_fp16)[name = string("op_4894_cast_fp16")];
            bool var_4897 = const()[name = string("op_4897"), val = bool(true)];
            tensor<int32, [1]> var_4898_axes_0 = const()[name = string("op_4898_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_4898_cast_fp16 = reduce_max(axes = var_4898_axes_0, keep_dims = var_4897, x = var_4894_cast_fp16)[name = string("op_4898_cast_fp16")];
            fp16 var_4899_to_fp16 = const()[name = string("op_4899_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_137_cast_fp16 = maximum(x = var_4898_cast_fp16, y = var_4899_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_137_cast_fp16 = real_div(x = x_135_cast_fp16, y = s_137_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4902_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4902_cast_fp16")];
            tensor<int32, [1]> var_4907_axes_0 = const()[name = string("op_4907_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4907_keep_dims_0 = const()[name = string("op_4907_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_4907_cast_fp16 = reduce_mean(axes = var_4907_axes_0, keep_dims = var_4907_keep_dims_0, x = var_4902_cast_fp16)[name = string("op_4907_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4908_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4908_cast_fp16")];
            fp32 var_4909_epsilon_0 = const()[name = string("op_4909_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_4909_cast_fp16 = inverse(epsilon = var_4909_epsilon_0, x = var_4908_cast_fp16)[name = string("op_4909_cast_fp16")];
            fp16 var_4910_to_fp16 = const()[name = string("op_4910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_4911_cast_fp16 = mul(x = var_4909_cast_fp16, y = var_4910_to_fp16)[name = string("op_4911_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_4913_cast_fp16 = add(x = var_4907_cast_fp16, y = var_4911_cast_fp16)[name = string("op_4913_cast_fp16")];
            fp32 var_4914_epsilon_0 = const()[name = string("op_4914_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_4914_cast_fp16 = rsqrt(epsilon = var_4914_epsilon_0, x = var_4913_cast_fp16)[name = string("op_4914_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_33_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4914_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4916_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4916_cast_fp16")];
            tensor<int32, [4]> var_4921_begin_0 = const()[name = string("op_4921_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4921_end_0 = const()[name = string("op_4921_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_4921_end_mask_0 = const()[name = string("op_4921_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_4921_cast_fp16 = slice_by_index(begin = var_4921_begin_0, end = var_4921_end_0, end_mask = var_4921_end_mask_0, x = q_33_cast_fp16)[name = string("op_4921_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_4922_cast_fp16 = mul(x = var_4921_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_4922_cast_fp16")];
            tensor<int32, [4]> var_4927_begin_0 = const()[name = string("op_4927_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4927_end_0 = const()[name = string("op_4927_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_4927_end_mask_0 = const()[name = string("op_4927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_4927_cast_fp16 = slice_by_index(begin = var_4927_begin_0, end = var_4927_end_0, end_mask = var_4927_end_mask_0, x = q_33_cast_fp16)[name = string("op_4927_cast_fp16")];
            int32 var_4929 = const()[name = string("op_4929"), val = int32(-1)];
            bool var_4930_interleave_0 = const()[name = string("op_4930_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_4930_cast_fp16 = concat(axis = var_4929, interleave = var_4930_interleave_0, values = (var_4922_cast_fp16, var_4927_cast_fp16))[name = string("op_4930_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4931_cast_fp16 = mul(x = var_4930_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4931_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_35_cast_fp16 = add(x = var_4916_cast_fp16, y = var_4931_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4934_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4934_cast_fp16")];
            tensor<int32, [4]> var_4939_begin_0 = const()[name = string("op_4939_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_4939_end_0 = const()[name = string("op_4939_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_4939_end_mask_0 = const()[name = string("op_4939_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_4939_cast_fp16 = slice_by_index(begin = var_4939_begin_0, end = var_4939_end_0, end_mask = var_4939_end_mask_0, x = k_49_cast_fp16)[name = string("op_4939_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_4940_cast_fp16 = mul(x = var_4939_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_4940_cast_fp16")];
            tensor<int32, [4]> var_4945_begin_0 = const()[name = string("op_4945_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4945_end_0 = const()[name = string("op_4945_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_4945_end_mask_0 = const()[name = string("op_4945_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_4945_cast_fp16 = slice_by_index(begin = var_4945_begin_0, end = var_4945_end_0, end_mask = var_4945_end_mask_0, x = k_49_cast_fp16)[name = string("op_4945_cast_fp16")];
            int32 var_4947 = const()[name = string("op_4947"), val = int32(-1)];
            bool var_4948_interleave_0 = const()[name = string("op_4948_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_4948_cast_fp16 = concat(axis = var_4947, interleave = var_4948_interleave_0, values = (var_4940_cast_fp16, var_4945_cast_fp16))[name = string("op_4948_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_4949_cast_fp16 = mul(x = var_4948_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4949_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_51_cast_fp16 = add(x = var_4934_cast_fp16, y = var_4949_cast_fp16)[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_16_reps_0 = const()[name = string("tile_16_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = k_51_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [4, 1, 256, 256]> tile_16_cast_fp16 = tile(reps = tile_16_reps_0, x = transpose_32_cast_fp16)[name = string("tile_16_cast_fp16")];
            tensor<int32, [5]> concat_32 = const()[name = string("concat_32"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_32_cast_fp16 = reshape(shape = concat_32, x = tile_16_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [5]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [4, 1, 256, 256]> reshape_33_cast_fp16 = reshape(shape = concat_33, x = transpose_33_cast_fp16)[name = string("reshape_33_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_17_reps_0 = const()[name = string("tile_17_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = v_33_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [4, 1, 256, 256]> tile_17_cast_fp16 = tile(reps = tile_17_reps_0, x = transpose_34_cast_fp16)[name = string("tile_17_cast_fp16")];
            tensor<int32, [5]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_34_cast_fp16 = reshape(shape = concat_34, x = tile_17_cast_fp16)[name = string("reshape_34_cast_fp16")];
            tensor<int32, [5]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = reshape_34_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [4, 1, 256, 256]> reshape_35_cast_fp16 = reshape(shape = concat_35, x = transpose_35_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_4963_transpose_x_0 = const()[name = string("op_4963_transpose_x_0"), val = bool(false)];
            bool var_4963_transpose_y_0 = const()[name = string("op_4963_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = reshape_33_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 4, 256, 256]> var_4963_cast_fp16 = matmul(transpose_x = var_4963_transpose_x_0, transpose_y = var_4963_transpose_y_0, x = q_35_cast_fp16, y = transpose_104_cast_fp16)[name = string("op_4963_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_4965_cast_fp16 = add(x = var_4963_cast_fp16, y = attention_bias)[name = string("op_4965_cast_fp16")];
            int32 var_4966 = const()[name = string("op_4966"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_4968_cast_fp16 = softmax(axis = var_4966, x = var_4965_cast_fp16)[name = string("op_4968_cast_fp16")];
            bool var_4969_transpose_x_0 = const()[name = string("op_4969_transpose_x_0"), val = bool(false)];
            bool var_4969_transpose_y_0 = const()[name = string("op_4969_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 4, 256, 256]> var_4969_cast_fp16 = matmul(transpose_x = var_4969_transpose_x_0, transpose_y = var_4969_transpose_y_0, x = var_4968_cast_fp16, y = v_35_cast_fp16)[name = string("op_4969_cast_fp16")];
            tensor<int32, [4]> var_4972_perm_0 = const()[name = string("op_4972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4976 = const()[name = string("op_4976"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_4972_cast_fp16 = transpose(perm = var_4972_perm_0, x = var_4969_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 256, 1024]> input_99_cast_fp16 = reshape(shape = var_4976, x = var_4972_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50592256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51116608))))[name = string("tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4981_cast_fp16 = abs(x = linear_76_cast_fp16)[name = string("op_4981_cast_fp16")];
            bool var_4984 = const()[name = string("op_4984"), val = bool(true)];
            tensor<int32, [1]> var_4985_axes_0 = const()[name = string("op_4985_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_4985_cast_fp16 = reduce_max(axes = var_4985_axes_0, keep_dims = var_4984, x = var_4981_cast_fp16)[name = string("op_4985_cast_fp16")];
            fp16 var_4986_to_fp16 = const()[name = string("op_4986_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_139_cast_fp16 = maximum(x = var_4985_cast_fp16, y = var_4986_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_139_cast_fp16 = real_div(x = linear_76_cast_fp16, y = s_139_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_4989_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_4989_cast_fp16")];
            tensor<int32, [1]> var_4994_axes_0 = const()[name = string("op_4994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4994_keep_dims_0 = const()[name = string("op_4994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_4994_cast_fp16 = reduce_mean(axes = var_4994_axes_0, keep_dims = var_4994_keep_dims_0, x = var_4989_cast_fp16)[name = string("op_4994_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_4995_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_4995_cast_fp16")];
            fp32 var_4996_epsilon_0 = const()[name = string("op_4996_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_4996_cast_fp16 = inverse(epsilon = var_4996_epsilon_0, x = var_4995_cast_fp16)[name = string("op_4996_cast_fp16")];
            fp16 var_4997_to_fp16 = const()[name = string("op_4997_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_4998_cast_fp16 = mul(x = var_4996_cast_fp16, y = var_4997_to_fp16)[name = string("op_4998_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5000_cast_fp16 = add(x = var_4994_cast_fp16, y = var_4998_cast_fp16)[name = string("op_5000_cast_fp16")];
            fp32 var_5001_epsilon_0 = const()[name = string("op_5001_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5001_cast_fp16 = rsqrt(epsilon = var_5001_epsilon_0, x = var_5000_cast_fp16)[name = string("op_5001_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_121_cast_fp16 = mul(x = y_139_cast_fp16, y = var_5001_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51117696)))];
            tensor<fp16, [1, 256, 512]> var_5003_cast_fp16 = mul(x = out_121_cast_fp16, y = tm_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_5003_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_139_cast_fp16 = add(x = x_129_cast_fp16, y = var_5003_cast_fp16)[name = string("x_139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5006_cast_fp16 = abs(x = x_139_cast_fp16)[name = string("op_5006_cast_fp16")];
            bool var_5009 = const()[name = string("op_5009"), val = bool(true)];
            tensor<int32, [1]> var_5010_axes_0 = const()[name = string("op_5010_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5010_cast_fp16 = reduce_max(axes = var_5010_axes_0, keep_dims = var_5009, x = var_5006_cast_fp16)[name = string("op_5010_cast_fp16")];
            fp16 var_5011_to_fp16 = const()[name = string("op_5011_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_141_cast_fp16 = maximum(x = var_5010_cast_fp16, y = var_5011_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_141_cast_fp16 = real_div(x = x_139_cast_fp16, y = s_141_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5014_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_5014_cast_fp16")];
            tensor<int32, [1]> var_5019_axes_0 = const()[name = string("op_5019_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5019_keep_dims_0 = const()[name = string("op_5019_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5019_cast_fp16 = reduce_mean(axes = var_5019_axes_0, keep_dims = var_5019_keep_dims_0, x = var_5014_cast_fp16)[name = string("op_5019_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5020_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_5020_cast_fp16")];
            fp32 var_5021_epsilon_0 = const()[name = string("op_5021_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5021_cast_fp16 = inverse(epsilon = var_5021_epsilon_0, x = var_5020_cast_fp16)[name = string("op_5021_cast_fp16")];
            fp16 var_5022_to_fp16 = const()[name = string("op_5022_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5023_cast_fp16 = mul(x = var_5021_cast_fp16, y = var_5022_to_fp16)[name = string("op_5023_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5025_cast_fp16 = add(x = var_5019_cast_fp16, y = var_5023_cast_fp16)[name = string("op_5025_cast_fp16")];
            fp32 var_5026_epsilon_0 = const()[name = string("op_5026_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5026_cast_fp16 = rsqrt(epsilon = var_5026_epsilon_0, x = var_5025_cast_fp16)[name = string("op_5026_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_123_cast_fp16 = mul(x = y_141_cast_fp16, y = var_5026_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51118784)))];
            tensor<fp16, [1, 256, 512]> input_101_cast_fp16 = mul(x = out_123_cast_fp16, y = tm_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51119872))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52168512))))[name = string("tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_77_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_77_cast_fp16")];
            string var_5033_mode_0 = const()[name = string("op_5033_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_5033_cast_fp16 = gelu(mode = var_5033_mode_0, x = linear_77_cast_fp16)[name = string("op_5033_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52172672))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53221312))))[name = string("tm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_78_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_103_cast_fp16 = mul(x = var_5033_cast_fp16, y = linear_78_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53225472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54274112))))[name = string("tm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5041_cast_fp16 = abs(x = linear_79_cast_fp16)[name = string("op_5041_cast_fp16")];
            bool var_5044 = const()[name = string("op_5044"), val = bool(true)];
            tensor<int32, [1]> var_5045_axes_0 = const()[name = string("op_5045_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5045_cast_fp16 = reduce_max(axes = var_5045_axes_0, keep_dims = var_5044, x = var_5041_cast_fp16)[name = string("op_5045_cast_fp16")];
            fp16 var_5046_to_fp16 = const()[name = string("op_5046_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_143_cast_fp16 = maximum(x = var_5045_cast_fp16, y = var_5046_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_143_cast_fp16 = real_div(x = linear_79_cast_fp16, y = s_143_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5049_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_5049_cast_fp16")];
            tensor<int32, [1]> var_5054_axes_0 = const()[name = string("op_5054_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5054_keep_dims_0 = const()[name = string("op_5054_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5054_cast_fp16 = reduce_mean(axes = var_5054_axes_0, keep_dims = var_5054_keep_dims_0, x = var_5049_cast_fp16)[name = string("op_5054_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5055_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_5055_cast_fp16")];
            fp32 var_5056_epsilon_0 = const()[name = string("op_5056_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5056_cast_fp16 = inverse(epsilon = var_5056_epsilon_0, x = var_5055_cast_fp16)[name = string("op_5056_cast_fp16")];
            fp16 var_5057_to_fp16 = const()[name = string("op_5057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5058_cast_fp16 = mul(x = var_5056_cast_fp16, y = var_5057_to_fp16)[name = string("op_5058_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5060_cast_fp16 = add(x = var_5054_cast_fp16, y = var_5058_cast_fp16)[name = string("op_5060_cast_fp16")];
            fp32 var_5061_epsilon_0 = const()[name = string("op_5061_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5061_cast_fp16 = rsqrt(epsilon = var_5061_epsilon_0, x = var_5060_cast_fp16)[name = string("op_5061_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_125_cast_fp16 = mul(x = y_143_cast_fp16, y = var_5061_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54275200)))];
            tensor<fp16, [1, 256, 512]> var_5063_cast_fp16 = mul(x = out_125_cast_fp16, y = tm_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5063_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_105_cast_fp16 = add(x = x_139_cast_fp16, y = var_5063_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54276288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54538496))))[name = string("tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_80_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_80_cast_fp16")];
            string var_5070_mode_0 = const()[name = string("op_5070_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_5070_cast_fp16 = gelu(mode = var_5070_mode_0, x = linear_80_cast_fp16)[name = string("op_5070_cast_fp16")];
            tensor<int32, [4]> var_5083_begin_0 = const()[name = string("op_5083_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_5083_end_0 = const()[name = string("op_5083_end_0"), val = tensor<int32, [4]>([1, 256, 9, 512])];
            tensor<bool, [4]> var_5083_end_mask_0 = const()[name = string("op_5083_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5083_squeeze_mask_0 = const()[name = string("op_5083_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_5083_cast_fp16 = slice_by_index(begin = var_5083_begin_0, end = var_5083_end_0, end_mask = var_5083_end_mask_0, squeeze_mask = var_5083_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_5083_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_107_cast_fp16 = mul(x = var_5070_cast_fp16, y = var_5083_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54539584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54801792))))[name = string("tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_8_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5093_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5093_cast_fp16")];
            bool var_5096 = const()[name = string("op_5096"), val = bool(true)];
            tensor<int32, [1]> var_5097_axes_0 = const()[name = string("op_5097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5097_cast_fp16 = reduce_max(axes = var_5097_axes_0, keep_dims = var_5096, x = var_5093_cast_fp16)[name = string("op_5097_cast_fp16")];
            fp16 var_5098_to_fp16 = const()[name = string("op_5098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_145_cast_fp16 = maximum(x = var_5097_cast_fp16, y = var_5098_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_145_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_145_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5101_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_5101_cast_fp16")];
            tensor<int32, [1]> var_5106_axes_0 = const()[name = string("op_5106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5106_keep_dims_0 = const()[name = string("op_5106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5106_cast_fp16 = reduce_mean(axes = var_5106_axes_0, keep_dims = var_5106_keep_dims_0, x = var_5101_cast_fp16)[name = string("op_5106_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5107_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_5107_cast_fp16")];
            fp32 var_5108_epsilon_0 = const()[name = string("op_5108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5108_cast_fp16 = inverse(epsilon = var_5108_epsilon_0, x = var_5107_cast_fp16)[name = string("op_5108_cast_fp16")];
            fp16 var_5109_to_fp16 = const()[name = string("op_5109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5110_cast_fp16 = mul(x = var_5108_cast_fp16, y = var_5109_to_fp16)[name = string("op_5110_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5112_cast_fp16 = add(x = var_5106_cast_fp16, y = var_5110_cast_fp16)[name = string("op_5112_cast_fp16")];
            fp32 var_5113_epsilon_0 = const()[name = string("op_5113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5113_cast_fp16 = rsqrt(epsilon = var_5113_epsilon_0, x = var_5112_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_127_cast_fp16 = mul(x = y_145_cast_fp16, y = var_5113_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [512]> tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54802880)))];
            tensor<fp16, [1, 256, 512]> var_5115_cast_fp16 = mul(x = out_127_cast_fp16, y = tm_layers_8_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5115_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_17_cast_fp16 = add(x = input_105_cast_fp16, y = var_5115_cast_fp16)[name = string("h_17_cast_fp16")];
            tensor<fp16, [1]> tm_layers_8_layer_scalar_to_fp16 = const()[name = string("tm_layers_8_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bcp-1])];
            tensor<fp16, [1, 256, 512]> x_145_cast_fp16 = mul(x = h_17_cast_fp16, y = tm_layers_8_layer_scalar_to_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5119_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_5119_cast_fp16")];
            bool var_5122 = const()[name = string("op_5122"), val = bool(true)];
            tensor<int32, [1]> var_5123_axes_0 = const()[name = string("op_5123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5123_cast_fp16 = reduce_max(axes = var_5123_axes_0, keep_dims = var_5122, x = var_5119_cast_fp16)[name = string("op_5123_cast_fp16")];
            fp16 var_5124_to_fp16 = const()[name = string("op_5124_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_147_cast_fp16 = maximum(x = var_5123_cast_fp16, y = var_5124_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_147_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_147_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5127_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_5127_cast_fp16")];
            tensor<int32, [1]> var_5132_axes_0 = const()[name = string("op_5132_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5132_keep_dims_0 = const()[name = string("op_5132_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5132_cast_fp16 = reduce_mean(axes = var_5132_axes_0, keep_dims = var_5132_keep_dims_0, x = var_5127_cast_fp16)[name = string("op_5132_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5133_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_5133_cast_fp16")];
            fp32 var_5134_epsilon_0 = const()[name = string("op_5134_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5134_cast_fp16 = inverse(epsilon = var_5134_epsilon_0, x = var_5133_cast_fp16)[name = string("op_5134_cast_fp16")];
            fp16 var_5135_to_fp16 = const()[name = string("op_5135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5136_cast_fp16 = mul(x = var_5134_cast_fp16, y = var_5135_to_fp16)[name = string("op_5136_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5138_cast_fp16 = add(x = var_5132_cast_fp16, y = var_5136_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp32 var_5139_epsilon_0 = const()[name = string("op_5139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5139_cast_fp16 = rsqrt(epsilon = var_5139_epsilon_0, x = var_5138_cast_fp16)[name = string("op_5139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_129_cast_fp16 = mul(x = y_147_cast_fp16, y = var_5139_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54803968)))];
            tensor<fp16, [1, 256, 512]> input_109_cast_fp16 = mul(x = out_129_cast_fp16, y = tm_layers_9_input_layernorm_weight_to_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54805056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55329408))))[name = string("tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [4]> var_5149 = const()[name = string("op_5149"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_5150_cast_fp16 = reshape(shape = var_5149, x = linear_82_cast_fp16)[name = string("op_5150_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55331520))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55593728))))[name = string("tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [4]> var_5161 = const()[name = string("op_5161"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_5162_cast_fp16 = reshape(shape = var_5161, x = linear_83_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55594816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55857024))))[name = string("tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5173 = const()[name = string("op_5173"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_5174_cast_fp16 = reshape(shape = var_5173, x = linear_84_cast_fp16)[name = string("op_5174_cast_fp16")];
            tensor<int32, [4]> x_151_perm_0 = const()[name = string("x_151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_5150_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 4, 256, 256]> var_5178_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_5178_cast_fp16")];
            bool var_5181 = const()[name = string("op_5181"), val = bool(true)];
            tensor<int32, [1]> var_5182_axes_0 = const()[name = string("op_5182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_5182_cast_fp16 = reduce_max(axes = var_5182_axes_0, keep_dims = var_5181, x = var_5178_cast_fp16)[name = string("op_5182_cast_fp16")];
            fp16 var_5183_to_fp16 = const()[name = string("op_5183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_149_cast_fp16 = maximum(x = var_5182_cast_fp16, y = var_5183_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_149_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_149_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5186_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5186_cast_fp16")];
            tensor<int32, [1]> var_5191_axes_0 = const()[name = string("op_5191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5191_keep_dims_0 = const()[name = string("op_5191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_5191_cast_fp16 = reduce_mean(axes = var_5191_axes_0, keep_dims = var_5191_keep_dims_0, x = var_5186_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5192_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_5192_cast_fp16")];
            fp32 var_5193_epsilon_0 = const()[name = string("op_5193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_5193_cast_fp16 = inverse(epsilon = var_5193_epsilon_0, x = var_5192_cast_fp16)[name = string("op_5193_cast_fp16")];
            fp16 var_5194_to_fp16 = const()[name = string("op_5194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_5195_cast_fp16 = mul(x = var_5193_cast_fp16, y = var_5194_to_fp16)[name = string("op_5195_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5197_cast_fp16 = add(x = var_5191_cast_fp16, y = var_5195_cast_fp16)[name = string("op_5197_cast_fp16")];
            fp32 var_5198_epsilon_0 = const()[name = string("op_5198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_5198_cast_fp16 = rsqrt(epsilon = var_5198_epsilon_0, x = var_5197_cast_fp16)[name = string("op_5198_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_131_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5198_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_5162_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 2, 256, 256]> var_5201_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_5201_cast_fp16")];
            bool var_5204 = const()[name = string("op_5204"), val = bool(true)];
            tensor<int32, [1]> var_5205_axes_0 = const()[name = string("op_5205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_5205_cast_fp16 = reduce_max(axes = var_5205_axes_0, keep_dims = var_5204, x = var_5201_cast_fp16)[name = string("op_5205_cast_fp16")];
            fp16 var_5206_to_fp16 = const()[name = string("op_5206_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_151_cast_fp16 = maximum(x = var_5205_cast_fp16, y = var_5206_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_151_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_151_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5209_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5209_cast_fp16")];
            tensor<int32, [1]> var_5214_axes_0 = const()[name = string("op_5214_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5214_keep_dims_0 = const()[name = string("op_5214_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_5214_cast_fp16 = reduce_mean(axes = var_5214_axes_0, keep_dims = var_5214_keep_dims_0, x = var_5209_cast_fp16)[name = string("op_5214_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5215_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_5215_cast_fp16")];
            fp32 var_5216_epsilon_0 = const()[name = string("op_5216_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_5216_cast_fp16 = inverse(epsilon = var_5216_epsilon_0, x = var_5215_cast_fp16)[name = string("op_5216_cast_fp16")];
            fp16 var_5217_to_fp16 = const()[name = string("op_5217_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_5218_cast_fp16 = mul(x = var_5216_cast_fp16, y = var_5217_to_fp16)[name = string("op_5218_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5220_cast_fp16 = add(x = var_5214_cast_fp16, y = var_5218_cast_fp16)[name = string("op_5220_cast_fp16")];
            fp32 var_5221_epsilon_0 = const()[name = string("op_5221_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_5221_cast_fp16 = rsqrt(epsilon = var_5221_epsilon_0, x = var_5220_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_133_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5221_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [256]> tm_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858112)))];
            tensor<fp16, [1, 2, 256, 256]> k_55_cast_fp16 = mul(x = out_133_cast_fp16, y = tm_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("k_55_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_151_cast_fp16 = transpose(perm = x_151_perm_0, x = var_5174_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 2, 256, 256]> var_5224_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_5224_cast_fp16")];
            bool var_5227 = const()[name = string("op_5227"), val = bool(true)];
            tensor<int32, [1]> var_5228_axes_0 = const()[name = string("op_5228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_5228_cast_fp16 = reduce_max(axes = var_5228_axes_0, keep_dims = var_5227, x = var_5224_cast_fp16)[name = string("op_5228_cast_fp16")];
            fp16 var_5229_to_fp16 = const()[name = string("op_5229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_153_cast_fp16 = maximum(x = var_5228_cast_fp16, y = var_5229_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_153_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_153_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5232_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5232_cast_fp16")];
            tensor<int32, [1]> var_5237_axes_0 = const()[name = string("op_5237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5237_keep_dims_0 = const()[name = string("op_5237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_5237_cast_fp16 = reduce_mean(axes = var_5237_axes_0, keep_dims = var_5237_keep_dims_0, x = var_5232_cast_fp16)[name = string("op_5237_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5238_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5238_cast_fp16")];
            fp32 var_5239_epsilon_0 = const()[name = string("op_5239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_5239_cast_fp16 = inverse(epsilon = var_5239_epsilon_0, x = var_5238_cast_fp16)[name = string("op_5239_cast_fp16")];
            fp16 var_5240_to_fp16 = const()[name = string("op_5240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_5241_cast_fp16 = mul(x = var_5239_cast_fp16, y = var_5240_to_fp16)[name = string("op_5241_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5243_cast_fp16 = add(x = var_5237_cast_fp16, y = var_5241_cast_fp16)[name = string("op_5243_cast_fp16")];
            fp32 var_5244_epsilon_0 = const()[name = string("op_5244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_5244_cast_fp16 = rsqrt(epsilon = var_5244_epsilon_0, x = var_5243_cast_fp16)[name = string("op_5244_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_37_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5244_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5246_cast_fp16 = mul(x = out_131_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5246_cast_fp16")];
            tensor<int32, [4]> var_5251_begin_0 = const()[name = string("op_5251_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5251_end_0 = const()[name = string("op_5251_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_5251_end_mask_0 = const()[name = string("op_5251_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_5251_cast_fp16 = slice_by_index(begin = var_5251_begin_0, end = var_5251_end_0, end_mask = var_5251_end_mask_0, x = out_131_cast_fp16)[name = string("op_5251_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_5252_cast_fp16 = mul(x = var_5251_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_5252_cast_fp16")];
            tensor<int32, [4]> var_5257_begin_0 = const()[name = string("op_5257_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5257_end_0 = const()[name = string("op_5257_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_5257_end_mask_0 = const()[name = string("op_5257_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_5257_cast_fp16 = slice_by_index(begin = var_5257_begin_0, end = var_5257_end_0, end_mask = var_5257_end_mask_0, x = out_131_cast_fp16)[name = string("op_5257_cast_fp16")];
            int32 var_5259 = const()[name = string("op_5259"), val = int32(-1)];
            bool var_5260_interleave_0 = const()[name = string("op_5260_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_5260_cast_fp16 = concat(axis = var_5259, interleave = var_5260_interleave_0, values = (var_5252_cast_fp16, var_5257_cast_fp16))[name = string("op_5260_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5261_cast_fp16 = mul(x = var_5260_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5261_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_39_cast_fp16 = add(x = var_5246_cast_fp16, y = var_5261_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5264_cast_fp16 = mul(x = k_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5264_cast_fp16")];
            tensor<int32, [4]> var_5269_begin_0 = const()[name = string("op_5269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5269_end_0 = const()[name = string("op_5269_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_5269_end_mask_0 = const()[name = string("op_5269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_5269_cast_fp16 = slice_by_index(begin = var_5269_begin_0, end = var_5269_end_0, end_mask = var_5269_end_mask_0, x = k_55_cast_fp16)[name = string("op_5269_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_5270_cast_fp16 = mul(x = var_5269_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_5270_cast_fp16")];
            tensor<int32, [4]> var_5275_begin_0 = const()[name = string("op_5275_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5275_end_0 = const()[name = string("op_5275_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_5275_end_mask_0 = const()[name = string("op_5275_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_5275_cast_fp16 = slice_by_index(begin = var_5275_begin_0, end = var_5275_end_0, end_mask = var_5275_end_mask_0, x = k_55_cast_fp16)[name = string("op_5275_cast_fp16")];
            int32 var_5277 = const()[name = string("op_5277"), val = int32(-1)];
            bool var_5278_interleave_0 = const()[name = string("op_5278_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_5278_cast_fp16 = concat(axis = var_5277, interleave = var_5278_interleave_0, values = (var_5270_cast_fp16, var_5275_cast_fp16))[name = string("op_5278_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5279_cast_fp16 = mul(x = var_5278_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5279_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_57_cast_fp16 = add(x = var_5264_cast_fp16, y = var_5279_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_18_reps_0 = const()[name = string("tile_18_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = k_57_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [4, 1, 256, 256]> tile_18_cast_fp16 = tile(reps = tile_18_reps_0, x = transpose_36_cast_fp16)[name = string("tile_18_cast_fp16")];
            tensor<int32, [5]> concat_36 = const()[name = string("concat_36"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_36_cast_fp16 = reshape(shape = concat_36, x = tile_18_cast_fp16)[name = string("reshape_36_cast_fp16")];
            tensor<int32, [5]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_37 = const()[name = string("concat_37"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = reshape_36_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [4, 1, 256, 256]> reshape_37_cast_fp16 = reshape(shape = concat_37, x = transpose_37_cast_fp16)[name = string("reshape_37_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_19_reps_0 = const()[name = string("tile_19_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = v_37_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [4, 1, 256, 256]> tile_19_cast_fp16 = tile(reps = tile_19_reps_0, x = transpose_38_cast_fp16)[name = string("tile_19_cast_fp16")];
            tensor<int32, [5]> concat_38 = const()[name = string("concat_38"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_38_cast_fp16 = reshape(shape = concat_38, x = tile_19_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [5]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [4, 1, 256, 256]> reshape_39_cast_fp16 = reshape(shape = concat_39, x = transpose_39_cast_fp16)[name = string("reshape_39_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5293_transpose_x_0 = const()[name = string("op_5293_transpose_x_0"), val = bool(false)];
            bool var_5293_transpose_y_0 = const()[name = string("op_5293_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = reshape_37_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 4, 256, 256]> var_5293_cast_fp16 = matmul(transpose_x = var_5293_transpose_x_0, transpose_y = var_5293_transpose_y_0, x = q_39_cast_fp16, y = transpose_105_cast_fp16)[name = string("op_5293_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5295_cast_fp16 = add(x = var_5293_cast_fp16, y = attention_bias)[name = string("op_5295_cast_fp16")];
            int32 var_5296 = const()[name = string("op_5296"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_5298_cast_fp16 = softmax(axis = var_5296, x = var_5295_cast_fp16)[name = string("op_5298_cast_fp16")];
            bool var_5299_transpose_x_0 = const()[name = string("op_5299_transpose_x_0"), val = bool(false)];
            bool var_5299_transpose_y_0 = const()[name = string("op_5299_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = reshape_39_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 4, 256, 256]> var_5299_cast_fp16 = matmul(transpose_x = var_5299_transpose_x_0, transpose_y = var_5299_transpose_y_0, x = var_5298_cast_fp16, y = v_39_cast_fp16)[name = string("op_5299_cast_fp16")];
            tensor<int32, [4]> var_5302_perm_0 = const()[name = string("op_5302_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5306 = const()[name = string("op_5306"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_5302_cast_fp16 = transpose(perm = var_5302_perm_0, x = var_5299_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 256, 1024]> input_111_cast_fp16 = reshape(shape = var_5306, x = var_5302_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55858688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56383040))))[name = string("tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5311_cast_fp16 = abs(x = linear_85_cast_fp16)[name = string("op_5311_cast_fp16")];
            bool var_5314 = const()[name = string("op_5314"), val = bool(true)];
            tensor<int32, [1]> var_5315_axes_0 = const()[name = string("op_5315_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5315_cast_fp16 = reduce_max(axes = var_5315_axes_0, keep_dims = var_5314, x = var_5311_cast_fp16)[name = string("op_5315_cast_fp16")];
            fp16 var_5316_to_fp16 = const()[name = string("op_5316_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_155_cast_fp16 = maximum(x = var_5315_cast_fp16, y = var_5316_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_155_cast_fp16 = real_div(x = linear_85_cast_fp16, y = s_155_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5319_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5319_cast_fp16")];
            tensor<int32, [1]> var_5324_axes_0 = const()[name = string("op_5324_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5324_keep_dims_0 = const()[name = string("op_5324_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5324_cast_fp16 = reduce_mean(axes = var_5324_axes_0, keep_dims = var_5324_keep_dims_0, x = var_5319_cast_fp16)[name = string("op_5324_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5325_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5325_cast_fp16")];
            fp32 var_5326_epsilon_0 = const()[name = string("op_5326_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5326_cast_fp16 = inverse(epsilon = var_5326_epsilon_0, x = var_5325_cast_fp16)[name = string("op_5326_cast_fp16")];
            fp16 var_5327_to_fp16 = const()[name = string("op_5327_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5328_cast_fp16 = mul(x = var_5326_cast_fp16, y = var_5327_to_fp16)[name = string("op_5328_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5330_cast_fp16 = add(x = var_5324_cast_fp16, y = var_5328_cast_fp16)[name = string("op_5330_cast_fp16")];
            fp32 var_5331_epsilon_0 = const()[name = string("op_5331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5331_cast_fp16 = rsqrt(epsilon = var_5331_epsilon_0, x = var_5330_cast_fp16)[name = string("op_5331_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_135_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5331_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56384128)))];
            tensor<fp16, [1, 256, 512]> var_5333_cast_fp16 = mul(x = out_135_cast_fp16, y = tm_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_5333_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = var_5333_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5336_cast_fp16 = abs(x = x_155_cast_fp16)[name = string("op_5336_cast_fp16")];
            bool var_5339 = const()[name = string("op_5339"), val = bool(true)];
            tensor<int32, [1]> var_5340_axes_0 = const()[name = string("op_5340_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5340_cast_fp16 = reduce_max(axes = var_5340_axes_0, keep_dims = var_5339, x = var_5336_cast_fp16)[name = string("op_5340_cast_fp16")];
            fp16 var_5341_to_fp16 = const()[name = string("op_5341_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_157_cast_fp16 = maximum(x = var_5340_cast_fp16, y = var_5341_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_157_cast_fp16 = real_div(x = x_155_cast_fp16, y = s_157_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5344_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5344_cast_fp16")];
            tensor<int32, [1]> var_5349_axes_0 = const()[name = string("op_5349_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5349_keep_dims_0 = const()[name = string("op_5349_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5349_cast_fp16 = reduce_mean(axes = var_5349_axes_0, keep_dims = var_5349_keep_dims_0, x = var_5344_cast_fp16)[name = string("op_5349_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5350_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5350_cast_fp16")];
            fp32 var_5351_epsilon_0 = const()[name = string("op_5351_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5351_cast_fp16 = inverse(epsilon = var_5351_epsilon_0, x = var_5350_cast_fp16)[name = string("op_5351_cast_fp16")];
            fp16 var_5352_to_fp16 = const()[name = string("op_5352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5353_cast_fp16 = mul(x = var_5351_cast_fp16, y = var_5352_to_fp16)[name = string("op_5353_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5355_cast_fp16 = add(x = var_5349_cast_fp16, y = var_5353_cast_fp16)[name = string("op_5355_cast_fp16")];
            fp32 var_5356_epsilon_0 = const()[name = string("op_5356_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5356_cast_fp16 = rsqrt(epsilon = var_5356_epsilon_0, x = var_5355_cast_fp16)[name = string("op_5356_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_137_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5356_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56385216)))];
            tensor<fp16, [1, 256, 512]> input_113_cast_fp16 = mul(x = out_137_cast_fp16, y = tm_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56386304))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57434944))))[name = string("tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_86_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_86_cast_fp16")];
            string var_5363_mode_0 = const()[name = string("op_5363_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_5363_cast_fp16 = gelu(mode = var_5363_mode_0, x = linear_86_cast_fp16)[name = string("op_5363_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57439104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58487744))))[name = string("tm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_87_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_115_cast_fp16 = mul(x = var_5363_cast_fp16, y = linear_87_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58491904))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59540544))))[name = string("tm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5371_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5371_cast_fp16")];
            bool var_5374 = const()[name = string("op_5374"), val = bool(true)];
            tensor<int32, [1]> var_5375_axes_0 = const()[name = string("op_5375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5375_cast_fp16 = reduce_max(axes = var_5375_axes_0, keep_dims = var_5374, x = var_5371_cast_fp16)[name = string("op_5375_cast_fp16")];
            fp16 var_5376_to_fp16 = const()[name = string("op_5376_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_159_cast_fp16 = maximum(x = var_5375_cast_fp16, y = var_5376_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_159_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_159_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5379_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5379_cast_fp16")];
            tensor<int32, [1]> var_5384_axes_0 = const()[name = string("op_5384_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5384_keep_dims_0 = const()[name = string("op_5384_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5384_cast_fp16 = reduce_mean(axes = var_5384_axes_0, keep_dims = var_5384_keep_dims_0, x = var_5379_cast_fp16)[name = string("op_5384_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5385_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5385_cast_fp16")];
            fp32 var_5386_epsilon_0 = const()[name = string("op_5386_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5386_cast_fp16 = inverse(epsilon = var_5386_epsilon_0, x = var_5385_cast_fp16)[name = string("op_5386_cast_fp16")];
            fp16 var_5387_to_fp16 = const()[name = string("op_5387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5388_cast_fp16 = mul(x = var_5386_cast_fp16, y = var_5387_to_fp16)[name = string("op_5388_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5390_cast_fp16 = add(x = var_5384_cast_fp16, y = var_5388_cast_fp16)[name = string("op_5390_cast_fp16")];
            fp32 var_5391_epsilon_0 = const()[name = string("op_5391_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5391_cast_fp16 = rsqrt(epsilon = var_5391_epsilon_0, x = var_5390_cast_fp16)[name = string("op_5391_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_139_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5391_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59541632)))];
            tensor<fp16, [1, 256, 512]> var_5393_cast_fp16 = mul(x = out_139_cast_fp16, y = tm_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_117_cast_fp16 = add(x = x_155_cast_fp16, y = var_5393_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59542720))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59804928))))[name = string("tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_89_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_89_cast_fp16")];
            string var_5400_mode_0 = const()[name = string("op_5400_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_5400_cast_fp16 = gelu(mode = var_5400_mode_0, x = linear_89_cast_fp16)[name = string("op_5400_cast_fp16")];
            tensor<int32, [4]> var_5413_begin_0 = const()[name = string("op_5413_begin_0"), val = tensor<int32, [4]>([0, 0, 9, 0])];
            tensor<int32, [4]> var_5413_end_0 = const()[name = string("op_5413_end_0"), val = tensor<int32, [4]>([1, 256, 10, 512])];
            tensor<bool, [4]> var_5413_end_mask_0 = const()[name = string("op_5413_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5413_squeeze_mask_0 = const()[name = string("op_5413_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_5413_cast_fp16 = slice_by_index(begin = var_5413_begin_0, end = var_5413_end_0, end_mask = var_5413_end_mask_0, squeeze_mask = var_5413_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_5413_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_119_cast_fp16 = mul(x = var_5400_cast_fp16, y = var_5413_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59806016))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60068224))))[name = string("tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_9_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5423_cast_fp16 = abs(x = linear_90_cast_fp16)[name = string("op_5423_cast_fp16")];
            bool var_5426 = const()[name = string("op_5426"), val = bool(true)];
            tensor<int32, [1]> var_5427_axes_0 = const()[name = string("op_5427_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5427_cast_fp16 = reduce_max(axes = var_5427_axes_0, keep_dims = var_5426, x = var_5423_cast_fp16)[name = string("op_5427_cast_fp16")];
            fp16 var_5428_to_fp16 = const()[name = string("op_5428_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_161_cast_fp16 = maximum(x = var_5427_cast_fp16, y = var_5428_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_161_cast_fp16 = real_div(x = linear_90_cast_fp16, y = s_161_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5431_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5431_cast_fp16")];
            tensor<int32, [1]> var_5436_axes_0 = const()[name = string("op_5436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5436_keep_dims_0 = const()[name = string("op_5436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5436_cast_fp16 = reduce_mean(axes = var_5436_axes_0, keep_dims = var_5436_keep_dims_0, x = var_5431_cast_fp16)[name = string("op_5436_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5437_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5437_cast_fp16")];
            fp32 var_5438_epsilon_0 = const()[name = string("op_5438_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5438_cast_fp16 = inverse(epsilon = var_5438_epsilon_0, x = var_5437_cast_fp16)[name = string("op_5438_cast_fp16")];
            fp16 var_5439_to_fp16 = const()[name = string("op_5439_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5440_cast_fp16 = mul(x = var_5438_cast_fp16, y = var_5439_to_fp16)[name = string("op_5440_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5442_cast_fp16 = add(x = var_5436_cast_fp16, y = var_5440_cast_fp16)[name = string("op_5442_cast_fp16")];
            fp32 var_5443_epsilon_0 = const()[name = string("op_5443_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5443_cast_fp16 = rsqrt(epsilon = var_5443_epsilon_0, x = var_5442_cast_fp16)[name = string("op_5443_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_141_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5443_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [512]> tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60069312)))];
            tensor<fp16, [1, 256, 512]> var_5445_cast_fp16 = mul(x = out_141_cast_fp16, y = tm_layers_9_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5445_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_19_cast_fp16 = add(x = input_117_cast_fp16, y = var_5445_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<fp16, [1]> tm_layers_9_layer_scalar_to_fp16 = const()[name = string("tm_layers_9_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9ep-1])];
            tensor<fp16, [1, 256, 512]> x_161_cast_fp16 = mul(x = h_19_cast_fp16, y = tm_layers_9_layer_scalar_to_fp16)[name = string("x_161_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5449_cast_fp16 = abs(x = x_161_cast_fp16)[name = string("op_5449_cast_fp16")];
            bool var_5452 = const()[name = string("op_5452"), val = bool(true)];
            tensor<int32, [1]> var_5453_axes_0 = const()[name = string("op_5453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5453_cast_fp16 = reduce_max(axes = var_5453_axes_0, keep_dims = var_5452, x = var_5449_cast_fp16)[name = string("op_5453_cast_fp16")];
            fp16 var_5454_to_fp16 = const()[name = string("op_5454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_163_cast_fp16 = maximum(x = var_5453_cast_fp16, y = var_5454_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_163_cast_fp16 = real_div(x = x_161_cast_fp16, y = s_163_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5457_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5457_cast_fp16")];
            tensor<int32, [1]> var_5462_axes_0 = const()[name = string("op_5462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5462_keep_dims_0 = const()[name = string("op_5462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5462_cast_fp16 = reduce_mean(axes = var_5462_axes_0, keep_dims = var_5462_keep_dims_0, x = var_5457_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5463_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5463_cast_fp16")];
            fp32 var_5464_epsilon_0 = const()[name = string("op_5464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5464_cast_fp16 = inverse(epsilon = var_5464_epsilon_0, x = var_5463_cast_fp16)[name = string("op_5464_cast_fp16")];
            fp16 var_5465_to_fp16 = const()[name = string("op_5465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5466_cast_fp16 = mul(x = var_5464_cast_fp16, y = var_5465_to_fp16)[name = string("op_5466_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5468_cast_fp16 = add(x = var_5462_cast_fp16, y = var_5466_cast_fp16)[name = string("op_5468_cast_fp16")];
            fp32 var_5469_epsilon_0 = const()[name = string("op_5469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5469_cast_fp16 = rsqrt(epsilon = var_5469_epsilon_0, x = var_5468_cast_fp16)[name = string("op_5469_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_143_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5469_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60070400)))];
            tensor<fp16, [1, 256, 512]> input_121_cast_fp16 = mul(x = out_143_cast_fp16, y = tm_layers_10_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60071488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60595840))))[name = string("tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5479 = const()[name = string("op_5479"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_5480_cast_fp16 = reshape(shape = var_5479, x = linear_91_cast_fp16)[name = string("op_5480_cast_fp16")];
            tensor<int32, [4]> x_163_perm_0 = const()[name = string("x_163_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60597952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60860160))))[name = string("tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5491 = const()[name = string("op_5491"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_5492_cast_fp16 = reshape(shape = var_5491, x = linear_92_cast_fp16)[name = string("op_5492_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60861248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61123456))))[name = string("tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5503 = const()[name = string("op_5503"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_5504_cast_fp16 = reshape(shape = var_5503, x = linear_93_cast_fp16)[name = string("op_5504_cast_fp16")];
            tensor<int32, [4]> x_167_perm_0 = const()[name = string("x_167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_163_cast_fp16 = transpose(perm = x_163_perm_0, x = var_5480_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 4, 256, 256]> var_5508_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_5508_cast_fp16")];
            bool var_5511 = const()[name = string("op_5511"), val = bool(true)];
            tensor<int32, [1]> var_5512_axes_0 = const()[name = string("op_5512_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_5512_cast_fp16 = reduce_max(axes = var_5512_axes_0, keep_dims = var_5511, x = var_5508_cast_fp16)[name = string("op_5512_cast_fp16")];
            fp16 var_5513_to_fp16 = const()[name = string("op_5513_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_165_cast_fp16 = maximum(x = var_5512_cast_fp16, y = var_5513_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_165_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_165_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5516_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5516_cast_fp16")];
            tensor<int32, [1]> var_5521_axes_0 = const()[name = string("op_5521_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5521_keep_dims_0 = const()[name = string("op_5521_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_5521_cast_fp16 = reduce_mean(axes = var_5521_axes_0, keep_dims = var_5521_keep_dims_0, x = var_5516_cast_fp16)[name = string("op_5521_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5522_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5522_cast_fp16")];
            fp32 var_5523_epsilon_0 = const()[name = string("op_5523_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_5523_cast_fp16 = inverse(epsilon = var_5523_epsilon_0, x = var_5522_cast_fp16)[name = string("op_5523_cast_fp16")];
            fp16 var_5524_to_fp16 = const()[name = string("op_5524_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_5525_cast_fp16 = mul(x = var_5523_cast_fp16, y = var_5524_to_fp16)[name = string("op_5525_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5527_cast_fp16 = add(x = var_5521_cast_fp16, y = var_5525_cast_fp16)[name = string("op_5527_cast_fp16")];
            fp32 var_5528_epsilon_0 = const()[name = string("op_5528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_5528_cast_fp16 = rsqrt(epsilon = var_5528_epsilon_0, x = var_5527_cast_fp16)[name = string("op_5528_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_145_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5528_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [256]> tm_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61124544)))];
            tensor<fp16, [1, 4, 256, 256]> q_41_cast_fp16 = mul(x = out_145_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_41_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_5492_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 2, 256, 256]> var_5531_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_5531_cast_fp16")];
            bool var_5534 = const()[name = string("op_5534"), val = bool(true)];
            tensor<int32, [1]> var_5535_axes_0 = const()[name = string("op_5535_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_5535_cast_fp16 = reduce_max(axes = var_5535_axes_0, keep_dims = var_5534, x = var_5531_cast_fp16)[name = string("op_5535_cast_fp16")];
            fp16 var_5536_to_fp16 = const()[name = string("op_5536_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_167_cast_fp16 = maximum(x = var_5535_cast_fp16, y = var_5536_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_167_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_167_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5539_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5539_cast_fp16")];
            tensor<int32, [1]> var_5544_axes_0 = const()[name = string("op_5544_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5544_keep_dims_0 = const()[name = string("op_5544_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_5544_cast_fp16 = reduce_mean(axes = var_5544_axes_0, keep_dims = var_5544_keep_dims_0, x = var_5539_cast_fp16)[name = string("op_5544_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5545_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5545_cast_fp16")];
            fp32 var_5546_epsilon_0 = const()[name = string("op_5546_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_5546_cast_fp16 = inverse(epsilon = var_5546_epsilon_0, x = var_5545_cast_fp16)[name = string("op_5546_cast_fp16")];
            fp16 var_5547_to_fp16 = const()[name = string("op_5547_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_5548_cast_fp16 = mul(x = var_5546_cast_fp16, y = var_5547_to_fp16)[name = string("op_5548_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5550_cast_fp16 = add(x = var_5544_cast_fp16, y = var_5548_cast_fp16)[name = string("op_5550_cast_fp16")];
            fp32 var_5551_epsilon_0 = const()[name = string("op_5551_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_5551_cast_fp16 = rsqrt(epsilon = var_5551_epsilon_0, x = var_5550_cast_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_147_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5551_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [256]> tm_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125120)))];
            tensor<fp16, [1, 2, 256, 256]> k_61_cast_fp16 = mul(x = out_147_cast_fp16, y = tm_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_167_cast_fp16 = transpose(perm = x_167_perm_0, x = var_5504_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 2, 256, 256]> var_5554_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_5554_cast_fp16")];
            bool var_5557 = const()[name = string("op_5557"), val = bool(true)];
            tensor<int32, [1]> var_5558_axes_0 = const()[name = string("op_5558_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_5558_cast_fp16 = reduce_max(axes = var_5558_axes_0, keep_dims = var_5557, x = var_5554_cast_fp16)[name = string("op_5558_cast_fp16")];
            fp16 var_5559_to_fp16 = const()[name = string("op_5559_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_169_cast_fp16 = maximum(x = var_5558_cast_fp16, y = var_5559_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_169_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_169_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5562_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5562_cast_fp16")];
            tensor<int32, [1]> var_5567_axes_0 = const()[name = string("op_5567_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5567_keep_dims_0 = const()[name = string("op_5567_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_5567_cast_fp16 = reduce_mean(axes = var_5567_axes_0, keep_dims = var_5567_keep_dims_0, x = var_5562_cast_fp16)[name = string("op_5567_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5568_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5568_cast_fp16")];
            fp32 var_5569_epsilon_0 = const()[name = string("op_5569_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_5569_cast_fp16 = inverse(epsilon = var_5569_epsilon_0, x = var_5568_cast_fp16)[name = string("op_5569_cast_fp16")];
            fp16 var_5570_to_fp16 = const()[name = string("op_5570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_5571_cast_fp16 = mul(x = var_5569_cast_fp16, y = var_5570_to_fp16)[name = string("op_5571_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_5573_cast_fp16 = add(x = var_5567_cast_fp16, y = var_5571_cast_fp16)[name = string("op_5573_cast_fp16")];
            fp32 var_5574_epsilon_0 = const()[name = string("op_5574_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_5574_cast_fp16 = rsqrt(epsilon = var_5574_epsilon_0, x = var_5573_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_41_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5574_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5576_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5576_cast_fp16")];
            tensor<int32, [4]> var_5581_begin_0 = const()[name = string("op_5581_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5581_end_0 = const()[name = string("op_5581_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_5581_end_mask_0 = const()[name = string("op_5581_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_5581_cast_fp16 = slice_by_index(begin = var_5581_begin_0, end = var_5581_end_0, end_mask = var_5581_end_mask_0, x = q_41_cast_fp16)[name = string("op_5581_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_5582_cast_fp16 = mul(x = var_5581_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_5582_cast_fp16")];
            tensor<int32, [4]> var_5587_begin_0 = const()[name = string("op_5587_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5587_end_0 = const()[name = string("op_5587_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_5587_end_mask_0 = const()[name = string("op_5587_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_5587_cast_fp16 = slice_by_index(begin = var_5587_begin_0, end = var_5587_end_0, end_mask = var_5587_end_mask_0, x = q_41_cast_fp16)[name = string("op_5587_cast_fp16")];
            int32 var_5589 = const()[name = string("op_5589"), val = int32(-1)];
            bool var_5590_interleave_0 = const()[name = string("op_5590_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_5590_cast_fp16 = concat(axis = var_5589, interleave = var_5590_interleave_0, values = (var_5582_cast_fp16, var_5587_cast_fp16))[name = string("op_5590_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5591_cast_fp16 = mul(x = var_5590_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5591_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_43_cast_fp16 = add(x = var_5576_cast_fp16, y = var_5591_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5594_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5594_cast_fp16")];
            tensor<int32, [4]> var_5599_begin_0 = const()[name = string("op_5599_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_5599_end_0 = const()[name = string("op_5599_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_5599_end_mask_0 = const()[name = string("op_5599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_5599_cast_fp16 = slice_by_index(begin = var_5599_begin_0, end = var_5599_end_0, end_mask = var_5599_end_mask_0, x = k_61_cast_fp16)[name = string("op_5599_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_5600_cast_fp16 = mul(x = var_5599_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_5600_cast_fp16")];
            tensor<int32, [4]> var_5605_begin_0 = const()[name = string("op_5605_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5605_end_0 = const()[name = string("op_5605_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_5605_end_mask_0 = const()[name = string("op_5605_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_5605_cast_fp16 = slice_by_index(begin = var_5605_begin_0, end = var_5605_end_0, end_mask = var_5605_end_mask_0, x = k_61_cast_fp16)[name = string("op_5605_cast_fp16")];
            int32 var_5607 = const()[name = string("op_5607"), val = int32(-1)];
            bool var_5608_interleave_0 = const()[name = string("op_5608_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_5608_cast_fp16 = concat(axis = var_5607, interleave = var_5608_interleave_0, values = (var_5600_cast_fp16, var_5605_cast_fp16))[name = string("op_5608_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_5609_cast_fp16 = mul(x = var_5608_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5609_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_63_cast_fp16 = add(x = var_5594_cast_fp16, y = var_5609_cast_fp16)[name = string("k_63_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_20_reps_0 = const()[name = string("tile_20_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = k_63_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [4, 1, 256, 256]> tile_20_cast_fp16 = tile(reps = tile_20_reps_0, x = transpose_40_cast_fp16)[name = string("tile_20_cast_fp16")];
            tensor<int32, [5]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_40_cast_fp16 = reshape(shape = concat_40, x = tile_20_cast_fp16)[name = string("reshape_40_cast_fp16")];
            tensor<int32, [5]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_41 = const()[name = string("concat_41"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = reshape_40_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [4, 1, 256, 256]> reshape_41_cast_fp16 = reshape(shape = concat_41, x = transpose_41_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_21_reps_0 = const()[name = string("tile_21_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = v_41_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [4, 1, 256, 256]> tile_21_cast_fp16 = tile(reps = tile_21_reps_0, x = transpose_42_cast_fp16)[name = string("tile_21_cast_fp16")];
            tensor<int32, [5]> concat_42 = const()[name = string("concat_42"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_42_cast_fp16 = reshape(shape = concat_42, x = tile_21_cast_fp16)[name = string("reshape_42_cast_fp16")];
            tensor<int32, [5]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = reshape_42_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [4, 1, 256, 256]> reshape_43_cast_fp16 = reshape(shape = concat_43, x = transpose_43_cast_fp16)[name = string("reshape_43_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5623_transpose_x_0 = const()[name = string("op_5623_transpose_x_0"), val = bool(false)];
            bool var_5623_transpose_y_0 = const()[name = string("op_5623_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 4, 256, 256]> var_5623_cast_fp16 = matmul(transpose_x = var_5623_transpose_x_0, transpose_y = var_5623_transpose_y_0, x = q_43_cast_fp16, y = transpose_106_cast_fp16)[name = string("op_5623_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5625_cast_fp16 = add(x = var_5623_cast_fp16, y = attention_bias)[name = string("op_5625_cast_fp16")];
            int32 var_5626 = const()[name = string("op_5626"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_5628_cast_fp16 = softmax(axis = var_5626, x = var_5625_cast_fp16)[name = string("op_5628_cast_fp16")];
            bool var_5629_transpose_x_0 = const()[name = string("op_5629_transpose_x_0"), val = bool(false)];
            bool var_5629_transpose_y_0 = const()[name = string("op_5629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = reshape_43_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 4, 256, 256]> var_5629_cast_fp16 = matmul(transpose_x = var_5629_transpose_x_0, transpose_y = var_5629_transpose_y_0, x = var_5628_cast_fp16, y = v_43_cast_fp16)[name = string("op_5629_cast_fp16")];
            tensor<int32, [4]> var_5632_perm_0 = const()[name = string("op_5632_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5636 = const()[name = string("op_5636"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_5632_cast_fp16 = transpose(perm = var_5632_perm_0, x = var_5629_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 256, 1024]> input_123_cast_fp16 = reshape(shape = var_5636, x = var_5632_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61125696))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61650048))))[name = string("tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5641_cast_fp16 = abs(x = linear_94_cast_fp16)[name = string("op_5641_cast_fp16")];
            bool var_5644 = const()[name = string("op_5644"), val = bool(true)];
            tensor<int32, [1]> var_5645_axes_0 = const()[name = string("op_5645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5645_cast_fp16 = reduce_max(axes = var_5645_axes_0, keep_dims = var_5644, x = var_5641_cast_fp16)[name = string("op_5645_cast_fp16")];
            fp16 var_5646_to_fp16 = const()[name = string("op_5646_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_171_cast_fp16 = maximum(x = var_5645_cast_fp16, y = var_5646_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_171_cast_fp16 = real_div(x = linear_94_cast_fp16, y = s_171_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5649_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5649_cast_fp16")];
            tensor<int32, [1]> var_5654_axes_0 = const()[name = string("op_5654_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5654_keep_dims_0 = const()[name = string("op_5654_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5654_cast_fp16 = reduce_mean(axes = var_5654_axes_0, keep_dims = var_5654_keep_dims_0, x = var_5649_cast_fp16)[name = string("op_5654_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5655_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5655_cast_fp16")];
            fp32 var_5656_epsilon_0 = const()[name = string("op_5656_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5656_cast_fp16 = inverse(epsilon = var_5656_epsilon_0, x = var_5655_cast_fp16)[name = string("op_5656_cast_fp16")];
            fp16 var_5657_to_fp16 = const()[name = string("op_5657_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5658_cast_fp16 = mul(x = var_5656_cast_fp16, y = var_5657_to_fp16)[name = string("op_5658_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5660_cast_fp16 = add(x = var_5654_cast_fp16, y = var_5658_cast_fp16)[name = string("op_5660_cast_fp16")];
            fp32 var_5661_epsilon_0 = const()[name = string("op_5661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5661_cast_fp16 = rsqrt(epsilon = var_5661_epsilon_0, x = var_5660_cast_fp16)[name = string("op_5661_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_149_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5661_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61651136)))];
            tensor<fp16, [1, 256, 512]> var_5663_cast_fp16 = mul(x = out_149_cast_fp16, y = tm_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5663_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_171_cast_fp16 = add(x = x_161_cast_fp16, y = var_5663_cast_fp16)[name = string("x_171_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5666_cast_fp16 = abs(x = x_171_cast_fp16)[name = string("op_5666_cast_fp16")];
            bool var_5669 = const()[name = string("op_5669"), val = bool(true)];
            tensor<int32, [1]> var_5670_axes_0 = const()[name = string("op_5670_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5670_cast_fp16 = reduce_max(axes = var_5670_axes_0, keep_dims = var_5669, x = var_5666_cast_fp16)[name = string("op_5670_cast_fp16")];
            fp16 var_5671_to_fp16 = const()[name = string("op_5671_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_173_cast_fp16 = maximum(x = var_5670_cast_fp16, y = var_5671_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_173_cast_fp16 = real_div(x = x_171_cast_fp16, y = s_173_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5674_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5674_cast_fp16")];
            tensor<int32, [1]> var_5679_axes_0 = const()[name = string("op_5679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5679_keep_dims_0 = const()[name = string("op_5679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5679_cast_fp16 = reduce_mean(axes = var_5679_axes_0, keep_dims = var_5679_keep_dims_0, x = var_5674_cast_fp16)[name = string("op_5679_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5680_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5680_cast_fp16")];
            fp32 var_5681_epsilon_0 = const()[name = string("op_5681_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5681_cast_fp16 = inverse(epsilon = var_5681_epsilon_0, x = var_5680_cast_fp16)[name = string("op_5681_cast_fp16")];
            fp16 var_5682_to_fp16 = const()[name = string("op_5682_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5683_cast_fp16 = mul(x = var_5681_cast_fp16, y = var_5682_to_fp16)[name = string("op_5683_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5685_cast_fp16 = add(x = var_5679_cast_fp16, y = var_5683_cast_fp16)[name = string("op_5685_cast_fp16")];
            fp32 var_5686_epsilon_0 = const()[name = string("op_5686_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5686_cast_fp16 = rsqrt(epsilon = var_5686_epsilon_0, x = var_5685_cast_fp16)[name = string("op_5686_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_151_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5686_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61652224)))];
            tensor<fp16, [1, 256, 512]> input_125_cast_fp16 = mul(x = out_151_cast_fp16, y = tm_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61653312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62701952))))[name = string("tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_95_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_95_cast_fp16")];
            string var_5693_mode_0 = const()[name = string("op_5693_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_5693_cast_fp16 = gelu(mode = var_5693_mode_0, x = linear_95_cast_fp16)[name = string("op_5693_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62706112))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63754752))))[name = string("tm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_127_cast_fp16 = mul(x = var_5693_cast_fp16, y = linear_96_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63758912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64807552))))[name = string("tm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5701_cast_fp16 = abs(x = linear_97_cast_fp16)[name = string("op_5701_cast_fp16")];
            bool var_5704 = const()[name = string("op_5704"), val = bool(true)];
            tensor<int32, [1]> var_5705_axes_0 = const()[name = string("op_5705_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5705_cast_fp16 = reduce_max(axes = var_5705_axes_0, keep_dims = var_5704, x = var_5701_cast_fp16)[name = string("op_5705_cast_fp16")];
            fp16 var_5706_to_fp16 = const()[name = string("op_5706_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_175_cast_fp16 = maximum(x = var_5705_cast_fp16, y = var_5706_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_175_cast_fp16 = real_div(x = linear_97_cast_fp16, y = s_175_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5709_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5709_cast_fp16")];
            tensor<int32, [1]> var_5714_axes_0 = const()[name = string("op_5714_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5714_keep_dims_0 = const()[name = string("op_5714_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5714_cast_fp16 = reduce_mean(axes = var_5714_axes_0, keep_dims = var_5714_keep_dims_0, x = var_5709_cast_fp16)[name = string("op_5714_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5715_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5715_cast_fp16")];
            fp32 var_5716_epsilon_0 = const()[name = string("op_5716_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5716_cast_fp16 = inverse(epsilon = var_5716_epsilon_0, x = var_5715_cast_fp16)[name = string("op_5716_cast_fp16")];
            fp16 var_5717_to_fp16 = const()[name = string("op_5717_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5718_cast_fp16 = mul(x = var_5716_cast_fp16, y = var_5717_to_fp16)[name = string("op_5718_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5720_cast_fp16 = add(x = var_5714_cast_fp16, y = var_5718_cast_fp16)[name = string("op_5720_cast_fp16")];
            fp32 var_5721_epsilon_0 = const()[name = string("op_5721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5721_cast_fp16 = rsqrt(epsilon = var_5721_epsilon_0, x = var_5720_cast_fp16)[name = string("op_5721_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_153_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5721_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64808640)))];
            tensor<fp16, [1, 256, 512]> var_5723_cast_fp16 = mul(x = out_153_cast_fp16, y = tm_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5723_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_129_cast_fp16 = add(x = x_171_cast_fp16, y = var_5723_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64809728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65071936))))[name = string("tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_98_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_98_cast_fp16")];
            string var_5730_mode_0 = const()[name = string("op_5730_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_5730_cast_fp16 = gelu(mode = var_5730_mode_0, x = linear_98_cast_fp16)[name = string("op_5730_cast_fp16")];
            tensor<int32, [4]> var_5743_begin_0 = const()[name = string("op_5743_begin_0"), val = tensor<int32, [4]>([0, 0, 10, 0])];
            tensor<int32, [4]> var_5743_end_0 = const()[name = string("op_5743_end_0"), val = tensor<int32, [4]>([1, 256, 11, 512])];
            tensor<bool, [4]> var_5743_end_mask_0 = const()[name = string("op_5743_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_5743_squeeze_mask_0 = const()[name = string("op_5743_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_5743_cast_fp16 = slice_by_index(begin = var_5743_begin_0, end = var_5743_end_0, end_mask = var_5743_end_mask_0, squeeze_mask = var_5743_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_5743_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_131_cast_fp16 = mul(x = var_5730_cast_fp16, y = var_5743_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65073024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65335232))))[name = string("tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_10_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5753_cast_fp16 = abs(x = linear_99_cast_fp16)[name = string("op_5753_cast_fp16")];
            bool var_5756 = const()[name = string("op_5756"), val = bool(true)];
            tensor<int32, [1]> var_5757_axes_0 = const()[name = string("op_5757_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5757_cast_fp16 = reduce_max(axes = var_5757_axes_0, keep_dims = var_5756, x = var_5753_cast_fp16)[name = string("op_5757_cast_fp16")];
            fp16 var_5758_to_fp16 = const()[name = string("op_5758_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_177_cast_fp16 = maximum(x = var_5757_cast_fp16, y = var_5758_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_177_cast_fp16 = real_div(x = linear_99_cast_fp16, y = s_177_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5761_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5761_cast_fp16")];
            tensor<int32, [1]> var_5766_axes_0 = const()[name = string("op_5766_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5766_keep_dims_0 = const()[name = string("op_5766_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5766_cast_fp16 = reduce_mean(axes = var_5766_axes_0, keep_dims = var_5766_keep_dims_0, x = var_5761_cast_fp16)[name = string("op_5766_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5767_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5767_cast_fp16")];
            fp32 var_5768_epsilon_0 = const()[name = string("op_5768_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5768_cast_fp16 = inverse(epsilon = var_5768_epsilon_0, x = var_5767_cast_fp16)[name = string("op_5768_cast_fp16")];
            fp16 var_5769_to_fp16 = const()[name = string("op_5769_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5770_cast_fp16 = mul(x = var_5768_cast_fp16, y = var_5769_to_fp16)[name = string("op_5770_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5772_cast_fp16 = add(x = var_5766_cast_fp16, y = var_5770_cast_fp16)[name = string("op_5772_cast_fp16")];
            fp32 var_5773_epsilon_0 = const()[name = string("op_5773_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5773_cast_fp16 = rsqrt(epsilon = var_5773_epsilon_0, x = var_5772_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_155_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5773_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [512]> tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65336320)))];
            tensor<fp16, [1, 256, 512]> var_5775_cast_fp16 = mul(x = out_155_cast_fp16, y = tm_layers_10_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_5775_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_21_cast_fp16 = add(x = input_129_cast_fp16, y = var_5775_cast_fp16)[name = string("h_21_cast_fp16")];
            tensor<fp16, [1]> tm_layers_10_layer_scalar_to_fp16 = const()[name = string("tm_layers_10_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.1cp-1])];
            tensor<fp16, [1, 256, 512]> x_177_cast_fp16 = mul(x = h_21_cast_fp16, y = tm_layers_10_layer_scalar_to_fp16)[name = string("x_177_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5779_cast_fp16 = abs(x = x_177_cast_fp16)[name = string("op_5779_cast_fp16")];
            bool var_5782 = const()[name = string("op_5782"), val = bool(true)];
            tensor<int32, [1]> var_5783_axes_0 = const()[name = string("op_5783_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5783_cast_fp16 = reduce_max(axes = var_5783_axes_0, keep_dims = var_5782, x = var_5779_cast_fp16)[name = string("op_5783_cast_fp16")];
            fp16 var_5784_to_fp16 = const()[name = string("op_5784_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_179_cast_fp16 = maximum(x = var_5783_cast_fp16, y = var_5784_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_179_cast_fp16 = real_div(x = x_177_cast_fp16, y = s_179_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5787_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5787_cast_fp16")];
            tensor<int32, [1]> var_5792_axes_0 = const()[name = string("op_5792_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5792_keep_dims_0 = const()[name = string("op_5792_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5792_cast_fp16 = reduce_mean(axes = var_5792_axes_0, keep_dims = var_5792_keep_dims_0, x = var_5787_cast_fp16)[name = string("op_5792_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5793_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5793_cast_fp16")];
            fp32 var_5794_epsilon_0 = const()[name = string("op_5794_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5794_cast_fp16 = inverse(epsilon = var_5794_epsilon_0, x = var_5793_cast_fp16)[name = string("op_5794_cast_fp16")];
            fp16 var_5795_to_fp16 = const()[name = string("op_5795_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5796_cast_fp16 = mul(x = var_5794_cast_fp16, y = var_5795_to_fp16)[name = string("op_5796_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5798_cast_fp16 = add(x = var_5792_cast_fp16, y = var_5796_cast_fp16)[name = string("op_5798_cast_fp16")];
            fp32 var_5799_epsilon_0 = const()[name = string("op_5799_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5799_cast_fp16 = rsqrt(epsilon = var_5799_epsilon_0, x = var_5798_cast_fp16)[name = string("op_5799_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_157_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5799_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65337408)))];
            tensor<fp16, [1, 256, 512]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = tm_layers_11_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65338496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66387136))))[name = string("tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_100_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5809 = const()[name = string("op_5809"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 256, 4, 512]> var_5810_cast_fp16 = reshape(shape = var_5809, x = linear_100_cast_fp16)[name = string("op_5810_cast_fp16")];
            tensor<int32, [4]> x_179_perm_0 = const()[name = string("x_179_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66391296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66653504))))[name = string("tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_5821 = const()[name = string("op_5821"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_5822_cast_fp16 = reshape(shape = var_5821, x = linear_101_cast_fp16)[name = string("op_5822_cast_fp16")];
            tensor<int32, [4]> x_181_perm_0 = const()[name = string("x_181_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66654592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66916800))))[name = string("tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<int32, [4]> var_5833 = const()[name = string("op_5833"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_5834_cast_fp16 = reshape(shape = var_5833, x = linear_102_cast_fp16)[name = string("op_5834_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 512]> x_179_cast_fp16 = transpose(perm = x_179_perm_0, x = var_5810_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 4, 256, 512]> var_5838_cast_fp16 = abs(x = x_179_cast_fp16)[name = string("op_5838_cast_fp16")];
            bool var_5841 = const()[name = string("op_5841"), val = bool(true)];
            tensor<int32, [1]> var_5842_axes_0 = const()[name = string("op_5842_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_5842_cast_fp16 = reduce_max(axes = var_5842_axes_0, keep_dims = var_5841, x = var_5838_cast_fp16)[name = string("op_5842_cast_fp16")];
            fp16 var_5843_to_fp16 = const()[name = string("op_5843_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_181_cast_fp16 = maximum(x = var_5842_cast_fp16, y = var_5843_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> y_181_cast_fp16 = real_div(x = x_179_cast_fp16, y = s_181_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_5846_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<int32, [1]> var_5851_axes_0 = const()[name = string("op_5851_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5851_keep_dims_0 = const()[name = string("op_5851_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_5851_cast_fp16 = reduce_mean(axes = var_5851_axes_0, keep_dims = var_5851_keep_dims_0, x = var_5846_cast_fp16)[name = string("op_5851_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5852_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5852_cast_fp16")];
            fp32 var_5853_epsilon_0 = const()[name = string("op_5853_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_5853_cast_fp16 = inverse(epsilon = var_5853_epsilon_0, x = var_5852_cast_fp16)[name = string("op_5853_cast_fp16")];
            fp16 var_5854_to_fp16 = const()[name = string("op_5854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_5855_cast_fp16 = mul(x = var_5853_cast_fp16, y = var_5854_to_fp16)[name = string("op_5855_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_5857_cast_fp16 = add(x = var_5851_cast_fp16, y = var_5855_cast_fp16)[name = string("op_5857_cast_fp16")];
            fp32 var_5858_epsilon_0 = const()[name = string("op_5858_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_5858_cast_fp16 = rsqrt(epsilon = var_5858_epsilon_0, x = var_5857_cast_fp16)[name = string("op_5858_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> out_159_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5858_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66917888)))];
            tensor<fp16, [1, 4, 256, 512]> q_45_cast_fp16 = mul(x = out_159_cast_fp16, y = tm_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("q_45_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_181_cast_fp16 = transpose(perm = x_181_perm_0, x = var_5822_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1, 256, 512]> var_5861_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_5861_cast_fp16")];
            bool var_5864 = const()[name = string("op_5864"), val = bool(true)];
            tensor<int32, [1]> var_5865_axes_0 = const()[name = string("op_5865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_5865_cast_fp16 = reduce_max(axes = var_5865_axes_0, keep_dims = var_5864, x = var_5861_cast_fp16)[name = string("op_5865_cast_fp16")];
            fp16 var_5866_to_fp16 = const()[name = string("op_5866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_183_cast_fp16 = maximum(x = var_5865_cast_fp16, y = var_5866_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_183_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_183_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_5869_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5869_cast_fp16")];
            tensor<int32, [1]> var_5874_axes_0 = const()[name = string("op_5874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5874_keep_dims_0 = const()[name = string("op_5874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_5874_cast_fp16 = reduce_mean(axes = var_5874_axes_0, keep_dims = var_5874_keep_dims_0, x = var_5869_cast_fp16)[name = string("op_5874_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_5875_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5875_cast_fp16")];
            fp32 var_5876_epsilon_0 = const()[name = string("op_5876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_5876_cast_fp16 = inverse(epsilon = var_5876_epsilon_0, x = var_5875_cast_fp16)[name = string("op_5876_cast_fp16")];
            fp16 var_5877_to_fp16 = const()[name = string("op_5877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_5878_cast_fp16 = mul(x = var_5876_cast_fp16, y = var_5877_to_fp16)[name = string("op_5878_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_5880_cast_fp16 = add(x = var_5874_cast_fp16, y = var_5878_cast_fp16)[name = string("op_5880_cast_fp16")];
            fp32 var_5881_epsilon_0 = const()[name = string("op_5881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_5881_cast_fp16 = rsqrt(epsilon = var_5881_epsilon_0, x = var_5880_cast_fp16)[name = string("op_5881_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> out_161_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5881_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66918976)))];
            tensor<fp16, [1, 1, 256, 512]> k_67_cast_fp16 = mul(x = out_161_cast_fp16, y = tm_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("k_67_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5834_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1, 256, 512]> var_5884_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_5884_cast_fp16")];
            bool var_5887 = const()[name = string("op_5887"), val = bool(true)];
            tensor<int32, [1]> var_5888_axes_0 = const()[name = string("op_5888_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_5888_cast_fp16 = reduce_max(axes = var_5888_axes_0, keep_dims = var_5887, x = var_5884_cast_fp16)[name = string("op_5888_cast_fp16")];
            fp16 var_5889_to_fp16 = const()[name = string("op_5889_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_185_cast_fp16 = maximum(x = var_5888_cast_fp16, y = var_5889_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_185_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_185_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_5892_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5892_cast_fp16")];
            tensor<int32, [1]> var_5897_axes_0 = const()[name = string("op_5897_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5897_keep_dims_0 = const()[name = string("op_5897_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_5897_cast_fp16 = reduce_mean(axes = var_5897_axes_0, keep_dims = var_5897_keep_dims_0, x = var_5892_cast_fp16)[name = string("op_5897_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_5898_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5898_cast_fp16")];
            fp32 var_5899_epsilon_0 = const()[name = string("op_5899_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_5899_cast_fp16 = inverse(epsilon = var_5899_epsilon_0, x = var_5898_cast_fp16)[name = string("op_5899_cast_fp16")];
            fp16 var_5900_to_fp16 = const()[name = string("op_5900_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_5901_cast_fp16 = mul(x = var_5899_cast_fp16, y = var_5900_to_fp16)[name = string("op_5901_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_5903_cast_fp16 = add(x = var_5897_cast_fp16, y = var_5901_cast_fp16)[name = string("op_5903_cast_fp16")];
            fp32 var_5904_epsilon_0 = const()[name = string("op_5904_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_5904_cast_fp16 = rsqrt(epsilon = var_5904_epsilon_0, x = var_5903_cast_fp16)[name = string("op_5904_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> v_45_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5904_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_5906_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_cast_fp16)[name = string("op_5906_cast_fp16")];
            tensor<int32, [4]> var_5911_begin_0 = const()[name = string("op_5911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5911_end_0 = const()[name = string("op_5911_end_0"), val = tensor<int32, [4]>([1, 4, 256, 512])];
            tensor<bool, [4]> var_5911_end_mask_0 = const()[name = string("op_5911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 256]> var_5911_cast_fp16 = slice_by_index(begin = var_5911_begin_0, end = var_5911_end_0, end_mask = var_5911_end_mask_0, x = q_45_cast_fp16)[name = string("op_5911_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 256]> var_5912_cast_fp16 = mul(x = var_5911_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_5912_cast_fp16")];
            tensor<int32, [4]> var_5917_begin_0 = const()[name = string("op_5917_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5917_end_0 = const()[name = string("op_5917_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_5917_end_mask_0 = const()[name = string("op_5917_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 256]> var_5917_cast_fp16 = slice_by_index(begin = var_5917_begin_0, end = var_5917_end_0, end_mask = var_5917_end_mask_0, x = q_45_cast_fp16)[name = string("op_5917_cast_fp16")];
            int32 var_5919 = const()[name = string("op_5919"), val = int32(-1)];
            bool var_5920_interleave_0 = const()[name = string("op_5920_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> var_5920_cast_fp16 = concat(axis = var_5919, interleave = var_5920_interleave_0, values = (var_5912_cast_fp16, var_5917_cast_fp16))[name = string("op_5920_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_5921_cast_fp16 = mul(x = var_5920_cast_fp16, y = sin_cast_fp16)[name = string("op_5921_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> q_47_cast_fp16 = add(x = var_5906_cast_fp16, y = var_5921_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_5924_cast_fp16 = mul(x = k_67_cast_fp16, y = cos_cast_fp16)[name = string("op_5924_cast_fp16")];
            tensor<int32, [4]> var_5929_begin_0 = const()[name = string("op_5929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_5929_end_0 = const()[name = string("op_5929_end_0"), val = tensor<int32, [4]>([1, 1, 256, 512])];
            tensor<bool, [4]> var_5929_end_mask_0 = const()[name = string("op_5929_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 256, 256]> var_5929_cast_fp16 = slice_by_index(begin = var_5929_begin_0, end = var_5929_end_0, end_mask = var_5929_end_mask_0, x = k_67_cast_fp16)[name = string("op_5929_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 256, 256]> var_5930_cast_fp16 = mul(x = var_5929_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_5930_cast_fp16")];
            tensor<int32, [4]> var_5935_begin_0 = const()[name = string("op_5935_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5935_end_0 = const()[name = string("op_5935_end_0"), val = tensor<int32, [4]>([1, 1, 256, 256])];
            tensor<bool, [4]> var_5935_end_mask_0 = const()[name = string("op_5935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 256, 256]> var_5935_cast_fp16 = slice_by_index(begin = var_5935_begin_0, end = var_5935_end_0, end_mask = var_5935_end_mask_0, x = k_67_cast_fp16)[name = string("op_5935_cast_fp16")];
            int32 var_5937 = const()[name = string("op_5937"), val = int32(-1)];
            bool var_5938_interleave_0 = const()[name = string("op_5938_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 256, 512]> var_5938_cast_fp16 = concat(axis = var_5937, interleave = var_5938_interleave_0, values = (var_5930_cast_fp16, var_5935_cast_fp16))[name = string("op_5938_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_5939_cast_fp16 = mul(x = var_5938_cast_fp16, y = sin_cast_fp16)[name = string("op_5939_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> k_69_cast_fp16 = add(x = var_5924_cast_fp16, y = var_5939_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_22_reps_0 = const()[name = string("tile_22_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = k_69_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [4, 1, 256, 512]> tile_22_cast_fp16 = tile(reps = tile_22_reps_0, x = transpose_44_cast_fp16)[name = string("tile_22_cast_fp16")];
            tensor<int32, [5]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_44_cast_fp16 = reshape(shape = concat_44, x = tile_22_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [5]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [4, 1, 256, 512]> reshape_45_cast_fp16 = reshape(shape = concat_45, x = transpose_45_cast_fp16)[name = string("reshape_45_cast_fp16")];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_23_reps_0 = const()[name = string("tile_23_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = v_45_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [4, 1, 256, 512]> tile_23_cast_fp16 = tile(reps = tile_23_reps_0, x = transpose_46_cast_fp16)[name = string("tile_23_cast_fp16")];
            tensor<int32, [5]> concat_46 = const()[name = string("concat_46"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_46_cast_fp16 = reshape(shape = concat_46, x = tile_23_cast_fp16)[name = string("reshape_46_cast_fp16")];
            tensor<int32, [5]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_47 = const()[name = string("concat_47"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = reshape_46_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [4, 1, 256, 512]> reshape_47_cast_fp16 = reshape(shape = concat_47, x = transpose_47_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_5953_transpose_x_0 = const()[name = string("op_5953_transpose_x_0"), val = bool(false)];
            bool var_5953_transpose_y_0 = const()[name = string("op_5953_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = reshape_45_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 4, 256, 256]> var_5953_cast_fp16 = matmul(transpose_x = var_5953_transpose_x_0, transpose_y = var_5953_transpose_y_0, x = q_47_cast_fp16, y = transpose_107_cast_fp16)[name = string("op_5953_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_5955_cast_fp16 = add(x = var_5953_cast_fp16, y = attention_bias)[name = string("op_5955_cast_fp16")];
            int32 var_5956 = const()[name = string("op_5956"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_5958_cast_fp16 = softmax(axis = var_5956, x = var_5955_cast_fp16)[name = string("op_5958_cast_fp16")];
            bool var_5959_transpose_x_0 = const()[name = string("op_5959_transpose_x_0"), val = bool(false)];
            bool var_5959_transpose_y_0 = const()[name = string("op_5959_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 4, 256, 512]> var_5959_cast_fp16 = matmul(transpose_x = var_5959_transpose_x_0, transpose_y = var_5959_transpose_y_0, x = var_5958_cast_fp16, y = v_47_cast_fp16)[name = string("op_5959_cast_fp16")];
            tensor<int32, [4]> var_5962_perm_0 = const()[name = string("op_5962_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5966 = const()[name = string("op_5966"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 256, 4, 512]> var_5962_cast_fp16 = transpose(perm = var_5962_perm_0, x = var_5959_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 256, 2048]> input_135_cast_fp16 = reshape(shape = var_5966, x = var_5962_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66920064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67968704))))[name = string("tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5971_cast_fp16 = abs(x = linear_103_cast_fp16)[name = string("op_5971_cast_fp16")];
            bool var_5974 = const()[name = string("op_5974"), val = bool(true)];
            tensor<int32, [1]> var_5975_axes_0 = const()[name = string("op_5975_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_5975_cast_fp16 = reduce_max(axes = var_5975_axes_0, keep_dims = var_5974, x = var_5971_cast_fp16)[name = string("op_5975_cast_fp16")];
            fp16 var_5976_to_fp16 = const()[name = string("op_5976_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_187_cast_fp16 = maximum(x = var_5975_cast_fp16, y = var_5976_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_187_cast_fp16 = real_div(x = linear_103_cast_fp16, y = s_187_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5979_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_5979_cast_fp16")];
            tensor<int32, [1]> var_5984_axes_0 = const()[name = string("op_5984_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5984_keep_dims_0 = const()[name = string("op_5984_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5984_cast_fp16 = reduce_mean(axes = var_5984_axes_0, keep_dims = var_5984_keep_dims_0, x = var_5979_cast_fp16)[name = string("op_5984_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5985_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_5985_cast_fp16")];
            fp32 var_5986_epsilon_0 = const()[name = string("op_5986_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_5986_cast_fp16 = inverse(epsilon = var_5986_epsilon_0, x = var_5985_cast_fp16)[name = string("op_5986_cast_fp16")];
            fp16 var_5987_to_fp16 = const()[name = string("op_5987_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_5988_cast_fp16 = mul(x = var_5986_cast_fp16, y = var_5987_to_fp16)[name = string("op_5988_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_5990_cast_fp16 = add(x = var_5984_cast_fp16, y = var_5988_cast_fp16)[name = string("op_5990_cast_fp16")];
            fp32 var_5991_epsilon_0 = const()[name = string("op_5991_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_5991_cast_fp16 = rsqrt(epsilon = var_5991_epsilon_0, x = var_5990_cast_fp16)[name = string("op_5991_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_163_cast_fp16 = mul(x = y_187_cast_fp16, y = var_5991_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969792)))];
            tensor<fp16, [1, 256, 512]> var_5993_cast_fp16 = mul(x = out_163_cast_fp16, y = tm_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_5993_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_187_cast_fp16 = add(x = x_177_cast_fp16, y = var_5993_cast_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_5996_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_5996_cast_fp16")];
            bool var_5999 = const()[name = string("op_5999"), val = bool(true)];
            tensor<int32, [1]> var_6000_axes_0 = const()[name = string("op_6000_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6000_cast_fp16 = reduce_max(axes = var_6000_axes_0, keep_dims = var_5999, x = var_5996_cast_fp16)[name = string("op_6000_cast_fp16")];
            fp16 var_6001_to_fp16 = const()[name = string("op_6001_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_189_cast_fp16 = maximum(x = var_6000_cast_fp16, y = var_6001_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_189_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_189_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6004_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_6004_cast_fp16")];
            tensor<int32, [1]> var_6009_axes_0 = const()[name = string("op_6009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6009_keep_dims_0 = const()[name = string("op_6009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6009_cast_fp16 = reduce_mean(axes = var_6009_axes_0, keep_dims = var_6009_keep_dims_0, x = var_6004_cast_fp16)[name = string("op_6009_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6010_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_6010_cast_fp16")];
            fp32 var_6011_epsilon_0 = const()[name = string("op_6011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6011_cast_fp16 = inverse(epsilon = var_6011_epsilon_0, x = var_6010_cast_fp16)[name = string("op_6011_cast_fp16")];
            fp16 var_6012_to_fp16 = const()[name = string("op_6012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6013_cast_fp16 = mul(x = var_6011_cast_fp16, y = var_6012_to_fp16)[name = string("op_6013_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6015_cast_fp16 = add(x = var_6009_cast_fp16, y = var_6013_cast_fp16)[name = string("op_6015_cast_fp16")];
            fp32 var_6016_epsilon_0 = const()[name = string("op_6016_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6016_cast_fp16 = rsqrt(epsilon = var_6016_epsilon_0, x = var_6015_cast_fp16)[name = string("op_6016_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_165_cast_fp16 = mul(x = y_189_cast_fp16, y = var_6016_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67970880)))];
            tensor<fp16, [1, 256, 512]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = tm_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971968))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69020608))))[name = string("tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_104_cast_fp16")];
            string var_6023_mode_0 = const()[name = string("op_6023_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_6023_cast_fp16 = gelu(mode = var_6023_mode_0, x = linear_104_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70073408))))[name = string("tm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_105_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_139_cast_fp16 = mul(x = var_6023_cast_fp16, y = linear_105_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70077568))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71126208))))[name = string("tm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6031_cast_fp16 = abs(x = linear_106_cast_fp16)[name = string("op_6031_cast_fp16")];
            bool var_6034 = const()[name = string("op_6034"), val = bool(true)];
            tensor<int32, [1]> var_6035_axes_0 = const()[name = string("op_6035_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6035_cast_fp16 = reduce_max(axes = var_6035_axes_0, keep_dims = var_6034, x = var_6031_cast_fp16)[name = string("op_6035_cast_fp16")];
            fp16 var_6036_to_fp16 = const()[name = string("op_6036_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_191_cast_fp16 = maximum(x = var_6035_cast_fp16, y = var_6036_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_191_cast_fp16 = real_div(x = linear_106_cast_fp16, y = s_191_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6039_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6039_cast_fp16")];
            tensor<int32, [1]> var_6044_axes_0 = const()[name = string("op_6044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6044_keep_dims_0 = const()[name = string("op_6044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6044_cast_fp16 = reduce_mean(axes = var_6044_axes_0, keep_dims = var_6044_keep_dims_0, x = var_6039_cast_fp16)[name = string("op_6044_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6045_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_6045_cast_fp16")];
            fp32 var_6046_epsilon_0 = const()[name = string("op_6046_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6046_cast_fp16 = inverse(epsilon = var_6046_epsilon_0, x = var_6045_cast_fp16)[name = string("op_6046_cast_fp16")];
            fp16 var_6047_to_fp16 = const()[name = string("op_6047_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6048_cast_fp16 = mul(x = var_6046_cast_fp16, y = var_6047_to_fp16)[name = string("op_6048_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6050_cast_fp16 = add(x = var_6044_cast_fp16, y = var_6048_cast_fp16)[name = string("op_6050_cast_fp16")];
            fp32 var_6051_epsilon_0 = const()[name = string("op_6051_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6051_cast_fp16 = rsqrt(epsilon = var_6051_epsilon_0, x = var_6050_cast_fp16)[name = string("op_6051_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_167_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6051_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71127296)))];
            tensor<fp16, [1, 256, 512]> var_6053_cast_fp16 = mul(x = out_167_cast_fp16, y = tm_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_141_cast_fp16 = add(x = x_187_cast_fp16, y = var_6053_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71128384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71390592))))[name = string("tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_107_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_107_cast_fp16")];
            string var_6060_mode_0 = const()[name = string("op_6060_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_6060_cast_fp16 = gelu(mode = var_6060_mode_0, x = linear_107_cast_fp16)[name = string("op_6060_cast_fp16")];
            tensor<int32, [4]> var_6073_begin_0 = const()[name = string("op_6073_begin_0"), val = tensor<int32, [4]>([0, 0, 11, 0])];
            tensor<int32, [4]> var_6073_end_0 = const()[name = string("op_6073_end_0"), val = tensor<int32, [4]>([1, 256, 12, 512])];
            tensor<bool, [4]> var_6073_end_mask_0 = const()[name = string("op_6073_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6073_squeeze_mask_0 = const()[name = string("op_6073_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_6073_cast_fp16 = slice_by_index(begin = var_6073_begin_0, end = var_6073_end_0, end_mask = var_6073_end_mask_0, squeeze_mask = var_6073_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_6073_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_143_cast_fp16 = mul(x = var_6060_cast_fp16, y = var_6073_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71391680))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71653888))))[name = string("tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_11_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6083_cast_fp16 = abs(x = linear_108_cast_fp16)[name = string("op_6083_cast_fp16")];
            bool var_6086 = const()[name = string("op_6086"), val = bool(true)];
            tensor<int32, [1]> var_6087_axes_0 = const()[name = string("op_6087_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6087_cast_fp16 = reduce_max(axes = var_6087_axes_0, keep_dims = var_6086, x = var_6083_cast_fp16)[name = string("op_6087_cast_fp16")];
            fp16 var_6088_to_fp16 = const()[name = string("op_6088_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_193_cast_fp16 = maximum(x = var_6087_cast_fp16, y = var_6088_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_193_cast_fp16 = real_div(x = linear_108_cast_fp16, y = s_193_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6091_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6091_cast_fp16")];
            tensor<int32, [1]> var_6096_axes_0 = const()[name = string("op_6096_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6096_keep_dims_0 = const()[name = string("op_6096_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6096_cast_fp16 = reduce_mean(axes = var_6096_axes_0, keep_dims = var_6096_keep_dims_0, x = var_6091_cast_fp16)[name = string("op_6096_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6097_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_6097_cast_fp16")];
            fp32 var_6098_epsilon_0 = const()[name = string("op_6098_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6098_cast_fp16 = inverse(epsilon = var_6098_epsilon_0, x = var_6097_cast_fp16)[name = string("op_6098_cast_fp16")];
            fp16 var_6099_to_fp16 = const()[name = string("op_6099_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6100_cast_fp16 = mul(x = var_6098_cast_fp16, y = var_6099_to_fp16)[name = string("op_6100_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6102_cast_fp16 = add(x = var_6096_cast_fp16, y = var_6100_cast_fp16)[name = string("op_6102_cast_fp16")];
            fp32 var_6103_epsilon_0 = const()[name = string("op_6103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6103_cast_fp16 = rsqrt(epsilon = var_6103_epsilon_0, x = var_6102_cast_fp16)[name = string("op_6103_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_169_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6103_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [512]> tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71654976)))];
            tensor<fp16, [1, 256, 512]> var_6105_cast_fp16 = mul(x = out_169_cast_fp16, y = tm_layers_11_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6105_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_23_cast_fp16 = add(x = input_141_cast_fp16, y = var_6105_cast_fp16)[name = string("h_23_cast_fp16")];
            tensor<fp16, [1]> tm_layers_11_layer_scalar_to_fp16 = const()[name = string("tm_layers_11_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.86p-2])];
            tensor<fp16, [1, 256, 512]> x_193_cast_fp16 = mul(x = h_23_cast_fp16, y = tm_layers_11_layer_scalar_to_fp16)[name = string("x_193_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6109_cast_fp16 = abs(x = x_193_cast_fp16)[name = string("op_6109_cast_fp16")];
            bool var_6112 = const()[name = string("op_6112"), val = bool(true)];
            tensor<int32, [1]> var_6113_axes_0 = const()[name = string("op_6113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6113_cast_fp16 = reduce_max(axes = var_6113_axes_0, keep_dims = var_6112, x = var_6109_cast_fp16)[name = string("op_6113_cast_fp16")];
            fp16 var_6114_to_fp16 = const()[name = string("op_6114_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_195_cast_fp16 = maximum(x = var_6113_cast_fp16, y = var_6114_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_195_cast_fp16 = real_div(x = x_193_cast_fp16, y = s_195_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6117_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6117_cast_fp16")];
            tensor<int32, [1]> var_6122_axes_0 = const()[name = string("op_6122_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6122_keep_dims_0 = const()[name = string("op_6122_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6122_cast_fp16 = reduce_mean(axes = var_6122_axes_0, keep_dims = var_6122_keep_dims_0, x = var_6117_cast_fp16)[name = string("op_6122_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6123_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6123_cast_fp16")];
            fp32 var_6124_epsilon_0 = const()[name = string("op_6124_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6124_cast_fp16 = inverse(epsilon = var_6124_epsilon_0, x = var_6123_cast_fp16)[name = string("op_6124_cast_fp16")];
            fp16 var_6125_to_fp16 = const()[name = string("op_6125_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6126_cast_fp16 = mul(x = var_6124_cast_fp16, y = var_6125_to_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6128_cast_fp16 = add(x = var_6122_cast_fp16, y = var_6126_cast_fp16)[name = string("op_6128_cast_fp16")];
            fp32 var_6129_epsilon_0 = const()[name = string("op_6129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6129_cast_fp16 = rsqrt(epsilon = var_6129_epsilon_0, x = var_6128_cast_fp16)[name = string("op_6129_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_171_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6129_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71656064)))];
            tensor<fp16, [1, 256, 512]> input_145_cast_fp16 = mul(x = out_171_cast_fp16, y = tm_layers_12_input_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71657152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72181504))))[name = string("tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<int32, [4]> var_6139 = const()[name = string("op_6139"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_6140_cast_fp16 = reshape(shape = var_6139, x = linear_109_cast_fp16)[name = string("op_6140_cast_fp16")];
            tensor<int32, [4]> x_195_perm_0 = const()[name = string("x_195_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72183616))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72445824))))[name = string("tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<int32, [4]> var_6151 = const()[name = string("op_6151"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6152_cast_fp16 = reshape(shape = var_6151, x = linear_110_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<int32, [4]> x_197_perm_0 = const()[name = string("x_197_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72446912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72709120))))[name = string("tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<int32, [4]> var_6163 = const()[name = string("op_6163"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6164_cast_fp16 = reshape(shape = var_6163, x = linear_111_cast_fp16)[name = string("op_6164_cast_fp16")];
            tensor<int32, [4]> x_199_perm_0 = const()[name = string("x_199_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_195_cast_fp16 = transpose(perm = x_195_perm_0, x = var_6140_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 4, 256, 256]> var_6168_cast_fp16 = abs(x = x_195_cast_fp16)[name = string("op_6168_cast_fp16")];
            bool var_6171 = const()[name = string("op_6171"), val = bool(true)];
            tensor<int32, [1]> var_6172_axes_0 = const()[name = string("op_6172_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_6172_cast_fp16 = reduce_max(axes = var_6172_axes_0, keep_dims = var_6171, x = var_6168_cast_fp16)[name = string("op_6172_cast_fp16")];
            fp16 var_6173_to_fp16 = const()[name = string("op_6173_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_197_cast_fp16 = maximum(x = var_6172_cast_fp16, y = var_6173_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_197_cast_fp16 = real_div(x = x_195_cast_fp16, y = s_197_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6176_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6176_cast_fp16")];
            tensor<int32, [1]> var_6181_axes_0 = const()[name = string("op_6181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6181_keep_dims_0 = const()[name = string("op_6181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_6181_cast_fp16 = reduce_mean(axes = var_6181_axes_0, keep_dims = var_6181_keep_dims_0, x = var_6176_cast_fp16)[name = string("op_6181_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6182_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6182_cast_fp16")];
            fp32 var_6183_epsilon_0 = const()[name = string("op_6183_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_6183_cast_fp16 = inverse(epsilon = var_6183_epsilon_0, x = var_6182_cast_fp16)[name = string("op_6183_cast_fp16")];
            fp16 var_6184_to_fp16 = const()[name = string("op_6184_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_6185_cast_fp16 = mul(x = var_6183_cast_fp16, y = var_6184_to_fp16)[name = string("op_6185_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6187_cast_fp16 = add(x = var_6181_cast_fp16, y = var_6185_cast_fp16)[name = string("op_6187_cast_fp16")];
            fp32 var_6188_epsilon_0 = const()[name = string("op_6188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_6188_cast_fp16 = rsqrt(epsilon = var_6188_epsilon_0, x = var_6187_cast_fp16)[name = string("op_6188_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_173_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6188_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_49_cast_fp16 = mul(x = out_173_cast_fp16, y = tm_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("q_49_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_197_cast_fp16 = transpose(perm = x_197_perm_0, x = var_6152_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 2, 256, 256]> var_6191_cast_fp16 = abs(x = x_197_cast_fp16)[name = string("op_6191_cast_fp16")];
            bool var_6194 = const()[name = string("op_6194"), val = bool(true)];
            tensor<int32, [1]> var_6195_axes_0 = const()[name = string("op_6195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6195_cast_fp16 = reduce_max(axes = var_6195_axes_0, keep_dims = var_6194, x = var_6191_cast_fp16)[name = string("op_6195_cast_fp16")];
            fp16 var_6196_to_fp16 = const()[name = string("op_6196_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_199_cast_fp16 = maximum(x = var_6195_cast_fp16, y = var_6196_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_199_cast_fp16 = real_div(x = x_197_cast_fp16, y = s_199_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6199_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6199_cast_fp16")];
            tensor<int32, [1]> var_6204_axes_0 = const()[name = string("op_6204_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6204_keep_dims_0 = const()[name = string("op_6204_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6204_cast_fp16 = reduce_mean(axes = var_6204_axes_0, keep_dims = var_6204_keep_dims_0, x = var_6199_cast_fp16)[name = string("op_6204_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6205_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6205_cast_fp16")];
            fp32 var_6206_epsilon_0 = const()[name = string("op_6206_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6206_cast_fp16 = inverse(epsilon = var_6206_epsilon_0, x = var_6205_cast_fp16)[name = string("op_6206_cast_fp16")];
            fp16 var_6207_to_fp16 = const()[name = string("op_6207_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6208_cast_fp16 = mul(x = var_6206_cast_fp16, y = var_6207_to_fp16)[name = string("op_6208_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6210_cast_fp16 = add(x = var_6204_cast_fp16, y = var_6208_cast_fp16)[name = string("op_6210_cast_fp16")];
            fp32 var_6211_epsilon_0 = const()[name = string("op_6211_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6211_cast_fp16 = rsqrt(epsilon = var_6211_epsilon_0, x = var_6210_cast_fp16)[name = string("op_6211_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_175_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6211_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_73_cast_fp16 = mul(x = out_175_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_199_cast_fp16 = transpose(perm = x_199_perm_0, x = var_6164_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 2, 256, 256]> var_6214_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_6214_cast_fp16")];
            bool var_6217 = const()[name = string("op_6217"), val = bool(true)];
            tensor<int32, [1]> var_6218_axes_0 = const()[name = string("op_6218_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6218_cast_fp16 = reduce_max(axes = var_6218_axes_0, keep_dims = var_6217, x = var_6214_cast_fp16)[name = string("op_6218_cast_fp16")];
            fp16 var_6219_to_fp16 = const()[name = string("op_6219_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_201_cast_fp16 = maximum(x = var_6218_cast_fp16, y = var_6219_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_201_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_201_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6222_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6222_cast_fp16")];
            tensor<int32, [1]> var_6227_axes_0 = const()[name = string("op_6227_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6227_keep_dims_0 = const()[name = string("op_6227_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6227_cast_fp16 = reduce_mean(axes = var_6227_axes_0, keep_dims = var_6227_keep_dims_0, x = var_6222_cast_fp16)[name = string("op_6227_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6228_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6228_cast_fp16")];
            fp32 var_6229_epsilon_0 = const()[name = string("op_6229_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6229_cast_fp16 = inverse(epsilon = var_6229_epsilon_0, x = var_6228_cast_fp16)[name = string("op_6229_cast_fp16")];
            fp16 var_6230_to_fp16 = const()[name = string("op_6230_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6231_cast_fp16 = mul(x = var_6229_cast_fp16, y = var_6230_to_fp16)[name = string("op_6231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6233_cast_fp16 = add(x = var_6227_cast_fp16, y = var_6231_cast_fp16)[name = string("op_6233_cast_fp16")];
            fp32 var_6234_epsilon_0 = const()[name = string("op_6234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6234_cast_fp16 = rsqrt(epsilon = var_6234_epsilon_0, x = var_6233_cast_fp16)[name = string("op_6234_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_49_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6234_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6236_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6236_cast_fp16")];
            tensor<int32, [4]> var_6241_begin_0 = const()[name = string("op_6241_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6241_end_0 = const()[name = string("op_6241_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_6241_end_mask_0 = const()[name = string("op_6241_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_6241_cast_fp16 = slice_by_index(begin = var_6241_begin_0, end = var_6241_end_0, end_mask = var_6241_end_mask_0, x = q_49_cast_fp16)[name = string("op_6241_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_6242_cast_fp16 = mul(x = var_6241_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_6242_cast_fp16")];
            tensor<int32, [4]> var_6247_begin_0 = const()[name = string("op_6247_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6247_end_0 = const()[name = string("op_6247_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_6247_end_mask_0 = const()[name = string("op_6247_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_6247_cast_fp16 = slice_by_index(begin = var_6247_begin_0, end = var_6247_end_0, end_mask = var_6247_end_mask_0, x = q_49_cast_fp16)[name = string("op_6247_cast_fp16")];
            int32 var_6249 = const()[name = string("op_6249"), val = int32(-1)];
            bool var_6250_interleave_0 = const()[name = string("op_6250_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_6250_cast_fp16 = concat(axis = var_6249, interleave = var_6250_interleave_0, values = (var_6242_cast_fp16, var_6247_cast_fp16))[name = string("op_6250_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6251_cast_fp16 = mul(x = var_6250_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6251_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_51_cast_fp16 = add(x = var_6236_cast_fp16, y = var_6251_cast_fp16)[name = string("q_51_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6254_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6254_cast_fp16")];
            tensor<int32, [4]> var_6259_begin_0 = const()[name = string("op_6259_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6259_end_0 = const()[name = string("op_6259_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_6259_end_mask_0 = const()[name = string("op_6259_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_6259_cast_fp16 = slice_by_index(begin = var_6259_begin_0, end = var_6259_end_0, end_mask = var_6259_end_mask_0, x = k_73_cast_fp16)[name = string("op_6259_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_6260_cast_fp16 = mul(x = var_6259_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_6260_cast_fp16")];
            tensor<int32, [4]> var_6265_begin_0 = const()[name = string("op_6265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6265_end_0 = const()[name = string("op_6265_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_6265_end_mask_0 = const()[name = string("op_6265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_6265_cast_fp16 = slice_by_index(begin = var_6265_begin_0, end = var_6265_end_0, end_mask = var_6265_end_mask_0, x = k_73_cast_fp16)[name = string("op_6265_cast_fp16")];
            int32 var_6267 = const()[name = string("op_6267"), val = int32(-1)];
            bool var_6268_interleave_0 = const()[name = string("op_6268_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_6268_cast_fp16 = concat(axis = var_6267, interleave = var_6268_interleave_0, values = (var_6260_cast_fp16, var_6265_cast_fp16))[name = string("op_6268_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6269_cast_fp16 = mul(x = var_6268_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6269_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_75_cast_fp16 = add(x = var_6254_cast_fp16, y = var_6269_cast_fp16)[name = string("k_75_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_24_reps_0 = const()[name = string("tile_24_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = k_75_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [4, 1, 256, 256]> tile_24_cast_fp16 = tile(reps = tile_24_reps_0, x = transpose_48_cast_fp16)[name = string("tile_24_cast_fp16")];
            tensor<int32, [5]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_48_cast_fp16 = reshape(shape = concat_48, x = tile_24_cast_fp16)[name = string("reshape_48_cast_fp16")];
            tensor<int32, [5]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = reshape_48_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [4, 1, 256, 256]> reshape_49_cast_fp16 = reshape(shape = concat_49, x = transpose_49_cast_fp16)[name = string("reshape_49_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_25_reps_0 = const()[name = string("tile_25_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = v_49_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [4, 1, 256, 256]> tile_25_cast_fp16 = tile(reps = tile_25_reps_0, x = transpose_50_cast_fp16)[name = string("tile_25_cast_fp16")];
            tensor<int32, [5]> concat_50 = const()[name = string("concat_50"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_50_cast_fp16 = reshape(shape = concat_50, x = tile_25_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [5]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [4, 1, 256, 256]> reshape_51_cast_fp16 = reshape(shape = concat_51, x = transpose_51_cast_fp16)[name = string("reshape_51_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6283_transpose_x_0 = const()[name = string("op_6283_transpose_x_0"), val = bool(false)];
            bool var_6283_transpose_y_0 = const()[name = string("op_6283_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = reshape_49_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 4, 256, 256]> var_6283_cast_fp16 = matmul(transpose_x = var_6283_transpose_x_0, transpose_y = var_6283_transpose_y_0, x = q_51_cast_fp16, y = transpose_108_cast_fp16)[name = string("op_6283_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6285_cast_fp16 = add(x = var_6283_cast_fp16, y = attention_bias)[name = string("op_6285_cast_fp16")];
            int32 var_6286 = const()[name = string("op_6286"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_6288_cast_fp16 = softmax(axis = var_6286, x = var_6285_cast_fp16)[name = string("op_6288_cast_fp16")];
            bool var_6289_transpose_x_0 = const()[name = string("op_6289_transpose_x_0"), val = bool(false)];
            bool var_6289_transpose_y_0 = const()[name = string("op_6289_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = reshape_51_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 4, 256, 256]> var_6289_cast_fp16 = matmul(transpose_x = var_6289_transpose_x_0, transpose_y = var_6289_transpose_y_0, x = var_6288_cast_fp16, y = v_51_cast_fp16)[name = string("op_6289_cast_fp16")];
            tensor<int32, [4]> var_6292_perm_0 = const()[name = string("op_6292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6296 = const()[name = string("op_6296"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_6292_cast_fp16 = transpose(perm = var_6292_perm_0, x = var_6289_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 256, 1024]> input_147_cast_fp16 = reshape(shape = var_6296, x = var_6292_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72710208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73234560))))[name = string("tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6301_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6301_cast_fp16")];
            bool var_6304 = const()[name = string("op_6304"), val = bool(true)];
            tensor<int32, [1]> var_6305_axes_0 = const()[name = string("op_6305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6305_cast_fp16 = reduce_max(axes = var_6305_axes_0, keep_dims = var_6304, x = var_6301_cast_fp16)[name = string("op_6305_cast_fp16")];
            fp16 var_6306_to_fp16 = const()[name = string("op_6306_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_203_cast_fp16 = maximum(x = var_6305_cast_fp16, y = var_6306_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_203_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_203_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6309_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6309_cast_fp16")];
            tensor<int32, [1]> var_6314_axes_0 = const()[name = string("op_6314_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6314_keep_dims_0 = const()[name = string("op_6314_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6314_cast_fp16 = reduce_mean(axes = var_6314_axes_0, keep_dims = var_6314_keep_dims_0, x = var_6309_cast_fp16)[name = string("op_6314_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6315_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6315_cast_fp16")];
            fp32 var_6316_epsilon_0 = const()[name = string("op_6316_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6316_cast_fp16 = inverse(epsilon = var_6316_epsilon_0, x = var_6315_cast_fp16)[name = string("op_6316_cast_fp16")];
            fp16 var_6317_to_fp16 = const()[name = string("op_6317_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6318_cast_fp16 = mul(x = var_6316_cast_fp16, y = var_6317_to_fp16)[name = string("op_6318_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6320_cast_fp16 = add(x = var_6314_cast_fp16, y = var_6318_cast_fp16)[name = string("op_6320_cast_fp16")];
            fp32 var_6321_epsilon_0 = const()[name = string("op_6321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6321_cast_fp16 = rsqrt(epsilon = var_6321_epsilon_0, x = var_6320_cast_fp16)[name = string("op_6321_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_177_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6321_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73235648)))];
            tensor<fp16, [1, 256, 512]> var_6323_cast_fp16 = mul(x = out_177_cast_fp16, y = tm_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_6323_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_203_cast_fp16 = add(x = x_193_cast_fp16, y = var_6323_cast_fp16)[name = string("x_203_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6326_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_6326_cast_fp16")];
            bool var_6329 = const()[name = string("op_6329"), val = bool(true)];
            tensor<int32, [1]> var_6330_axes_0 = const()[name = string("op_6330_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6330_cast_fp16 = reduce_max(axes = var_6330_axes_0, keep_dims = var_6329, x = var_6326_cast_fp16)[name = string("op_6330_cast_fp16")];
            fp16 var_6331_to_fp16 = const()[name = string("op_6331_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_205_cast_fp16 = maximum(x = var_6330_cast_fp16, y = var_6331_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_205_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_205_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6334_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6334_cast_fp16")];
            tensor<int32, [1]> var_6339_axes_0 = const()[name = string("op_6339_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6339_keep_dims_0 = const()[name = string("op_6339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6339_cast_fp16 = reduce_mean(axes = var_6339_axes_0, keep_dims = var_6339_keep_dims_0, x = var_6334_cast_fp16)[name = string("op_6339_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6340_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6340_cast_fp16")];
            fp32 var_6341_epsilon_0 = const()[name = string("op_6341_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6341_cast_fp16 = inverse(epsilon = var_6341_epsilon_0, x = var_6340_cast_fp16)[name = string("op_6341_cast_fp16")];
            fp16 var_6342_to_fp16 = const()[name = string("op_6342_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6343_cast_fp16 = mul(x = var_6341_cast_fp16, y = var_6342_to_fp16)[name = string("op_6343_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6345_cast_fp16 = add(x = var_6339_cast_fp16, y = var_6343_cast_fp16)[name = string("op_6345_cast_fp16")];
            fp32 var_6346_epsilon_0 = const()[name = string("op_6346_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6346_cast_fp16 = rsqrt(epsilon = var_6346_epsilon_0, x = var_6345_cast_fp16)[name = string("op_6346_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_179_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6346_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73236736)))];
            tensor<fp16, [1, 256, 512]> input_149_cast_fp16 = mul(x = out_179_cast_fp16, y = tm_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73237824))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74286464))))[name = string("tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_113_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_113_cast_fp16")];
            string var_6353_mode_0 = const()[name = string("op_6353_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_6353_cast_fp16 = gelu(mode = var_6353_mode_0, x = linear_113_cast_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74290624))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339264))))[name = string("tm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_114_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_151_cast_fp16 = mul(x = var_6353_cast_fp16, y = linear_114_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75343424))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76392064))))[name = string("tm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6361_cast_fp16 = abs(x = linear_115_cast_fp16)[name = string("op_6361_cast_fp16")];
            bool var_6364 = const()[name = string("op_6364"), val = bool(true)];
            tensor<int32, [1]> var_6365_axes_0 = const()[name = string("op_6365_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6365_cast_fp16 = reduce_max(axes = var_6365_axes_0, keep_dims = var_6364, x = var_6361_cast_fp16)[name = string("op_6365_cast_fp16")];
            fp16 var_6366_to_fp16 = const()[name = string("op_6366_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_207_cast_fp16 = maximum(x = var_6365_cast_fp16, y = var_6366_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_207_cast_fp16 = real_div(x = linear_115_cast_fp16, y = s_207_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6369_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<int32, [1]> var_6374_axes_0 = const()[name = string("op_6374_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6374_keep_dims_0 = const()[name = string("op_6374_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6374_cast_fp16 = reduce_mean(axes = var_6374_axes_0, keep_dims = var_6374_keep_dims_0, x = var_6369_cast_fp16)[name = string("op_6374_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6375_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6375_cast_fp16")];
            fp32 var_6376_epsilon_0 = const()[name = string("op_6376_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6376_cast_fp16 = inverse(epsilon = var_6376_epsilon_0, x = var_6375_cast_fp16)[name = string("op_6376_cast_fp16")];
            fp16 var_6377_to_fp16 = const()[name = string("op_6377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6378_cast_fp16 = mul(x = var_6376_cast_fp16, y = var_6377_to_fp16)[name = string("op_6378_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6380_cast_fp16 = add(x = var_6374_cast_fp16, y = var_6378_cast_fp16)[name = string("op_6380_cast_fp16")];
            fp32 var_6381_epsilon_0 = const()[name = string("op_6381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6381_cast_fp16 = rsqrt(epsilon = var_6381_epsilon_0, x = var_6380_cast_fp16)[name = string("op_6381_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_181_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6381_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76393152)))];
            tensor<fp16, [1, 256, 512]> var_6383_cast_fp16 = mul(x = out_181_cast_fp16, y = tm_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6383_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_153_cast_fp16 = add(x = x_203_cast_fp16, y = var_6383_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76394240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76656448))))[name = string("tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_116_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_116_cast_fp16")];
            string var_6390_mode_0 = const()[name = string("op_6390_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_6390_cast_fp16 = gelu(mode = var_6390_mode_0, x = linear_116_cast_fp16)[name = string("op_6390_cast_fp16")];
            tensor<int32, [4]> var_6403_begin_0 = const()[name = string("op_6403_begin_0"), val = tensor<int32, [4]>([0, 0, 12, 0])];
            tensor<int32, [4]> var_6403_end_0 = const()[name = string("op_6403_end_0"), val = tensor<int32, [4]>([1, 256, 13, 512])];
            tensor<bool, [4]> var_6403_end_mask_0 = const()[name = string("op_6403_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6403_squeeze_mask_0 = const()[name = string("op_6403_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_6403_cast_fp16 = slice_by_index(begin = var_6403_begin_0, end = var_6403_end_0, end_mask = var_6403_end_mask_0, squeeze_mask = var_6403_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_6403_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_155_cast_fp16 = mul(x = var_6390_cast_fp16, y = var_6403_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76657536))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76919744))))[name = string("tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_12_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6413_cast_fp16 = abs(x = linear_117_cast_fp16)[name = string("op_6413_cast_fp16")];
            bool var_6416 = const()[name = string("op_6416"), val = bool(true)];
            tensor<int32, [1]> var_6417_axes_0 = const()[name = string("op_6417_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6417_cast_fp16 = reduce_max(axes = var_6417_axes_0, keep_dims = var_6416, x = var_6413_cast_fp16)[name = string("op_6417_cast_fp16")];
            fp16 var_6418_to_fp16 = const()[name = string("op_6418_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_209_cast_fp16 = maximum(x = var_6417_cast_fp16, y = var_6418_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_209_cast_fp16 = real_div(x = linear_117_cast_fp16, y = s_209_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6421_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6421_cast_fp16")];
            tensor<int32, [1]> var_6426_axes_0 = const()[name = string("op_6426_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6426_keep_dims_0 = const()[name = string("op_6426_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6426_cast_fp16 = reduce_mean(axes = var_6426_axes_0, keep_dims = var_6426_keep_dims_0, x = var_6421_cast_fp16)[name = string("op_6426_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6427_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6427_cast_fp16")];
            fp32 var_6428_epsilon_0 = const()[name = string("op_6428_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6428_cast_fp16 = inverse(epsilon = var_6428_epsilon_0, x = var_6427_cast_fp16)[name = string("op_6428_cast_fp16")];
            fp16 var_6429_to_fp16 = const()[name = string("op_6429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6430_cast_fp16 = mul(x = var_6428_cast_fp16, y = var_6429_to_fp16)[name = string("op_6430_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6432_cast_fp16 = add(x = var_6426_cast_fp16, y = var_6430_cast_fp16)[name = string("op_6432_cast_fp16")];
            fp32 var_6433_epsilon_0 = const()[name = string("op_6433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6433_cast_fp16 = rsqrt(epsilon = var_6433_epsilon_0, x = var_6432_cast_fp16)[name = string("op_6433_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_183_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6433_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [512]> tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76920832)))];
            tensor<fp16, [1, 256, 512]> var_6435_cast_fp16 = mul(x = out_183_cast_fp16, y = tm_layers_12_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6435_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_25_cast_fp16 = add(x = input_153_cast_fp16, y = var_6435_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<fp16, [1]> tm_layers_12_layer_scalar_to_fp16 = const()[name = string("tm_layers_12_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.8cp-1])];
            tensor<fp16, [1, 256, 512]> x_209_cast_fp16 = mul(x = h_25_cast_fp16, y = tm_layers_12_layer_scalar_to_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6439_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_6439_cast_fp16")];
            bool var_6442 = const()[name = string("op_6442"), val = bool(true)];
            tensor<int32, [1]> var_6443_axes_0 = const()[name = string("op_6443_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6443_cast_fp16 = reduce_max(axes = var_6443_axes_0, keep_dims = var_6442, x = var_6439_cast_fp16)[name = string("op_6443_cast_fp16")];
            fp16 var_6444_to_fp16 = const()[name = string("op_6444_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_211_cast_fp16 = maximum(x = var_6443_cast_fp16, y = var_6444_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_211_cast_fp16 = real_div(x = x_209_cast_fp16, y = s_211_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6447_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6447_cast_fp16")];
            tensor<int32, [1]> var_6452_axes_0 = const()[name = string("op_6452_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6452_keep_dims_0 = const()[name = string("op_6452_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6452_cast_fp16 = reduce_mean(axes = var_6452_axes_0, keep_dims = var_6452_keep_dims_0, x = var_6447_cast_fp16)[name = string("op_6452_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6453_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6453_cast_fp16")];
            fp32 var_6454_epsilon_0 = const()[name = string("op_6454_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6454_cast_fp16 = inverse(epsilon = var_6454_epsilon_0, x = var_6453_cast_fp16)[name = string("op_6454_cast_fp16")];
            fp16 var_6455_to_fp16 = const()[name = string("op_6455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6456_cast_fp16 = mul(x = var_6454_cast_fp16, y = var_6455_to_fp16)[name = string("op_6456_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6458_cast_fp16 = add(x = var_6452_cast_fp16, y = var_6456_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp32 var_6459_epsilon_0 = const()[name = string("op_6459_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6459_cast_fp16 = rsqrt(epsilon = var_6459_epsilon_0, x = var_6458_cast_fp16)[name = string("op_6459_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_185_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6459_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76921920)))];
            tensor<fp16, [1, 256, 512]> input_157_cast_fp16 = mul(x = out_185_cast_fp16, y = tm_layers_13_input_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76923008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77447360))))[name = string("tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<int32, [4]> var_6469 = const()[name = string("op_6469"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_6470_cast_fp16 = reshape(shape = var_6469, x = linear_118_cast_fp16)[name = string("op_6470_cast_fp16")];
            tensor<int32, [4]> x_211_perm_0 = const()[name = string("x_211_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77449472))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77711680))))[name = string("tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6481 = const()[name = string("op_6481"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6482_cast_fp16 = reshape(shape = var_6481, x = linear_119_cast_fp16)[name = string("op_6482_cast_fp16")];
            tensor<int32, [4]> x_213_perm_0 = const()[name = string("x_213_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77712768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77974976))))[name = string("tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6493 = const()[name = string("op_6493"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6494_cast_fp16 = reshape(shape = var_6493, x = linear_120_cast_fp16)[name = string("op_6494_cast_fp16")];
            tensor<int32, [4]> x_215_perm_0 = const()[name = string("x_215_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_211_cast_fp16 = transpose(perm = x_211_perm_0, x = var_6470_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 4, 256, 256]> var_6498_cast_fp16 = abs(x = x_211_cast_fp16)[name = string("op_6498_cast_fp16")];
            bool var_6501 = const()[name = string("op_6501"), val = bool(true)];
            tensor<int32, [1]> var_6502_axes_0 = const()[name = string("op_6502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_6502_cast_fp16 = reduce_max(axes = var_6502_axes_0, keep_dims = var_6501, x = var_6498_cast_fp16)[name = string("op_6502_cast_fp16")];
            fp16 var_6503_to_fp16 = const()[name = string("op_6503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_213_cast_fp16 = maximum(x = var_6502_cast_fp16, y = var_6503_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_213_cast_fp16 = real_div(x = x_211_cast_fp16, y = s_213_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6506_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6506_cast_fp16")];
            tensor<int32, [1]> var_6511_axes_0 = const()[name = string("op_6511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6511_keep_dims_0 = const()[name = string("op_6511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_6511_cast_fp16 = reduce_mean(axes = var_6511_axes_0, keep_dims = var_6511_keep_dims_0, x = var_6506_cast_fp16)[name = string("op_6511_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6512_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6512_cast_fp16")];
            fp32 var_6513_epsilon_0 = const()[name = string("op_6513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_6513_cast_fp16 = inverse(epsilon = var_6513_epsilon_0, x = var_6512_cast_fp16)[name = string("op_6513_cast_fp16")];
            fp16 var_6514_to_fp16 = const()[name = string("op_6514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_6515_cast_fp16 = mul(x = var_6513_cast_fp16, y = var_6514_to_fp16)[name = string("op_6515_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6517_cast_fp16 = add(x = var_6511_cast_fp16, y = var_6515_cast_fp16)[name = string("op_6517_cast_fp16")];
            fp32 var_6518_epsilon_0 = const()[name = string("op_6518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_6518_cast_fp16 = rsqrt(epsilon = var_6518_epsilon_0, x = var_6517_cast_fp16)[name = string("op_6518_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_187_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6518_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976064)))];
            tensor<fp16, [1, 4, 256, 256]> q_53_cast_fp16 = mul(x = out_187_cast_fp16, y = tm_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("q_53_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_213_cast_fp16 = transpose(perm = x_213_perm_0, x = var_6482_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2, 256, 256]> var_6521_cast_fp16 = abs(x = x_213_cast_fp16)[name = string("op_6521_cast_fp16")];
            bool var_6524 = const()[name = string("op_6524"), val = bool(true)];
            tensor<int32, [1]> var_6525_axes_0 = const()[name = string("op_6525_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6525_cast_fp16 = reduce_max(axes = var_6525_axes_0, keep_dims = var_6524, x = var_6521_cast_fp16)[name = string("op_6525_cast_fp16")];
            fp16 var_6526_to_fp16 = const()[name = string("op_6526_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_215_cast_fp16 = maximum(x = var_6525_cast_fp16, y = var_6526_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_215_cast_fp16 = real_div(x = x_213_cast_fp16, y = s_215_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6529_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6529_cast_fp16")];
            tensor<int32, [1]> var_6534_axes_0 = const()[name = string("op_6534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6534_keep_dims_0 = const()[name = string("op_6534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6534_cast_fp16 = reduce_mean(axes = var_6534_axes_0, keep_dims = var_6534_keep_dims_0, x = var_6529_cast_fp16)[name = string("op_6534_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6535_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6535_cast_fp16")];
            fp32 var_6536_epsilon_0 = const()[name = string("op_6536_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6536_cast_fp16 = inverse(epsilon = var_6536_epsilon_0, x = var_6535_cast_fp16)[name = string("op_6536_cast_fp16")];
            fp16 var_6537_to_fp16 = const()[name = string("op_6537_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6538_cast_fp16 = mul(x = var_6536_cast_fp16, y = var_6537_to_fp16)[name = string("op_6538_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6540_cast_fp16 = add(x = var_6534_cast_fp16, y = var_6538_cast_fp16)[name = string("op_6540_cast_fp16")];
            fp32 var_6541_epsilon_0 = const()[name = string("op_6541_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6541_cast_fp16 = rsqrt(epsilon = var_6541_epsilon_0, x = var_6540_cast_fp16)[name = string("op_6541_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_189_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6541_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [256]> tm_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77976640)))];
            tensor<fp16, [1, 2, 256, 256]> k_79_cast_fp16 = mul(x = out_189_cast_fp16, y = tm_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("k_79_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_215_cast_fp16 = transpose(perm = x_215_perm_0, x = var_6494_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 2, 256, 256]> var_6544_cast_fp16 = abs(x = x_215_cast_fp16)[name = string("op_6544_cast_fp16")];
            bool var_6547 = const()[name = string("op_6547"), val = bool(true)];
            tensor<int32, [1]> var_6548_axes_0 = const()[name = string("op_6548_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6548_cast_fp16 = reduce_max(axes = var_6548_axes_0, keep_dims = var_6547, x = var_6544_cast_fp16)[name = string("op_6548_cast_fp16")];
            fp16 var_6549_to_fp16 = const()[name = string("op_6549_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_217_cast_fp16 = maximum(x = var_6548_cast_fp16, y = var_6549_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_217_cast_fp16 = real_div(x = x_215_cast_fp16, y = s_217_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6552_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6552_cast_fp16")];
            tensor<int32, [1]> var_6557_axes_0 = const()[name = string("op_6557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6557_keep_dims_0 = const()[name = string("op_6557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6557_cast_fp16 = reduce_mean(axes = var_6557_axes_0, keep_dims = var_6557_keep_dims_0, x = var_6552_cast_fp16)[name = string("op_6557_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6558_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6558_cast_fp16")];
            fp32 var_6559_epsilon_0 = const()[name = string("op_6559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6559_cast_fp16 = inverse(epsilon = var_6559_epsilon_0, x = var_6558_cast_fp16)[name = string("op_6559_cast_fp16")];
            fp16 var_6560_to_fp16 = const()[name = string("op_6560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6561_cast_fp16 = mul(x = var_6559_cast_fp16, y = var_6560_to_fp16)[name = string("op_6561_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6563_cast_fp16 = add(x = var_6557_cast_fp16, y = var_6561_cast_fp16)[name = string("op_6563_cast_fp16")];
            fp32 var_6564_epsilon_0 = const()[name = string("op_6564_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6564_cast_fp16 = rsqrt(epsilon = var_6564_epsilon_0, x = var_6563_cast_fp16)[name = string("op_6564_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_53_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6564_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6566_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6566_cast_fp16")];
            tensor<int32, [4]> var_6571_begin_0 = const()[name = string("op_6571_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6571_end_0 = const()[name = string("op_6571_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_6571_end_mask_0 = const()[name = string("op_6571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_6571_cast_fp16 = slice_by_index(begin = var_6571_begin_0, end = var_6571_end_0, end_mask = var_6571_end_mask_0, x = q_53_cast_fp16)[name = string("op_6571_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_6572_cast_fp16 = mul(x = var_6571_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_6572_cast_fp16")];
            tensor<int32, [4]> var_6577_begin_0 = const()[name = string("op_6577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6577_end_0 = const()[name = string("op_6577_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_6577_end_mask_0 = const()[name = string("op_6577_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_6577_cast_fp16 = slice_by_index(begin = var_6577_begin_0, end = var_6577_end_0, end_mask = var_6577_end_mask_0, x = q_53_cast_fp16)[name = string("op_6577_cast_fp16")];
            int32 var_6579 = const()[name = string("op_6579"), val = int32(-1)];
            bool var_6580_interleave_0 = const()[name = string("op_6580_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_6580_cast_fp16 = concat(axis = var_6579, interleave = var_6580_interleave_0, values = (var_6572_cast_fp16, var_6577_cast_fp16))[name = string("op_6580_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6581_cast_fp16 = mul(x = var_6580_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6581_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_55_cast_fp16 = add(x = var_6566_cast_fp16, y = var_6581_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6584_cast_fp16 = mul(x = k_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6584_cast_fp16")];
            tensor<int32, [4]> var_6589_begin_0 = const()[name = string("op_6589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6589_end_0 = const()[name = string("op_6589_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_6589_end_mask_0 = const()[name = string("op_6589_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_6589_cast_fp16 = slice_by_index(begin = var_6589_begin_0, end = var_6589_end_0, end_mask = var_6589_end_mask_0, x = k_79_cast_fp16)[name = string("op_6589_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_6590_cast_fp16 = mul(x = var_6589_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_6590_cast_fp16")];
            tensor<int32, [4]> var_6595_begin_0 = const()[name = string("op_6595_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6595_end_0 = const()[name = string("op_6595_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_6595_end_mask_0 = const()[name = string("op_6595_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_6595_cast_fp16 = slice_by_index(begin = var_6595_begin_0, end = var_6595_end_0, end_mask = var_6595_end_mask_0, x = k_79_cast_fp16)[name = string("op_6595_cast_fp16")];
            int32 var_6597 = const()[name = string("op_6597"), val = int32(-1)];
            bool var_6598_interleave_0 = const()[name = string("op_6598_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_6598_cast_fp16 = concat(axis = var_6597, interleave = var_6598_interleave_0, values = (var_6590_cast_fp16, var_6595_cast_fp16))[name = string("op_6598_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6599_cast_fp16 = mul(x = var_6598_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6599_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_81_cast_fp16 = add(x = var_6584_cast_fp16, y = var_6599_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_26_reps_0 = const()[name = string("tile_26_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = k_81_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [4, 1, 256, 256]> tile_26_cast_fp16 = tile(reps = tile_26_reps_0, x = transpose_52_cast_fp16)[name = string("tile_26_cast_fp16")];
            tensor<int32, [5]> concat_52 = const()[name = string("concat_52"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_52_cast_fp16 = reshape(shape = concat_52, x = tile_26_cast_fp16)[name = string("reshape_52_cast_fp16")];
            tensor<int32, [5]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_53 = const()[name = string("concat_53"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = reshape_52_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [4, 1, 256, 256]> reshape_53_cast_fp16 = reshape(shape = concat_53, x = transpose_53_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_27_reps_0 = const()[name = string("tile_27_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = v_53_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [4, 1, 256, 256]> tile_27_cast_fp16 = tile(reps = tile_27_reps_0, x = transpose_54_cast_fp16)[name = string("tile_27_cast_fp16")];
            tensor<int32, [5]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_54_cast_fp16 = reshape(shape = concat_54, x = tile_27_cast_fp16)[name = string("reshape_54_cast_fp16")];
            tensor<int32, [5]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = reshape_54_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [4, 1, 256, 256]> reshape_55_cast_fp16 = reshape(shape = concat_55, x = transpose_55_cast_fp16)[name = string("reshape_55_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6613_transpose_x_0 = const()[name = string("op_6613_transpose_x_0"), val = bool(false)];
            bool var_6613_transpose_y_0 = const()[name = string("op_6613_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 4, 256, 256]> var_6613_cast_fp16 = matmul(transpose_x = var_6613_transpose_x_0, transpose_y = var_6613_transpose_y_0, x = q_55_cast_fp16, y = transpose_109_cast_fp16)[name = string("op_6613_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6615_cast_fp16 = add(x = var_6613_cast_fp16, y = attention_bias)[name = string("op_6615_cast_fp16")];
            int32 var_6616 = const()[name = string("op_6616"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_6618_cast_fp16 = softmax(axis = var_6616, x = var_6615_cast_fp16)[name = string("op_6618_cast_fp16")];
            bool var_6619_transpose_x_0 = const()[name = string("op_6619_transpose_x_0"), val = bool(false)];
            bool var_6619_transpose_y_0 = const()[name = string("op_6619_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = reshape_55_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 4, 256, 256]> var_6619_cast_fp16 = matmul(transpose_x = var_6619_transpose_x_0, transpose_y = var_6619_transpose_y_0, x = var_6618_cast_fp16, y = v_55_cast_fp16)[name = string("op_6619_cast_fp16")];
            tensor<int32, [4]> var_6622_perm_0 = const()[name = string("op_6622_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6626 = const()[name = string("op_6626"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_6622_cast_fp16 = transpose(perm = var_6622_perm_0, x = var_6619_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 256, 1024]> input_159_cast_fp16 = reshape(shape = var_6626, x = var_6622_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77977216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78501568))))[name = string("tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6631_cast_fp16 = abs(x = linear_121_cast_fp16)[name = string("op_6631_cast_fp16")];
            bool var_6634 = const()[name = string("op_6634"), val = bool(true)];
            tensor<int32, [1]> var_6635_axes_0 = const()[name = string("op_6635_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6635_cast_fp16 = reduce_max(axes = var_6635_axes_0, keep_dims = var_6634, x = var_6631_cast_fp16)[name = string("op_6635_cast_fp16")];
            fp16 var_6636_to_fp16 = const()[name = string("op_6636_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_219_cast_fp16 = maximum(x = var_6635_cast_fp16, y = var_6636_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_219_cast_fp16 = real_div(x = linear_121_cast_fp16, y = s_219_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6639_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6639_cast_fp16")];
            tensor<int32, [1]> var_6644_axes_0 = const()[name = string("op_6644_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6644_keep_dims_0 = const()[name = string("op_6644_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6644_cast_fp16 = reduce_mean(axes = var_6644_axes_0, keep_dims = var_6644_keep_dims_0, x = var_6639_cast_fp16)[name = string("op_6644_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6645_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6645_cast_fp16")];
            fp32 var_6646_epsilon_0 = const()[name = string("op_6646_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6646_cast_fp16 = inverse(epsilon = var_6646_epsilon_0, x = var_6645_cast_fp16)[name = string("op_6646_cast_fp16")];
            fp16 var_6647_to_fp16 = const()[name = string("op_6647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6648_cast_fp16 = mul(x = var_6646_cast_fp16, y = var_6647_to_fp16)[name = string("op_6648_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6650_cast_fp16 = add(x = var_6644_cast_fp16, y = var_6648_cast_fp16)[name = string("op_6650_cast_fp16")];
            fp32 var_6651_epsilon_0 = const()[name = string("op_6651_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6651_cast_fp16 = rsqrt(epsilon = var_6651_epsilon_0, x = var_6650_cast_fp16)[name = string("op_6651_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_191_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6651_cast_fp16)[name = string("out_191_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78502656)))];
            tensor<fp16, [1, 256, 512]> var_6653_cast_fp16 = mul(x = out_191_cast_fp16, y = tm_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6653_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_219_cast_fp16 = add(x = x_209_cast_fp16, y = var_6653_cast_fp16)[name = string("x_219_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6656_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_6656_cast_fp16")];
            bool var_6659 = const()[name = string("op_6659"), val = bool(true)];
            tensor<int32, [1]> var_6660_axes_0 = const()[name = string("op_6660_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6660_cast_fp16 = reduce_max(axes = var_6660_axes_0, keep_dims = var_6659, x = var_6656_cast_fp16)[name = string("op_6660_cast_fp16")];
            fp16 var_6661_to_fp16 = const()[name = string("op_6661_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_221_cast_fp16 = maximum(x = var_6660_cast_fp16, y = var_6661_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_221_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_221_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6664_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6664_cast_fp16")];
            tensor<int32, [1]> var_6669_axes_0 = const()[name = string("op_6669_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6669_keep_dims_0 = const()[name = string("op_6669_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6669_cast_fp16 = reduce_mean(axes = var_6669_axes_0, keep_dims = var_6669_keep_dims_0, x = var_6664_cast_fp16)[name = string("op_6669_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6670_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6670_cast_fp16")];
            fp32 var_6671_epsilon_0 = const()[name = string("op_6671_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6671_cast_fp16 = inverse(epsilon = var_6671_epsilon_0, x = var_6670_cast_fp16)[name = string("op_6671_cast_fp16")];
            fp16 var_6672_to_fp16 = const()[name = string("op_6672_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6673_cast_fp16 = mul(x = var_6671_cast_fp16, y = var_6672_to_fp16)[name = string("op_6673_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6675_cast_fp16 = add(x = var_6669_cast_fp16, y = var_6673_cast_fp16)[name = string("op_6675_cast_fp16")];
            fp32 var_6676_epsilon_0 = const()[name = string("op_6676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6676_cast_fp16 = rsqrt(epsilon = var_6676_epsilon_0, x = var_6675_cast_fp16)[name = string("op_6676_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_193_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6676_cast_fp16)[name = string("out_193_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78503744)))];
            tensor<fp16, [1, 256, 512]> input_161_cast_fp16 = mul(x = out_193_cast_fp16, y = tm_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78504832))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79553472))))[name = string("tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_122_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_122_cast_fp16")];
            string var_6683_mode_0 = const()[name = string("op_6683_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_6683_cast_fp16 = gelu(mode = var_6683_mode_0, x = linear_122_cast_fp16)[name = string("op_6683_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79557632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80606272))))[name = string("tm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_123_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_163_cast_fp16 = mul(x = var_6683_cast_fp16, y = linear_123_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80610432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81659072))))[name = string("tm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6691_cast_fp16 = abs(x = linear_124_cast_fp16)[name = string("op_6691_cast_fp16")];
            bool var_6694 = const()[name = string("op_6694"), val = bool(true)];
            tensor<int32, [1]> var_6695_axes_0 = const()[name = string("op_6695_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6695_cast_fp16 = reduce_max(axes = var_6695_axes_0, keep_dims = var_6694, x = var_6691_cast_fp16)[name = string("op_6695_cast_fp16")];
            fp16 var_6696_to_fp16 = const()[name = string("op_6696_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_223_cast_fp16 = maximum(x = var_6695_cast_fp16, y = var_6696_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_223_cast_fp16 = real_div(x = linear_124_cast_fp16, y = s_223_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6699_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6699_cast_fp16")];
            tensor<int32, [1]> var_6704_axes_0 = const()[name = string("op_6704_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6704_keep_dims_0 = const()[name = string("op_6704_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6704_cast_fp16 = reduce_mean(axes = var_6704_axes_0, keep_dims = var_6704_keep_dims_0, x = var_6699_cast_fp16)[name = string("op_6704_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6705_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6705_cast_fp16")];
            fp32 var_6706_epsilon_0 = const()[name = string("op_6706_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6706_cast_fp16 = inverse(epsilon = var_6706_epsilon_0, x = var_6705_cast_fp16)[name = string("op_6706_cast_fp16")];
            fp16 var_6707_to_fp16 = const()[name = string("op_6707_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6708_cast_fp16 = mul(x = var_6706_cast_fp16, y = var_6707_to_fp16)[name = string("op_6708_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6710_cast_fp16 = add(x = var_6704_cast_fp16, y = var_6708_cast_fp16)[name = string("op_6710_cast_fp16")];
            fp32 var_6711_epsilon_0 = const()[name = string("op_6711_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6711_cast_fp16 = rsqrt(epsilon = var_6711_epsilon_0, x = var_6710_cast_fp16)[name = string("op_6711_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_195_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6711_cast_fp16)[name = string("out_195_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81660160)))];
            tensor<fp16, [1, 256, 512]> var_6713_cast_fp16 = mul(x = out_195_cast_fp16, y = tm_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6713_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_165_cast_fp16 = add(x = x_219_cast_fp16, y = var_6713_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81661248))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81923456))))[name = string("tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_125_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_125_cast_fp16")];
            string var_6720_mode_0 = const()[name = string("op_6720_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_6720_cast_fp16 = gelu(mode = var_6720_mode_0, x = linear_125_cast_fp16)[name = string("op_6720_cast_fp16")];
            tensor<int32, [4]> var_6733_begin_0 = const()[name = string("op_6733_begin_0"), val = tensor<int32, [4]>([0, 0, 13, 0])];
            tensor<int32, [4]> var_6733_end_0 = const()[name = string("op_6733_end_0"), val = tensor<int32, [4]>([1, 256, 14, 512])];
            tensor<bool, [4]> var_6733_end_mask_0 = const()[name = string("op_6733_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_6733_squeeze_mask_0 = const()[name = string("op_6733_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_6733_cast_fp16 = slice_by_index(begin = var_6733_begin_0, end = var_6733_end_0, end_mask = var_6733_end_mask_0, squeeze_mask = var_6733_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_6733_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_167_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6733_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81924544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82186752))))[name = string("tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_13_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6743_cast_fp16 = abs(x = linear_126_cast_fp16)[name = string("op_6743_cast_fp16")];
            bool var_6746 = const()[name = string("op_6746"), val = bool(true)];
            tensor<int32, [1]> var_6747_axes_0 = const()[name = string("op_6747_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6747_cast_fp16 = reduce_max(axes = var_6747_axes_0, keep_dims = var_6746, x = var_6743_cast_fp16)[name = string("op_6747_cast_fp16")];
            fp16 var_6748_to_fp16 = const()[name = string("op_6748_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_225_cast_fp16 = maximum(x = var_6747_cast_fp16, y = var_6748_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_225_cast_fp16 = real_div(x = linear_126_cast_fp16, y = s_225_cast_fp16)[name = string("y_225_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6751_cast_fp16 = mul(x = y_225_cast_fp16, y = y_225_cast_fp16)[name = string("op_6751_cast_fp16")];
            tensor<int32, [1]> var_6756_axes_0 = const()[name = string("op_6756_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6756_keep_dims_0 = const()[name = string("op_6756_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6756_cast_fp16 = reduce_mean(axes = var_6756_axes_0, keep_dims = var_6756_keep_dims_0, x = var_6751_cast_fp16)[name = string("op_6756_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6757_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6757_cast_fp16")];
            fp32 var_6758_epsilon_0 = const()[name = string("op_6758_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6758_cast_fp16 = inverse(epsilon = var_6758_epsilon_0, x = var_6757_cast_fp16)[name = string("op_6758_cast_fp16")];
            fp16 var_6759_to_fp16 = const()[name = string("op_6759_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6760_cast_fp16 = mul(x = var_6758_cast_fp16, y = var_6759_to_fp16)[name = string("op_6760_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6762_cast_fp16 = add(x = var_6756_cast_fp16, y = var_6760_cast_fp16)[name = string("op_6762_cast_fp16")];
            fp32 var_6763_epsilon_0 = const()[name = string("op_6763_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6763_cast_fp16 = rsqrt(epsilon = var_6763_epsilon_0, x = var_6762_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_197_cast_fp16 = mul(x = y_225_cast_fp16, y = var_6763_cast_fp16)[name = string("out_197_cast_fp16")];
            tensor<fp16, [512]> tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82187840)))];
            tensor<fp16, [1, 256, 512]> var_6765_cast_fp16 = mul(x = out_197_cast_fp16, y = tm_layers_13_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_6765_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_27_cast_fp16 = add(x = input_165_cast_fp16, y = var_6765_cast_fp16)[name = string("h_27_cast_fp16")];
            tensor<fp16, [1]> tm_layers_13_layer_scalar_to_fp16 = const()[name = string("tm_layers_13_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.9cp-1])];
            tensor<fp16, [1, 256, 512]> x_225_cast_fp16 = mul(x = h_27_cast_fp16, y = tm_layers_13_layer_scalar_to_fp16)[name = string("x_225_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6769_cast_fp16 = abs(x = x_225_cast_fp16)[name = string("op_6769_cast_fp16")];
            bool var_6772 = const()[name = string("op_6772"), val = bool(true)];
            tensor<int32, [1]> var_6773_axes_0 = const()[name = string("op_6773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6773_cast_fp16 = reduce_max(axes = var_6773_axes_0, keep_dims = var_6772, x = var_6769_cast_fp16)[name = string("op_6773_cast_fp16")];
            fp16 var_6774_to_fp16 = const()[name = string("op_6774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_227_cast_fp16 = maximum(x = var_6773_cast_fp16, y = var_6774_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_227_cast_fp16 = real_div(x = x_225_cast_fp16, y = s_227_cast_fp16)[name = string("y_227_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6777_cast_fp16 = mul(x = y_227_cast_fp16, y = y_227_cast_fp16)[name = string("op_6777_cast_fp16")];
            tensor<int32, [1]> var_6782_axes_0 = const()[name = string("op_6782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6782_keep_dims_0 = const()[name = string("op_6782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6782_cast_fp16 = reduce_mean(axes = var_6782_axes_0, keep_dims = var_6782_keep_dims_0, x = var_6777_cast_fp16)[name = string("op_6782_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6783_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6783_cast_fp16")];
            fp32 var_6784_epsilon_0 = const()[name = string("op_6784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6784_cast_fp16 = inverse(epsilon = var_6784_epsilon_0, x = var_6783_cast_fp16)[name = string("op_6784_cast_fp16")];
            fp16 var_6785_to_fp16 = const()[name = string("op_6785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6786_cast_fp16 = mul(x = var_6784_cast_fp16, y = var_6785_to_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6788_cast_fp16 = add(x = var_6782_cast_fp16, y = var_6786_cast_fp16)[name = string("op_6788_cast_fp16")];
            fp32 var_6789_epsilon_0 = const()[name = string("op_6789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6789_cast_fp16 = rsqrt(epsilon = var_6789_epsilon_0, x = var_6788_cast_fp16)[name = string("op_6789_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_199_cast_fp16 = mul(x = y_227_cast_fp16, y = var_6789_cast_fp16)[name = string("out_199_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82188928)))];
            tensor<fp16, [1, 256, 512]> input_169_cast_fp16 = mul(x = out_199_cast_fp16, y = tm_layers_14_input_layernorm_weight_to_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82190016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82714368))))[name = string("tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6799 = const()[name = string("op_6799"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_6800_cast_fp16 = reshape(shape = var_6799, x = linear_127_cast_fp16)[name = string("op_6800_cast_fp16")];
            tensor<int32, [4]> x_227_perm_0 = const()[name = string("x_227_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82716480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82978688))))[name = string("tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6811 = const()[name = string("op_6811"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6812_cast_fp16 = reshape(shape = var_6811, x = linear_128_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<int32, [4]> x_229_perm_0 = const()[name = string("x_229_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82979776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83241984))))[name = string("tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_6823 = const()[name = string("op_6823"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_6824_cast_fp16 = reshape(shape = var_6823, x = linear_129_cast_fp16)[name = string("op_6824_cast_fp16")];
            tensor<int32, [4]> x_231_perm_0 = const()[name = string("x_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_227_cast_fp16 = transpose(perm = x_227_perm_0, x = var_6800_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 4, 256, 256]> var_6828_cast_fp16 = abs(x = x_227_cast_fp16)[name = string("op_6828_cast_fp16")];
            bool var_6831 = const()[name = string("op_6831"), val = bool(true)];
            tensor<int32, [1]> var_6832_axes_0 = const()[name = string("op_6832_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_6832_cast_fp16 = reduce_max(axes = var_6832_axes_0, keep_dims = var_6831, x = var_6828_cast_fp16)[name = string("op_6832_cast_fp16")];
            fp16 var_6833_to_fp16 = const()[name = string("op_6833_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_229_cast_fp16 = maximum(x = var_6832_cast_fp16, y = var_6833_to_fp16)[name = string("s_229_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_229_cast_fp16 = real_div(x = x_227_cast_fp16, y = s_229_cast_fp16)[name = string("y_229_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6836_cast_fp16 = mul(x = y_229_cast_fp16, y = y_229_cast_fp16)[name = string("op_6836_cast_fp16")];
            tensor<int32, [1]> var_6841_axes_0 = const()[name = string("op_6841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6841_keep_dims_0 = const()[name = string("op_6841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_6841_cast_fp16 = reduce_mean(axes = var_6841_axes_0, keep_dims = var_6841_keep_dims_0, x = var_6836_cast_fp16)[name = string("op_6841_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6842_cast_fp16 = mul(x = s_229_cast_fp16, y = s_229_cast_fp16)[name = string("op_6842_cast_fp16")];
            fp32 var_6843_epsilon_0 = const()[name = string("op_6843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_6843_cast_fp16 = inverse(epsilon = var_6843_epsilon_0, x = var_6842_cast_fp16)[name = string("op_6843_cast_fp16")];
            fp16 var_6844_to_fp16 = const()[name = string("op_6844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_6845_cast_fp16 = mul(x = var_6843_cast_fp16, y = var_6844_to_fp16)[name = string("op_6845_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_6847_cast_fp16 = add(x = var_6841_cast_fp16, y = var_6845_cast_fp16)[name = string("op_6847_cast_fp16")];
            fp32 var_6848_epsilon_0 = const()[name = string("op_6848_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_6848_cast_fp16 = rsqrt(epsilon = var_6848_epsilon_0, x = var_6847_cast_fp16)[name = string("op_6848_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_201_cast_fp16 = mul(x = y_229_cast_fp16, y = var_6848_cast_fp16)[name = string("out_201_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243072)))];
            tensor<fp16, [1, 4, 256, 256]> q_57_cast_fp16 = mul(x = out_201_cast_fp16, y = tm_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("q_57_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_229_cast_fp16 = transpose(perm = x_229_perm_0, x = var_6812_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 2, 256, 256]> var_6851_cast_fp16 = abs(x = x_229_cast_fp16)[name = string("op_6851_cast_fp16")];
            bool var_6854 = const()[name = string("op_6854"), val = bool(true)];
            tensor<int32, [1]> var_6855_axes_0 = const()[name = string("op_6855_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6855_cast_fp16 = reduce_max(axes = var_6855_axes_0, keep_dims = var_6854, x = var_6851_cast_fp16)[name = string("op_6855_cast_fp16")];
            fp16 var_6856_to_fp16 = const()[name = string("op_6856_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_231_cast_fp16 = maximum(x = var_6855_cast_fp16, y = var_6856_to_fp16)[name = string("s_231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_231_cast_fp16 = real_div(x = x_229_cast_fp16, y = s_231_cast_fp16)[name = string("y_231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6859_cast_fp16 = mul(x = y_231_cast_fp16, y = y_231_cast_fp16)[name = string("op_6859_cast_fp16")];
            tensor<int32, [1]> var_6864_axes_0 = const()[name = string("op_6864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6864_keep_dims_0 = const()[name = string("op_6864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6864_cast_fp16 = reduce_mean(axes = var_6864_axes_0, keep_dims = var_6864_keep_dims_0, x = var_6859_cast_fp16)[name = string("op_6864_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6865_cast_fp16 = mul(x = s_231_cast_fp16, y = s_231_cast_fp16)[name = string("op_6865_cast_fp16")];
            fp32 var_6866_epsilon_0 = const()[name = string("op_6866_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6866_cast_fp16 = inverse(epsilon = var_6866_epsilon_0, x = var_6865_cast_fp16)[name = string("op_6866_cast_fp16")];
            fp16 var_6867_to_fp16 = const()[name = string("op_6867_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6868_cast_fp16 = mul(x = var_6866_cast_fp16, y = var_6867_to_fp16)[name = string("op_6868_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6870_cast_fp16 = add(x = var_6864_cast_fp16, y = var_6868_cast_fp16)[name = string("op_6870_cast_fp16")];
            fp32 var_6871_epsilon_0 = const()[name = string("op_6871_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6871_cast_fp16 = rsqrt(epsilon = var_6871_epsilon_0, x = var_6870_cast_fp16)[name = string("op_6871_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_203_cast_fp16 = mul(x = y_231_cast_fp16, y = var_6871_cast_fp16)[name = string("out_203_cast_fp16")];
            tensor<fp16, [256]> tm_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83243648)))];
            tensor<fp16, [1, 2, 256, 256]> k_85_cast_fp16 = mul(x = out_203_cast_fp16, y = tm_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("k_85_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_231_cast_fp16 = transpose(perm = x_231_perm_0, x = var_6824_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 2, 256, 256]> var_6874_cast_fp16 = abs(x = x_231_cast_fp16)[name = string("op_6874_cast_fp16")];
            bool var_6877 = const()[name = string("op_6877"), val = bool(true)];
            tensor<int32, [1]> var_6878_axes_0 = const()[name = string("op_6878_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_6878_cast_fp16 = reduce_max(axes = var_6878_axes_0, keep_dims = var_6877, x = var_6874_cast_fp16)[name = string("op_6878_cast_fp16")];
            fp16 var_6879_to_fp16 = const()[name = string("op_6879_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_233_cast_fp16 = maximum(x = var_6878_cast_fp16, y = var_6879_to_fp16)[name = string("s_233_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_233_cast_fp16 = real_div(x = x_231_cast_fp16, y = s_233_cast_fp16)[name = string("y_233_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6882_cast_fp16 = mul(x = y_233_cast_fp16, y = y_233_cast_fp16)[name = string("op_6882_cast_fp16")];
            tensor<int32, [1]> var_6887_axes_0 = const()[name = string("op_6887_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6887_keep_dims_0 = const()[name = string("op_6887_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_6887_cast_fp16 = reduce_mean(axes = var_6887_axes_0, keep_dims = var_6887_keep_dims_0, x = var_6882_cast_fp16)[name = string("op_6887_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6888_cast_fp16 = mul(x = s_233_cast_fp16, y = s_233_cast_fp16)[name = string("op_6888_cast_fp16")];
            fp32 var_6889_epsilon_0 = const()[name = string("op_6889_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_6889_cast_fp16 = inverse(epsilon = var_6889_epsilon_0, x = var_6888_cast_fp16)[name = string("op_6889_cast_fp16")];
            fp16 var_6890_to_fp16 = const()[name = string("op_6890_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_6891_cast_fp16 = mul(x = var_6889_cast_fp16, y = var_6890_to_fp16)[name = string("op_6891_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_6893_cast_fp16 = add(x = var_6887_cast_fp16, y = var_6891_cast_fp16)[name = string("op_6893_cast_fp16")];
            fp32 var_6894_epsilon_0 = const()[name = string("op_6894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_6894_cast_fp16 = rsqrt(epsilon = var_6894_epsilon_0, x = var_6893_cast_fp16)[name = string("op_6894_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_57_cast_fp16 = mul(x = y_233_cast_fp16, y = var_6894_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6896_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6896_cast_fp16")];
            tensor<int32, [4]> var_6901_begin_0 = const()[name = string("op_6901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6901_end_0 = const()[name = string("op_6901_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_6901_end_mask_0 = const()[name = string("op_6901_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_6901_cast_fp16 = slice_by_index(begin = var_6901_begin_0, end = var_6901_end_0, end_mask = var_6901_end_mask_0, x = q_57_cast_fp16)[name = string("op_6901_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_6902_cast_fp16 = mul(x = var_6901_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_6902_cast_fp16")];
            tensor<int32, [4]> var_6907_begin_0 = const()[name = string("op_6907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6907_end_0 = const()[name = string("op_6907_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_6907_end_mask_0 = const()[name = string("op_6907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_6907_cast_fp16 = slice_by_index(begin = var_6907_begin_0, end = var_6907_end_0, end_mask = var_6907_end_mask_0, x = q_57_cast_fp16)[name = string("op_6907_cast_fp16")];
            int32 var_6909 = const()[name = string("op_6909"), val = int32(-1)];
            bool var_6910_interleave_0 = const()[name = string("op_6910_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_6910_cast_fp16 = concat(axis = var_6909, interleave = var_6910_interleave_0, values = (var_6902_cast_fp16, var_6907_cast_fp16))[name = string("op_6910_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6911_cast_fp16 = mul(x = var_6910_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6911_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_59_cast_fp16 = add(x = var_6896_cast_fp16, y = var_6911_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6914_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6914_cast_fp16")];
            tensor<int32, [4]> var_6919_begin_0 = const()[name = string("op_6919_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_6919_end_0 = const()[name = string("op_6919_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_6919_end_mask_0 = const()[name = string("op_6919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_6919_cast_fp16 = slice_by_index(begin = var_6919_begin_0, end = var_6919_end_0, end_mask = var_6919_end_mask_0, x = k_85_cast_fp16)[name = string("op_6919_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_6920_cast_fp16 = mul(x = var_6919_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_6920_cast_fp16")];
            tensor<int32, [4]> var_6925_begin_0 = const()[name = string("op_6925_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6925_end_0 = const()[name = string("op_6925_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_6925_end_mask_0 = const()[name = string("op_6925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_6925_cast_fp16 = slice_by_index(begin = var_6925_begin_0, end = var_6925_end_0, end_mask = var_6925_end_mask_0, x = k_85_cast_fp16)[name = string("op_6925_cast_fp16")];
            int32 var_6927 = const()[name = string("op_6927"), val = int32(-1)];
            bool var_6928_interleave_0 = const()[name = string("op_6928_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_6928_cast_fp16 = concat(axis = var_6927, interleave = var_6928_interleave_0, values = (var_6920_cast_fp16, var_6925_cast_fp16))[name = string("op_6928_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_6929_cast_fp16 = mul(x = var_6928_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6929_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_87_cast_fp16 = add(x = var_6914_cast_fp16, y = var_6929_cast_fp16)[name = string("k_87_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_28_reps_0 = const()[name = string("tile_28_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = k_87_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 1, 256, 256]> tile_28_cast_fp16 = tile(reps = tile_28_reps_0, x = transpose_56_cast_fp16)[name = string("tile_28_cast_fp16")];
            tensor<int32, [5]> concat_56 = const()[name = string("concat_56"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_56_cast_fp16 = reshape(shape = concat_56, x = tile_28_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [5]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_57 = const()[name = string("concat_57"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 1, 256, 256]> reshape_57_cast_fp16 = reshape(shape = concat_57, x = transpose_57_cast_fp16)[name = string("reshape_57_cast_fp16")];
            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_29_reps_0 = const()[name = string("tile_29_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = v_57_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 1, 256, 256]> tile_29_cast_fp16 = tile(reps = tile_29_reps_0, x = transpose_58_cast_fp16)[name = string("tile_29_cast_fp16")];
            tensor<int32, [5]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_58_cast_fp16 = reshape(shape = concat_58, x = tile_29_cast_fp16)[name = string("reshape_58_cast_fp16")];
            tensor<int32, [5]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = reshape_58_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 1, 256, 256]> reshape_59_cast_fp16 = reshape(shape = concat_59, x = transpose_59_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_6943_transpose_x_0 = const()[name = string("op_6943_transpose_x_0"), val = bool(false)];
            bool var_6943_transpose_y_0 = const()[name = string("op_6943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_110_cast_fp16 = transpose(perm = transpose_110_perm_0, x = reshape_57_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 4, 256, 256]> var_6943_cast_fp16 = matmul(transpose_x = var_6943_transpose_x_0, transpose_y = var_6943_transpose_y_0, x = q_59_cast_fp16, y = transpose_110_cast_fp16)[name = string("op_6943_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_6945_cast_fp16 = add(x = var_6943_cast_fp16, y = attention_bias)[name = string("op_6945_cast_fp16")];
            int32 var_6946 = const()[name = string("op_6946"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_6948_cast_fp16 = softmax(axis = var_6946, x = var_6945_cast_fp16)[name = string("op_6948_cast_fp16")];
            bool var_6949_transpose_x_0 = const()[name = string("op_6949_transpose_x_0"), val = bool(false)];
            bool var_6949_transpose_y_0 = const()[name = string("op_6949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 4, 256, 256]> var_6949_cast_fp16 = matmul(transpose_x = var_6949_transpose_x_0, transpose_y = var_6949_transpose_y_0, x = var_6948_cast_fp16, y = v_59_cast_fp16)[name = string("op_6949_cast_fp16")];
            tensor<int32, [4]> var_6952_perm_0 = const()[name = string("op_6952_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6956 = const()[name = string("op_6956"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_6952_cast_fp16 = transpose(perm = var_6952_perm_0, x = var_6949_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 256, 1024]> input_171_cast_fp16 = reshape(shape = var_6956, x = var_6952_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83244224))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83768576))))[name = string("tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6961_cast_fp16 = abs(x = linear_130_cast_fp16)[name = string("op_6961_cast_fp16")];
            bool var_6964 = const()[name = string("op_6964"), val = bool(true)];
            tensor<int32, [1]> var_6965_axes_0 = const()[name = string("op_6965_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6965_cast_fp16 = reduce_max(axes = var_6965_axes_0, keep_dims = var_6964, x = var_6961_cast_fp16)[name = string("op_6965_cast_fp16")];
            fp16 var_6966_to_fp16 = const()[name = string("op_6966_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_235_cast_fp16 = maximum(x = var_6965_cast_fp16, y = var_6966_to_fp16)[name = string("s_235_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_235_cast_fp16 = real_div(x = linear_130_cast_fp16, y = s_235_cast_fp16)[name = string("y_235_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6969_cast_fp16 = mul(x = y_235_cast_fp16, y = y_235_cast_fp16)[name = string("op_6969_cast_fp16")];
            tensor<int32, [1]> var_6974_axes_0 = const()[name = string("op_6974_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6974_keep_dims_0 = const()[name = string("op_6974_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6974_cast_fp16 = reduce_mean(axes = var_6974_axes_0, keep_dims = var_6974_keep_dims_0, x = var_6969_cast_fp16)[name = string("op_6974_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6975_cast_fp16 = mul(x = s_235_cast_fp16, y = s_235_cast_fp16)[name = string("op_6975_cast_fp16")];
            fp32 var_6976_epsilon_0 = const()[name = string("op_6976_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_6976_cast_fp16 = inverse(epsilon = var_6976_epsilon_0, x = var_6975_cast_fp16)[name = string("op_6976_cast_fp16")];
            fp16 var_6977_to_fp16 = const()[name = string("op_6977_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_6978_cast_fp16 = mul(x = var_6976_cast_fp16, y = var_6977_to_fp16)[name = string("op_6978_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_6980_cast_fp16 = add(x = var_6974_cast_fp16, y = var_6978_cast_fp16)[name = string("op_6980_cast_fp16")];
            fp32 var_6981_epsilon_0 = const()[name = string("op_6981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_6981_cast_fp16 = rsqrt(epsilon = var_6981_epsilon_0, x = var_6980_cast_fp16)[name = string("op_6981_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_205_cast_fp16 = mul(x = y_235_cast_fp16, y = var_6981_cast_fp16)[name = string("out_205_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83769664)))];
            tensor<fp16, [1, 256, 512]> var_6983_cast_fp16 = mul(x = out_205_cast_fp16, y = tm_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_6983_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_235_cast_fp16 = add(x = x_225_cast_fp16, y = var_6983_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6986_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_6986_cast_fp16")];
            bool var_6989 = const()[name = string("op_6989"), val = bool(true)];
            tensor<int32, [1]> var_6990_axes_0 = const()[name = string("op_6990_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_6990_cast_fp16 = reduce_max(axes = var_6990_axes_0, keep_dims = var_6989, x = var_6986_cast_fp16)[name = string("op_6990_cast_fp16")];
            fp16 var_6991_to_fp16 = const()[name = string("op_6991_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_237_cast_fp16 = maximum(x = var_6990_cast_fp16, y = var_6991_to_fp16)[name = string("s_237_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_237_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_237_cast_fp16)[name = string("y_237_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_6994_cast_fp16 = mul(x = y_237_cast_fp16, y = y_237_cast_fp16)[name = string("op_6994_cast_fp16")];
            tensor<int32, [1]> var_6999_axes_0 = const()[name = string("op_6999_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6999_keep_dims_0 = const()[name = string("op_6999_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_6999_cast_fp16 = reduce_mean(axes = var_6999_axes_0, keep_dims = var_6999_keep_dims_0, x = var_6994_cast_fp16)[name = string("op_6999_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7000_cast_fp16 = mul(x = s_237_cast_fp16, y = s_237_cast_fp16)[name = string("op_7000_cast_fp16")];
            fp32 var_7001_epsilon_0 = const()[name = string("op_7001_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7001_cast_fp16 = inverse(epsilon = var_7001_epsilon_0, x = var_7000_cast_fp16)[name = string("op_7001_cast_fp16")];
            fp16 var_7002_to_fp16 = const()[name = string("op_7002_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7003_cast_fp16 = mul(x = var_7001_cast_fp16, y = var_7002_to_fp16)[name = string("op_7003_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7005_cast_fp16 = add(x = var_6999_cast_fp16, y = var_7003_cast_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7006_cast_fp16 = rsqrt(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_207_cast_fp16 = mul(x = y_237_cast_fp16, y = var_7006_cast_fp16)[name = string("out_207_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83770752)))];
            tensor<fp16, [1, 256, 512]> input_173_cast_fp16 = mul(x = out_207_cast_fp16, y = tm_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83771840))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84820480))))[name = string("tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_131_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_131_cast_fp16")];
            string var_7013_mode_0 = const()[name = string("op_7013_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_7013_cast_fp16 = gelu(mode = var_7013_mode_0, x = linear_131_cast_fp16)[name = string("op_7013_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84824640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85873280))))[name = string("tm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_132_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_175_cast_fp16 = mul(x = var_7013_cast_fp16, y = linear_132_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85877440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86926080))))[name = string("tm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7021_cast_fp16 = abs(x = linear_133_cast_fp16)[name = string("op_7021_cast_fp16")];
            bool var_7024 = const()[name = string("op_7024"), val = bool(true)];
            tensor<int32, [1]> var_7025_axes_0 = const()[name = string("op_7025_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7025_cast_fp16 = reduce_max(axes = var_7025_axes_0, keep_dims = var_7024, x = var_7021_cast_fp16)[name = string("op_7025_cast_fp16")];
            fp16 var_7026_to_fp16 = const()[name = string("op_7026_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_239_cast_fp16 = maximum(x = var_7025_cast_fp16, y = var_7026_to_fp16)[name = string("s_239_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_239_cast_fp16 = real_div(x = linear_133_cast_fp16, y = s_239_cast_fp16)[name = string("y_239_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7029_cast_fp16 = mul(x = y_239_cast_fp16, y = y_239_cast_fp16)[name = string("op_7029_cast_fp16")];
            tensor<int32, [1]> var_7034_axes_0 = const()[name = string("op_7034_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7034_keep_dims_0 = const()[name = string("op_7034_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7034_cast_fp16 = reduce_mean(axes = var_7034_axes_0, keep_dims = var_7034_keep_dims_0, x = var_7029_cast_fp16)[name = string("op_7034_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7035_cast_fp16 = mul(x = s_239_cast_fp16, y = s_239_cast_fp16)[name = string("op_7035_cast_fp16")];
            fp32 var_7036_epsilon_0 = const()[name = string("op_7036_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7036_cast_fp16 = inverse(epsilon = var_7036_epsilon_0, x = var_7035_cast_fp16)[name = string("op_7036_cast_fp16")];
            fp16 var_7037_to_fp16 = const()[name = string("op_7037_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7038_cast_fp16 = mul(x = var_7036_cast_fp16, y = var_7037_to_fp16)[name = string("op_7038_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7040_cast_fp16 = add(x = var_7034_cast_fp16, y = var_7038_cast_fp16)[name = string("op_7040_cast_fp16")];
            fp32 var_7041_epsilon_0 = const()[name = string("op_7041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7041_cast_fp16 = rsqrt(epsilon = var_7041_epsilon_0, x = var_7040_cast_fp16)[name = string("op_7041_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_209_cast_fp16 = mul(x = y_239_cast_fp16, y = var_7041_cast_fp16)[name = string("out_209_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86927168)))];
            tensor<fp16, [1, 256, 512]> var_7043_cast_fp16 = mul(x = out_209_cast_fp16, y = tm_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7043_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_177_cast_fp16 = add(x = x_235_cast_fp16, y = var_7043_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86928256))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87190464))))[name = string("tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_134_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_134_cast_fp16")];
            string var_7050_mode_0 = const()[name = string("op_7050_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_7050_cast_fp16 = gelu(mode = var_7050_mode_0, x = linear_134_cast_fp16)[name = string("op_7050_cast_fp16")];
            tensor<int32, [4]> var_7063_begin_0 = const()[name = string("op_7063_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
            tensor<int32, [4]> var_7063_end_0 = const()[name = string("op_7063_end_0"), val = tensor<int32, [4]>([1, 256, 15, 512])];
            tensor<bool, [4]> var_7063_end_mask_0 = const()[name = string("op_7063_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7063_squeeze_mask_0 = const()[name = string("op_7063_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_7063_cast_fp16 = slice_by_index(begin = var_7063_begin_0, end = var_7063_end_0, end_mask = var_7063_end_mask_0, squeeze_mask = var_7063_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_7063_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_179_cast_fp16 = mul(x = var_7050_cast_fp16, y = var_7063_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87191552))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87453760))))[name = string("tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_14_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7073_cast_fp16 = abs(x = linear_135_cast_fp16)[name = string("op_7073_cast_fp16")];
            bool var_7076 = const()[name = string("op_7076"), val = bool(true)];
            tensor<int32, [1]> var_7077_axes_0 = const()[name = string("op_7077_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7077_cast_fp16 = reduce_max(axes = var_7077_axes_0, keep_dims = var_7076, x = var_7073_cast_fp16)[name = string("op_7077_cast_fp16")];
            fp16 var_7078_to_fp16 = const()[name = string("op_7078_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_241_cast_fp16 = maximum(x = var_7077_cast_fp16, y = var_7078_to_fp16)[name = string("s_241_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_241_cast_fp16 = real_div(x = linear_135_cast_fp16, y = s_241_cast_fp16)[name = string("y_241_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7081_cast_fp16 = mul(x = y_241_cast_fp16, y = y_241_cast_fp16)[name = string("op_7081_cast_fp16")];
            tensor<int32, [1]> var_7086_axes_0 = const()[name = string("op_7086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7086_keep_dims_0 = const()[name = string("op_7086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7086_cast_fp16 = reduce_mean(axes = var_7086_axes_0, keep_dims = var_7086_keep_dims_0, x = var_7081_cast_fp16)[name = string("op_7086_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7087_cast_fp16 = mul(x = s_241_cast_fp16, y = s_241_cast_fp16)[name = string("op_7087_cast_fp16")];
            fp32 var_7088_epsilon_0 = const()[name = string("op_7088_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7088_cast_fp16 = inverse(epsilon = var_7088_epsilon_0, x = var_7087_cast_fp16)[name = string("op_7088_cast_fp16")];
            fp16 var_7089_to_fp16 = const()[name = string("op_7089_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7090_cast_fp16 = mul(x = var_7088_cast_fp16, y = var_7089_to_fp16)[name = string("op_7090_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7092_cast_fp16 = add(x = var_7086_cast_fp16, y = var_7090_cast_fp16)[name = string("op_7092_cast_fp16")];
            fp32 var_7093_epsilon_0 = const()[name = string("op_7093_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7093_cast_fp16 = rsqrt(epsilon = var_7093_epsilon_0, x = var_7092_cast_fp16)[name = string("op_7093_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_211_cast_fp16 = mul(x = y_241_cast_fp16, y = var_7093_cast_fp16)[name = string("out_211_cast_fp16")];
            tensor<fp16, [512]> tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87454848)))];
            tensor<fp16, [1, 256, 512]> var_7095_cast_fp16 = mul(x = out_211_cast_fp16, y = tm_layers_14_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7095_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_29_cast_fp16 = add(x = input_177_cast_fp16, y = var_7095_cast_fp16)[name = string("h_29_cast_fp16")];
            tensor<fp16, [1]> tm_layers_14_layer_scalar_to_fp16 = const()[name = string("tm_layers_14_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1p-1])];
            tensor<fp16, [1, 256, 512]> x_241_cast_fp16 = mul(x = h_29_cast_fp16, y = tm_layers_14_layer_scalar_to_fp16)[name = string("x_241_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7099_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_7099_cast_fp16")];
            bool var_7102 = const()[name = string("op_7102"), val = bool(true)];
            tensor<int32, [1]> var_7103_axes_0 = const()[name = string("op_7103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7103_cast_fp16 = reduce_max(axes = var_7103_axes_0, keep_dims = var_7102, x = var_7099_cast_fp16)[name = string("op_7103_cast_fp16")];
            fp16 var_7104_to_fp16 = const()[name = string("op_7104_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_243_cast_fp16 = maximum(x = var_7103_cast_fp16, y = var_7104_to_fp16)[name = string("s_243_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_243_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_243_cast_fp16)[name = string("y_243_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7107_cast_fp16 = mul(x = y_243_cast_fp16, y = y_243_cast_fp16)[name = string("op_7107_cast_fp16")];
            tensor<int32, [1]> var_7112_axes_0 = const()[name = string("op_7112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7112_keep_dims_0 = const()[name = string("op_7112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7112_cast_fp16 = reduce_mean(axes = var_7112_axes_0, keep_dims = var_7112_keep_dims_0, x = var_7107_cast_fp16)[name = string("op_7112_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7113_cast_fp16 = mul(x = s_243_cast_fp16, y = s_243_cast_fp16)[name = string("op_7113_cast_fp16")];
            fp32 var_7114_epsilon_0 = const()[name = string("op_7114_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7114_cast_fp16 = inverse(epsilon = var_7114_epsilon_0, x = var_7113_cast_fp16)[name = string("op_7114_cast_fp16")];
            fp16 var_7115_to_fp16 = const()[name = string("op_7115_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7116_cast_fp16 = mul(x = var_7114_cast_fp16, y = var_7115_to_fp16)[name = string("op_7116_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7118_cast_fp16 = add(x = var_7112_cast_fp16, y = var_7116_cast_fp16)[name = string("op_7118_cast_fp16")];
            fp32 var_7119_epsilon_0 = const()[name = string("op_7119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7119_cast_fp16 = rsqrt(epsilon = var_7119_epsilon_0, x = var_7118_cast_fp16)[name = string("op_7119_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_213_cast_fp16 = mul(x = y_243_cast_fp16, y = var_7119_cast_fp16)[name = string("out_213_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87455936)))];
            tensor<fp16, [1, 256, 512]> input_181_cast_fp16 = mul(x = out_213_cast_fp16, y = tm_layers_15_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87457024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87981376))))[name = string("tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<int32, [4]> var_7129 = const()[name = string("op_7129"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_7130_cast_fp16 = reshape(shape = var_7129, x = linear_136_cast_fp16)[name = string("op_7130_cast_fp16")];
            tensor<int32, [4]> x_243_perm_0 = const()[name = string("x_243_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87983488))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88245696))))[name = string("tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<int32, [4]> var_7141 = const()[name = string("op_7141"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_7142_cast_fp16 = reshape(shape = var_7141, x = linear_137_cast_fp16)[name = string("op_7142_cast_fp16")];
            tensor<int32, [4]> x_245_perm_0 = const()[name = string("x_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88246784))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88508992))))[name = string("tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_7153 = const()[name = string("op_7153"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_7154_cast_fp16 = reshape(shape = var_7153, x = linear_138_cast_fp16)[name = string("op_7154_cast_fp16")];
            tensor<int32, [4]> x_247_perm_0 = const()[name = string("x_247_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_243_cast_fp16 = transpose(perm = x_243_perm_0, x = var_7130_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 4, 256, 256]> var_7158_cast_fp16 = abs(x = x_243_cast_fp16)[name = string("op_7158_cast_fp16")];
            bool var_7161 = const()[name = string("op_7161"), val = bool(true)];
            tensor<int32, [1]> var_7162_axes_0 = const()[name = string("op_7162_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_7162_cast_fp16 = reduce_max(axes = var_7162_axes_0, keep_dims = var_7161, x = var_7158_cast_fp16)[name = string("op_7162_cast_fp16")];
            fp16 var_7163_to_fp16 = const()[name = string("op_7163_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_245_cast_fp16 = maximum(x = var_7162_cast_fp16, y = var_7163_to_fp16)[name = string("s_245_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_245_cast_fp16 = real_div(x = x_243_cast_fp16, y = s_245_cast_fp16)[name = string("y_245_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7166_cast_fp16 = mul(x = y_245_cast_fp16, y = y_245_cast_fp16)[name = string("op_7166_cast_fp16")];
            tensor<int32, [1]> var_7171_axes_0 = const()[name = string("op_7171_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7171_keep_dims_0 = const()[name = string("op_7171_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_7171_cast_fp16 = reduce_mean(axes = var_7171_axes_0, keep_dims = var_7171_keep_dims_0, x = var_7166_cast_fp16)[name = string("op_7171_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7172_cast_fp16 = mul(x = s_245_cast_fp16, y = s_245_cast_fp16)[name = string("op_7172_cast_fp16")];
            fp32 var_7173_epsilon_0 = const()[name = string("op_7173_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_7173_cast_fp16 = inverse(epsilon = var_7173_epsilon_0, x = var_7172_cast_fp16)[name = string("op_7173_cast_fp16")];
            fp16 var_7174_to_fp16 = const()[name = string("op_7174_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_7175_cast_fp16 = mul(x = var_7173_cast_fp16, y = var_7174_to_fp16)[name = string("op_7175_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7177_cast_fp16 = add(x = var_7171_cast_fp16, y = var_7175_cast_fp16)[name = string("op_7177_cast_fp16")];
            fp32 var_7178_epsilon_0 = const()[name = string("op_7178_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_7178_cast_fp16 = rsqrt(epsilon = var_7178_epsilon_0, x = var_7177_cast_fp16)[name = string("op_7178_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_215_cast_fp16 = mul(x = y_245_cast_fp16, y = var_7178_cast_fp16)[name = string("out_215_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_61_cast_fp16 = mul(x = out_215_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_61_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_245_cast_fp16 = transpose(perm = x_245_perm_0, x = var_7142_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 2, 256, 256]> var_7181_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_7181_cast_fp16")];
            bool var_7184 = const()[name = string("op_7184"), val = bool(true)];
            tensor<int32, [1]> var_7185_axes_0 = const()[name = string("op_7185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_7185_cast_fp16 = reduce_max(axes = var_7185_axes_0, keep_dims = var_7184, x = var_7181_cast_fp16)[name = string("op_7185_cast_fp16")];
            fp16 var_7186_to_fp16 = const()[name = string("op_7186_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_247_cast_fp16 = maximum(x = var_7185_cast_fp16, y = var_7186_to_fp16)[name = string("s_247_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_247_cast_fp16 = real_div(x = x_245_cast_fp16, y = s_247_cast_fp16)[name = string("y_247_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7189_cast_fp16 = mul(x = y_247_cast_fp16, y = y_247_cast_fp16)[name = string("op_7189_cast_fp16")];
            tensor<int32, [1]> var_7194_axes_0 = const()[name = string("op_7194_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7194_keep_dims_0 = const()[name = string("op_7194_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_7194_cast_fp16 = reduce_mean(axes = var_7194_axes_0, keep_dims = var_7194_keep_dims_0, x = var_7189_cast_fp16)[name = string("op_7194_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7195_cast_fp16 = mul(x = s_247_cast_fp16, y = s_247_cast_fp16)[name = string("op_7195_cast_fp16")];
            fp32 var_7196_epsilon_0 = const()[name = string("op_7196_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_7196_cast_fp16 = inverse(epsilon = var_7196_epsilon_0, x = var_7195_cast_fp16)[name = string("op_7196_cast_fp16")];
            fp16 var_7197_to_fp16 = const()[name = string("op_7197_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_7198_cast_fp16 = mul(x = var_7196_cast_fp16, y = var_7197_to_fp16)[name = string("op_7198_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7200_cast_fp16 = add(x = var_7194_cast_fp16, y = var_7198_cast_fp16)[name = string("op_7200_cast_fp16")];
            fp32 var_7201_epsilon_0 = const()[name = string("op_7201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_7201_cast_fp16 = rsqrt(epsilon = var_7201_epsilon_0, x = var_7200_cast_fp16)[name = string("op_7201_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_217_cast_fp16 = mul(x = y_247_cast_fp16, y = var_7201_cast_fp16)[name = string("out_217_cast_fp16")];
            tensor<fp16, [256]> tm_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510080)))];
            tensor<fp16, [1, 2, 256, 256]> k_91_cast_fp16 = mul(x = out_217_cast_fp16, y = tm_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("k_91_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_247_cast_fp16 = transpose(perm = x_247_perm_0, x = var_7154_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 2, 256, 256]> var_7204_cast_fp16 = abs(x = x_247_cast_fp16)[name = string("op_7204_cast_fp16")];
            bool var_7207 = const()[name = string("op_7207"), val = bool(true)];
            tensor<int32, [1]> var_7208_axes_0 = const()[name = string("op_7208_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_7208_cast_fp16 = reduce_max(axes = var_7208_axes_0, keep_dims = var_7207, x = var_7204_cast_fp16)[name = string("op_7208_cast_fp16")];
            fp16 var_7209_to_fp16 = const()[name = string("op_7209_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_249_cast_fp16 = maximum(x = var_7208_cast_fp16, y = var_7209_to_fp16)[name = string("s_249_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_249_cast_fp16 = real_div(x = x_247_cast_fp16, y = s_249_cast_fp16)[name = string("y_249_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7212_cast_fp16 = mul(x = y_249_cast_fp16, y = y_249_cast_fp16)[name = string("op_7212_cast_fp16")];
            tensor<int32, [1]> var_7217_axes_0 = const()[name = string("op_7217_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7217_keep_dims_0 = const()[name = string("op_7217_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_7217_cast_fp16 = reduce_mean(axes = var_7217_axes_0, keep_dims = var_7217_keep_dims_0, x = var_7212_cast_fp16)[name = string("op_7217_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7218_cast_fp16 = mul(x = s_249_cast_fp16, y = s_249_cast_fp16)[name = string("op_7218_cast_fp16")];
            fp32 var_7219_epsilon_0 = const()[name = string("op_7219_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_7219_cast_fp16 = inverse(epsilon = var_7219_epsilon_0, x = var_7218_cast_fp16)[name = string("op_7219_cast_fp16")];
            fp16 var_7220_to_fp16 = const()[name = string("op_7220_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_7221_cast_fp16 = mul(x = var_7219_cast_fp16, y = var_7220_to_fp16)[name = string("op_7221_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7223_cast_fp16 = add(x = var_7217_cast_fp16, y = var_7221_cast_fp16)[name = string("op_7223_cast_fp16")];
            fp32 var_7224_epsilon_0 = const()[name = string("op_7224_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_7224_cast_fp16 = rsqrt(epsilon = var_7224_epsilon_0, x = var_7223_cast_fp16)[name = string("op_7224_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_61_cast_fp16 = mul(x = y_249_cast_fp16, y = var_7224_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7226_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7226_cast_fp16")];
            tensor<int32, [4]> var_7231_begin_0 = const()[name = string("op_7231_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7231_end_0 = const()[name = string("op_7231_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_7231_end_mask_0 = const()[name = string("op_7231_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_7231_cast_fp16 = slice_by_index(begin = var_7231_begin_0, end = var_7231_end_0, end_mask = var_7231_end_mask_0, x = q_61_cast_fp16)[name = string("op_7231_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_7232_cast_fp16 = mul(x = var_7231_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_7232_cast_fp16")];
            tensor<int32, [4]> var_7237_begin_0 = const()[name = string("op_7237_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7237_end_0 = const()[name = string("op_7237_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_7237_end_mask_0 = const()[name = string("op_7237_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_7237_cast_fp16 = slice_by_index(begin = var_7237_begin_0, end = var_7237_end_0, end_mask = var_7237_end_mask_0, x = q_61_cast_fp16)[name = string("op_7237_cast_fp16")];
            int32 var_7239 = const()[name = string("op_7239"), val = int32(-1)];
            bool var_7240_interleave_0 = const()[name = string("op_7240_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_7240_cast_fp16 = concat(axis = var_7239, interleave = var_7240_interleave_0, values = (var_7232_cast_fp16, var_7237_cast_fp16))[name = string("op_7240_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7241_cast_fp16 = mul(x = var_7240_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7241_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_63_cast_fp16 = add(x = var_7226_cast_fp16, y = var_7241_cast_fp16)[name = string("q_63_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7244_cast_fp16 = mul(x = k_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7244_cast_fp16")];
            tensor<int32, [4]> var_7249_begin_0 = const()[name = string("op_7249_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7249_end_0 = const()[name = string("op_7249_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_7249_end_mask_0 = const()[name = string("op_7249_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_7249_cast_fp16 = slice_by_index(begin = var_7249_begin_0, end = var_7249_end_0, end_mask = var_7249_end_mask_0, x = k_91_cast_fp16)[name = string("op_7249_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_7250_cast_fp16 = mul(x = var_7249_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_7250_cast_fp16")];
            tensor<int32, [4]> var_7255_begin_0 = const()[name = string("op_7255_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7255_end_0 = const()[name = string("op_7255_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_7255_end_mask_0 = const()[name = string("op_7255_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_7255_cast_fp16 = slice_by_index(begin = var_7255_begin_0, end = var_7255_end_0, end_mask = var_7255_end_mask_0, x = k_91_cast_fp16)[name = string("op_7255_cast_fp16")];
            int32 var_7257 = const()[name = string("op_7257"), val = int32(-1)];
            bool var_7258_interleave_0 = const()[name = string("op_7258_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_7258_cast_fp16 = concat(axis = var_7257, interleave = var_7258_interleave_0, values = (var_7250_cast_fp16, var_7255_cast_fp16))[name = string("op_7258_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7259_cast_fp16 = mul(x = var_7258_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7259_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_93_cast_fp16 = add(x = var_7244_cast_fp16, y = var_7259_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_30_reps_0 = const()[name = string("tile_30_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = k_93_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 1, 256, 256]> tile_30_cast_fp16 = tile(reps = tile_30_reps_0, x = transpose_60_cast_fp16)[name = string("tile_30_cast_fp16")];
            tensor<int32, [5]> concat_60 = const()[name = string("concat_60"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_60_cast_fp16 = reshape(shape = concat_60, x = tile_30_cast_fp16)[name = string("reshape_60_cast_fp16")];
            tensor<int32, [5]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = reshape_60_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 1, 256, 256]> reshape_61_cast_fp16 = reshape(shape = concat_61, x = transpose_61_cast_fp16)[name = string("reshape_61_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_31_reps_0 = const()[name = string("tile_31_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = v_61_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 1, 256, 256]> tile_31_cast_fp16 = tile(reps = tile_31_reps_0, x = transpose_62_cast_fp16)[name = string("tile_31_cast_fp16")];
            tensor<int32, [5]> concat_62 = const()[name = string("concat_62"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_62_cast_fp16 = reshape(shape = concat_62, x = tile_31_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [5]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_63 = const()[name = string("concat_63"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 1, 256, 256]> reshape_63_cast_fp16 = reshape(shape = concat_63, x = transpose_63_cast_fp16)[name = string("reshape_63_cast_fp16")];
            tensor<int32, [4]> v_63_perm_0 = const()[name = string("v_63_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7273_transpose_x_0 = const()[name = string("op_7273_transpose_x_0"), val = bool(false)];
            bool var_7273_transpose_y_0 = const()[name = string("op_7273_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = reshape_61_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 4, 256, 256]> var_7273_cast_fp16 = matmul(transpose_x = var_7273_transpose_x_0, transpose_y = var_7273_transpose_y_0, x = q_63_cast_fp16, y = transpose_111_cast_fp16)[name = string("op_7273_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7275_cast_fp16 = add(x = var_7273_cast_fp16, y = attention_bias)[name = string("op_7275_cast_fp16")];
            int32 var_7276 = const()[name = string("op_7276"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_7278_cast_fp16 = softmax(axis = var_7276, x = var_7275_cast_fp16)[name = string("op_7278_cast_fp16")];
            bool var_7279_transpose_x_0 = const()[name = string("op_7279_transpose_x_0"), val = bool(false)];
            bool var_7279_transpose_y_0 = const()[name = string("op_7279_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = reshape_63_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 4, 256, 256]> var_7279_cast_fp16 = matmul(transpose_x = var_7279_transpose_x_0, transpose_y = var_7279_transpose_y_0, x = var_7278_cast_fp16, y = v_63_cast_fp16)[name = string("op_7279_cast_fp16")];
            tensor<int32, [4]> var_7282_perm_0 = const()[name = string("op_7282_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7286 = const()[name = string("op_7286"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_7282_cast_fp16 = transpose(perm = var_7282_perm_0, x = var_7279_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 256, 1024]> input_183_cast_fp16 = reshape(shape = var_7286, x = var_7282_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88510656))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89035008))))[name = string("tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7291_cast_fp16 = abs(x = linear_139_cast_fp16)[name = string("op_7291_cast_fp16")];
            bool var_7294 = const()[name = string("op_7294"), val = bool(true)];
            tensor<int32, [1]> var_7295_axes_0 = const()[name = string("op_7295_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7295_cast_fp16 = reduce_max(axes = var_7295_axes_0, keep_dims = var_7294, x = var_7291_cast_fp16)[name = string("op_7295_cast_fp16")];
            fp16 var_7296_to_fp16 = const()[name = string("op_7296_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_251_cast_fp16 = maximum(x = var_7295_cast_fp16, y = var_7296_to_fp16)[name = string("s_251_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_251_cast_fp16 = real_div(x = linear_139_cast_fp16, y = s_251_cast_fp16)[name = string("y_251_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7299_cast_fp16 = mul(x = y_251_cast_fp16, y = y_251_cast_fp16)[name = string("op_7299_cast_fp16")];
            tensor<int32, [1]> var_7304_axes_0 = const()[name = string("op_7304_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7304_keep_dims_0 = const()[name = string("op_7304_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7304_cast_fp16 = reduce_mean(axes = var_7304_axes_0, keep_dims = var_7304_keep_dims_0, x = var_7299_cast_fp16)[name = string("op_7304_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7305_cast_fp16 = mul(x = s_251_cast_fp16, y = s_251_cast_fp16)[name = string("op_7305_cast_fp16")];
            fp32 var_7306_epsilon_0 = const()[name = string("op_7306_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7306_cast_fp16 = inverse(epsilon = var_7306_epsilon_0, x = var_7305_cast_fp16)[name = string("op_7306_cast_fp16")];
            fp16 var_7307_to_fp16 = const()[name = string("op_7307_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7308_cast_fp16 = mul(x = var_7306_cast_fp16, y = var_7307_to_fp16)[name = string("op_7308_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7310_cast_fp16 = add(x = var_7304_cast_fp16, y = var_7308_cast_fp16)[name = string("op_7310_cast_fp16")];
            fp32 var_7311_epsilon_0 = const()[name = string("op_7311_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7311_cast_fp16 = rsqrt(epsilon = var_7311_epsilon_0, x = var_7310_cast_fp16)[name = string("op_7311_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_219_cast_fp16 = mul(x = y_251_cast_fp16, y = var_7311_cast_fp16)[name = string("out_219_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89036096)))];
            tensor<fp16, [1, 256, 512]> var_7313_cast_fp16 = mul(x = out_219_cast_fp16, y = tm_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_7313_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_251_cast_fp16 = add(x = x_241_cast_fp16, y = var_7313_cast_fp16)[name = string("x_251_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7316_cast_fp16 = abs(x = x_251_cast_fp16)[name = string("op_7316_cast_fp16")];
            bool var_7319 = const()[name = string("op_7319"), val = bool(true)];
            tensor<int32, [1]> var_7320_axes_0 = const()[name = string("op_7320_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7320_cast_fp16 = reduce_max(axes = var_7320_axes_0, keep_dims = var_7319, x = var_7316_cast_fp16)[name = string("op_7320_cast_fp16")];
            fp16 var_7321_to_fp16 = const()[name = string("op_7321_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_253_cast_fp16 = maximum(x = var_7320_cast_fp16, y = var_7321_to_fp16)[name = string("s_253_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_253_cast_fp16 = real_div(x = x_251_cast_fp16, y = s_253_cast_fp16)[name = string("y_253_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7324_cast_fp16 = mul(x = y_253_cast_fp16, y = y_253_cast_fp16)[name = string("op_7324_cast_fp16")];
            tensor<int32, [1]> var_7329_axes_0 = const()[name = string("op_7329_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7329_keep_dims_0 = const()[name = string("op_7329_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7329_cast_fp16 = reduce_mean(axes = var_7329_axes_0, keep_dims = var_7329_keep_dims_0, x = var_7324_cast_fp16)[name = string("op_7329_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7330_cast_fp16 = mul(x = s_253_cast_fp16, y = s_253_cast_fp16)[name = string("op_7330_cast_fp16")];
            fp32 var_7331_epsilon_0 = const()[name = string("op_7331_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7331_cast_fp16 = inverse(epsilon = var_7331_epsilon_0, x = var_7330_cast_fp16)[name = string("op_7331_cast_fp16")];
            fp16 var_7332_to_fp16 = const()[name = string("op_7332_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7333_cast_fp16 = mul(x = var_7331_cast_fp16, y = var_7332_to_fp16)[name = string("op_7333_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7335_cast_fp16 = add(x = var_7329_cast_fp16, y = var_7333_cast_fp16)[name = string("op_7335_cast_fp16")];
            fp32 var_7336_epsilon_0 = const()[name = string("op_7336_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7336_cast_fp16 = rsqrt(epsilon = var_7336_epsilon_0, x = var_7335_cast_fp16)[name = string("op_7336_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_221_cast_fp16 = mul(x = y_253_cast_fp16, y = var_7336_cast_fp16)[name = string("out_221_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89037184)))];
            tensor<fp16, [1, 256, 512]> input_185_cast_fp16 = mul(x = out_221_cast_fp16, y = tm_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89038272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90086912))))[name = string("tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_140_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_140_cast_fp16")];
            string var_7343_mode_0 = const()[name = string("op_7343_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_7343_cast_fp16 = gelu(mode = var_7343_mode_0, x = linear_140_cast_fp16)[name = string("op_7343_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90091072))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91139712))))[name = string("tm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_141_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_187_cast_fp16 = mul(x = var_7343_cast_fp16, y = linear_141_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91143872))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92192512))))[name = string("tm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7351_cast_fp16 = abs(x = linear_142_cast_fp16)[name = string("op_7351_cast_fp16")];
            bool var_7354 = const()[name = string("op_7354"), val = bool(true)];
            tensor<int32, [1]> var_7355_axes_0 = const()[name = string("op_7355_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7355_cast_fp16 = reduce_max(axes = var_7355_axes_0, keep_dims = var_7354, x = var_7351_cast_fp16)[name = string("op_7355_cast_fp16")];
            fp16 var_7356_to_fp16 = const()[name = string("op_7356_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_255_cast_fp16 = maximum(x = var_7355_cast_fp16, y = var_7356_to_fp16)[name = string("s_255_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_255_cast_fp16 = real_div(x = linear_142_cast_fp16, y = s_255_cast_fp16)[name = string("y_255_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7359_cast_fp16 = mul(x = y_255_cast_fp16, y = y_255_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<int32, [1]> var_7364_axes_0 = const()[name = string("op_7364_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7364_keep_dims_0 = const()[name = string("op_7364_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7364_cast_fp16 = reduce_mean(axes = var_7364_axes_0, keep_dims = var_7364_keep_dims_0, x = var_7359_cast_fp16)[name = string("op_7364_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7365_cast_fp16 = mul(x = s_255_cast_fp16, y = s_255_cast_fp16)[name = string("op_7365_cast_fp16")];
            fp32 var_7366_epsilon_0 = const()[name = string("op_7366_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7366_cast_fp16 = inverse(epsilon = var_7366_epsilon_0, x = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            fp16 var_7367_to_fp16 = const()[name = string("op_7367_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7368_cast_fp16 = mul(x = var_7366_cast_fp16, y = var_7367_to_fp16)[name = string("op_7368_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7370_cast_fp16 = add(x = var_7364_cast_fp16, y = var_7368_cast_fp16)[name = string("op_7370_cast_fp16")];
            fp32 var_7371_epsilon_0 = const()[name = string("op_7371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7371_cast_fp16 = rsqrt(epsilon = var_7371_epsilon_0, x = var_7370_cast_fp16)[name = string("op_7371_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_223_cast_fp16 = mul(x = y_255_cast_fp16, y = var_7371_cast_fp16)[name = string("out_223_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92193600)))];
            tensor<fp16, [1, 256, 512]> var_7373_cast_fp16 = mul(x = out_223_cast_fp16, y = tm_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7373_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_189_cast_fp16 = add(x = x_251_cast_fp16, y = var_7373_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92194688))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92456896))))[name = string("tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_143_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_143_cast_fp16")];
            string var_7380_mode_0 = const()[name = string("op_7380_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_7380_cast_fp16 = gelu(mode = var_7380_mode_0, x = linear_143_cast_fp16)[name = string("op_7380_cast_fp16")];
            tensor<int32, [4]> var_7393_begin_0 = const()[name = string("op_7393_begin_0"), val = tensor<int32, [4]>([0, 0, 15, 0])];
            tensor<int32, [4]> var_7393_end_0 = const()[name = string("op_7393_end_0"), val = tensor<int32, [4]>([1, 256, 16, 512])];
            tensor<bool, [4]> var_7393_end_mask_0 = const()[name = string("op_7393_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7393_squeeze_mask_0 = const()[name = string("op_7393_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_7393_cast_fp16 = slice_by_index(begin = var_7393_begin_0, end = var_7393_end_0, end_mask = var_7393_end_mask_0, squeeze_mask = var_7393_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_7393_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_191_cast_fp16 = mul(x = var_7380_cast_fp16, y = var_7393_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92457984))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92720192))))[name = string("tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_15_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7403_cast_fp16 = abs(x = linear_144_cast_fp16)[name = string("op_7403_cast_fp16")];
            bool var_7406 = const()[name = string("op_7406"), val = bool(true)];
            tensor<int32, [1]> var_7407_axes_0 = const()[name = string("op_7407_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7407_cast_fp16 = reduce_max(axes = var_7407_axes_0, keep_dims = var_7406, x = var_7403_cast_fp16)[name = string("op_7407_cast_fp16")];
            fp16 var_7408_to_fp16 = const()[name = string("op_7408_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_257_cast_fp16 = maximum(x = var_7407_cast_fp16, y = var_7408_to_fp16)[name = string("s_257_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_257_cast_fp16 = real_div(x = linear_144_cast_fp16, y = s_257_cast_fp16)[name = string("y_257_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7411_cast_fp16 = mul(x = y_257_cast_fp16, y = y_257_cast_fp16)[name = string("op_7411_cast_fp16")];
            tensor<int32, [1]> var_7416_axes_0 = const()[name = string("op_7416_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7416_keep_dims_0 = const()[name = string("op_7416_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7416_cast_fp16 = reduce_mean(axes = var_7416_axes_0, keep_dims = var_7416_keep_dims_0, x = var_7411_cast_fp16)[name = string("op_7416_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7417_cast_fp16 = mul(x = s_257_cast_fp16, y = s_257_cast_fp16)[name = string("op_7417_cast_fp16")];
            fp32 var_7418_epsilon_0 = const()[name = string("op_7418_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7418_cast_fp16 = inverse(epsilon = var_7418_epsilon_0, x = var_7417_cast_fp16)[name = string("op_7418_cast_fp16")];
            fp16 var_7419_to_fp16 = const()[name = string("op_7419_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7420_cast_fp16 = mul(x = var_7418_cast_fp16, y = var_7419_to_fp16)[name = string("op_7420_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7422_cast_fp16 = add(x = var_7416_cast_fp16, y = var_7420_cast_fp16)[name = string("op_7422_cast_fp16")];
            fp32 var_7423_epsilon_0 = const()[name = string("op_7423_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7423_cast_fp16 = rsqrt(epsilon = var_7423_epsilon_0, x = var_7422_cast_fp16)[name = string("op_7423_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_225_cast_fp16 = mul(x = y_257_cast_fp16, y = var_7423_cast_fp16)[name = string("out_225_cast_fp16")];
            tensor<fp16, [512]> tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92721280)))];
            tensor<fp16, [1, 256, 512]> var_7425_cast_fp16 = mul(x = out_225_cast_fp16, y = tm_layers_15_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7425_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_31_cast_fp16 = add(x = input_189_cast_fp16, y = var_7425_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<fp16, [1]> tm_layers_15_layer_scalar_to_fp16 = const()[name = string("tm_layers_15_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7ap-1])];
            tensor<fp16, [1, 256, 512]> x_257_cast_fp16 = mul(x = h_31_cast_fp16, y = tm_layers_15_layer_scalar_to_fp16)[name = string("x_257_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7429_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_7429_cast_fp16")];
            bool var_7432 = const()[name = string("op_7432"), val = bool(true)];
            tensor<int32, [1]> var_7433_axes_0 = const()[name = string("op_7433_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7433_cast_fp16 = reduce_max(axes = var_7433_axes_0, keep_dims = var_7432, x = var_7429_cast_fp16)[name = string("op_7433_cast_fp16")];
            fp16 var_7434_to_fp16 = const()[name = string("op_7434_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_259_cast_fp16 = maximum(x = var_7433_cast_fp16, y = var_7434_to_fp16)[name = string("s_259_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_259_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_259_cast_fp16)[name = string("y_259_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7437_cast_fp16 = mul(x = y_259_cast_fp16, y = y_259_cast_fp16)[name = string("op_7437_cast_fp16")];
            tensor<int32, [1]> var_7442_axes_0 = const()[name = string("op_7442_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7442_keep_dims_0 = const()[name = string("op_7442_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7442_cast_fp16 = reduce_mean(axes = var_7442_axes_0, keep_dims = var_7442_keep_dims_0, x = var_7437_cast_fp16)[name = string("op_7442_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7443_cast_fp16 = mul(x = s_259_cast_fp16, y = s_259_cast_fp16)[name = string("op_7443_cast_fp16")];
            fp32 var_7444_epsilon_0 = const()[name = string("op_7444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7444_cast_fp16 = inverse(epsilon = var_7444_epsilon_0, x = var_7443_cast_fp16)[name = string("op_7444_cast_fp16")];
            fp16 var_7445_to_fp16 = const()[name = string("op_7445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7446_cast_fp16 = mul(x = var_7444_cast_fp16, y = var_7445_to_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7448_cast_fp16 = add(x = var_7442_cast_fp16, y = var_7446_cast_fp16)[name = string("op_7448_cast_fp16")];
            fp32 var_7449_epsilon_0 = const()[name = string("op_7449_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7449_cast_fp16 = rsqrt(epsilon = var_7449_epsilon_0, x = var_7448_cast_fp16)[name = string("op_7449_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_227_cast_fp16 = mul(x = y_259_cast_fp16, y = var_7449_cast_fp16)[name = string("out_227_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92722368)))];
            tensor<fp16, [1, 256, 512]> input_193_cast_fp16 = mul(x = out_227_cast_fp16, y = tm_layers_16_input_layernorm_weight_to_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92723456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93247808))))[name = string("tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<int32, [4]> var_7459 = const()[name = string("op_7459"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_7460_cast_fp16 = reshape(shape = var_7459, x = linear_145_cast_fp16)[name = string("op_7460_cast_fp16")];
            tensor<int32, [4]> x_259_perm_0 = const()[name = string("x_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93249920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93512128))))[name = string("tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<int32, [4]> var_7471 = const()[name = string("op_7471"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_7472_cast_fp16 = reshape(shape = var_7471, x = linear_146_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<int32, [4]> x_261_perm_0 = const()[name = string("x_261_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93513216))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93775424))))[name = string("tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7483 = const()[name = string("op_7483"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_7484_cast_fp16 = reshape(shape = var_7483, x = linear_147_cast_fp16)[name = string("op_7484_cast_fp16")];
            tensor<int32, [4]> x_263_perm_0 = const()[name = string("x_263_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_259_cast_fp16 = transpose(perm = x_259_perm_0, x = var_7460_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 4, 256, 256]> var_7488_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_7488_cast_fp16")];
            bool var_7491 = const()[name = string("op_7491"), val = bool(true)];
            tensor<int32, [1]> var_7492_axes_0 = const()[name = string("op_7492_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_7492_cast_fp16 = reduce_max(axes = var_7492_axes_0, keep_dims = var_7491, x = var_7488_cast_fp16)[name = string("op_7492_cast_fp16")];
            fp16 var_7493_to_fp16 = const()[name = string("op_7493_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_261_cast_fp16 = maximum(x = var_7492_cast_fp16, y = var_7493_to_fp16)[name = string("s_261_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_261_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_261_cast_fp16)[name = string("y_261_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7496_cast_fp16 = mul(x = y_261_cast_fp16, y = y_261_cast_fp16)[name = string("op_7496_cast_fp16")];
            tensor<int32, [1]> var_7501_axes_0 = const()[name = string("op_7501_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7501_keep_dims_0 = const()[name = string("op_7501_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_7501_cast_fp16 = reduce_mean(axes = var_7501_axes_0, keep_dims = var_7501_keep_dims_0, x = var_7496_cast_fp16)[name = string("op_7501_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7502_cast_fp16 = mul(x = s_261_cast_fp16, y = s_261_cast_fp16)[name = string("op_7502_cast_fp16")];
            fp32 var_7503_epsilon_0 = const()[name = string("op_7503_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_7503_cast_fp16 = inverse(epsilon = var_7503_epsilon_0, x = var_7502_cast_fp16)[name = string("op_7503_cast_fp16")];
            fp16 var_7504_to_fp16 = const()[name = string("op_7504_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_7505_cast_fp16 = mul(x = var_7503_cast_fp16, y = var_7504_to_fp16)[name = string("op_7505_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7507_cast_fp16 = add(x = var_7501_cast_fp16, y = var_7505_cast_fp16)[name = string("op_7507_cast_fp16")];
            fp32 var_7508_epsilon_0 = const()[name = string("op_7508_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_7508_cast_fp16 = rsqrt(epsilon = var_7508_epsilon_0, x = var_7507_cast_fp16)[name = string("op_7508_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_229_cast_fp16 = mul(x = y_261_cast_fp16, y = var_7508_cast_fp16)[name = string("out_229_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_65_cast_fp16 = mul(x = out_229_cast_fp16, y = tm_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("q_65_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_261_cast_fp16 = transpose(perm = x_261_perm_0, x = var_7472_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 256, 256]> var_7511_cast_fp16 = abs(x = x_261_cast_fp16)[name = string("op_7511_cast_fp16")];
            bool var_7514 = const()[name = string("op_7514"), val = bool(true)];
            tensor<int32, [1]> var_7515_axes_0 = const()[name = string("op_7515_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_7515_cast_fp16 = reduce_max(axes = var_7515_axes_0, keep_dims = var_7514, x = var_7511_cast_fp16)[name = string("op_7515_cast_fp16")];
            fp16 var_7516_to_fp16 = const()[name = string("op_7516_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_263_cast_fp16 = maximum(x = var_7515_cast_fp16, y = var_7516_to_fp16)[name = string("s_263_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_263_cast_fp16 = real_div(x = x_261_cast_fp16, y = s_263_cast_fp16)[name = string("y_263_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7519_cast_fp16 = mul(x = y_263_cast_fp16, y = y_263_cast_fp16)[name = string("op_7519_cast_fp16")];
            tensor<int32, [1]> var_7524_axes_0 = const()[name = string("op_7524_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7524_keep_dims_0 = const()[name = string("op_7524_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_7524_cast_fp16 = reduce_mean(axes = var_7524_axes_0, keep_dims = var_7524_keep_dims_0, x = var_7519_cast_fp16)[name = string("op_7524_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7525_cast_fp16 = mul(x = s_263_cast_fp16, y = s_263_cast_fp16)[name = string("op_7525_cast_fp16")];
            fp32 var_7526_epsilon_0 = const()[name = string("op_7526_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_7526_cast_fp16 = inverse(epsilon = var_7526_epsilon_0, x = var_7525_cast_fp16)[name = string("op_7526_cast_fp16")];
            fp16 var_7527_to_fp16 = const()[name = string("op_7527_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_7528_cast_fp16 = mul(x = var_7526_cast_fp16, y = var_7527_to_fp16)[name = string("op_7528_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7530_cast_fp16 = add(x = var_7524_cast_fp16, y = var_7528_cast_fp16)[name = string("op_7530_cast_fp16")];
            fp32 var_7531_epsilon_0 = const()[name = string("op_7531_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_7531_cast_fp16 = rsqrt(epsilon = var_7531_epsilon_0, x = var_7530_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_231_cast_fp16 = mul(x = y_263_cast_fp16, y = var_7531_cast_fp16)[name = string("out_231_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_97_cast_fp16 = mul(x = out_231_cast_fp16, y = tm_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("k_97_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_263_cast_fp16 = transpose(perm = x_263_perm_0, x = var_7484_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 256, 256]> var_7534_cast_fp16 = abs(x = x_263_cast_fp16)[name = string("op_7534_cast_fp16")];
            bool var_7537 = const()[name = string("op_7537"), val = bool(true)];
            tensor<int32, [1]> var_7538_axes_0 = const()[name = string("op_7538_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_7538_cast_fp16 = reduce_max(axes = var_7538_axes_0, keep_dims = var_7537, x = var_7534_cast_fp16)[name = string("op_7538_cast_fp16")];
            fp16 var_7539_to_fp16 = const()[name = string("op_7539_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_265_cast_fp16 = maximum(x = var_7538_cast_fp16, y = var_7539_to_fp16)[name = string("s_265_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_265_cast_fp16 = real_div(x = x_263_cast_fp16, y = s_265_cast_fp16)[name = string("y_265_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7542_cast_fp16 = mul(x = y_265_cast_fp16, y = y_265_cast_fp16)[name = string("op_7542_cast_fp16")];
            tensor<int32, [1]> var_7547_axes_0 = const()[name = string("op_7547_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7547_keep_dims_0 = const()[name = string("op_7547_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_7547_cast_fp16 = reduce_mean(axes = var_7547_axes_0, keep_dims = var_7547_keep_dims_0, x = var_7542_cast_fp16)[name = string("op_7547_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7548_cast_fp16 = mul(x = s_265_cast_fp16, y = s_265_cast_fp16)[name = string("op_7548_cast_fp16")];
            fp32 var_7549_epsilon_0 = const()[name = string("op_7549_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_7549_cast_fp16 = inverse(epsilon = var_7549_epsilon_0, x = var_7548_cast_fp16)[name = string("op_7549_cast_fp16")];
            fp16 var_7550_to_fp16 = const()[name = string("op_7550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_7551_cast_fp16 = mul(x = var_7549_cast_fp16, y = var_7550_to_fp16)[name = string("op_7551_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_7553_cast_fp16 = add(x = var_7547_cast_fp16, y = var_7551_cast_fp16)[name = string("op_7553_cast_fp16")];
            fp32 var_7554_epsilon_0 = const()[name = string("op_7554_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_7554_cast_fp16 = rsqrt(epsilon = var_7554_epsilon_0, x = var_7553_cast_fp16)[name = string("op_7554_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_65_cast_fp16 = mul(x = y_265_cast_fp16, y = var_7554_cast_fp16)[name = string("v_65_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7556_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7556_cast_fp16")];
            tensor<int32, [4]> var_7561_begin_0 = const()[name = string("op_7561_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7561_end_0 = const()[name = string("op_7561_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_7561_end_mask_0 = const()[name = string("op_7561_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_7561_cast_fp16 = slice_by_index(begin = var_7561_begin_0, end = var_7561_end_0, end_mask = var_7561_end_mask_0, x = q_65_cast_fp16)[name = string("op_7561_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_7562_cast_fp16 = mul(x = var_7561_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_7562_cast_fp16")];
            tensor<int32, [4]> var_7567_begin_0 = const()[name = string("op_7567_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7567_end_0 = const()[name = string("op_7567_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_7567_end_mask_0 = const()[name = string("op_7567_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_7567_cast_fp16 = slice_by_index(begin = var_7567_begin_0, end = var_7567_end_0, end_mask = var_7567_end_mask_0, x = q_65_cast_fp16)[name = string("op_7567_cast_fp16")];
            int32 var_7569 = const()[name = string("op_7569"), val = int32(-1)];
            bool var_7570_interleave_0 = const()[name = string("op_7570_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_7570_cast_fp16 = concat(axis = var_7569, interleave = var_7570_interleave_0, values = (var_7562_cast_fp16, var_7567_cast_fp16))[name = string("op_7570_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7571_cast_fp16 = mul(x = var_7570_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7571_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_67_cast_fp16 = add(x = var_7556_cast_fp16, y = var_7571_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7574_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7574_cast_fp16")];
            tensor<int32, [4]> var_7579_begin_0 = const()[name = string("op_7579_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_7579_end_0 = const()[name = string("op_7579_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_7579_end_mask_0 = const()[name = string("op_7579_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_7579_cast_fp16 = slice_by_index(begin = var_7579_begin_0, end = var_7579_end_0, end_mask = var_7579_end_mask_0, x = k_97_cast_fp16)[name = string("op_7579_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_7580_cast_fp16 = mul(x = var_7579_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_7580_cast_fp16")];
            tensor<int32, [4]> var_7585_begin_0 = const()[name = string("op_7585_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7585_end_0 = const()[name = string("op_7585_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_7585_end_mask_0 = const()[name = string("op_7585_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_7585_cast_fp16 = slice_by_index(begin = var_7585_begin_0, end = var_7585_end_0, end_mask = var_7585_end_mask_0, x = k_97_cast_fp16)[name = string("op_7585_cast_fp16")];
            int32 var_7587 = const()[name = string("op_7587"), val = int32(-1)];
            bool var_7588_interleave_0 = const()[name = string("op_7588_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_7588_cast_fp16 = concat(axis = var_7587, interleave = var_7588_interleave_0, values = (var_7580_cast_fp16, var_7585_cast_fp16))[name = string("op_7588_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_7589_cast_fp16 = mul(x = var_7588_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7589_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_99_cast_fp16 = add(x = var_7574_cast_fp16, y = var_7589_cast_fp16)[name = string("k_99_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_32_reps_0 = const()[name = string("tile_32_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = k_99_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 1, 256, 256]> tile_32_cast_fp16 = tile(reps = tile_32_reps_0, x = transpose_64_cast_fp16)[name = string("tile_32_cast_fp16")];
            tensor<int32, [5]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_64_cast_fp16 = reshape(shape = concat_64, x = tile_32_cast_fp16)[name = string("reshape_64_cast_fp16")];
            tensor<int32, [5]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = reshape_64_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 1, 256, 256]> reshape_65_cast_fp16 = reshape(shape = concat_65, x = transpose_65_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_33_reps_0 = const()[name = string("tile_33_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = v_65_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 1, 256, 256]> tile_33_cast_fp16 = tile(reps = tile_33_reps_0, x = transpose_66_cast_fp16)[name = string("tile_33_cast_fp16")];
            tensor<int32, [5]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_66_cast_fp16 = reshape(shape = concat_66, x = tile_33_cast_fp16)[name = string("reshape_66_cast_fp16")];
            tensor<int32, [5]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_67 = const()[name = string("concat_67"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = reshape_66_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 1, 256, 256]> reshape_67_cast_fp16 = reshape(shape = concat_67, x = transpose_67_cast_fp16)[name = string("reshape_67_cast_fp16")];
            tensor<int32, [4]> v_67_perm_0 = const()[name = string("v_67_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7603_transpose_x_0 = const()[name = string("op_7603_transpose_x_0"), val = bool(false)];
            bool var_7603_transpose_y_0 = const()[name = string("op_7603_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_112_cast_fp16 = transpose(perm = transpose_112_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 4, 256, 256]> var_7603_cast_fp16 = matmul(transpose_x = var_7603_transpose_x_0, transpose_y = var_7603_transpose_y_0, x = q_67_cast_fp16, y = transpose_112_cast_fp16)[name = string("op_7603_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7605_cast_fp16 = add(x = var_7603_cast_fp16, y = attention_bias)[name = string("op_7605_cast_fp16")];
            int32 var_7606 = const()[name = string("op_7606"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_7608_cast_fp16 = softmax(axis = var_7606, x = var_7605_cast_fp16)[name = string("op_7608_cast_fp16")];
            bool var_7609_transpose_x_0 = const()[name = string("op_7609_transpose_x_0"), val = bool(false)];
            bool var_7609_transpose_y_0 = const()[name = string("op_7609_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_67_cast_fp16 = transpose(perm = v_67_perm_0, x = reshape_67_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 4, 256, 256]> var_7609_cast_fp16 = matmul(transpose_x = var_7609_transpose_x_0, transpose_y = var_7609_transpose_y_0, x = var_7608_cast_fp16, y = v_67_cast_fp16)[name = string("op_7609_cast_fp16")];
            tensor<int32, [4]> var_7612_perm_0 = const()[name = string("op_7612_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7616 = const()[name = string("op_7616"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_7612_cast_fp16 = transpose(perm = var_7612_perm_0, x = var_7609_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 256, 1024]> input_195_cast_fp16 = reshape(shape = var_7616, x = var_7612_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93776512))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94300864))))[name = string("tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7621_cast_fp16 = abs(x = linear_148_cast_fp16)[name = string("op_7621_cast_fp16")];
            bool var_7624 = const()[name = string("op_7624"), val = bool(true)];
            tensor<int32, [1]> var_7625_axes_0 = const()[name = string("op_7625_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7625_cast_fp16 = reduce_max(axes = var_7625_axes_0, keep_dims = var_7624, x = var_7621_cast_fp16)[name = string("op_7625_cast_fp16")];
            fp16 var_7626_to_fp16 = const()[name = string("op_7626_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_267_cast_fp16 = maximum(x = var_7625_cast_fp16, y = var_7626_to_fp16)[name = string("s_267_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_267_cast_fp16 = real_div(x = linear_148_cast_fp16, y = s_267_cast_fp16)[name = string("y_267_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7629_cast_fp16 = mul(x = y_267_cast_fp16, y = y_267_cast_fp16)[name = string("op_7629_cast_fp16")];
            tensor<int32, [1]> var_7634_axes_0 = const()[name = string("op_7634_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7634_keep_dims_0 = const()[name = string("op_7634_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7634_cast_fp16 = reduce_mean(axes = var_7634_axes_0, keep_dims = var_7634_keep_dims_0, x = var_7629_cast_fp16)[name = string("op_7634_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7635_cast_fp16 = mul(x = s_267_cast_fp16, y = s_267_cast_fp16)[name = string("op_7635_cast_fp16")];
            fp32 var_7636_epsilon_0 = const()[name = string("op_7636_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7636_cast_fp16 = inverse(epsilon = var_7636_epsilon_0, x = var_7635_cast_fp16)[name = string("op_7636_cast_fp16")];
            fp16 var_7637_to_fp16 = const()[name = string("op_7637_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7638_cast_fp16 = mul(x = var_7636_cast_fp16, y = var_7637_to_fp16)[name = string("op_7638_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7640_cast_fp16 = add(x = var_7634_cast_fp16, y = var_7638_cast_fp16)[name = string("op_7640_cast_fp16")];
            fp32 var_7641_epsilon_0 = const()[name = string("op_7641_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7641_cast_fp16 = rsqrt(epsilon = var_7641_epsilon_0, x = var_7640_cast_fp16)[name = string("op_7641_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_233_cast_fp16 = mul(x = y_267_cast_fp16, y = var_7641_cast_fp16)[name = string("out_233_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94301952)))];
            tensor<fp16, [1, 256, 512]> var_7643_cast_fp16 = mul(x = out_233_cast_fp16, y = tm_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("op_7643_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_267_cast_fp16 = add(x = x_257_cast_fp16, y = var_7643_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7646_cast_fp16 = abs(x = x_267_cast_fp16)[name = string("op_7646_cast_fp16")];
            bool var_7649 = const()[name = string("op_7649"), val = bool(true)];
            tensor<int32, [1]> var_7650_axes_0 = const()[name = string("op_7650_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7650_cast_fp16 = reduce_max(axes = var_7650_axes_0, keep_dims = var_7649, x = var_7646_cast_fp16)[name = string("op_7650_cast_fp16")];
            fp16 var_7651_to_fp16 = const()[name = string("op_7651_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_269_cast_fp16 = maximum(x = var_7650_cast_fp16, y = var_7651_to_fp16)[name = string("s_269_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_269_cast_fp16 = real_div(x = x_267_cast_fp16, y = s_269_cast_fp16)[name = string("y_269_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7654_cast_fp16 = mul(x = y_269_cast_fp16, y = y_269_cast_fp16)[name = string("op_7654_cast_fp16")];
            tensor<int32, [1]> var_7659_axes_0 = const()[name = string("op_7659_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7659_keep_dims_0 = const()[name = string("op_7659_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7659_cast_fp16 = reduce_mean(axes = var_7659_axes_0, keep_dims = var_7659_keep_dims_0, x = var_7654_cast_fp16)[name = string("op_7659_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7660_cast_fp16 = mul(x = s_269_cast_fp16, y = s_269_cast_fp16)[name = string("op_7660_cast_fp16")];
            fp32 var_7661_epsilon_0 = const()[name = string("op_7661_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7661_cast_fp16 = inverse(epsilon = var_7661_epsilon_0, x = var_7660_cast_fp16)[name = string("op_7661_cast_fp16")];
            fp16 var_7662_to_fp16 = const()[name = string("op_7662_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7663_cast_fp16 = mul(x = var_7661_cast_fp16, y = var_7662_to_fp16)[name = string("op_7663_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7665_cast_fp16 = add(x = var_7659_cast_fp16, y = var_7663_cast_fp16)[name = string("op_7665_cast_fp16")];
            fp32 var_7666_epsilon_0 = const()[name = string("op_7666_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7666_cast_fp16 = rsqrt(epsilon = var_7666_epsilon_0, x = var_7665_cast_fp16)[name = string("op_7666_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_235_cast_fp16 = mul(x = y_269_cast_fp16, y = var_7666_cast_fp16)[name = string("out_235_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94303040)))];
            tensor<fp16, [1, 256, 512]> input_197_cast_fp16 = mul(x = out_235_cast_fp16, y = tm_layers_16_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94304128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95352768))))[name = string("tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_149_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_149_cast_fp16")];
            string var_7673_mode_0 = const()[name = string("op_7673_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_7673_cast_fp16 = gelu(mode = var_7673_mode_0, x = linear_149_cast_fp16)[name = string("op_7673_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95356928))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96405568))))[name = string("tm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_150_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_199_cast_fp16 = mul(x = var_7673_cast_fp16, y = linear_150_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96409728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97458368))))[name = string("tm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7681_cast_fp16 = abs(x = linear_151_cast_fp16)[name = string("op_7681_cast_fp16")];
            bool var_7684 = const()[name = string("op_7684"), val = bool(true)];
            tensor<int32, [1]> var_7685_axes_0 = const()[name = string("op_7685_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7685_cast_fp16 = reduce_max(axes = var_7685_axes_0, keep_dims = var_7684, x = var_7681_cast_fp16)[name = string("op_7685_cast_fp16")];
            fp16 var_7686_to_fp16 = const()[name = string("op_7686_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_271_cast_fp16 = maximum(x = var_7685_cast_fp16, y = var_7686_to_fp16)[name = string("s_271_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_271_cast_fp16 = real_div(x = linear_151_cast_fp16, y = s_271_cast_fp16)[name = string("y_271_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7689_cast_fp16 = mul(x = y_271_cast_fp16, y = y_271_cast_fp16)[name = string("op_7689_cast_fp16")];
            tensor<int32, [1]> var_7694_axes_0 = const()[name = string("op_7694_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7694_keep_dims_0 = const()[name = string("op_7694_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7694_cast_fp16 = reduce_mean(axes = var_7694_axes_0, keep_dims = var_7694_keep_dims_0, x = var_7689_cast_fp16)[name = string("op_7694_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7695_cast_fp16 = mul(x = s_271_cast_fp16, y = s_271_cast_fp16)[name = string("op_7695_cast_fp16")];
            fp32 var_7696_epsilon_0 = const()[name = string("op_7696_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7696_cast_fp16 = inverse(epsilon = var_7696_epsilon_0, x = var_7695_cast_fp16)[name = string("op_7696_cast_fp16")];
            fp16 var_7697_to_fp16 = const()[name = string("op_7697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7698_cast_fp16 = mul(x = var_7696_cast_fp16, y = var_7697_to_fp16)[name = string("op_7698_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7700_cast_fp16 = add(x = var_7694_cast_fp16, y = var_7698_cast_fp16)[name = string("op_7700_cast_fp16")];
            fp32 var_7701_epsilon_0 = const()[name = string("op_7701_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7701_cast_fp16 = rsqrt(epsilon = var_7701_epsilon_0, x = var_7700_cast_fp16)[name = string("op_7701_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_237_cast_fp16 = mul(x = y_271_cast_fp16, y = var_7701_cast_fp16)[name = string("out_237_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_16_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97459456)))];
            tensor<fp16, [1, 256, 512]> var_7703_cast_fp16 = mul(x = out_237_cast_fp16, y = tm_layers_16_post_feedforward_layernorm_weight_to_fp16)[name = string("op_7703_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_201_cast_fp16 = add(x = x_267_cast_fp16, y = var_7703_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97460544))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97722752))))[name = string("tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_152_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_152_cast_fp16")];
            string var_7710_mode_0 = const()[name = string("op_7710_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_7710_cast_fp16 = gelu(mode = var_7710_mode_0, x = linear_152_cast_fp16)[name = string("op_7710_cast_fp16")];
            tensor<int32, [4]> var_7723_begin_0 = const()[name = string("op_7723_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7723_end_0 = const()[name = string("op_7723_end_0"), val = tensor<int32, [4]>([1, 256, 17, 512])];
            tensor<bool, [4]> var_7723_end_mask_0 = const()[name = string("op_7723_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_7723_squeeze_mask_0 = const()[name = string("op_7723_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_7723_cast_fp16 = slice_by_index(begin = var_7723_begin_0, end = var_7723_end_0, end_mask = var_7723_end_mask_0, squeeze_mask = var_7723_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_7723_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_203_cast_fp16 = mul(x = var_7710_cast_fp16, y = var_7723_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97723840))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97986048))))[name = string("tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_16_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7733_cast_fp16 = abs(x = linear_153_cast_fp16)[name = string("op_7733_cast_fp16")];
            bool var_7736 = const()[name = string("op_7736"), val = bool(true)];
            tensor<int32, [1]> var_7737_axes_0 = const()[name = string("op_7737_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7737_cast_fp16 = reduce_max(axes = var_7737_axes_0, keep_dims = var_7736, x = var_7733_cast_fp16)[name = string("op_7737_cast_fp16")];
            fp16 var_7738_to_fp16 = const()[name = string("op_7738_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_273_cast_fp16 = maximum(x = var_7737_cast_fp16, y = var_7738_to_fp16)[name = string("s_273_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_273_cast_fp16 = real_div(x = linear_153_cast_fp16, y = s_273_cast_fp16)[name = string("y_273_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7741_cast_fp16 = mul(x = y_273_cast_fp16, y = y_273_cast_fp16)[name = string("op_7741_cast_fp16")];
            tensor<int32, [1]> var_7746_axes_0 = const()[name = string("op_7746_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7746_keep_dims_0 = const()[name = string("op_7746_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7746_cast_fp16 = reduce_mean(axes = var_7746_axes_0, keep_dims = var_7746_keep_dims_0, x = var_7741_cast_fp16)[name = string("op_7746_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7747_cast_fp16 = mul(x = s_273_cast_fp16, y = s_273_cast_fp16)[name = string("op_7747_cast_fp16")];
            fp32 var_7748_epsilon_0 = const()[name = string("op_7748_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7748_cast_fp16 = inverse(epsilon = var_7748_epsilon_0, x = var_7747_cast_fp16)[name = string("op_7748_cast_fp16")];
            fp16 var_7749_to_fp16 = const()[name = string("op_7749_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7750_cast_fp16 = mul(x = var_7748_cast_fp16, y = var_7749_to_fp16)[name = string("op_7750_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7752_cast_fp16 = add(x = var_7746_cast_fp16, y = var_7750_cast_fp16)[name = string("op_7752_cast_fp16")];
            fp32 var_7753_epsilon_0 = const()[name = string("op_7753_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7753_cast_fp16 = rsqrt(epsilon = var_7753_epsilon_0, x = var_7752_cast_fp16)[name = string("op_7753_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_239_cast_fp16 = mul(x = y_273_cast_fp16, y = var_7753_cast_fp16)[name = string("out_239_cast_fp16")];
            tensor<fp16, [512]> tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97987136)))];
            tensor<fp16, [1, 256, 512]> var_7755_cast_fp16 = mul(x = out_239_cast_fp16, y = tm_layers_16_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_7755_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_33_cast_fp16 = add(x = input_201_cast_fp16, y = var_7755_cast_fp16)[name = string("h_33_cast_fp16")];
            tensor<fp16, [1]> tm_layers_16_layer_scalar_to_fp16 = const()[name = string("tm_layers_16_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.7p-2])];
            tensor<fp16, [1, 256, 512]> x_273_cast_fp16 = mul(x = h_33_cast_fp16, y = tm_layers_16_layer_scalar_to_fp16)[name = string("x_273_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7759_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_7759_cast_fp16")];
            bool var_7762 = const()[name = string("op_7762"), val = bool(true)];
            tensor<int32, [1]> var_7763_axes_0 = const()[name = string("op_7763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7763_cast_fp16 = reduce_max(axes = var_7763_axes_0, keep_dims = var_7762, x = var_7759_cast_fp16)[name = string("op_7763_cast_fp16")];
            fp16 var_7764_to_fp16 = const()[name = string("op_7764_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_275_cast_fp16 = maximum(x = var_7763_cast_fp16, y = var_7764_to_fp16)[name = string("s_275_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_275_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_275_cast_fp16)[name = string("y_275_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7767_cast_fp16 = mul(x = y_275_cast_fp16, y = y_275_cast_fp16)[name = string("op_7767_cast_fp16")];
            tensor<int32, [1]> var_7772_axes_0 = const()[name = string("op_7772_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7772_keep_dims_0 = const()[name = string("op_7772_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7772_cast_fp16 = reduce_mean(axes = var_7772_axes_0, keep_dims = var_7772_keep_dims_0, x = var_7767_cast_fp16)[name = string("op_7772_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7773_cast_fp16 = mul(x = s_275_cast_fp16, y = s_275_cast_fp16)[name = string("op_7773_cast_fp16")];
            fp32 var_7774_epsilon_0 = const()[name = string("op_7774_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7774_cast_fp16 = inverse(epsilon = var_7774_epsilon_0, x = var_7773_cast_fp16)[name = string("op_7774_cast_fp16")];
            fp16 var_7775_to_fp16 = const()[name = string("op_7775_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7776_cast_fp16 = mul(x = var_7774_cast_fp16, y = var_7775_to_fp16)[name = string("op_7776_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7778_cast_fp16 = add(x = var_7772_cast_fp16, y = var_7776_cast_fp16)[name = string("op_7778_cast_fp16")];
            fp32 var_7779_epsilon_0 = const()[name = string("op_7779_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7779_cast_fp16 = rsqrt(epsilon = var_7779_epsilon_0, x = var_7778_cast_fp16)[name = string("op_7779_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_241_cast_fp16 = mul(x = y_275_cast_fp16, y = var_7779_cast_fp16)[name = string("out_241_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97988224)))];
            tensor<fp16, [1, 256, 512]> input_205_cast_fp16 = mul(x = out_241_cast_fp16, y = tm_layers_17_input_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97989312))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99037952))))[name = string("tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_154_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_7789 = const()[name = string("op_7789"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 256, 4, 512]> var_7790_cast_fp16 = reshape(shape = var_7789, x = linear_154_cast_fp16)[name = string("op_7790_cast_fp16")];
            tensor<int32, [4]> x_275_perm_0 = const()[name = string("x_275_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99042112))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99304320))))[name = string("tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_7801 = const()[name = string("op_7801"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_7802_cast_fp16 = reshape(shape = var_7801, x = linear_155_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<int32, [4]> x_277_perm_0 = const()[name = string("x_277_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99305408))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99567616))))[name = string("tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_7813 = const()[name = string("op_7813"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_7814_cast_fp16 = reshape(shape = var_7813, x = linear_156_cast_fp16)[name = string("op_7814_cast_fp16")];
            tensor<int32, [4]> x_279_perm_0 = const()[name = string("x_279_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 512]> x_275_cast_fp16 = transpose(perm = x_275_perm_0, x = var_7790_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 4, 256, 512]> var_7818_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_7818_cast_fp16")];
            bool var_7821 = const()[name = string("op_7821"), val = bool(true)];
            tensor<int32, [1]> var_7822_axes_0 = const()[name = string("op_7822_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_7822_cast_fp16 = reduce_max(axes = var_7822_axes_0, keep_dims = var_7821, x = var_7818_cast_fp16)[name = string("op_7822_cast_fp16")];
            fp16 var_7823_to_fp16 = const()[name = string("op_7823_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_277_cast_fp16 = maximum(x = var_7822_cast_fp16, y = var_7823_to_fp16)[name = string("s_277_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> y_277_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_277_cast_fp16)[name = string("y_277_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_7826_cast_fp16 = mul(x = y_277_cast_fp16, y = y_277_cast_fp16)[name = string("op_7826_cast_fp16")];
            tensor<int32, [1]> var_7831_axes_0 = const()[name = string("op_7831_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7831_keep_dims_0 = const()[name = string("op_7831_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_7831_cast_fp16 = reduce_mean(axes = var_7831_axes_0, keep_dims = var_7831_keep_dims_0, x = var_7826_cast_fp16)[name = string("op_7831_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7832_cast_fp16 = mul(x = s_277_cast_fp16, y = s_277_cast_fp16)[name = string("op_7832_cast_fp16")];
            fp32 var_7833_epsilon_0 = const()[name = string("op_7833_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_7833_cast_fp16 = inverse(epsilon = var_7833_epsilon_0, x = var_7832_cast_fp16)[name = string("op_7833_cast_fp16")];
            fp16 var_7834_to_fp16 = const()[name = string("op_7834_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_7835_cast_fp16 = mul(x = var_7833_cast_fp16, y = var_7834_to_fp16)[name = string("op_7835_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_7837_cast_fp16 = add(x = var_7831_cast_fp16, y = var_7835_cast_fp16)[name = string("op_7837_cast_fp16")];
            fp32 var_7838_epsilon_0 = const()[name = string("op_7838_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_7838_cast_fp16 = rsqrt(epsilon = var_7838_epsilon_0, x = var_7837_cast_fp16)[name = string("op_7838_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> out_243_cast_fp16 = mul(x = y_277_cast_fp16, y = var_7838_cast_fp16)[name = string("out_243_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99568704)))];
            tensor<fp16, [1, 4, 256, 512]> q_69_cast_fp16 = mul(x = out_243_cast_fp16, y = tm_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("q_69_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_277_cast_fp16 = transpose(perm = x_277_perm_0, x = var_7802_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 1, 256, 512]> var_7841_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_7841_cast_fp16")];
            bool var_7844 = const()[name = string("op_7844"), val = bool(true)];
            tensor<int32, [1]> var_7845_axes_0 = const()[name = string("op_7845_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_7845_cast_fp16 = reduce_max(axes = var_7845_axes_0, keep_dims = var_7844, x = var_7841_cast_fp16)[name = string("op_7845_cast_fp16")];
            fp16 var_7846_to_fp16 = const()[name = string("op_7846_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_279_cast_fp16 = maximum(x = var_7845_cast_fp16, y = var_7846_to_fp16)[name = string("s_279_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_279_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_279_cast_fp16)[name = string("y_279_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_7849_cast_fp16 = mul(x = y_279_cast_fp16, y = y_279_cast_fp16)[name = string("op_7849_cast_fp16")];
            tensor<int32, [1]> var_7854_axes_0 = const()[name = string("op_7854_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7854_keep_dims_0 = const()[name = string("op_7854_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_7854_cast_fp16 = reduce_mean(axes = var_7854_axes_0, keep_dims = var_7854_keep_dims_0, x = var_7849_cast_fp16)[name = string("op_7854_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_7855_cast_fp16 = mul(x = s_279_cast_fp16, y = s_279_cast_fp16)[name = string("op_7855_cast_fp16")];
            fp32 var_7856_epsilon_0 = const()[name = string("op_7856_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_7856_cast_fp16 = inverse(epsilon = var_7856_epsilon_0, x = var_7855_cast_fp16)[name = string("op_7856_cast_fp16")];
            fp16 var_7857_to_fp16 = const()[name = string("op_7857_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_7858_cast_fp16 = mul(x = var_7856_cast_fp16, y = var_7857_to_fp16)[name = string("op_7858_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_7860_cast_fp16 = add(x = var_7854_cast_fp16, y = var_7858_cast_fp16)[name = string("op_7860_cast_fp16")];
            fp32 var_7861_epsilon_0 = const()[name = string("op_7861_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_7861_cast_fp16 = rsqrt(epsilon = var_7861_epsilon_0, x = var_7860_cast_fp16)[name = string("op_7861_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> out_245_cast_fp16 = mul(x = y_279_cast_fp16, y = var_7861_cast_fp16)[name = string("out_245_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99569792)))];
            tensor<fp16, [1, 1, 256, 512]> k_103_cast_fp16 = mul(x = out_245_cast_fp16, y = tm_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("k_103_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_279_cast_fp16 = transpose(perm = x_279_perm_0, x = var_7814_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1, 256, 512]> var_7864_cast_fp16 = abs(x = x_279_cast_fp16)[name = string("op_7864_cast_fp16")];
            bool var_7867 = const()[name = string("op_7867"), val = bool(true)];
            tensor<int32, [1]> var_7868_axes_0 = const()[name = string("op_7868_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_7868_cast_fp16 = reduce_max(axes = var_7868_axes_0, keep_dims = var_7867, x = var_7864_cast_fp16)[name = string("op_7868_cast_fp16")];
            fp16 var_7869_to_fp16 = const()[name = string("op_7869_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_281_cast_fp16 = maximum(x = var_7868_cast_fp16, y = var_7869_to_fp16)[name = string("s_281_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_281_cast_fp16 = real_div(x = x_279_cast_fp16, y = s_281_cast_fp16)[name = string("y_281_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_7872_cast_fp16 = mul(x = y_281_cast_fp16, y = y_281_cast_fp16)[name = string("op_7872_cast_fp16")];
            tensor<int32, [1]> var_7877_axes_0 = const()[name = string("op_7877_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7877_keep_dims_0 = const()[name = string("op_7877_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_7877_cast_fp16 = reduce_mean(axes = var_7877_axes_0, keep_dims = var_7877_keep_dims_0, x = var_7872_cast_fp16)[name = string("op_7877_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_7878_cast_fp16 = mul(x = s_281_cast_fp16, y = s_281_cast_fp16)[name = string("op_7878_cast_fp16")];
            fp32 var_7879_epsilon_0 = const()[name = string("op_7879_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_7879_cast_fp16 = inverse(epsilon = var_7879_epsilon_0, x = var_7878_cast_fp16)[name = string("op_7879_cast_fp16")];
            fp16 var_7880_to_fp16 = const()[name = string("op_7880_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_7881_cast_fp16 = mul(x = var_7879_cast_fp16, y = var_7880_to_fp16)[name = string("op_7881_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_7883_cast_fp16 = add(x = var_7877_cast_fp16, y = var_7881_cast_fp16)[name = string("op_7883_cast_fp16")];
            fp32 var_7884_epsilon_0 = const()[name = string("op_7884_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_7884_cast_fp16 = rsqrt(epsilon = var_7884_epsilon_0, x = var_7883_cast_fp16)[name = string("op_7884_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> v_69_cast_fp16 = mul(x = y_281_cast_fp16, y = var_7884_cast_fp16)[name = string("v_69_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_7886_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_cast_fp16)[name = string("op_7886_cast_fp16")];
            tensor<int32, [4]> var_7891_begin_0 = const()[name = string("op_7891_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7891_end_0 = const()[name = string("op_7891_end_0"), val = tensor<int32, [4]>([1, 4, 256, 512])];
            tensor<bool, [4]> var_7891_end_mask_0 = const()[name = string("op_7891_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 256]> var_7891_cast_fp16 = slice_by_index(begin = var_7891_begin_0, end = var_7891_end_0, end_mask = var_7891_end_mask_0, x = q_69_cast_fp16)[name = string("op_7891_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 256]> var_7892_cast_fp16 = mul(x = var_7891_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_7892_cast_fp16")];
            tensor<int32, [4]> var_7897_begin_0 = const()[name = string("op_7897_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7897_end_0 = const()[name = string("op_7897_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_7897_end_mask_0 = const()[name = string("op_7897_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 256]> var_7897_cast_fp16 = slice_by_index(begin = var_7897_begin_0, end = var_7897_end_0, end_mask = var_7897_end_mask_0, x = q_69_cast_fp16)[name = string("op_7897_cast_fp16")];
            int32 var_7899 = const()[name = string("op_7899"), val = int32(-1)];
            bool var_7900_interleave_0 = const()[name = string("op_7900_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> var_7900_cast_fp16 = concat(axis = var_7899, interleave = var_7900_interleave_0, values = (var_7892_cast_fp16, var_7897_cast_fp16))[name = string("op_7900_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_7901_cast_fp16 = mul(x = var_7900_cast_fp16, y = sin_cast_fp16)[name = string("op_7901_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> q_71_cast_fp16 = add(x = var_7886_cast_fp16, y = var_7901_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_7904_cast_fp16 = mul(x = k_103_cast_fp16, y = cos_cast_fp16)[name = string("op_7904_cast_fp16")];
            tensor<int32, [4]> var_7909_begin_0 = const()[name = string("op_7909_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_7909_end_0 = const()[name = string("op_7909_end_0"), val = tensor<int32, [4]>([1, 1, 256, 512])];
            tensor<bool, [4]> var_7909_end_mask_0 = const()[name = string("op_7909_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 256, 256]> var_7909_cast_fp16 = slice_by_index(begin = var_7909_begin_0, end = var_7909_end_0, end_mask = var_7909_end_mask_0, x = k_103_cast_fp16)[name = string("op_7909_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 256, 256]> var_7910_cast_fp16 = mul(x = var_7909_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_7910_cast_fp16")];
            tensor<int32, [4]> var_7915_begin_0 = const()[name = string("op_7915_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7915_end_0 = const()[name = string("op_7915_end_0"), val = tensor<int32, [4]>([1, 1, 256, 256])];
            tensor<bool, [4]> var_7915_end_mask_0 = const()[name = string("op_7915_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 256, 256]> var_7915_cast_fp16 = slice_by_index(begin = var_7915_begin_0, end = var_7915_end_0, end_mask = var_7915_end_mask_0, x = k_103_cast_fp16)[name = string("op_7915_cast_fp16")];
            int32 var_7917 = const()[name = string("op_7917"), val = int32(-1)];
            bool var_7918_interleave_0 = const()[name = string("op_7918_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 256, 512]> var_7918_cast_fp16 = concat(axis = var_7917, interleave = var_7918_interleave_0, values = (var_7910_cast_fp16, var_7915_cast_fp16))[name = string("op_7918_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_7919_cast_fp16 = mul(x = var_7918_cast_fp16, y = sin_cast_fp16)[name = string("op_7919_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> k_105_cast_fp16 = add(x = var_7904_cast_fp16, y = var_7919_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_34_reps_0 = const()[name = string("tile_34_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = k_105_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 1, 256, 512]> tile_34_cast_fp16 = tile(reps = tile_34_reps_0, x = transpose_68_cast_fp16)[name = string("tile_34_cast_fp16")];
            tensor<int32, [5]> concat_68 = const()[name = string("concat_68"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_68_cast_fp16 = reshape(shape = concat_68, x = tile_34_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [5]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 1, 256, 512]> reshape_69_cast_fp16 = reshape(shape = concat_69, x = transpose_69_cast_fp16)[name = string("reshape_69_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_35_reps_0 = const()[name = string("tile_35_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = v_69_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 1, 256, 512]> tile_35_cast_fp16 = tile(reps = tile_35_reps_0, x = transpose_70_cast_fp16)[name = string("tile_35_cast_fp16")];
            tensor<int32, [5]> concat_70 = const()[name = string("concat_70"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_70_cast_fp16 = reshape(shape = concat_70, x = tile_35_cast_fp16)[name = string("reshape_70_cast_fp16")];
            tensor<int32, [5]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_71 = const()[name = string("concat_71"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = reshape_70_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 1, 256, 512]> reshape_71_cast_fp16 = reshape(shape = concat_71, x = transpose_71_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_7933_transpose_x_0 = const()[name = string("op_7933_transpose_x_0"), val = bool(false)];
            bool var_7933_transpose_y_0 = const()[name = string("op_7933_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = reshape_69_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 4, 256, 256]> var_7933_cast_fp16 = matmul(transpose_x = var_7933_transpose_x_0, transpose_y = var_7933_transpose_y_0, x = q_71_cast_fp16, y = transpose_113_cast_fp16)[name = string("op_7933_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_7935_cast_fp16 = add(x = var_7933_cast_fp16, y = attention_bias)[name = string("op_7935_cast_fp16")];
            int32 var_7936 = const()[name = string("op_7936"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_7938_cast_fp16 = softmax(axis = var_7936, x = var_7935_cast_fp16)[name = string("op_7938_cast_fp16")];
            bool var_7939_transpose_x_0 = const()[name = string("op_7939_transpose_x_0"), val = bool(false)];
            bool var_7939_transpose_y_0 = const()[name = string("op_7939_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> v_71_cast_fp16 = transpose(perm = v_71_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 4, 256, 512]> var_7939_cast_fp16 = matmul(transpose_x = var_7939_transpose_x_0, transpose_y = var_7939_transpose_y_0, x = var_7938_cast_fp16, y = v_71_cast_fp16)[name = string("op_7939_cast_fp16")];
            tensor<int32, [4]> var_7942_perm_0 = const()[name = string("op_7942_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7946 = const()[name = string("op_7946"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 256, 4, 512]> var_7942_cast_fp16 = transpose(perm = var_7942_perm_0, x = var_7939_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 256, 2048]> input_207_cast_fp16 = reshape(shape = var_7946, x = var_7942_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99570880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619520))))[name = string("tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7951_cast_fp16 = abs(x = linear_157_cast_fp16)[name = string("op_7951_cast_fp16")];
            bool var_7954 = const()[name = string("op_7954"), val = bool(true)];
            tensor<int32, [1]> var_7955_axes_0 = const()[name = string("op_7955_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7955_cast_fp16 = reduce_max(axes = var_7955_axes_0, keep_dims = var_7954, x = var_7951_cast_fp16)[name = string("op_7955_cast_fp16")];
            fp16 var_7956_to_fp16 = const()[name = string("op_7956_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_283_cast_fp16 = maximum(x = var_7955_cast_fp16, y = var_7956_to_fp16)[name = string("s_283_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_283_cast_fp16 = real_div(x = linear_157_cast_fp16, y = s_283_cast_fp16)[name = string("y_283_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7959_cast_fp16 = mul(x = y_283_cast_fp16, y = y_283_cast_fp16)[name = string("op_7959_cast_fp16")];
            tensor<int32, [1]> var_7964_axes_0 = const()[name = string("op_7964_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7964_keep_dims_0 = const()[name = string("op_7964_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7964_cast_fp16 = reduce_mean(axes = var_7964_axes_0, keep_dims = var_7964_keep_dims_0, x = var_7959_cast_fp16)[name = string("op_7964_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7965_cast_fp16 = mul(x = s_283_cast_fp16, y = s_283_cast_fp16)[name = string("op_7965_cast_fp16")];
            fp32 var_7966_epsilon_0 = const()[name = string("op_7966_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7966_cast_fp16 = inverse(epsilon = var_7966_epsilon_0, x = var_7965_cast_fp16)[name = string("op_7966_cast_fp16")];
            fp16 var_7967_to_fp16 = const()[name = string("op_7967_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7968_cast_fp16 = mul(x = var_7966_cast_fp16, y = var_7967_to_fp16)[name = string("op_7968_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7970_cast_fp16 = add(x = var_7964_cast_fp16, y = var_7968_cast_fp16)[name = string("op_7970_cast_fp16")];
            fp32 var_7971_epsilon_0 = const()[name = string("op_7971_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7971_cast_fp16 = rsqrt(epsilon = var_7971_epsilon_0, x = var_7970_cast_fp16)[name = string("op_7971_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_247_cast_fp16 = mul(x = y_283_cast_fp16, y = var_7971_cast_fp16)[name = string("out_247_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100620608)))];
            tensor<fp16, [1, 256, 512]> var_7973_cast_fp16 = mul(x = out_247_cast_fp16, y = tm_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("op_7973_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_283_cast_fp16 = add(x = x_273_cast_fp16, y = var_7973_cast_fp16)[name = string("x_283_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7976_cast_fp16 = abs(x = x_283_cast_fp16)[name = string("op_7976_cast_fp16")];
            bool var_7979 = const()[name = string("op_7979"), val = bool(true)];
            tensor<int32, [1]> var_7980_axes_0 = const()[name = string("op_7980_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_7980_cast_fp16 = reduce_max(axes = var_7980_axes_0, keep_dims = var_7979, x = var_7976_cast_fp16)[name = string("op_7980_cast_fp16")];
            fp16 var_7981_to_fp16 = const()[name = string("op_7981_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_285_cast_fp16 = maximum(x = var_7980_cast_fp16, y = var_7981_to_fp16)[name = string("s_285_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_285_cast_fp16 = real_div(x = x_283_cast_fp16, y = s_285_cast_fp16)[name = string("y_285_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_7984_cast_fp16 = mul(x = y_285_cast_fp16, y = y_285_cast_fp16)[name = string("op_7984_cast_fp16")];
            tensor<int32, [1]> var_7989_axes_0 = const()[name = string("op_7989_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7989_keep_dims_0 = const()[name = string("op_7989_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7989_cast_fp16 = reduce_mean(axes = var_7989_axes_0, keep_dims = var_7989_keep_dims_0, x = var_7984_cast_fp16)[name = string("op_7989_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7990_cast_fp16 = mul(x = s_285_cast_fp16, y = s_285_cast_fp16)[name = string("op_7990_cast_fp16")];
            fp32 var_7991_epsilon_0 = const()[name = string("op_7991_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_7991_cast_fp16 = inverse(epsilon = var_7991_epsilon_0, x = var_7990_cast_fp16)[name = string("op_7991_cast_fp16")];
            fp16 var_7992_to_fp16 = const()[name = string("op_7992_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_7993_cast_fp16 = mul(x = var_7991_cast_fp16, y = var_7992_to_fp16)[name = string("op_7993_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_7995_cast_fp16 = add(x = var_7989_cast_fp16, y = var_7993_cast_fp16)[name = string("op_7995_cast_fp16")];
            fp32 var_7996_epsilon_0 = const()[name = string("op_7996_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_7996_cast_fp16 = rsqrt(epsilon = var_7996_epsilon_0, x = var_7995_cast_fp16)[name = string("op_7996_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_249_cast_fp16 = mul(x = y_285_cast_fp16, y = var_7996_cast_fp16)[name = string("out_249_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621696)))];
            tensor<fp16, [1, 256, 512]> input_209_cast_fp16 = mul(x = out_249_cast_fp16, y = tm_layers_17_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100622784))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101671424))))[name = string("tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_158_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_158_cast_fp16")];
            string var_8003_mode_0 = const()[name = string("op_8003_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_8003_cast_fp16 = gelu(mode = var_8003_mode_0, x = linear_158_cast_fp16)[name = string("op_8003_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101675584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102724224))))[name = string("tm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_159_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_211_cast_fp16 = mul(x = var_8003_cast_fp16, y = linear_159_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102728384))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103777024))))[name = string("tm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8011_cast_fp16 = abs(x = linear_160_cast_fp16)[name = string("op_8011_cast_fp16")];
            bool var_8014 = const()[name = string("op_8014"), val = bool(true)];
            tensor<int32, [1]> var_8015_axes_0 = const()[name = string("op_8015_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8015_cast_fp16 = reduce_max(axes = var_8015_axes_0, keep_dims = var_8014, x = var_8011_cast_fp16)[name = string("op_8015_cast_fp16")];
            fp16 var_8016_to_fp16 = const()[name = string("op_8016_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_287_cast_fp16 = maximum(x = var_8015_cast_fp16, y = var_8016_to_fp16)[name = string("s_287_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_287_cast_fp16 = real_div(x = linear_160_cast_fp16, y = s_287_cast_fp16)[name = string("y_287_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8019_cast_fp16 = mul(x = y_287_cast_fp16, y = y_287_cast_fp16)[name = string("op_8019_cast_fp16")];
            tensor<int32, [1]> var_8024_axes_0 = const()[name = string("op_8024_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8024_keep_dims_0 = const()[name = string("op_8024_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8024_cast_fp16 = reduce_mean(axes = var_8024_axes_0, keep_dims = var_8024_keep_dims_0, x = var_8019_cast_fp16)[name = string("op_8024_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8025_cast_fp16 = mul(x = s_287_cast_fp16, y = s_287_cast_fp16)[name = string("op_8025_cast_fp16")];
            fp32 var_8026_epsilon_0 = const()[name = string("op_8026_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8026_cast_fp16 = inverse(epsilon = var_8026_epsilon_0, x = var_8025_cast_fp16)[name = string("op_8026_cast_fp16")];
            fp16 var_8027_to_fp16 = const()[name = string("op_8027_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8028_cast_fp16 = mul(x = var_8026_cast_fp16, y = var_8027_to_fp16)[name = string("op_8028_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8030_cast_fp16 = add(x = var_8024_cast_fp16, y = var_8028_cast_fp16)[name = string("op_8030_cast_fp16")];
            fp32 var_8031_epsilon_0 = const()[name = string("op_8031_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8031_cast_fp16 = rsqrt(epsilon = var_8031_epsilon_0, x = var_8030_cast_fp16)[name = string("op_8031_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_251_cast_fp16 = mul(x = y_287_cast_fp16, y = var_8031_cast_fp16)[name = string("out_251_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_17_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103778112)))];
            tensor<fp16, [1, 256, 512]> var_8033_cast_fp16 = mul(x = out_251_cast_fp16, y = tm_layers_17_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8033_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_213_cast_fp16 = add(x = x_283_cast_fp16, y = var_8033_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103779200))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041408))))[name = string("tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_161_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_161_cast_fp16")];
            string var_8040_mode_0 = const()[name = string("op_8040_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_8040_cast_fp16 = gelu(mode = var_8040_mode_0, x = linear_161_cast_fp16)[name = string("op_8040_cast_fp16")];
            tensor<int32, [4]> var_8053_begin_0 = const()[name = string("op_8053_begin_0"), val = tensor<int32, [4]>([0, 0, 17, 0])];
            tensor<int32, [4]> var_8053_end_0 = const()[name = string("op_8053_end_0"), val = tensor<int32, [4]>([1, 256, 18, 512])];
            tensor<bool, [4]> var_8053_end_mask_0 = const()[name = string("op_8053_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8053_squeeze_mask_0 = const()[name = string("op_8053_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_8053_cast_fp16 = slice_by_index(begin = var_8053_begin_0, end = var_8053_end_0, end_mask = var_8053_end_mask_0, squeeze_mask = var_8053_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_8053_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_215_cast_fp16 = mul(x = var_8040_cast_fp16, y = var_8053_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104042496))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104304704))))[name = string("tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_17_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8063_cast_fp16 = abs(x = linear_162_cast_fp16)[name = string("op_8063_cast_fp16")];
            bool var_8066 = const()[name = string("op_8066"), val = bool(true)];
            tensor<int32, [1]> var_8067_axes_0 = const()[name = string("op_8067_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8067_cast_fp16 = reduce_max(axes = var_8067_axes_0, keep_dims = var_8066, x = var_8063_cast_fp16)[name = string("op_8067_cast_fp16")];
            fp16 var_8068_to_fp16 = const()[name = string("op_8068_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_289_cast_fp16 = maximum(x = var_8067_cast_fp16, y = var_8068_to_fp16)[name = string("s_289_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_289_cast_fp16 = real_div(x = linear_162_cast_fp16, y = s_289_cast_fp16)[name = string("y_289_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8071_cast_fp16 = mul(x = y_289_cast_fp16, y = y_289_cast_fp16)[name = string("op_8071_cast_fp16")];
            tensor<int32, [1]> var_8076_axes_0 = const()[name = string("op_8076_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8076_keep_dims_0 = const()[name = string("op_8076_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8076_cast_fp16 = reduce_mean(axes = var_8076_axes_0, keep_dims = var_8076_keep_dims_0, x = var_8071_cast_fp16)[name = string("op_8076_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8077_cast_fp16 = mul(x = s_289_cast_fp16, y = s_289_cast_fp16)[name = string("op_8077_cast_fp16")];
            fp32 var_8078_epsilon_0 = const()[name = string("op_8078_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8078_cast_fp16 = inverse(epsilon = var_8078_epsilon_0, x = var_8077_cast_fp16)[name = string("op_8078_cast_fp16")];
            fp16 var_8079_to_fp16 = const()[name = string("op_8079_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8080_cast_fp16 = mul(x = var_8078_cast_fp16, y = var_8079_to_fp16)[name = string("op_8080_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8082_cast_fp16 = add(x = var_8076_cast_fp16, y = var_8080_cast_fp16)[name = string("op_8082_cast_fp16")];
            fp32 var_8083_epsilon_0 = const()[name = string("op_8083_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8083_cast_fp16 = rsqrt(epsilon = var_8083_epsilon_0, x = var_8082_cast_fp16)[name = string("op_8083_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_253_cast_fp16 = mul(x = y_289_cast_fp16, y = var_8083_cast_fp16)[name = string("out_253_cast_fp16")];
            tensor<fp16, [512]> tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104305792)))];
            tensor<fp16, [1, 256, 512]> var_8085_cast_fp16 = mul(x = out_253_cast_fp16, y = tm_layers_17_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8085_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_35_cast_fp16 = add(x = input_213_cast_fp16, y = var_8085_cast_fp16)[name = string("h_35_cast_fp16")];
            tensor<fp16, [1]> tm_layers_17_layer_scalar_to_fp16 = const()[name = string("tm_layers_17_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e6p-1])];
            tensor<fp16, [1, 256, 512]> x_289_cast_fp16 = mul(x = h_35_cast_fp16, y = tm_layers_17_layer_scalar_to_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8089_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_8089_cast_fp16")];
            bool var_8092 = const()[name = string("op_8092"), val = bool(true)];
            tensor<int32, [1]> var_8093_axes_0 = const()[name = string("op_8093_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8093_cast_fp16 = reduce_max(axes = var_8093_axes_0, keep_dims = var_8092, x = var_8089_cast_fp16)[name = string("op_8093_cast_fp16")];
            fp16 var_8094_to_fp16 = const()[name = string("op_8094_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_291_cast_fp16 = maximum(x = var_8093_cast_fp16, y = var_8094_to_fp16)[name = string("s_291_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_291_cast_fp16 = real_div(x = x_289_cast_fp16, y = s_291_cast_fp16)[name = string("y_291_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8097_cast_fp16 = mul(x = y_291_cast_fp16, y = y_291_cast_fp16)[name = string("op_8097_cast_fp16")];
            tensor<int32, [1]> var_8102_axes_0 = const()[name = string("op_8102_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8102_keep_dims_0 = const()[name = string("op_8102_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8102_cast_fp16 = reduce_mean(axes = var_8102_axes_0, keep_dims = var_8102_keep_dims_0, x = var_8097_cast_fp16)[name = string("op_8102_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8103_cast_fp16 = mul(x = s_291_cast_fp16, y = s_291_cast_fp16)[name = string("op_8103_cast_fp16")];
            fp32 var_8104_epsilon_0 = const()[name = string("op_8104_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8104_cast_fp16 = inverse(epsilon = var_8104_epsilon_0, x = var_8103_cast_fp16)[name = string("op_8104_cast_fp16")];
            fp16 var_8105_to_fp16 = const()[name = string("op_8105_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8106_cast_fp16 = mul(x = var_8104_cast_fp16, y = var_8105_to_fp16)[name = string("op_8106_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8108_cast_fp16 = add(x = var_8102_cast_fp16, y = var_8106_cast_fp16)[name = string("op_8108_cast_fp16")];
            fp32 var_8109_epsilon_0 = const()[name = string("op_8109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8109_cast_fp16 = rsqrt(epsilon = var_8109_epsilon_0, x = var_8108_cast_fp16)[name = string("op_8109_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_255_cast_fp16 = mul(x = y_291_cast_fp16, y = var_8109_cast_fp16)[name = string("out_255_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104306880)))];
            tensor<fp16, [1, 256, 512]> input_217_cast_fp16 = mul(x = out_255_cast_fp16, y = tm_layers_18_input_layernorm_weight_to_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104307968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104832320))))[name = string("tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8119 = const()[name = string("op_8119"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_8120_cast_fp16 = reshape(shape = var_8119, x = linear_163_cast_fp16)[name = string("op_8120_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104834432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105096640))))[name = string("tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<int32, [4]> var_8131 = const()[name = string("op_8131"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8132_cast_fp16 = reshape(shape = var_8131, x = linear_164_cast_fp16)[name = string("op_8132_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105097728))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105359936))))[name = string("tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<int32, [4]> var_8143 = const()[name = string("op_8143"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8144_cast_fp16 = reshape(shape = var_8143, x = linear_165_cast_fp16)[name = string("op_8144_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_8120_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 4, 256, 256]> var_8148_cast_fp16 = abs(x = x_291_cast_fp16)[name = string("op_8148_cast_fp16")];
            bool var_8151 = const()[name = string("op_8151"), val = bool(true)];
            tensor<int32, [1]> var_8152_axes_0 = const()[name = string("op_8152_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_8152_cast_fp16 = reduce_max(axes = var_8152_axes_0, keep_dims = var_8151, x = var_8148_cast_fp16)[name = string("op_8152_cast_fp16")];
            fp16 var_8153_to_fp16 = const()[name = string("op_8153_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_293_cast_fp16 = maximum(x = var_8152_cast_fp16, y = var_8153_to_fp16)[name = string("s_293_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_293_cast_fp16 = real_div(x = x_291_cast_fp16, y = s_293_cast_fp16)[name = string("y_293_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8156_cast_fp16 = mul(x = y_293_cast_fp16, y = y_293_cast_fp16)[name = string("op_8156_cast_fp16")];
            tensor<int32, [1]> var_8161_axes_0 = const()[name = string("op_8161_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8161_keep_dims_0 = const()[name = string("op_8161_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_8161_cast_fp16 = reduce_mean(axes = var_8161_axes_0, keep_dims = var_8161_keep_dims_0, x = var_8156_cast_fp16)[name = string("op_8161_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8162_cast_fp16 = mul(x = s_293_cast_fp16, y = s_293_cast_fp16)[name = string("op_8162_cast_fp16")];
            fp32 var_8163_epsilon_0 = const()[name = string("op_8163_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_8163_cast_fp16 = inverse(epsilon = var_8163_epsilon_0, x = var_8162_cast_fp16)[name = string("op_8163_cast_fp16")];
            fp16 var_8164_to_fp16 = const()[name = string("op_8164_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_8165_cast_fp16 = mul(x = var_8163_cast_fp16, y = var_8164_to_fp16)[name = string("op_8165_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8167_cast_fp16 = add(x = var_8161_cast_fp16, y = var_8165_cast_fp16)[name = string("op_8167_cast_fp16")];
            fp32 var_8168_epsilon_0 = const()[name = string("op_8168_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_8168_cast_fp16 = rsqrt(epsilon = var_8168_epsilon_0, x = var_8167_cast_fp16)[name = string("op_8168_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_257_cast_fp16 = mul(x = y_293_cast_fp16, y = var_8168_cast_fp16)[name = string("out_257_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_73_cast_fp16 = mul(x = out_257_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_73_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_8132_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 256, 256]> var_8171_cast_fp16 = abs(x = x_293_cast_fp16)[name = string("op_8171_cast_fp16")];
            bool var_8174 = const()[name = string("op_8174"), val = bool(true)];
            tensor<int32, [1]> var_8175_axes_0 = const()[name = string("op_8175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8175_cast_fp16 = reduce_max(axes = var_8175_axes_0, keep_dims = var_8174, x = var_8171_cast_fp16)[name = string("op_8175_cast_fp16")];
            fp16 var_8176_to_fp16 = const()[name = string("op_8176_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_295_cast_fp16 = maximum(x = var_8175_cast_fp16, y = var_8176_to_fp16)[name = string("s_295_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_295_cast_fp16 = real_div(x = x_293_cast_fp16, y = s_295_cast_fp16)[name = string("y_295_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8179_cast_fp16 = mul(x = y_295_cast_fp16, y = y_295_cast_fp16)[name = string("op_8179_cast_fp16")];
            tensor<int32, [1]> var_8184_axes_0 = const()[name = string("op_8184_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8184_keep_dims_0 = const()[name = string("op_8184_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8184_cast_fp16 = reduce_mean(axes = var_8184_axes_0, keep_dims = var_8184_keep_dims_0, x = var_8179_cast_fp16)[name = string("op_8184_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8185_cast_fp16 = mul(x = s_295_cast_fp16, y = s_295_cast_fp16)[name = string("op_8185_cast_fp16")];
            fp32 var_8186_epsilon_0 = const()[name = string("op_8186_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8186_cast_fp16 = inverse(epsilon = var_8186_epsilon_0, x = var_8185_cast_fp16)[name = string("op_8186_cast_fp16")];
            fp16 var_8187_to_fp16 = const()[name = string("op_8187_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8188_cast_fp16 = mul(x = var_8186_cast_fp16, y = var_8187_to_fp16)[name = string("op_8188_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8190_cast_fp16 = add(x = var_8184_cast_fp16, y = var_8188_cast_fp16)[name = string("op_8190_cast_fp16")];
            fp32 var_8191_epsilon_0 = const()[name = string("op_8191_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8191_cast_fp16 = rsqrt(epsilon = var_8191_epsilon_0, x = var_8190_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_259_cast_fp16 = mul(x = y_295_cast_fp16, y = var_8191_cast_fp16)[name = string("out_259_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_109_cast_fp16 = mul(x = out_259_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_109_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_8144_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 256, 256]> var_8194_cast_fp16 = abs(x = x_295_cast_fp16)[name = string("op_8194_cast_fp16")];
            bool var_8197 = const()[name = string("op_8197"), val = bool(true)];
            tensor<int32, [1]> var_8198_axes_0 = const()[name = string("op_8198_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8198_cast_fp16 = reduce_max(axes = var_8198_axes_0, keep_dims = var_8197, x = var_8194_cast_fp16)[name = string("op_8198_cast_fp16")];
            fp16 var_8199_to_fp16 = const()[name = string("op_8199_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_297_cast_fp16 = maximum(x = var_8198_cast_fp16, y = var_8199_to_fp16)[name = string("s_297_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_297_cast_fp16 = real_div(x = x_295_cast_fp16, y = s_297_cast_fp16)[name = string("y_297_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8202_cast_fp16 = mul(x = y_297_cast_fp16, y = y_297_cast_fp16)[name = string("op_8202_cast_fp16")];
            tensor<int32, [1]> var_8207_axes_0 = const()[name = string("op_8207_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8207_keep_dims_0 = const()[name = string("op_8207_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8207_cast_fp16 = reduce_mean(axes = var_8207_axes_0, keep_dims = var_8207_keep_dims_0, x = var_8202_cast_fp16)[name = string("op_8207_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8208_cast_fp16 = mul(x = s_297_cast_fp16, y = s_297_cast_fp16)[name = string("op_8208_cast_fp16")];
            fp32 var_8209_epsilon_0 = const()[name = string("op_8209_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8209_cast_fp16 = inverse(epsilon = var_8209_epsilon_0, x = var_8208_cast_fp16)[name = string("op_8209_cast_fp16")];
            fp16 var_8210_to_fp16 = const()[name = string("op_8210_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8211_cast_fp16 = mul(x = var_8209_cast_fp16, y = var_8210_to_fp16)[name = string("op_8211_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8213_cast_fp16 = add(x = var_8207_cast_fp16, y = var_8211_cast_fp16)[name = string("op_8213_cast_fp16")];
            fp32 var_8214_epsilon_0 = const()[name = string("op_8214_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8214_cast_fp16 = rsqrt(epsilon = var_8214_epsilon_0, x = var_8213_cast_fp16)[name = string("op_8214_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_73_cast_fp16 = mul(x = y_297_cast_fp16, y = var_8214_cast_fp16)[name = string("v_73_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8216_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8216_cast_fp16")];
            tensor<int32, [4]> var_8221_begin_0 = const()[name = string("op_8221_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8221_end_0 = const()[name = string("op_8221_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_8221_end_mask_0 = const()[name = string("op_8221_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_8221_cast_fp16 = slice_by_index(begin = var_8221_begin_0, end = var_8221_end_0, end_mask = var_8221_end_mask_0, x = q_73_cast_fp16)[name = string("op_8221_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_8222_cast_fp16 = mul(x = var_8221_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_8222_cast_fp16")];
            tensor<int32, [4]> var_8227_begin_0 = const()[name = string("op_8227_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8227_end_0 = const()[name = string("op_8227_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_8227_end_mask_0 = const()[name = string("op_8227_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_8227_cast_fp16 = slice_by_index(begin = var_8227_begin_0, end = var_8227_end_0, end_mask = var_8227_end_mask_0, x = q_73_cast_fp16)[name = string("op_8227_cast_fp16")];
            int32 var_8229 = const()[name = string("op_8229"), val = int32(-1)];
            bool var_8230_interleave_0 = const()[name = string("op_8230_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_8230_cast_fp16 = concat(axis = var_8229, interleave = var_8230_interleave_0, values = (var_8222_cast_fp16, var_8227_cast_fp16))[name = string("op_8230_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8231_cast_fp16 = mul(x = var_8230_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8231_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_75_cast_fp16 = add(x = var_8216_cast_fp16, y = var_8231_cast_fp16)[name = string("q_75_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8234_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8234_cast_fp16")];
            tensor<int32, [4]> var_8239_begin_0 = const()[name = string("op_8239_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8239_end_0 = const()[name = string("op_8239_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_8239_end_mask_0 = const()[name = string("op_8239_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_8239_cast_fp16 = slice_by_index(begin = var_8239_begin_0, end = var_8239_end_0, end_mask = var_8239_end_mask_0, x = k_109_cast_fp16)[name = string("op_8239_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_8240_cast_fp16 = mul(x = var_8239_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_8240_cast_fp16")];
            tensor<int32, [4]> var_8245_begin_0 = const()[name = string("op_8245_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8245_end_0 = const()[name = string("op_8245_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_8245_end_mask_0 = const()[name = string("op_8245_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_8245_cast_fp16 = slice_by_index(begin = var_8245_begin_0, end = var_8245_end_0, end_mask = var_8245_end_mask_0, x = k_109_cast_fp16)[name = string("op_8245_cast_fp16")];
            int32 var_8247 = const()[name = string("op_8247"), val = int32(-1)];
            bool var_8248_interleave_0 = const()[name = string("op_8248_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_8248_cast_fp16 = concat(axis = var_8247, interleave = var_8248_interleave_0, values = (var_8240_cast_fp16, var_8245_cast_fp16))[name = string("op_8248_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8249_cast_fp16 = mul(x = var_8248_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8249_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_111_cast_fp16 = add(x = var_8234_cast_fp16, y = var_8249_cast_fp16)[name = string("k_111_cast_fp16")];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_36_reps_0 = const()[name = string("tile_36_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_72_cast_fp16 = transpose(perm = transpose_72_perm_0, x = k_111_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 1, 256, 256]> tile_36_cast_fp16 = tile(reps = tile_36_reps_0, x = transpose_72_cast_fp16)[name = string("tile_36_cast_fp16")];
            tensor<int32, [5]> concat_72 = const()[name = string("concat_72"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_72_cast_fp16 = reshape(shape = concat_72, x = tile_36_cast_fp16)[name = string("reshape_72_cast_fp16")];
            tensor<int32, [5]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_73 = const()[name = string("concat_73"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = reshape_72_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 1, 256, 256]> reshape_73_cast_fp16 = reshape(shape = concat_73, x = transpose_73_cast_fp16)[name = string("reshape_73_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_37_reps_0 = const()[name = string("tile_37_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = v_73_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 1, 256, 256]> tile_37_cast_fp16 = tile(reps = tile_37_reps_0, x = transpose_74_cast_fp16)[name = string("tile_37_cast_fp16")];
            tensor<int32, [5]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_74_cast_fp16 = reshape(shape = concat_74, x = tile_37_cast_fp16)[name = string("reshape_74_cast_fp16")];
            tensor<int32, [5]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = reshape_74_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 1, 256, 256]> reshape_75_cast_fp16 = reshape(shape = concat_75, x = transpose_75_cast_fp16)[name = string("reshape_75_cast_fp16")];
            tensor<int32, [4]> v_75_perm_0 = const()[name = string("v_75_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8263_transpose_x_0 = const()[name = string("op_8263_transpose_x_0"), val = bool(false)];
            bool var_8263_transpose_y_0 = const()[name = string("op_8263_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = reshape_73_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 4, 256, 256]> var_8263_cast_fp16 = matmul(transpose_x = var_8263_transpose_x_0, transpose_y = var_8263_transpose_y_0, x = q_75_cast_fp16, y = transpose_114_cast_fp16)[name = string("op_8263_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8265_cast_fp16 = add(x = var_8263_cast_fp16, y = attention_bias)[name = string("op_8265_cast_fp16")];
            int32 var_8266 = const()[name = string("op_8266"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_8268_cast_fp16 = softmax(axis = var_8266, x = var_8265_cast_fp16)[name = string("op_8268_cast_fp16")];
            bool var_8269_transpose_x_0 = const()[name = string("op_8269_transpose_x_0"), val = bool(false)];
            bool var_8269_transpose_y_0 = const()[name = string("op_8269_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_75_cast_fp16 = transpose(perm = v_75_perm_0, x = reshape_75_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 4, 256, 256]> var_8269_cast_fp16 = matmul(transpose_x = var_8269_transpose_x_0, transpose_y = var_8269_transpose_y_0, x = var_8268_cast_fp16, y = v_75_cast_fp16)[name = string("op_8269_cast_fp16")];
            tensor<int32, [4]> var_8272_perm_0 = const()[name = string("op_8272_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8276 = const()[name = string("op_8276"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_8272_cast_fp16 = transpose(perm = var_8272_perm_0, x = var_8269_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 256, 1024]> input_219_cast_fp16 = reshape(shape = var_8276, x = var_8272_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105361024))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105885376))))[name = string("tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8281_cast_fp16 = abs(x = linear_166_cast_fp16)[name = string("op_8281_cast_fp16")];
            bool var_8284 = const()[name = string("op_8284"), val = bool(true)];
            tensor<int32, [1]> var_8285_axes_0 = const()[name = string("op_8285_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8285_cast_fp16 = reduce_max(axes = var_8285_axes_0, keep_dims = var_8284, x = var_8281_cast_fp16)[name = string("op_8285_cast_fp16")];
            fp16 var_8286_to_fp16 = const()[name = string("op_8286_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_299_cast_fp16 = maximum(x = var_8285_cast_fp16, y = var_8286_to_fp16)[name = string("s_299_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_299_cast_fp16 = real_div(x = linear_166_cast_fp16, y = s_299_cast_fp16)[name = string("y_299_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8289_cast_fp16 = mul(x = y_299_cast_fp16, y = y_299_cast_fp16)[name = string("op_8289_cast_fp16")];
            tensor<int32, [1]> var_8294_axes_0 = const()[name = string("op_8294_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8294_keep_dims_0 = const()[name = string("op_8294_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8294_cast_fp16 = reduce_mean(axes = var_8294_axes_0, keep_dims = var_8294_keep_dims_0, x = var_8289_cast_fp16)[name = string("op_8294_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8295_cast_fp16 = mul(x = s_299_cast_fp16, y = s_299_cast_fp16)[name = string("op_8295_cast_fp16")];
            fp32 var_8296_epsilon_0 = const()[name = string("op_8296_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8296_cast_fp16 = inverse(epsilon = var_8296_epsilon_0, x = var_8295_cast_fp16)[name = string("op_8296_cast_fp16")];
            fp16 var_8297_to_fp16 = const()[name = string("op_8297_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8298_cast_fp16 = mul(x = var_8296_cast_fp16, y = var_8297_to_fp16)[name = string("op_8298_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8300_cast_fp16 = add(x = var_8294_cast_fp16, y = var_8298_cast_fp16)[name = string("op_8300_cast_fp16")];
            fp32 var_8301_epsilon_0 = const()[name = string("op_8301_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8301_cast_fp16 = rsqrt(epsilon = var_8301_epsilon_0, x = var_8300_cast_fp16)[name = string("op_8301_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_261_cast_fp16 = mul(x = y_299_cast_fp16, y = var_8301_cast_fp16)[name = string("out_261_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105886464)))];
            tensor<fp16, [1, 256, 512]> var_8303_cast_fp16 = mul(x = out_261_cast_fp16, y = tm_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("op_8303_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = var_8303_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8306_cast_fp16 = abs(x = x_299_cast_fp16)[name = string("op_8306_cast_fp16")];
            bool var_8309 = const()[name = string("op_8309"), val = bool(true)];
            tensor<int32, [1]> var_8310_axes_0 = const()[name = string("op_8310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8310_cast_fp16 = reduce_max(axes = var_8310_axes_0, keep_dims = var_8309, x = var_8306_cast_fp16)[name = string("op_8310_cast_fp16")];
            fp16 var_8311_to_fp16 = const()[name = string("op_8311_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_301_cast_fp16 = maximum(x = var_8310_cast_fp16, y = var_8311_to_fp16)[name = string("s_301_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_301_cast_fp16 = real_div(x = x_299_cast_fp16, y = s_301_cast_fp16)[name = string("y_301_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8314_cast_fp16 = mul(x = y_301_cast_fp16, y = y_301_cast_fp16)[name = string("op_8314_cast_fp16")];
            tensor<int32, [1]> var_8319_axes_0 = const()[name = string("op_8319_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8319_keep_dims_0 = const()[name = string("op_8319_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8319_cast_fp16 = reduce_mean(axes = var_8319_axes_0, keep_dims = var_8319_keep_dims_0, x = var_8314_cast_fp16)[name = string("op_8319_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8320_cast_fp16 = mul(x = s_301_cast_fp16, y = s_301_cast_fp16)[name = string("op_8320_cast_fp16")];
            fp32 var_8321_epsilon_0 = const()[name = string("op_8321_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8321_cast_fp16 = inverse(epsilon = var_8321_epsilon_0, x = var_8320_cast_fp16)[name = string("op_8321_cast_fp16")];
            fp16 var_8322_to_fp16 = const()[name = string("op_8322_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8323_cast_fp16 = mul(x = var_8321_cast_fp16, y = var_8322_to_fp16)[name = string("op_8323_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8325_cast_fp16 = add(x = var_8319_cast_fp16, y = var_8323_cast_fp16)[name = string("op_8325_cast_fp16")];
            fp32 var_8326_epsilon_0 = const()[name = string("op_8326_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8326_cast_fp16 = rsqrt(epsilon = var_8326_epsilon_0, x = var_8325_cast_fp16)[name = string("op_8326_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_263_cast_fp16 = mul(x = y_301_cast_fp16, y = var_8326_cast_fp16)[name = string("out_263_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105887552)))];
            tensor<fp16, [1, 256, 512]> input_221_cast_fp16 = mul(x = out_263_cast_fp16, y = tm_layers_18_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105888640))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106937280))))[name = string("tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_167_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_167_cast_fp16")];
            string var_8333_mode_0 = const()[name = string("op_8333_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_8333_cast_fp16 = gelu(mode = var_8333_mode_0, x = linear_167_cast_fp16)[name = string("op_8333_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106941440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107990080))))[name = string("tm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_168_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_223_cast_fp16 = mul(x = var_8333_cast_fp16, y = linear_168_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107994240))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109042880))))[name = string("tm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8341_cast_fp16 = abs(x = linear_169_cast_fp16)[name = string("op_8341_cast_fp16")];
            bool var_8344 = const()[name = string("op_8344"), val = bool(true)];
            tensor<int32, [1]> var_8345_axes_0 = const()[name = string("op_8345_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8345_cast_fp16 = reduce_max(axes = var_8345_axes_0, keep_dims = var_8344, x = var_8341_cast_fp16)[name = string("op_8345_cast_fp16")];
            fp16 var_8346_to_fp16 = const()[name = string("op_8346_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_303_cast_fp16 = maximum(x = var_8345_cast_fp16, y = var_8346_to_fp16)[name = string("s_303_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_303_cast_fp16 = real_div(x = linear_169_cast_fp16, y = s_303_cast_fp16)[name = string("y_303_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8349_cast_fp16 = mul(x = y_303_cast_fp16, y = y_303_cast_fp16)[name = string("op_8349_cast_fp16")];
            tensor<int32, [1]> var_8354_axes_0 = const()[name = string("op_8354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8354_keep_dims_0 = const()[name = string("op_8354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8354_cast_fp16 = reduce_mean(axes = var_8354_axes_0, keep_dims = var_8354_keep_dims_0, x = var_8349_cast_fp16)[name = string("op_8354_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8355_cast_fp16 = mul(x = s_303_cast_fp16, y = s_303_cast_fp16)[name = string("op_8355_cast_fp16")];
            fp32 var_8356_epsilon_0 = const()[name = string("op_8356_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8356_cast_fp16 = inverse(epsilon = var_8356_epsilon_0, x = var_8355_cast_fp16)[name = string("op_8356_cast_fp16")];
            fp16 var_8357_to_fp16 = const()[name = string("op_8357_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8358_cast_fp16 = mul(x = var_8356_cast_fp16, y = var_8357_to_fp16)[name = string("op_8358_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8360_cast_fp16 = add(x = var_8354_cast_fp16, y = var_8358_cast_fp16)[name = string("op_8360_cast_fp16")];
            fp32 var_8361_epsilon_0 = const()[name = string("op_8361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8361_cast_fp16 = rsqrt(epsilon = var_8361_epsilon_0, x = var_8360_cast_fp16)[name = string("op_8361_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_265_cast_fp16 = mul(x = y_303_cast_fp16, y = var_8361_cast_fp16)[name = string("out_265_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_18_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109043968)))];
            tensor<fp16, [1, 256, 512]> var_8363_cast_fp16 = mul(x = out_265_cast_fp16, y = tm_layers_18_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8363_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_225_cast_fp16 = add(x = x_299_cast_fp16, y = var_8363_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109045056))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109307264))))[name = string("tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_170_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_170_cast_fp16")];
            string var_8370_mode_0 = const()[name = string("op_8370_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_8370_cast_fp16 = gelu(mode = var_8370_mode_0, x = linear_170_cast_fp16)[name = string("op_8370_cast_fp16")];
            tensor<int32, [4]> var_8383_begin_0 = const()[name = string("op_8383_begin_0"), val = tensor<int32, [4]>([0, 0, 18, 0])];
            tensor<int32, [4]> var_8383_end_0 = const()[name = string("op_8383_end_0"), val = tensor<int32, [4]>([1, 256, 19, 512])];
            tensor<bool, [4]> var_8383_end_mask_0 = const()[name = string("op_8383_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8383_squeeze_mask_0 = const()[name = string("op_8383_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_8383_cast_fp16 = slice_by_index(begin = var_8383_begin_0, end = var_8383_end_0, end_mask = var_8383_end_mask_0, squeeze_mask = var_8383_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_8383_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_227_cast_fp16 = mul(x = var_8370_cast_fp16, y = var_8383_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109308352))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109570560))))[name = string("tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_18_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8393_cast_fp16 = abs(x = linear_171_cast_fp16)[name = string("op_8393_cast_fp16")];
            bool var_8396 = const()[name = string("op_8396"), val = bool(true)];
            tensor<int32, [1]> var_8397_axes_0 = const()[name = string("op_8397_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8397_cast_fp16 = reduce_max(axes = var_8397_axes_0, keep_dims = var_8396, x = var_8393_cast_fp16)[name = string("op_8397_cast_fp16")];
            fp16 var_8398_to_fp16 = const()[name = string("op_8398_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_305_cast_fp16 = maximum(x = var_8397_cast_fp16, y = var_8398_to_fp16)[name = string("s_305_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_305_cast_fp16 = real_div(x = linear_171_cast_fp16, y = s_305_cast_fp16)[name = string("y_305_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8401_cast_fp16 = mul(x = y_305_cast_fp16, y = y_305_cast_fp16)[name = string("op_8401_cast_fp16")];
            tensor<int32, [1]> var_8406_axes_0 = const()[name = string("op_8406_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8406_keep_dims_0 = const()[name = string("op_8406_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8406_cast_fp16 = reduce_mean(axes = var_8406_axes_0, keep_dims = var_8406_keep_dims_0, x = var_8401_cast_fp16)[name = string("op_8406_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8407_cast_fp16 = mul(x = s_305_cast_fp16, y = s_305_cast_fp16)[name = string("op_8407_cast_fp16")];
            fp32 var_8408_epsilon_0 = const()[name = string("op_8408_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8408_cast_fp16 = inverse(epsilon = var_8408_epsilon_0, x = var_8407_cast_fp16)[name = string("op_8408_cast_fp16")];
            fp16 var_8409_to_fp16 = const()[name = string("op_8409_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8410_cast_fp16 = mul(x = var_8408_cast_fp16, y = var_8409_to_fp16)[name = string("op_8410_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8412_cast_fp16 = add(x = var_8406_cast_fp16, y = var_8410_cast_fp16)[name = string("op_8412_cast_fp16")];
            fp32 var_8413_epsilon_0 = const()[name = string("op_8413_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8413_cast_fp16 = rsqrt(epsilon = var_8413_epsilon_0, x = var_8412_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_267_cast_fp16 = mul(x = y_305_cast_fp16, y = var_8413_cast_fp16)[name = string("out_267_cast_fp16")];
            tensor<fp16, [512]> tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109571648)))];
            tensor<fp16, [1, 256, 512]> var_8415_cast_fp16 = mul(x = out_267_cast_fp16, y = tm_layers_18_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8415_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_37_cast_fp16 = add(x = input_225_cast_fp16, y = var_8415_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<fp16, [1]> tm_layers_18_layer_scalar_to_fp16 = const()[name = string("tm_layers_18_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.68p-1])];
            tensor<fp16, [1, 256, 512]> x_305_cast_fp16 = mul(x = h_37_cast_fp16, y = tm_layers_18_layer_scalar_to_fp16)[name = string("x_305_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8419_cast_fp16 = abs(x = x_305_cast_fp16)[name = string("op_8419_cast_fp16")];
            bool var_8422 = const()[name = string("op_8422"), val = bool(true)];
            tensor<int32, [1]> var_8423_axes_0 = const()[name = string("op_8423_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8423_cast_fp16 = reduce_max(axes = var_8423_axes_0, keep_dims = var_8422, x = var_8419_cast_fp16)[name = string("op_8423_cast_fp16")];
            fp16 var_8424_to_fp16 = const()[name = string("op_8424_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_307_cast_fp16 = maximum(x = var_8423_cast_fp16, y = var_8424_to_fp16)[name = string("s_307_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_307_cast_fp16 = real_div(x = x_305_cast_fp16, y = s_307_cast_fp16)[name = string("y_307_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8427_cast_fp16 = mul(x = y_307_cast_fp16, y = y_307_cast_fp16)[name = string("op_8427_cast_fp16")];
            tensor<int32, [1]> var_8432_axes_0 = const()[name = string("op_8432_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8432_keep_dims_0 = const()[name = string("op_8432_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8432_cast_fp16 = reduce_mean(axes = var_8432_axes_0, keep_dims = var_8432_keep_dims_0, x = var_8427_cast_fp16)[name = string("op_8432_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8433_cast_fp16 = mul(x = s_307_cast_fp16, y = s_307_cast_fp16)[name = string("op_8433_cast_fp16")];
            fp32 var_8434_epsilon_0 = const()[name = string("op_8434_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8434_cast_fp16 = inverse(epsilon = var_8434_epsilon_0, x = var_8433_cast_fp16)[name = string("op_8434_cast_fp16")];
            fp16 var_8435_to_fp16 = const()[name = string("op_8435_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8436_cast_fp16 = mul(x = var_8434_cast_fp16, y = var_8435_to_fp16)[name = string("op_8436_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8438_cast_fp16 = add(x = var_8432_cast_fp16, y = var_8436_cast_fp16)[name = string("op_8438_cast_fp16")];
            fp32 var_8439_epsilon_0 = const()[name = string("op_8439_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8439_cast_fp16 = rsqrt(epsilon = var_8439_epsilon_0, x = var_8438_cast_fp16)[name = string("op_8439_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_269_cast_fp16 = mul(x = y_307_cast_fp16, y = var_8439_cast_fp16)[name = string("out_269_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109572736)))];
            tensor<fp16, [1, 256, 512]> input_229_cast_fp16 = mul(x = out_269_cast_fp16, y = tm_layers_19_input_layernorm_weight_to_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109573824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110098176))))[name = string("tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [4]> var_8449 = const()[name = string("op_8449"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_8450_cast_fp16 = reshape(shape = var_8449, x = linear_172_cast_fp16)[name = string("op_8450_cast_fp16")];
            tensor<int32, [4]> x_307_perm_0 = const()[name = string("x_307_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110100288))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110362496))))[name = string("tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<int32, [4]> var_8461 = const()[name = string("op_8461"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8462_cast_fp16 = reshape(shape = var_8461, x = linear_173_cast_fp16)[name = string("op_8462_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110363584))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110625792))))[name = string("tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<int32, [4]> var_8473 = const()[name = string("op_8473"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8474_cast_fp16 = reshape(shape = var_8473, x = linear_174_cast_fp16)[name = string("op_8474_cast_fp16")];
            tensor<int32, [4]> x_311_perm_0 = const()[name = string("x_311_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_307_cast_fp16 = transpose(perm = x_307_perm_0, x = var_8450_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 4, 256, 256]> var_8478_cast_fp16 = abs(x = x_307_cast_fp16)[name = string("op_8478_cast_fp16")];
            bool var_8481 = const()[name = string("op_8481"), val = bool(true)];
            tensor<int32, [1]> var_8482_axes_0 = const()[name = string("op_8482_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_8482_cast_fp16 = reduce_max(axes = var_8482_axes_0, keep_dims = var_8481, x = var_8478_cast_fp16)[name = string("op_8482_cast_fp16")];
            fp16 var_8483_to_fp16 = const()[name = string("op_8483_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_309_cast_fp16 = maximum(x = var_8482_cast_fp16, y = var_8483_to_fp16)[name = string("s_309_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_309_cast_fp16 = real_div(x = x_307_cast_fp16, y = s_309_cast_fp16)[name = string("y_309_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8486_cast_fp16 = mul(x = y_309_cast_fp16, y = y_309_cast_fp16)[name = string("op_8486_cast_fp16")];
            tensor<int32, [1]> var_8491_axes_0 = const()[name = string("op_8491_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8491_keep_dims_0 = const()[name = string("op_8491_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_8491_cast_fp16 = reduce_mean(axes = var_8491_axes_0, keep_dims = var_8491_keep_dims_0, x = var_8486_cast_fp16)[name = string("op_8491_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8492_cast_fp16 = mul(x = s_309_cast_fp16, y = s_309_cast_fp16)[name = string("op_8492_cast_fp16")];
            fp32 var_8493_epsilon_0 = const()[name = string("op_8493_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_8493_cast_fp16 = inverse(epsilon = var_8493_epsilon_0, x = var_8492_cast_fp16)[name = string("op_8493_cast_fp16")];
            fp16 var_8494_to_fp16 = const()[name = string("op_8494_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_8495_cast_fp16 = mul(x = var_8493_cast_fp16, y = var_8494_to_fp16)[name = string("op_8495_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8497_cast_fp16 = add(x = var_8491_cast_fp16, y = var_8495_cast_fp16)[name = string("op_8497_cast_fp16")];
            fp32 var_8498_epsilon_0 = const()[name = string("op_8498_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_8498_cast_fp16 = rsqrt(epsilon = var_8498_epsilon_0, x = var_8497_cast_fp16)[name = string("op_8498_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_271_cast_fp16 = mul(x = y_309_cast_fp16, y = var_8498_cast_fp16)[name = string("out_271_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_77_cast_fp16 = mul(x = out_271_cast_fp16, y = tm_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("q_77_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_8462_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2, 256, 256]> var_8501_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_8501_cast_fp16")];
            bool var_8504 = const()[name = string("op_8504"), val = bool(true)];
            tensor<int32, [1]> var_8505_axes_0 = const()[name = string("op_8505_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8505_cast_fp16 = reduce_max(axes = var_8505_axes_0, keep_dims = var_8504, x = var_8501_cast_fp16)[name = string("op_8505_cast_fp16")];
            fp16 var_8506_to_fp16 = const()[name = string("op_8506_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_311_cast_fp16 = maximum(x = var_8505_cast_fp16, y = var_8506_to_fp16)[name = string("s_311_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_311_cast_fp16 = real_div(x = x_309_cast_fp16, y = s_311_cast_fp16)[name = string("y_311_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8509_cast_fp16 = mul(x = y_311_cast_fp16, y = y_311_cast_fp16)[name = string("op_8509_cast_fp16")];
            tensor<int32, [1]> var_8514_axes_0 = const()[name = string("op_8514_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8514_keep_dims_0 = const()[name = string("op_8514_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8514_cast_fp16 = reduce_mean(axes = var_8514_axes_0, keep_dims = var_8514_keep_dims_0, x = var_8509_cast_fp16)[name = string("op_8514_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8515_cast_fp16 = mul(x = s_311_cast_fp16, y = s_311_cast_fp16)[name = string("op_8515_cast_fp16")];
            fp32 var_8516_epsilon_0 = const()[name = string("op_8516_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8516_cast_fp16 = inverse(epsilon = var_8516_epsilon_0, x = var_8515_cast_fp16)[name = string("op_8516_cast_fp16")];
            fp16 var_8517_to_fp16 = const()[name = string("op_8517_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8518_cast_fp16 = mul(x = var_8516_cast_fp16, y = var_8517_to_fp16)[name = string("op_8518_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8520_cast_fp16 = add(x = var_8514_cast_fp16, y = var_8518_cast_fp16)[name = string("op_8520_cast_fp16")];
            fp32 var_8521_epsilon_0 = const()[name = string("op_8521_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8521_cast_fp16 = rsqrt(epsilon = var_8521_epsilon_0, x = var_8520_cast_fp16)[name = string("op_8521_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_273_cast_fp16 = mul(x = y_311_cast_fp16, y = var_8521_cast_fp16)[name = string("out_273_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_115_cast_fp16 = mul(x = out_273_cast_fp16, y = tm_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("k_115_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_311_cast_fp16 = transpose(perm = x_311_perm_0, x = var_8474_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 2, 256, 256]> var_8524_cast_fp16 = abs(x = x_311_cast_fp16)[name = string("op_8524_cast_fp16")];
            bool var_8527 = const()[name = string("op_8527"), val = bool(true)];
            tensor<int32, [1]> var_8528_axes_0 = const()[name = string("op_8528_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8528_cast_fp16 = reduce_max(axes = var_8528_axes_0, keep_dims = var_8527, x = var_8524_cast_fp16)[name = string("op_8528_cast_fp16")];
            fp16 var_8529_to_fp16 = const()[name = string("op_8529_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_313_cast_fp16 = maximum(x = var_8528_cast_fp16, y = var_8529_to_fp16)[name = string("s_313_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_313_cast_fp16 = real_div(x = x_311_cast_fp16, y = s_313_cast_fp16)[name = string("y_313_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8532_cast_fp16 = mul(x = y_313_cast_fp16, y = y_313_cast_fp16)[name = string("op_8532_cast_fp16")];
            tensor<int32, [1]> var_8537_axes_0 = const()[name = string("op_8537_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8537_keep_dims_0 = const()[name = string("op_8537_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8537_cast_fp16 = reduce_mean(axes = var_8537_axes_0, keep_dims = var_8537_keep_dims_0, x = var_8532_cast_fp16)[name = string("op_8537_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8538_cast_fp16 = mul(x = s_313_cast_fp16, y = s_313_cast_fp16)[name = string("op_8538_cast_fp16")];
            fp32 var_8539_epsilon_0 = const()[name = string("op_8539_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8539_cast_fp16 = inverse(epsilon = var_8539_epsilon_0, x = var_8538_cast_fp16)[name = string("op_8539_cast_fp16")];
            fp16 var_8540_to_fp16 = const()[name = string("op_8540_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8541_cast_fp16 = mul(x = var_8539_cast_fp16, y = var_8540_to_fp16)[name = string("op_8541_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8543_cast_fp16 = add(x = var_8537_cast_fp16, y = var_8541_cast_fp16)[name = string("op_8543_cast_fp16")];
            fp32 var_8544_epsilon_0 = const()[name = string("op_8544_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8544_cast_fp16 = rsqrt(epsilon = var_8544_epsilon_0, x = var_8543_cast_fp16)[name = string("op_8544_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_77_cast_fp16 = mul(x = y_313_cast_fp16, y = var_8544_cast_fp16)[name = string("v_77_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8546_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8546_cast_fp16")];
            tensor<int32, [4]> var_8551_begin_0 = const()[name = string("op_8551_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8551_end_0 = const()[name = string("op_8551_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_8551_end_mask_0 = const()[name = string("op_8551_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_8551_cast_fp16 = slice_by_index(begin = var_8551_begin_0, end = var_8551_end_0, end_mask = var_8551_end_mask_0, x = q_77_cast_fp16)[name = string("op_8551_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_8552_cast_fp16 = mul(x = var_8551_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_8552_cast_fp16")];
            tensor<int32, [4]> var_8557_begin_0 = const()[name = string("op_8557_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8557_end_0 = const()[name = string("op_8557_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_8557_end_mask_0 = const()[name = string("op_8557_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_8557_cast_fp16 = slice_by_index(begin = var_8557_begin_0, end = var_8557_end_0, end_mask = var_8557_end_mask_0, x = q_77_cast_fp16)[name = string("op_8557_cast_fp16")];
            int32 var_8559 = const()[name = string("op_8559"), val = int32(-1)];
            bool var_8560_interleave_0 = const()[name = string("op_8560_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_8560_cast_fp16 = concat(axis = var_8559, interleave = var_8560_interleave_0, values = (var_8552_cast_fp16, var_8557_cast_fp16))[name = string("op_8560_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8561_cast_fp16 = mul(x = var_8560_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8561_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_79_cast_fp16 = add(x = var_8546_cast_fp16, y = var_8561_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8564_cast_fp16 = mul(x = k_115_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8564_cast_fp16")];
            tensor<int32, [4]> var_8569_begin_0 = const()[name = string("op_8569_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8569_end_0 = const()[name = string("op_8569_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_8569_end_mask_0 = const()[name = string("op_8569_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_8569_cast_fp16 = slice_by_index(begin = var_8569_begin_0, end = var_8569_end_0, end_mask = var_8569_end_mask_0, x = k_115_cast_fp16)[name = string("op_8569_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_8570_cast_fp16 = mul(x = var_8569_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_8570_cast_fp16")];
            tensor<int32, [4]> var_8575_begin_0 = const()[name = string("op_8575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8575_end_0 = const()[name = string("op_8575_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_8575_end_mask_0 = const()[name = string("op_8575_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_8575_cast_fp16 = slice_by_index(begin = var_8575_begin_0, end = var_8575_end_0, end_mask = var_8575_end_mask_0, x = k_115_cast_fp16)[name = string("op_8575_cast_fp16")];
            int32 var_8577 = const()[name = string("op_8577"), val = int32(-1)];
            bool var_8578_interleave_0 = const()[name = string("op_8578_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_8578_cast_fp16 = concat(axis = var_8577, interleave = var_8578_interleave_0, values = (var_8570_cast_fp16, var_8575_cast_fp16))[name = string("op_8578_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8579_cast_fp16 = mul(x = var_8578_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8579_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_117_cast_fp16 = add(x = var_8564_cast_fp16, y = var_8579_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_38_reps_0 = const()[name = string("tile_38_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = k_117_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 1, 256, 256]> tile_38_cast_fp16 = tile(reps = tile_38_reps_0, x = transpose_76_cast_fp16)[name = string("tile_38_cast_fp16")];
            tensor<int32, [5]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_76_cast_fp16 = reshape(shape = concat_76, x = tile_38_cast_fp16)[name = string("reshape_76_cast_fp16")];
            tensor<int32, [5]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_77 = const()[name = string("concat_77"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_77_cast_fp16 = transpose(perm = transpose_77_perm_0, x = reshape_76_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 1, 256, 256]> reshape_77_cast_fp16 = reshape(shape = concat_77, x = transpose_77_cast_fp16)[name = string("reshape_77_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_39_reps_0 = const()[name = string("tile_39_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_78_cast_fp16 = transpose(perm = transpose_78_perm_0, x = v_77_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 1, 256, 256]> tile_39_cast_fp16 = tile(reps = tile_39_reps_0, x = transpose_78_cast_fp16)[name = string("tile_39_cast_fp16")];
            tensor<int32, [5]> concat_78 = const()[name = string("concat_78"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_78_cast_fp16 = reshape(shape = concat_78, x = tile_39_cast_fp16)[name = string("reshape_78_cast_fp16")];
            tensor<int32, [5]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = reshape_78_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 1, 256, 256]> reshape_79_cast_fp16 = reshape(shape = concat_79, x = transpose_79_cast_fp16)[name = string("reshape_79_cast_fp16")];
            tensor<int32, [4]> v_79_perm_0 = const()[name = string("v_79_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8593_transpose_x_0 = const()[name = string("op_8593_transpose_x_0"), val = bool(false)];
            bool var_8593_transpose_y_0 = const()[name = string("op_8593_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = reshape_77_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4, 256, 256]> var_8593_cast_fp16 = matmul(transpose_x = var_8593_transpose_x_0, transpose_y = var_8593_transpose_y_0, x = q_79_cast_fp16, y = transpose_115_cast_fp16)[name = string("op_8593_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8595_cast_fp16 = add(x = var_8593_cast_fp16, y = attention_bias)[name = string("op_8595_cast_fp16")];
            int32 var_8596 = const()[name = string("op_8596"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_8598_cast_fp16 = softmax(axis = var_8596, x = var_8595_cast_fp16)[name = string("op_8598_cast_fp16")];
            bool var_8599_transpose_x_0 = const()[name = string("op_8599_transpose_x_0"), val = bool(false)];
            bool var_8599_transpose_y_0 = const()[name = string("op_8599_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_79_cast_fp16 = transpose(perm = v_79_perm_0, x = reshape_79_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 4, 256, 256]> var_8599_cast_fp16 = matmul(transpose_x = var_8599_transpose_x_0, transpose_y = var_8599_transpose_y_0, x = var_8598_cast_fp16, y = v_79_cast_fp16)[name = string("op_8599_cast_fp16")];
            tensor<int32, [4]> var_8602_perm_0 = const()[name = string("op_8602_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8606 = const()[name = string("op_8606"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_8602_cast_fp16 = transpose(perm = var_8602_perm_0, x = var_8599_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 256, 1024]> input_231_cast_fp16 = reshape(shape = var_8606, x = var_8602_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110626880))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111151232))))[name = string("tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8611_cast_fp16 = abs(x = linear_175_cast_fp16)[name = string("op_8611_cast_fp16")];
            bool var_8614 = const()[name = string("op_8614"), val = bool(true)];
            tensor<int32, [1]> var_8615_axes_0 = const()[name = string("op_8615_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8615_cast_fp16 = reduce_max(axes = var_8615_axes_0, keep_dims = var_8614, x = var_8611_cast_fp16)[name = string("op_8615_cast_fp16")];
            fp16 var_8616_to_fp16 = const()[name = string("op_8616_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_315_cast_fp16 = maximum(x = var_8615_cast_fp16, y = var_8616_to_fp16)[name = string("s_315_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_315_cast_fp16 = real_div(x = linear_175_cast_fp16, y = s_315_cast_fp16)[name = string("y_315_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8619_cast_fp16 = mul(x = y_315_cast_fp16, y = y_315_cast_fp16)[name = string("op_8619_cast_fp16")];
            tensor<int32, [1]> var_8624_axes_0 = const()[name = string("op_8624_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8624_keep_dims_0 = const()[name = string("op_8624_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8624_cast_fp16 = reduce_mean(axes = var_8624_axes_0, keep_dims = var_8624_keep_dims_0, x = var_8619_cast_fp16)[name = string("op_8624_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8625_cast_fp16 = mul(x = s_315_cast_fp16, y = s_315_cast_fp16)[name = string("op_8625_cast_fp16")];
            fp32 var_8626_epsilon_0 = const()[name = string("op_8626_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8626_cast_fp16 = inverse(epsilon = var_8626_epsilon_0, x = var_8625_cast_fp16)[name = string("op_8626_cast_fp16")];
            fp16 var_8627_to_fp16 = const()[name = string("op_8627_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8628_cast_fp16 = mul(x = var_8626_cast_fp16, y = var_8627_to_fp16)[name = string("op_8628_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8630_cast_fp16 = add(x = var_8624_cast_fp16, y = var_8628_cast_fp16)[name = string("op_8630_cast_fp16")];
            fp32 var_8631_epsilon_0 = const()[name = string("op_8631_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8631_cast_fp16 = rsqrt(epsilon = var_8631_epsilon_0, x = var_8630_cast_fp16)[name = string("op_8631_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_275_cast_fp16 = mul(x = y_315_cast_fp16, y = var_8631_cast_fp16)[name = string("out_275_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111152320)))];
            tensor<fp16, [1, 256, 512]> var_8633_cast_fp16 = mul(x = out_275_cast_fp16, y = tm_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("op_8633_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_315_cast_fp16 = add(x = x_305_cast_fp16, y = var_8633_cast_fp16)[name = string("x_315_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8636_cast_fp16 = abs(x = x_315_cast_fp16)[name = string("op_8636_cast_fp16")];
            bool var_8639 = const()[name = string("op_8639"), val = bool(true)];
            tensor<int32, [1]> var_8640_axes_0 = const()[name = string("op_8640_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8640_cast_fp16 = reduce_max(axes = var_8640_axes_0, keep_dims = var_8639, x = var_8636_cast_fp16)[name = string("op_8640_cast_fp16")];
            fp16 var_8641_to_fp16 = const()[name = string("op_8641_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_317_cast_fp16 = maximum(x = var_8640_cast_fp16, y = var_8641_to_fp16)[name = string("s_317_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_317_cast_fp16 = real_div(x = x_315_cast_fp16, y = s_317_cast_fp16)[name = string("y_317_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8644_cast_fp16 = mul(x = y_317_cast_fp16, y = y_317_cast_fp16)[name = string("op_8644_cast_fp16")];
            tensor<int32, [1]> var_8649_axes_0 = const()[name = string("op_8649_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8649_keep_dims_0 = const()[name = string("op_8649_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8649_cast_fp16 = reduce_mean(axes = var_8649_axes_0, keep_dims = var_8649_keep_dims_0, x = var_8644_cast_fp16)[name = string("op_8649_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8650_cast_fp16 = mul(x = s_317_cast_fp16, y = s_317_cast_fp16)[name = string("op_8650_cast_fp16")];
            fp32 var_8651_epsilon_0 = const()[name = string("op_8651_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8651_cast_fp16 = inverse(epsilon = var_8651_epsilon_0, x = var_8650_cast_fp16)[name = string("op_8651_cast_fp16")];
            fp16 var_8652_to_fp16 = const()[name = string("op_8652_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8653_cast_fp16 = mul(x = var_8651_cast_fp16, y = var_8652_to_fp16)[name = string("op_8653_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8655_cast_fp16 = add(x = var_8649_cast_fp16, y = var_8653_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp32 var_8656_epsilon_0 = const()[name = string("op_8656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8656_cast_fp16 = rsqrt(epsilon = var_8656_epsilon_0, x = var_8655_cast_fp16)[name = string("op_8656_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_277_cast_fp16 = mul(x = y_317_cast_fp16, y = var_8656_cast_fp16)[name = string("out_277_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111153408)))];
            tensor<fp16, [1, 256, 512]> input_233_cast_fp16 = mul(x = out_277_cast_fp16, y = tm_layers_19_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111154496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112203136))))[name = string("tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_176_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_176_cast_fp16")];
            string var_8663_mode_0 = const()[name = string("op_8663_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_8663_cast_fp16 = gelu(mode = var_8663_mode_0, x = linear_176_cast_fp16)[name = string("op_8663_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112207296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113255936))))[name = string("tm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_177_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_235_cast_fp16 = mul(x = var_8663_cast_fp16, y = linear_177_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113260096))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114308736))))[name = string("tm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8671_cast_fp16 = abs(x = linear_178_cast_fp16)[name = string("op_8671_cast_fp16")];
            bool var_8674 = const()[name = string("op_8674"), val = bool(true)];
            tensor<int32, [1]> var_8675_axes_0 = const()[name = string("op_8675_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8675_cast_fp16 = reduce_max(axes = var_8675_axes_0, keep_dims = var_8674, x = var_8671_cast_fp16)[name = string("op_8675_cast_fp16")];
            fp16 var_8676_to_fp16 = const()[name = string("op_8676_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_319_cast_fp16 = maximum(x = var_8675_cast_fp16, y = var_8676_to_fp16)[name = string("s_319_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_319_cast_fp16 = real_div(x = linear_178_cast_fp16, y = s_319_cast_fp16)[name = string("y_319_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8679_cast_fp16 = mul(x = y_319_cast_fp16, y = y_319_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<int32, [1]> var_8684_axes_0 = const()[name = string("op_8684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8684_keep_dims_0 = const()[name = string("op_8684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8684_cast_fp16 = reduce_mean(axes = var_8684_axes_0, keep_dims = var_8684_keep_dims_0, x = var_8679_cast_fp16)[name = string("op_8684_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8685_cast_fp16 = mul(x = s_319_cast_fp16, y = s_319_cast_fp16)[name = string("op_8685_cast_fp16")];
            fp32 var_8686_epsilon_0 = const()[name = string("op_8686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8686_cast_fp16 = inverse(epsilon = var_8686_epsilon_0, x = var_8685_cast_fp16)[name = string("op_8686_cast_fp16")];
            fp16 var_8687_to_fp16 = const()[name = string("op_8687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8688_cast_fp16 = mul(x = var_8686_cast_fp16, y = var_8687_to_fp16)[name = string("op_8688_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8690_cast_fp16 = add(x = var_8684_cast_fp16, y = var_8688_cast_fp16)[name = string("op_8690_cast_fp16")];
            fp32 var_8691_epsilon_0 = const()[name = string("op_8691_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8691_cast_fp16 = rsqrt(epsilon = var_8691_epsilon_0, x = var_8690_cast_fp16)[name = string("op_8691_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_279_cast_fp16 = mul(x = y_319_cast_fp16, y = var_8691_cast_fp16)[name = string("out_279_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_19_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114309824)))];
            tensor<fp16, [1, 256, 512]> var_8693_cast_fp16 = mul(x = out_279_cast_fp16, y = tm_layers_19_post_feedforward_layernorm_weight_to_fp16)[name = string("op_8693_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_237_cast_fp16 = add(x = x_315_cast_fp16, y = var_8693_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114310912))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114573120))))[name = string("tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_179_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_237_cast_fp16)[name = string("linear_179_cast_fp16")];
            string var_8700_mode_0 = const()[name = string("op_8700_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_8700_cast_fp16 = gelu(mode = var_8700_mode_0, x = linear_179_cast_fp16)[name = string("op_8700_cast_fp16")];
            tensor<int32, [4]> var_8713_begin_0 = const()[name = string("op_8713_begin_0"), val = tensor<int32, [4]>([0, 0, 19, 0])];
            tensor<int32, [4]> var_8713_end_0 = const()[name = string("op_8713_end_0"), val = tensor<int32, [4]>([1, 256, 20, 512])];
            tensor<bool, [4]> var_8713_end_mask_0 = const()[name = string("op_8713_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_8713_squeeze_mask_0 = const()[name = string("op_8713_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_8713_cast_fp16 = slice_by_index(begin = var_8713_begin_0, end = var_8713_end_0, end_mask = var_8713_end_mask_0, squeeze_mask = var_8713_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_8713_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_239_cast_fp16 = mul(x = var_8700_cast_fp16, y = var_8713_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114574208))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114836416))))[name = string("tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_19_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8723_cast_fp16 = abs(x = linear_180_cast_fp16)[name = string("op_8723_cast_fp16")];
            bool var_8726 = const()[name = string("op_8726"), val = bool(true)];
            tensor<int32, [1]> var_8727_axes_0 = const()[name = string("op_8727_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8727_cast_fp16 = reduce_max(axes = var_8727_axes_0, keep_dims = var_8726, x = var_8723_cast_fp16)[name = string("op_8727_cast_fp16")];
            fp16 var_8728_to_fp16 = const()[name = string("op_8728_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_321_cast_fp16 = maximum(x = var_8727_cast_fp16, y = var_8728_to_fp16)[name = string("s_321_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_321_cast_fp16 = real_div(x = linear_180_cast_fp16, y = s_321_cast_fp16)[name = string("y_321_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8731_cast_fp16 = mul(x = y_321_cast_fp16, y = y_321_cast_fp16)[name = string("op_8731_cast_fp16")];
            tensor<int32, [1]> var_8736_axes_0 = const()[name = string("op_8736_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8736_keep_dims_0 = const()[name = string("op_8736_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8736_cast_fp16 = reduce_mean(axes = var_8736_axes_0, keep_dims = var_8736_keep_dims_0, x = var_8731_cast_fp16)[name = string("op_8736_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8737_cast_fp16 = mul(x = s_321_cast_fp16, y = s_321_cast_fp16)[name = string("op_8737_cast_fp16")];
            fp32 var_8738_epsilon_0 = const()[name = string("op_8738_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8738_cast_fp16 = inverse(epsilon = var_8738_epsilon_0, x = var_8737_cast_fp16)[name = string("op_8738_cast_fp16")];
            fp16 var_8739_to_fp16 = const()[name = string("op_8739_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8740_cast_fp16 = mul(x = var_8738_cast_fp16, y = var_8739_to_fp16)[name = string("op_8740_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8742_cast_fp16 = add(x = var_8736_cast_fp16, y = var_8740_cast_fp16)[name = string("op_8742_cast_fp16")];
            fp32 var_8743_epsilon_0 = const()[name = string("op_8743_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8743_cast_fp16 = rsqrt(epsilon = var_8743_epsilon_0, x = var_8742_cast_fp16)[name = string("op_8743_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_281_cast_fp16 = mul(x = y_321_cast_fp16, y = var_8743_cast_fp16)[name = string("out_281_cast_fp16")];
            tensor<fp16, [512]> tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114837504)))];
            tensor<fp16, [1, 256, 512]> var_8745_cast_fp16 = mul(x = out_281_cast_fp16, y = tm_layers_19_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_8745_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_39_cast_fp16 = add(x = input_237_cast_fp16, y = var_8745_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<fp16, [1]> tm_layers_19_layer_scalar_to_fp16 = const()[name = string("tm_layers_19_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.fp-1])];
            tensor<fp16, [1, 256, 512]> x_321_cast_fp16 = mul(x = h_39_cast_fp16, y = tm_layers_19_layer_scalar_to_fp16)[name = string("x_321_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8749_cast_fp16 = abs(x = x_321_cast_fp16)[name = string("op_8749_cast_fp16")];
            bool var_8752 = const()[name = string("op_8752"), val = bool(true)];
            tensor<int32, [1]> var_8753_axes_0 = const()[name = string("op_8753_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8753_cast_fp16 = reduce_max(axes = var_8753_axes_0, keep_dims = var_8752, x = var_8749_cast_fp16)[name = string("op_8753_cast_fp16")];
            fp16 var_8754_to_fp16 = const()[name = string("op_8754_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_323_cast_fp16 = maximum(x = var_8753_cast_fp16, y = var_8754_to_fp16)[name = string("s_323_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_323_cast_fp16 = real_div(x = x_321_cast_fp16, y = s_323_cast_fp16)[name = string("y_323_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8757_cast_fp16 = mul(x = y_323_cast_fp16, y = y_323_cast_fp16)[name = string("op_8757_cast_fp16")];
            tensor<int32, [1]> var_8762_axes_0 = const()[name = string("op_8762_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8762_keep_dims_0 = const()[name = string("op_8762_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8762_cast_fp16 = reduce_mean(axes = var_8762_axes_0, keep_dims = var_8762_keep_dims_0, x = var_8757_cast_fp16)[name = string("op_8762_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8763_cast_fp16 = mul(x = s_323_cast_fp16, y = s_323_cast_fp16)[name = string("op_8763_cast_fp16")];
            fp32 var_8764_epsilon_0 = const()[name = string("op_8764_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8764_cast_fp16 = inverse(epsilon = var_8764_epsilon_0, x = var_8763_cast_fp16)[name = string("op_8764_cast_fp16")];
            fp16 var_8765_to_fp16 = const()[name = string("op_8765_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8766_cast_fp16 = mul(x = var_8764_cast_fp16, y = var_8765_to_fp16)[name = string("op_8766_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8768_cast_fp16 = add(x = var_8762_cast_fp16, y = var_8766_cast_fp16)[name = string("op_8768_cast_fp16")];
            fp32 var_8769_epsilon_0 = const()[name = string("op_8769_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8769_cast_fp16 = rsqrt(epsilon = var_8769_epsilon_0, x = var_8768_cast_fp16)[name = string("op_8769_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_283_cast_fp16 = mul(x = y_323_cast_fp16, y = var_8769_cast_fp16)[name = string("out_283_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114838592)))];
            tensor<fp16, [1, 256, 512]> input_241_cast_fp16 = mul(x = out_283_cast_fp16, y = tm_layers_20_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114839680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115364032))))[name = string("tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<int32, [4]> var_8779 = const()[name = string("op_8779"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_8780_cast_fp16 = reshape(shape = var_8779, x = linear_181_cast_fp16)[name = string("op_8780_cast_fp16")];
            tensor<int32, [4]> x_323_perm_0 = const()[name = string("x_323_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115366144))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115628352))))[name = string("tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_8791 = const()[name = string("op_8791"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8792_cast_fp16 = reshape(shape = var_8791, x = linear_182_cast_fp16)[name = string("op_8792_cast_fp16")];
            tensor<int32, [4]> x_325_perm_0 = const()[name = string("x_325_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115629440))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115891648))))[name = string("tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_8803 = const()[name = string("op_8803"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_8804_cast_fp16 = reshape(shape = var_8803, x = linear_183_cast_fp16)[name = string("op_8804_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_323_cast_fp16 = transpose(perm = x_323_perm_0, x = var_8780_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 4, 256, 256]> var_8808_cast_fp16 = abs(x = x_323_cast_fp16)[name = string("op_8808_cast_fp16")];
            bool var_8811 = const()[name = string("op_8811"), val = bool(true)];
            tensor<int32, [1]> var_8812_axes_0 = const()[name = string("op_8812_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_8812_cast_fp16 = reduce_max(axes = var_8812_axes_0, keep_dims = var_8811, x = var_8808_cast_fp16)[name = string("op_8812_cast_fp16")];
            fp16 var_8813_to_fp16 = const()[name = string("op_8813_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_325_cast_fp16 = maximum(x = var_8812_cast_fp16, y = var_8813_to_fp16)[name = string("s_325_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_325_cast_fp16 = real_div(x = x_323_cast_fp16, y = s_325_cast_fp16)[name = string("y_325_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8816_cast_fp16 = mul(x = y_325_cast_fp16, y = y_325_cast_fp16)[name = string("op_8816_cast_fp16")];
            tensor<int32, [1]> var_8821_axes_0 = const()[name = string("op_8821_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8821_keep_dims_0 = const()[name = string("op_8821_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_8821_cast_fp16 = reduce_mean(axes = var_8821_axes_0, keep_dims = var_8821_keep_dims_0, x = var_8816_cast_fp16)[name = string("op_8821_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8822_cast_fp16 = mul(x = s_325_cast_fp16, y = s_325_cast_fp16)[name = string("op_8822_cast_fp16")];
            fp32 var_8823_epsilon_0 = const()[name = string("op_8823_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_8823_cast_fp16 = inverse(epsilon = var_8823_epsilon_0, x = var_8822_cast_fp16)[name = string("op_8823_cast_fp16")];
            fp16 var_8824_to_fp16 = const()[name = string("op_8824_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_8825_cast_fp16 = mul(x = var_8823_cast_fp16, y = var_8824_to_fp16)[name = string("op_8825_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_8827_cast_fp16 = add(x = var_8821_cast_fp16, y = var_8825_cast_fp16)[name = string("op_8827_cast_fp16")];
            fp32 var_8828_epsilon_0 = const()[name = string("op_8828_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_8828_cast_fp16 = rsqrt(epsilon = var_8828_epsilon_0, x = var_8827_cast_fp16)[name = string("op_8828_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_285_cast_fp16 = mul(x = y_325_cast_fp16, y = var_8828_cast_fp16)[name = string("out_285_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_81_cast_fp16 = mul(x = out_285_cast_fp16, y = tm_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("q_81_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_325_cast_fp16 = transpose(perm = x_325_perm_0, x = var_8792_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 256, 256]> var_8831_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_8831_cast_fp16")];
            bool var_8834 = const()[name = string("op_8834"), val = bool(true)];
            tensor<int32, [1]> var_8835_axes_0 = const()[name = string("op_8835_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8835_cast_fp16 = reduce_max(axes = var_8835_axes_0, keep_dims = var_8834, x = var_8831_cast_fp16)[name = string("op_8835_cast_fp16")];
            fp16 var_8836_to_fp16 = const()[name = string("op_8836_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_327_cast_fp16 = maximum(x = var_8835_cast_fp16, y = var_8836_to_fp16)[name = string("s_327_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_327_cast_fp16 = real_div(x = x_325_cast_fp16, y = s_327_cast_fp16)[name = string("y_327_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8839_cast_fp16 = mul(x = y_327_cast_fp16, y = y_327_cast_fp16)[name = string("op_8839_cast_fp16")];
            tensor<int32, [1]> var_8844_axes_0 = const()[name = string("op_8844_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8844_keep_dims_0 = const()[name = string("op_8844_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8844_cast_fp16 = reduce_mean(axes = var_8844_axes_0, keep_dims = var_8844_keep_dims_0, x = var_8839_cast_fp16)[name = string("op_8844_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8845_cast_fp16 = mul(x = s_327_cast_fp16, y = s_327_cast_fp16)[name = string("op_8845_cast_fp16")];
            fp32 var_8846_epsilon_0 = const()[name = string("op_8846_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8846_cast_fp16 = inverse(epsilon = var_8846_epsilon_0, x = var_8845_cast_fp16)[name = string("op_8846_cast_fp16")];
            fp16 var_8847_to_fp16 = const()[name = string("op_8847_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8848_cast_fp16 = mul(x = var_8846_cast_fp16, y = var_8847_to_fp16)[name = string("op_8848_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8850_cast_fp16 = add(x = var_8844_cast_fp16, y = var_8848_cast_fp16)[name = string("op_8850_cast_fp16")];
            fp32 var_8851_epsilon_0 = const()[name = string("op_8851_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8851_cast_fp16 = rsqrt(epsilon = var_8851_epsilon_0, x = var_8850_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_287_cast_fp16 = mul(x = y_327_cast_fp16, y = var_8851_cast_fp16)[name = string("out_287_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_121_cast_fp16 = mul(x = out_287_cast_fp16, y = tm_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("k_121_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_8804_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 256, 256]> var_8854_cast_fp16 = abs(x = x_327_cast_fp16)[name = string("op_8854_cast_fp16")];
            bool var_8857 = const()[name = string("op_8857"), val = bool(true)];
            tensor<int32, [1]> var_8858_axes_0 = const()[name = string("op_8858_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_8858_cast_fp16 = reduce_max(axes = var_8858_axes_0, keep_dims = var_8857, x = var_8854_cast_fp16)[name = string("op_8858_cast_fp16")];
            fp16 var_8859_to_fp16 = const()[name = string("op_8859_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_329_cast_fp16 = maximum(x = var_8858_cast_fp16, y = var_8859_to_fp16)[name = string("s_329_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_329_cast_fp16 = real_div(x = x_327_cast_fp16, y = s_329_cast_fp16)[name = string("y_329_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8862_cast_fp16 = mul(x = y_329_cast_fp16, y = y_329_cast_fp16)[name = string("op_8862_cast_fp16")];
            tensor<int32, [1]> var_8867_axes_0 = const()[name = string("op_8867_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8867_keep_dims_0 = const()[name = string("op_8867_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_8867_cast_fp16 = reduce_mean(axes = var_8867_axes_0, keep_dims = var_8867_keep_dims_0, x = var_8862_cast_fp16)[name = string("op_8867_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8868_cast_fp16 = mul(x = s_329_cast_fp16, y = s_329_cast_fp16)[name = string("op_8868_cast_fp16")];
            fp32 var_8869_epsilon_0 = const()[name = string("op_8869_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_8869_cast_fp16 = inverse(epsilon = var_8869_epsilon_0, x = var_8868_cast_fp16)[name = string("op_8869_cast_fp16")];
            fp16 var_8870_to_fp16 = const()[name = string("op_8870_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_8871_cast_fp16 = mul(x = var_8869_cast_fp16, y = var_8870_to_fp16)[name = string("op_8871_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_8873_cast_fp16 = add(x = var_8867_cast_fp16, y = var_8871_cast_fp16)[name = string("op_8873_cast_fp16")];
            fp32 var_8874_epsilon_0 = const()[name = string("op_8874_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_8874_cast_fp16 = rsqrt(epsilon = var_8874_epsilon_0, x = var_8873_cast_fp16)[name = string("op_8874_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_81_cast_fp16 = mul(x = y_329_cast_fp16, y = var_8874_cast_fp16)[name = string("v_81_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8876_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8876_cast_fp16")];
            tensor<int32, [4]> var_8881_begin_0 = const()[name = string("op_8881_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8881_end_0 = const()[name = string("op_8881_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_8881_end_mask_0 = const()[name = string("op_8881_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_8881_cast_fp16 = slice_by_index(begin = var_8881_begin_0, end = var_8881_end_0, end_mask = var_8881_end_mask_0, x = q_81_cast_fp16)[name = string("op_8881_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_8882_cast_fp16 = mul(x = var_8881_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_8882_cast_fp16")];
            tensor<int32, [4]> var_8887_begin_0 = const()[name = string("op_8887_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8887_end_0 = const()[name = string("op_8887_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_8887_end_mask_0 = const()[name = string("op_8887_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_8887_cast_fp16 = slice_by_index(begin = var_8887_begin_0, end = var_8887_end_0, end_mask = var_8887_end_mask_0, x = q_81_cast_fp16)[name = string("op_8887_cast_fp16")];
            int32 var_8889 = const()[name = string("op_8889"), val = int32(-1)];
            bool var_8890_interleave_0 = const()[name = string("op_8890_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_8890_cast_fp16 = concat(axis = var_8889, interleave = var_8890_interleave_0, values = (var_8882_cast_fp16, var_8887_cast_fp16))[name = string("op_8890_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8891_cast_fp16 = mul(x = var_8890_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8891_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_83_cast_fp16 = add(x = var_8876_cast_fp16, y = var_8891_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8894_cast_fp16 = mul(x = k_121_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8894_cast_fp16")];
            tensor<int32, [4]> var_8899_begin_0 = const()[name = string("op_8899_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_8899_end_0 = const()[name = string("op_8899_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_8899_end_mask_0 = const()[name = string("op_8899_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_8899_cast_fp16 = slice_by_index(begin = var_8899_begin_0, end = var_8899_end_0, end_mask = var_8899_end_mask_0, x = k_121_cast_fp16)[name = string("op_8899_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_8900_cast_fp16 = mul(x = var_8899_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_8900_cast_fp16")];
            tensor<int32, [4]> var_8905_begin_0 = const()[name = string("op_8905_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8905_end_0 = const()[name = string("op_8905_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_8905_end_mask_0 = const()[name = string("op_8905_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_8905_cast_fp16 = slice_by_index(begin = var_8905_begin_0, end = var_8905_end_0, end_mask = var_8905_end_mask_0, x = k_121_cast_fp16)[name = string("op_8905_cast_fp16")];
            int32 var_8907 = const()[name = string("op_8907"), val = int32(-1)];
            bool var_8908_interleave_0 = const()[name = string("op_8908_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_8908_cast_fp16 = concat(axis = var_8907, interleave = var_8908_interleave_0, values = (var_8900_cast_fp16, var_8905_cast_fp16))[name = string("op_8908_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_8909_cast_fp16 = mul(x = var_8908_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8909_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_123_cast_fp16 = add(x = var_8894_cast_fp16, y = var_8909_cast_fp16)[name = string("k_123_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_40_reps_0 = const()[name = string("tile_40_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = k_123_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 1, 256, 256]> tile_40_cast_fp16 = tile(reps = tile_40_reps_0, x = transpose_80_cast_fp16)[name = string("tile_40_cast_fp16")];
            tensor<int32, [5]> concat_80 = const()[name = string("concat_80"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_80_cast_fp16 = reshape(shape = concat_80, x = tile_40_cast_fp16)[name = string("reshape_80_cast_fp16")];
            tensor<int32, [5]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_81 = const()[name = string("concat_81"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = reshape_80_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 1, 256, 256]> reshape_81_cast_fp16 = reshape(shape = concat_81, x = transpose_81_cast_fp16)[name = string("reshape_81_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_41_reps_0 = const()[name = string("tile_41_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_82_cast_fp16 = transpose(perm = transpose_82_perm_0, x = v_81_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 1, 256, 256]> tile_41_cast_fp16 = tile(reps = tile_41_reps_0, x = transpose_82_cast_fp16)[name = string("tile_41_cast_fp16")];
            tensor<int32, [5]> concat_82 = const()[name = string("concat_82"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_82_cast_fp16 = reshape(shape = concat_82, x = tile_41_cast_fp16)[name = string("reshape_82_cast_fp16")];
            tensor<int32, [5]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_83 = const()[name = string("concat_83"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = reshape_82_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 1, 256, 256]> reshape_83_cast_fp16 = reshape(shape = concat_83, x = transpose_83_cast_fp16)[name = string("reshape_83_cast_fp16")];
            tensor<int32, [4]> v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_8923_transpose_x_0 = const()[name = string("op_8923_transpose_x_0"), val = bool(false)];
            bool var_8923_transpose_y_0 = const()[name = string("op_8923_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = reshape_81_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 4, 256, 256]> var_8923_cast_fp16 = matmul(transpose_x = var_8923_transpose_x_0, transpose_y = var_8923_transpose_y_0, x = q_83_cast_fp16, y = transpose_116_cast_fp16)[name = string("op_8923_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_8925_cast_fp16 = add(x = var_8923_cast_fp16, y = attention_bias)[name = string("op_8925_cast_fp16")];
            int32 var_8926 = const()[name = string("op_8926"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_8928_cast_fp16 = softmax(axis = var_8926, x = var_8925_cast_fp16)[name = string("op_8928_cast_fp16")];
            bool var_8929_transpose_x_0 = const()[name = string("op_8929_transpose_x_0"), val = bool(false)];
            bool var_8929_transpose_y_0 = const()[name = string("op_8929_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_83_cast_fp16 = transpose(perm = v_83_perm_0, x = reshape_83_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 4, 256, 256]> var_8929_cast_fp16 = matmul(transpose_x = var_8929_transpose_x_0, transpose_y = var_8929_transpose_y_0, x = var_8928_cast_fp16, y = v_83_cast_fp16)[name = string("op_8929_cast_fp16")];
            tensor<int32, [4]> var_8932_perm_0 = const()[name = string("op_8932_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8936 = const()[name = string("op_8936"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_8932_cast_fp16 = transpose(perm = var_8932_perm_0, x = var_8929_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 256, 1024]> input_243_cast_fp16 = reshape(shape = var_8936, x = var_8932_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115892736))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116417088))))[name = string("tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8941_cast_fp16 = abs(x = linear_184_cast_fp16)[name = string("op_8941_cast_fp16")];
            bool var_8944 = const()[name = string("op_8944"), val = bool(true)];
            tensor<int32, [1]> var_8945_axes_0 = const()[name = string("op_8945_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8945_cast_fp16 = reduce_max(axes = var_8945_axes_0, keep_dims = var_8944, x = var_8941_cast_fp16)[name = string("op_8945_cast_fp16")];
            fp16 var_8946_to_fp16 = const()[name = string("op_8946_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_331_cast_fp16 = maximum(x = var_8945_cast_fp16, y = var_8946_to_fp16)[name = string("s_331_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_331_cast_fp16 = real_div(x = linear_184_cast_fp16, y = s_331_cast_fp16)[name = string("y_331_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8949_cast_fp16 = mul(x = y_331_cast_fp16, y = y_331_cast_fp16)[name = string("op_8949_cast_fp16")];
            tensor<int32, [1]> var_8954_axes_0 = const()[name = string("op_8954_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8954_keep_dims_0 = const()[name = string("op_8954_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8954_cast_fp16 = reduce_mean(axes = var_8954_axes_0, keep_dims = var_8954_keep_dims_0, x = var_8949_cast_fp16)[name = string("op_8954_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8955_cast_fp16 = mul(x = s_331_cast_fp16, y = s_331_cast_fp16)[name = string("op_8955_cast_fp16")];
            fp32 var_8956_epsilon_0 = const()[name = string("op_8956_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8956_cast_fp16 = inverse(epsilon = var_8956_epsilon_0, x = var_8955_cast_fp16)[name = string("op_8956_cast_fp16")];
            fp16 var_8957_to_fp16 = const()[name = string("op_8957_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8958_cast_fp16 = mul(x = var_8956_cast_fp16, y = var_8957_to_fp16)[name = string("op_8958_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8960_cast_fp16 = add(x = var_8954_cast_fp16, y = var_8958_cast_fp16)[name = string("op_8960_cast_fp16")];
            fp32 var_8961_epsilon_0 = const()[name = string("op_8961_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8961_cast_fp16 = rsqrt(epsilon = var_8961_epsilon_0, x = var_8960_cast_fp16)[name = string("op_8961_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_289_cast_fp16 = mul(x = y_331_cast_fp16, y = var_8961_cast_fp16)[name = string("out_289_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116418176)))];
            tensor<fp16, [1, 256, 512]> var_8963_cast_fp16 = mul(x = out_289_cast_fp16, y = tm_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("op_8963_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_331_cast_fp16 = add(x = x_321_cast_fp16, y = var_8963_cast_fp16)[name = string("x_331_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8966_cast_fp16 = abs(x = x_331_cast_fp16)[name = string("op_8966_cast_fp16")];
            bool var_8969 = const()[name = string("op_8969"), val = bool(true)];
            tensor<int32, [1]> var_8970_axes_0 = const()[name = string("op_8970_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_8970_cast_fp16 = reduce_max(axes = var_8970_axes_0, keep_dims = var_8969, x = var_8966_cast_fp16)[name = string("op_8970_cast_fp16")];
            fp16 var_8971_to_fp16 = const()[name = string("op_8971_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_333_cast_fp16 = maximum(x = var_8970_cast_fp16, y = var_8971_to_fp16)[name = string("s_333_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_333_cast_fp16 = real_div(x = x_331_cast_fp16, y = s_333_cast_fp16)[name = string("y_333_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_8974_cast_fp16 = mul(x = y_333_cast_fp16, y = y_333_cast_fp16)[name = string("op_8974_cast_fp16")];
            tensor<int32, [1]> var_8979_axes_0 = const()[name = string("op_8979_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8979_keep_dims_0 = const()[name = string("op_8979_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_8979_cast_fp16 = reduce_mean(axes = var_8979_axes_0, keep_dims = var_8979_keep_dims_0, x = var_8974_cast_fp16)[name = string("op_8979_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8980_cast_fp16 = mul(x = s_333_cast_fp16, y = s_333_cast_fp16)[name = string("op_8980_cast_fp16")];
            fp32 var_8981_epsilon_0 = const()[name = string("op_8981_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_8981_cast_fp16 = inverse(epsilon = var_8981_epsilon_0, x = var_8980_cast_fp16)[name = string("op_8981_cast_fp16")];
            fp16 var_8982_to_fp16 = const()[name = string("op_8982_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_8983_cast_fp16 = mul(x = var_8981_cast_fp16, y = var_8982_to_fp16)[name = string("op_8983_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_8985_cast_fp16 = add(x = var_8979_cast_fp16, y = var_8983_cast_fp16)[name = string("op_8985_cast_fp16")];
            fp32 var_8986_epsilon_0 = const()[name = string("op_8986_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_8986_cast_fp16 = rsqrt(epsilon = var_8986_epsilon_0, x = var_8985_cast_fp16)[name = string("op_8986_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_291_cast_fp16 = mul(x = y_333_cast_fp16, y = var_8986_cast_fp16)[name = string("out_291_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116419264)))];
            tensor<fp16, [1, 256, 512]> input_245_cast_fp16 = mul(x = out_291_cast_fp16, y = tm_layers_20_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116420352))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117468992))))[name = string("tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_185_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_185_cast_fp16")];
            string var_8993_mode_0 = const()[name = string("op_8993_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_8993_cast_fp16 = gelu(mode = var_8993_mode_0, x = linear_185_cast_fp16)[name = string("op_8993_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117473152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118521792))))[name = string("tm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_186_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_247_cast_fp16 = mul(x = var_8993_cast_fp16, y = linear_186_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118525952))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119574592))))[name = string("tm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9001_cast_fp16 = abs(x = linear_187_cast_fp16)[name = string("op_9001_cast_fp16")];
            bool var_9004 = const()[name = string("op_9004"), val = bool(true)];
            tensor<int32, [1]> var_9005_axes_0 = const()[name = string("op_9005_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9005_cast_fp16 = reduce_max(axes = var_9005_axes_0, keep_dims = var_9004, x = var_9001_cast_fp16)[name = string("op_9005_cast_fp16")];
            fp16 var_9006_to_fp16 = const()[name = string("op_9006_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_335_cast_fp16 = maximum(x = var_9005_cast_fp16, y = var_9006_to_fp16)[name = string("s_335_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_335_cast_fp16 = real_div(x = linear_187_cast_fp16, y = s_335_cast_fp16)[name = string("y_335_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9009_cast_fp16 = mul(x = y_335_cast_fp16, y = y_335_cast_fp16)[name = string("op_9009_cast_fp16")];
            tensor<int32, [1]> var_9014_axes_0 = const()[name = string("op_9014_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9014_keep_dims_0 = const()[name = string("op_9014_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9014_cast_fp16 = reduce_mean(axes = var_9014_axes_0, keep_dims = var_9014_keep_dims_0, x = var_9009_cast_fp16)[name = string("op_9014_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9015_cast_fp16 = mul(x = s_335_cast_fp16, y = s_335_cast_fp16)[name = string("op_9015_cast_fp16")];
            fp32 var_9016_epsilon_0 = const()[name = string("op_9016_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9016_cast_fp16 = inverse(epsilon = var_9016_epsilon_0, x = var_9015_cast_fp16)[name = string("op_9016_cast_fp16")];
            fp16 var_9017_to_fp16 = const()[name = string("op_9017_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9018_cast_fp16 = mul(x = var_9016_cast_fp16, y = var_9017_to_fp16)[name = string("op_9018_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9020_cast_fp16 = add(x = var_9014_cast_fp16, y = var_9018_cast_fp16)[name = string("op_9020_cast_fp16")];
            fp32 var_9021_epsilon_0 = const()[name = string("op_9021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9021_cast_fp16 = rsqrt(epsilon = var_9021_epsilon_0, x = var_9020_cast_fp16)[name = string("op_9021_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_293_cast_fp16 = mul(x = y_335_cast_fp16, y = var_9021_cast_fp16)[name = string("out_293_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_20_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119575680)))];
            tensor<fp16, [1, 256, 512]> var_9023_cast_fp16 = mul(x = out_293_cast_fp16, y = tm_layers_20_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_249_cast_fp16 = add(x = x_331_cast_fp16, y = var_9023_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119576768))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119838976))))[name = string("tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_188_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_188_cast_fp16")];
            string var_9030_mode_0 = const()[name = string("op_9030_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_9030_cast_fp16 = gelu(mode = var_9030_mode_0, x = linear_188_cast_fp16)[name = string("op_9030_cast_fp16")];
            tensor<int32, [4]> var_9043_begin_0 = const()[name = string("op_9043_begin_0"), val = tensor<int32, [4]>([0, 0, 20, 0])];
            tensor<int32, [4]> var_9043_end_0 = const()[name = string("op_9043_end_0"), val = tensor<int32, [4]>([1, 256, 21, 512])];
            tensor<bool, [4]> var_9043_end_mask_0 = const()[name = string("op_9043_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9043_squeeze_mask_0 = const()[name = string("op_9043_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_9043_cast_fp16 = slice_by_index(begin = var_9043_begin_0, end = var_9043_end_0, end_mask = var_9043_end_mask_0, squeeze_mask = var_9043_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_9043_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_251_cast_fp16 = mul(x = var_9030_cast_fp16, y = var_9043_cast_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119840064))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120102272))))[name = string("tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_20_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9053_cast_fp16 = abs(x = linear_189_cast_fp16)[name = string("op_9053_cast_fp16")];
            bool var_9056 = const()[name = string("op_9056"), val = bool(true)];
            tensor<int32, [1]> var_9057_axes_0 = const()[name = string("op_9057_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9057_cast_fp16 = reduce_max(axes = var_9057_axes_0, keep_dims = var_9056, x = var_9053_cast_fp16)[name = string("op_9057_cast_fp16")];
            fp16 var_9058_to_fp16 = const()[name = string("op_9058_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_337_cast_fp16 = maximum(x = var_9057_cast_fp16, y = var_9058_to_fp16)[name = string("s_337_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_337_cast_fp16 = real_div(x = linear_189_cast_fp16, y = s_337_cast_fp16)[name = string("y_337_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9061_cast_fp16 = mul(x = y_337_cast_fp16, y = y_337_cast_fp16)[name = string("op_9061_cast_fp16")];
            tensor<int32, [1]> var_9066_axes_0 = const()[name = string("op_9066_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9066_keep_dims_0 = const()[name = string("op_9066_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9066_cast_fp16 = reduce_mean(axes = var_9066_axes_0, keep_dims = var_9066_keep_dims_0, x = var_9061_cast_fp16)[name = string("op_9066_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9067_cast_fp16 = mul(x = s_337_cast_fp16, y = s_337_cast_fp16)[name = string("op_9067_cast_fp16")];
            fp32 var_9068_epsilon_0 = const()[name = string("op_9068_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9068_cast_fp16 = inverse(epsilon = var_9068_epsilon_0, x = var_9067_cast_fp16)[name = string("op_9068_cast_fp16")];
            fp16 var_9069_to_fp16 = const()[name = string("op_9069_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9070_cast_fp16 = mul(x = var_9068_cast_fp16, y = var_9069_to_fp16)[name = string("op_9070_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9072_cast_fp16 = add(x = var_9066_cast_fp16, y = var_9070_cast_fp16)[name = string("op_9072_cast_fp16")];
            fp32 var_9073_epsilon_0 = const()[name = string("op_9073_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9073_cast_fp16 = rsqrt(epsilon = var_9073_epsilon_0, x = var_9072_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_295_cast_fp16 = mul(x = y_337_cast_fp16, y = var_9073_cast_fp16)[name = string("out_295_cast_fp16")];
            tensor<fp16, [512]> tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120103360)))];
            tensor<fp16, [1, 256, 512]> var_9075_cast_fp16 = mul(x = out_295_cast_fp16, y = tm_layers_20_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9075_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_41_cast_fp16 = add(x = input_249_cast_fp16, y = var_9075_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<fp16, [1]> tm_layers_20_layer_scalar_to_fp16 = const()[name = string("tm_layers_20_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.ecp-1])];
            tensor<fp16, [1, 256, 512]> x_337_cast_fp16 = mul(x = h_41_cast_fp16, y = tm_layers_20_layer_scalar_to_fp16)[name = string("x_337_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9079_cast_fp16 = abs(x = x_337_cast_fp16)[name = string("op_9079_cast_fp16")];
            bool var_9082 = const()[name = string("op_9082"), val = bool(true)];
            tensor<int32, [1]> var_9083_axes_0 = const()[name = string("op_9083_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9083_cast_fp16 = reduce_max(axes = var_9083_axes_0, keep_dims = var_9082, x = var_9079_cast_fp16)[name = string("op_9083_cast_fp16")];
            fp16 var_9084_to_fp16 = const()[name = string("op_9084_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_339_cast_fp16 = maximum(x = var_9083_cast_fp16, y = var_9084_to_fp16)[name = string("s_339_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_339_cast_fp16 = real_div(x = x_337_cast_fp16, y = s_339_cast_fp16)[name = string("y_339_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9087_cast_fp16 = mul(x = y_339_cast_fp16, y = y_339_cast_fp16)[name = string("op_9087_cast_fp16")];
            tensor<int32, [1]> var_9092_axes_0 = const()[name = string("op_9092_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9092_keep_dims_0 = const()[name = string("op_9092_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9092_cast_fp16 = reduce_mean(axes = var_9092_axes_0, keep_dims = var_9092_keep_dims_0, x = var_9087_cast_fp16)[name = string("op_9092_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9093_cast_fp16 = mul(x = s_339_cast_fp16, y = s_339_cast_fp16)[name = string("op_9093_cast_fp16")];
            fp32 var_9094_epsilon_0 = const()[name = string("op_9094_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9094_cast_fp16 = inverse(epsilon = var_9094_epsilon_0, x = var_9093_cast_fp16)[name = string("op_9094_cast_fp16")];
            fp16 var_9095_to_fp16 = const()[name = string("op_9095_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9096_cast_fp16 = mul(x = var_9094_cast_fp16, y = var_9095_to_fp16)[name = string("op_9096_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9098_cast_fp16 = add(x = var_9092_cast_fp16, y = var_9096_cast_fp16)[name = string("op_9098_cast_fp16")];
            fp32 var_9099_epsilon_0 = const()[name = string("op_9099_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9099_cast_fp16 = rsqrt(epsilon = var_9099_epsilon_0, x = var_9098_cast_fp16)[name = string("op_9099_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_297_cast_fp16 = mul(x = y_339_cast_fp16, y = var_9099_cast_fp16)[name = string("out_297_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120104448)))];
            tensor<fp16, [1, 256, 512]> input_253_cast_fp16 = mul(x = out_297_cast_fp16, y = tm_layers_21_input_layernorm_weight_to_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120105536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120629888))))[name = string("tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9109 = const()[name = string("op_9109"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_9110_cast_fp16 = reshape(shape = var_9109, x = linear_190_cast_fp16)[name = string("op_9110_cast_fp16")];
            tensor<int32, [4]> x_339_perm_0 = const()[name = string("x_339_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120632000))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120894208))))[name = string("tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9121 = const()[name = string("op_9121"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_9122_cast_fp16 = reshape(shape = var_9121, x = linear_191_cast_fp16)[name = string("op_9122_cast_fp16")];
            tensor<int32, [4]> x_341_perm_0 = const()[name = string("x_341_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120895296))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121157504))))[name = string("tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_9133 = const()[name = string("op_9133"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_9134_cast_fp16 = reshape(shape = var_9133, x = linear_192_cast_fp16)[name = string("op_9134_cast_fp16")];
            tensor<int32, [4]> x_343_perm_0 = const()[name = string("x_343_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_339_cast_fp16 = transpose(perm = x_339_perm_0, x = var_9110_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 4, 256, 256]> var_9138_cast_fp16 = abs(x = x_339_cast_fp16)[name = string("op_9138_cast_fp16")];
            bool var_9141 = const()[name = string("op_9141"), val = bool(true)];
            tensor<int32, [1]> var_9142_axes_0 = const()[name = string("op_9142_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_9142_cast_fp16 = reduce_max(axes = var_9142_axes_0, keep_dims = var_9141, x = var_9138_cast_fp16)[name = string("op_9142_cast_fp16")];
            fp16 var_9143_to_fp16 = const()[name = string("op_9143_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_341_cast_fp16 = maximum(x = var_9142_cast_fp16, y = var_9143_to_fp16)[name = string("s_341_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_341_cast_fp16 = real_div(x = x_339_cast_fp16, y = s_341_cast_fp16)[name = string("y_341_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9146_cast_fp16 = mul(x = y_341_cast_fp16, y = y_341_cast_fp16)[name = string("op_9146_cast_fp16")];
            tensor<int32, [1]> var_9151_axes_0 = const()[name = string("op_9151_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9151_keep_dims_0 = const()[name = string("op_9151_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_9151_cast_fp16 = reduce_mean(axes = var_9151_axes_0, keep_dims = var_9151_keep_dims_0, x = var_9146_cast_fp16)[name = string("op_9151_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9152_cast_fp16 = mul(x = s_341_cast_fp16, y = s_341_cast_fp16)[name = string("op_9152_cast_fp16")];
            fp32 var_9153_epsilon_0 = const()[name = string("op_9153_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_9153_cast_fp16 = inverse(epsilon = var_9153_epsilon_0, x = var_9152_cast_fp16)[name = string("op_9153_cast_fp16")];
            fp16 var_9154_to_fp16 = const()[name = string("op_9154_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_9155_cast_fp16 = mul(x = var_9153_cast_fp16, y = var_9154_to_fp16)[name = string("op_9155_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9157_cast_fp16 = add(x = var_9151_cast_fp16, y = var_9155_cast_fp16)[name = string("op_9157_cast_fp16")];
            fp32 var_9158_epsilon_0 = const()[name = string("op_9158_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_9158_cast_fp16 = rsqrt(epsilon = var_9158_epsilon_0, x = var_9157_cast_fp16)[name = string("op_9158_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_299_cast_fp16 = mul(x = y_341_cast_fp16, y = var_9158_cast_fp16)[name = string("out_299_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_85_cast_fp16 = mul(x = out_299_cast_fp16, y = tm_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("q_85_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_341_cast_fp16 = transpose(perm = x_341_perm_0, x = var_9122_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2, 256, 256]> var_9161_cast_fp16 = abs(x = x_341_cast_fp16)[name = string("op_9161_cast_fp16")];
            bool var_9164 = const()[name = string("op_9164"), val = bool(true)];
            tensor<int32, [1]> var_9165_axes_0 = const()[name = string("op_9165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_9165_cast_fp16 = reduce_max(axes = var_9165_axes_0, keep_dims = var_9164, x = var_9161_cast_fp16)[name = string("op_9165_cast_fp16")];
            fp16 var_9166_to_fp16 = const()[name = string("op_9166_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_343_cast_fp16 = maximum(x = var_9165_cast_fp16, y = var_9166_to_fp16)[name = string("s_343_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_343_cast_fp16 = real_div(x = x_341_cast_fp16, y = s_343_cast_fp16)[name = string("y_343_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9169_cast_fp16 = mul(x = y_343_cast_fp16, y = y_343_cast_fp16)[name = string("op_9169_cast_fp16")];
            tensor<int32, [1]> var_9174_axes_0 = const()[name = string("op_9174_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9174_keep_dims_0 = const()[name = string("op_9174_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_9174_cast_fp16 = reduce_mean(axes = var_9174_axes_0, keep_dims = var_9174_keep_dims_0, x = var_9169_cast_fp16)[name = string("op_9174_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9175_cast_fp16 = mul(x = s_343_cast_fp16, y = s_343_cast_fp16)[name = string("op_9175_cast_fp16")];
            fp32 var_9176_epsilon_0 = const()[name = string("op_9176_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_9176_cast_fp16 = inverse(epsilon = var_9176_epsilon_0, x = var_9175_cast_fp16)[name = string("op_9176_cast_fp16")];
            fp16 var_9177_to_fp16 = const()[name = string("op_9177_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_9178_cast_fp16 = mul(x = var_9176_cast_fp16, y = var_9177_to_fp16)[name = string("op_9178_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9180_cast_fp16 = add(x = var_9174_cast_fp16, y = var_9178_cast_fp16)[name = string("op_9180_cast_fp16")];
            fp32 var_9181_epsilon_0 = const()[name = string("op_9181_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_9181_cast_fp16 = rsqrt(epsilon = var_9181_epsilon_0, x = var_9180_cast_fp16)[name = string("op_9181_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_301_cast_fp16 = mul(x = y_343_cast_fp16, y = var_9181_cast_fp16)[name = string("out_301_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_127_cast_fp16 = mul(x = out_301_cast_fp16, y = tm_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("k_127_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_343_cast_fp16 = transpose(perm = x_343_perm_0, x = var_9134_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 2, 256, 256]> var_9184_cast_fp16 = abs(x = x_343_cast_fp16)[name = string("op_9184_cast_fp16")];
            bool var_9187 = const()[name = string("op_9187"), val = bool(true)];
            tensor<int32, [1]> var_9188_axes_0 = const()[name = string("op_9188_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_9188_cast_fp16 = reduce_max(axes = var_9188_axes_0, keep_dims = var_9187, x = var_9184_cast_fp16)[name = string("op_9188_cast_fp16")];
            fp16 var_9189_to_fp16 = const()[name = string("op_9189_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_345_cast_fp16 = maximum(x = var_9188_cast_fp16, y = var_9189_to_fp16)[name = string("s_345_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_345_cast_fp16 = real_div(x = x_343_cast_fp16, y = s_345_cast_fp16)[name = string("y_345_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9192_cast_fp16 = mul(x = y_345_cast_fp16, y = y_345_cast_fp16)[name = string("op_9192_cast_fp16")];
            tensor<int32, [1]> var_9197_axes_0 = const()[name = string("op_9197_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9197_keep_dims_0 = const()[name = string("op_9197_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_9197_cast_fp16 = reduce_mean(axes = var_9197_axes_0, keep_dims = var_9197_keep_dims_0, x = var_9192_cast_fp16)[name = string("op_9197_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9198_cast_fp16 = mul(x = s_345_cast_fp16, y = s_345_cast_fp16)[name = string("op_9198_cast_fp16")];
            fp32 var_9199_epsilon_0 = const()[name = string("op_9199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_9199_cast_fp16 = inverse(epsilon = var_9199_epsilon_0, x = var_9198_cast_fp16)[name = string("op_9199_cast_fp16")];
            fp16 var_9200_to_fp16 = const()[name = string("op_9200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_9201_cast_fp16 = mul(x = var_9199_cast_fp16, y = var_9200_to_fp16)[name = string("op_9201_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9203_cast_fp16 = add(x = var_9197_cast_fp16, y = var_9201_cast_fp16)[name = string("op_9203_cast_fp16")];
            fp32 var_9204_epsilon_0 = const()[name = string("op_9204_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_9204_cast_fp16 = rsqrt(epsilon = var_9204_epsilon_0, x = var_9203_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_85_cast_fp16 = mul(x = y_345_cast_fp16, y = var_9204_cast_fp16)[name = string("v_85_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9206_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9206_cast_fp16")];
            tensor<int32, [4]> var_9211_begin_0 = const()[name = string("op_9211_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9211_end_0 = const()[name = string("op_9211_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_9211_end_mask_0 = const()[name = string("op_9211_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_9211_cast_fp16 = slice_by_index(begin = var_9211_begin_0, end = var_9211_end_0, end_mask = var_9211_end_mask_0, x = q_85_cast_fp16)[name = string("op_9211_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_9212_cast_fp16 = mul(x = var_9211_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_9212_cast_fp16")];
            tensor<int32, [4]> var_9217_begin_0 = const()[name = string("op_9217_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9217_end_0 = const()[name = string("op_9217_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_9217_end_mask_0 = const()[name = string("op_9217_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_9217_cast_fp16 = slice_by_index(begin = var_9217_begin_0, end = var_9217_end_0, end_mask = var_9217_end_mask_0, x = q_85_cast_fp16)[name = string("op_9217_cast_fp16")];
            int32 var_9219 = const()[name = string("op_9219"), val = int32(-1)];
            bool var_9220_interleave_0 = const()[name = string("op_9220_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_9220_cast_fp16 = concat(axis = var_9219, interleave = var_9220_interleave_0, values = (var_9212_cast_fp16, var_9217_cast_fp16))[name = string("op_9220_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9221_cast_fp16 = mul(x = var_9220_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9221_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_87_cast_fp16 = add(x = var_9206_cast_fp16, y = var_9221_cast_fp16)[name = string("q_87_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9224_cast_fp16 = mul(x = k_127_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9224_cast_fp16")];
            tensor<int32, [4]> var_9229_begin_0 = const()[name = string("op_9229_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9229_end_0 = const()[name = string("op_9229_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_9229_end_mask_0 = const()[name = string("op_9229_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_9229_cast_fp16 = slice_by_index(begin = var_9229_begin_0, end = var_9229_end_0, end_mask = var_9229_end_mask_0, x = k_127_cast_fp16)[name = string("op_9229_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_9230_cast_fp16 = mul(x = var_9229_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_9230_cast_fp16")];
            tensor<int32, [4]> var_9235_begin_0 = const()[name = string("op_9235_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9235_end_0 = const()[name = string("op_9235_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_9235_end_mask_0 = const()[name = string("op_9235_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_9235_cast_fp16 = slice_by_index(begin = var_9235_begin_0, end = var_9235_end_0, end_mask = var_9235_end_mask_0, x = k_127_cast_fp16)[name = string("op_9235_cast_fp16")];
            int32 var_9237 = const()[name = string("op_9237"), val = int32(-1)];
            bool var_9238_interleave_0 = const()[name = string("op_9238_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_9238_cast_fp16 = concat(axis = var_9237, interleave = var_9238_interleave_0, values = (var_9230_cast_fp16, var_9235_cast_fp16))[name = string("op_9238_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9239_cast_fp16 = mul(x = var_9238_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9239_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_129_cast_fp16 = add(x = var_9224_cast_fp16, y = var_9239_cast_fp16)[name = string("k_129_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_42_reps_0 = const()[name = string("tile_42_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = k_129_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 1, 256, 256]> tile_42_cast_fp16 = tile(reps = tile_42_reps_0, x = transpose_84_cast_fp16)[name = string("tile_42_cast_fp16")];
            tensor<int32, [5]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_84_cast_fp16 = reshape(shape = concat_84, x = tile_42_cast_fp16)[name = string("reshape_84_cast_fp16")];
            tensor<int32, [5]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = reshape_84_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 1, 256, 256]> reshape_85_cast_fp16 = reshape(shape = concat_85, x = transpose_85_cast_fp16)[name = string("reshape_85_cast_fp16")];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_43_reps_0 = const()[name = string("tile_43_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_86_cast_fp16 = transpose(perm = transpose_86_perm_0, x = v_85_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 1, 256, 256]> tile_43_cast_fp16 = tile(reps = tile_43_reps_0, x = transpose_86_cast_fp16)[name = string("tile_43_cast_fp16")];
            tensor<int32, [5]> concat_86 = const()[name = string("concat_86"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_86_cast_fp16 = reshape(shape = concat_86, x = tile_43_cast_fp16)[name = string("reshape_86_cast_fp16")];
            tensor<int32, [5]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_87_cast_fp16 = transpose(perm = transpose_87_perm_0, x = reshape_86_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 1, 256, 256]> reshape_87_cast_fp16 = reshape(shape = concat_87, x = transpose_87_cast_fp16)[name = string("reshape_87_cast_fp16")];
            tensor<int32, [4]> v_87_perm_0 = const()[name = string("v_87_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9253_transpose_x_0 = const()[name = string("op_9253_transpose_x_0"), val = bool(false)];
            bool var_9253_transpose_y_0 = const()[name = string("op_9253_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_117_cast_fp16 = transpose(perm = transpose_117_perm_0, x = reshape_85_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 4, 256, 256]> var_9253_cast_fp16 = matmul(transpose_x = var_9253_transpose_x_0, transpose_y = var_9253_transpose_y_0, x = q_87_cast_fp16, y = transpose_117_cast_fp16)[name = string("op_9253_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9255_cast_fp16 = add(x = var_9253_cast_fp16, y = attention_bias)[name = string("op_9255_cast_fp16")];
            int32 var_9256 = const()[name = string("op_9256"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_9258_cast_fp16 = softmax(axis = var_9256, x = var_9255_cast_fp16)[name = string("op_9258_cast_fp16")];
            bool var_9259_transpose_x_0 = const()[name = string("op_9259_transpose_x_0"), val = bool(false)];
            bool var_9259_transpose_y_0 = const()[name = string("op_9259_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_87_cast_fp16 = transpose(perm = v_87_perm_0, x = reshape_87_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 4, 256, 256]> var_9259_cast_fp16 = matmul(transpose_x = var_9259_transpose_x_0, transpose_y = var_9259_transpose_y_0, x = var_9258_cast_fp16, y = v_87_cast_fp16)[name = string("op_9259_cast_fp16")];
            tensor<int32, [4]> var_9262_perm_0 = const()[name = string("op_9262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9266 = const()[name = string("op_9266"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_9262_cast_fp16 = transpose(perm = var_9262_perm_0, x = var_9259_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 256, 1024]> input_255_cast_fp16 = reshape(shape = var_9266, x = var_9262_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121158592))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121682944))))[name = string("tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9271_cast_fp16 = abs(x = linear_193_cast_fp16)[name = string("op_9271_cast_fp16")];
            bool var_9274 = const()[name = string("op_9274"), val = bool(true)];
            tensor<int32, [1]> var_9275_axes_0 = const()[name = string("op_9275_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9275_cast_fp16 = reduce_max(axes = var_9275_axes_0, keep_dims = var_9274, x = var_9271_cast_fp16)[name = string("op_9275_cast_fp16")];
            fp16 var_9276_to_fp16 = const()[name = string("op_9276_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_347_cast_fp16 = maximum(x = var_9275_cast_fp16, y = var_9276_to_fp16)[name = string("s_347_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_347_cast_fp16 = real_div(x = linear_193_cast_fp16, y = s_347_cast_fp16)[name = string("y_347_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9279_cast_fp16 = mul(x = y_347_cast_fp16, y = y_347_cast_fp16)[name = string("op_9279_cast_fp16")];
            tensor<int32, [1]> var_9284_axes_0 = const()[name = string("op_9284_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9284_keep_dims_0 = const()[name = string("op_9284_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9284_cast_fp16 = reduce_mean(axes = var_9284_axes_0, keep_dims = var_9284_keep_dims_0, x = var_9279_cast_fp16)[name = string("op_9284_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9285_cast_fp16 = mul(x = s_347_cast_fp16, y = s_347_cast_fp16)[name = string("op_9285_cast_fp16")];
            fp32 var_9286_epsilon_0 = const()[name = string("op_9286_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9286_cast_fp16 = inverse(epsilon = var_9286_epsilon_0, x = var_9285_cast_fp16)[name = string("op_9286_cast_fp16")];
            fp16 var_9287_to_fp16 = const()[name = string("op_9287_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9288_cast_fp16 = mul(x = var_9286_cast_fp16, y = var_9287_to_fp16)[name = string("op_9288_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9290_cast_fp16 = add(x = var_9284_cast_fp16, y = var_9288_cast_fp16)[name = string("op_9290_cast_fp16")];
            fp32 var_9291_epsilon_0 = const()[name = string("op_9291_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9291_cast_fp16 = rsqrt(epsilon = var_9291_epsilon_0, x = var_9290_cast_fp16)[name = string("op_9291_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_303_cast_fp16 = mul(x = y_347_cast_fp16, y = var_9291_cast_fp16)[name = string("out_303_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121684032)))];
            tensor<fp16, [1, 256, 512]> var_9293_cast_fp16 = mul(x = out_303_cast_fp16, y = tm_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("op_9293_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_347_cast_fp16 = add(x = x_337_cast_fp16, y = var_9293_cast_fp16)[name = string("x_347_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9296_cast_fp16 = abs(x = x_347_cast_fp16)[name = string("op_9296_cast_fp16")];
            bool var_9299 = const()[name = string("op_9299"), val = bool(true)];
            tensor<int32, [1]> var_9300_axes_0 = const()[name = string("op_9300_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9300_cast_fp16 = reduce_max(axes = var_9300_axes_0, keep_dims = var_9299, x = var_9296_cast_fp16)[name = string("op_9300_cast_fp16")];
            fp16 var_9301_to_fp16 = const()[name = string("op_9301_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_349_cast_fp16 = maximum(x = var_9300_cast_fp16, y = var_9301_to_fp16)[name = string("s_349_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_349_cast_fp16 = real_div(x = x_347_cast_fp16, y = s_349_cast_fp16)[name = string("y_349_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9304_cast_fp16 = mul(x = y_349_cast_fp16, y = y_349_cast_fp16)[name = string("op_9304_cast_fp16")];
            tensor<int32, [1]> var_9309_axes_0 = const()[name = string("op_9309_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9309_keep_dims_0 = const()[name = string("op_9309_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9309_cast_fp16 = reduce_mean(axes = var_9309_axes_0, keep_dims = var_9309_keep_dims_0, x = var_9304_cast_fp16)[name = string("op_9309_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9310_cast_fp16 = mul(x = s_349_cast_fp16, y = s_349_cast_fp16)[name = string("op_9310_cast_fp16")];
            fp32 var_9311_epsilon_0 = const()[name = string("op_9311_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9311_cast_fp16 = inverse(epsilon = var_9311_epsilon_0, x = var_9310_cast_fp16)[name = string("op_9311_cast_fp16")];
            fp16 var_9312_to_fp16 = const()[name = string("op_9312_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9313_cast_fp16 = mul(x = var_9311_cast_fp16, y = var_9312_to_fp16)[name = string("op_9313_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9315_cast_fp16 = add(x = var_9309_cast_fp16, y = var_9313_cast_fp16)[name = string("op_9315_cast_fp16")];
            fp32 var_9316_epsilon_0 = const()[name = string("op_9316_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9316_cast_fp16 = rsqrt(epsilon = var_9316_epsilon_0, x = var_9315_cast_fp16)[name = string("op_9316_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_305_cast_fp16 = mul(x = y_349_cast_fp16, y = var_9316_cast_fp16)[name = string("out_305_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121685120)))];
            tensor<fp16, [1, 256, 512]> input_257_cast_fp16 = mul(x = out_305_cast_fp16, y = tm_layers_21_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_257_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121686208))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122734848))))[name = string("tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_194_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_194_cast_fp16")];
            string var_9323_mode_0 = const()[name = string("op_9323_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_9323_cast_fp16 = gelu(mode = var_9323_mode_0, x = linear_194_cast_fp16)[name = string("op_9323_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122739008))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123787648))))[name = string("tm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_195_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_259_cast_fp16 = mul(x = var_9323_cast_fp16, y = linear_195_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123791808))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124840448))))[name = string("tm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9331_cast_fp16 = abs(x = linear_196_cast_fp16)[name = string("op_9331_cast_fp16")];
            bool var_9334 = const()[name = string("op_9334"), val = bool(true)];
            tensor<int32, [1]> var_9335_axes_0 = const()[name = string("op_9335_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9335_cast_fp16 = reduce_max(axes = var_9335_axes_0, keep_dims = var_9334, x = var_9331_cast_fp16)[name = string("op_9335_cast_fp16")];
            fp16 var_9336_to_fp16 = const()[name = string("op_9336_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_351_cast_fp16 = maximum(x = var_9335_cast_fp16, y = var_9336_to_fp16)[name = string("s_351_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_351_cast_fp16 = real_div(x = linear_196_cast_fp16, y = s_351_cast_fp16)[name = string("y_351_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9339_cast_fp16 = mul(x = y_351_cast_fp16, y = y_351_cast_fp16)[name = string("op_9339_cast_fp16")];
            tensor<int32, [1]> var_9344_axes_0 = const()[name = string("op_9344_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9344_keep_dims_0 = const()[name = string("op_9344_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9344_cast_fp16 = reduce_mean(axes = var_9344_axes_0, keep_dims = var_9344_keep_dims_0, x = var_9339_cast_fp16)[name = string("op_9344_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9345_cast_fp16 = mul(x = s_351_cast_fp16, y = s_351_cast_fp16)[name = string("op_9345_cast_fp16")];
            fp32 var_9346_epsilon_0 = const()[name = string("op_9346_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9346_cast_fp16 = inverse(epsilon = var_9346_epsilon_0, x = var_9345_cast_fp16)[name = string("op_9346_cast_fp16")];
            fp16 var_9347_to_fp16 = const()[name = string("op_9347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9348_cast_fp16 = mul(x = var_9346_cast_fp16, y = var_9347_to_fp16)[name = string("op_9348_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9350_cast_fp16 = add(x = var_9344_cast_fp16, y = var_9348_cast_fp16)[name = string("op_9350_cast_fp16")];
            fp32 var_9351_epsilon_0 = const()[name = string("op_9351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9351_cast_fp16 = rsqrt(epsilon = var_9351_epsilon_0, x = var_9350_cast_fp16)[name = string("op_9351_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_307_cast_fp16 = mul(x = y_351_cast_fp16, y = var_9351_cast_fp16)[name = string("out_307_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_21_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124841536)))];
            tensor<fp16, [1, 256, 512]> var_9353_cast_fp16 = mul(x = out_307_cast_fp16, y = tm_layers_21_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_261_cast_fp16 = add(x = x_347_cast_fp16, y = var_9353_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124842624))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125104832))))[name = string("tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_197_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_197_cast_fp16")];
            string var_9360_mode_0 = const()[name = string("op_9360_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_9360_cast_fp16 = gelu(mode = var_9360_mode_0, x = linear_197_cast_fp16)[name = string("op_9360_cast_fp16")];
            tensor<int32, [4]> var_9373_begin_0 = const()[name = string("op_9373_begin_0"), val = tensor<int32, [4]>([0, 0, 21, 0])];
            tensor<int32, [4]> var_9373_end_0 = const()[name = string("op_9373_end_0"), val = tensor<int32, [4]>([1, 256, 22, 512])];
            tensor<bool, [4]> var_9373_end_mask_0 = const()[name = string("op_9373_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9373_squeeze_mask_0 = const()[name = string("op_9373_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_9373_cast_fp16 = slice_by_index(begin = var_9373_begin_0, end = var_9373_end_0, end_mask = var_9373_end_mask_0, squeeze_mask = var_9373_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_9373_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_263_cast_fp16 = mul(x = var_9360_cast_fp16, y = var_9373_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125105920))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125368128))))[name = string("tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_21_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9383_cast_fp16 = abs(x = linear_198_cast_fp16)[name = string("op_9383_cast_fp16")];
            bool var_9386 = const()[name = string("op_9386"), val = bool(true)];
            tensor<int32, [1]> var_9387_axes_0 = const()[name = string("op_9387_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9387_cast_fp16 = reduce_max(axes = var_9387_axes_0, keep_dims = var_9386, x = var_9383_cast_fp16)[name = string("op_9387_cast_fp16")];
            fp16 var_9388_to_fp16 = const()[name = string("op_9388_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_353_cast_fp16 = maximum(x = var_9387_cast_fp16, y = var_9388_to_fp16)[name = string("s_353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_353_cast_fp16 = real_div(x = linear_198_cast_fp16, y = s_353_cast_fp16)[name = string("y_353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9391_cast_fp16 = mul(x = y_353_cast_fp16, y = y_353_cast_fp16)[name = string("op_9391_cast_fp16")];
            tensor<int32, [1]> var_9396_axes_0 = const()[name = string("op_9396_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9396_keep_dims_0 = const()[name = string("op_9396_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9396_cast_fp16 = reduce_mean(axes = var_9396_axes_0, keep_dims = var_9396_keep_dims_0, x = var_9391_cast_fp16)[name = string("op_9396_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9397_cast_fp16 = mul(x = s_353_cast_fp16, y = s_353_cast_fp16)[name = string("op_9397_cast_fp16")];
            fp32 var_9398_epsilon_0 = const()[name = string("op_9398_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9398_cast_fp16 = inverse(epsilon = var_9398_epsilon_0, x = var_9397_cast_fp16)[name = string("op_9398_cast_fp16")];
            fp16 var_9399_to_fp16 = const()[name = string("op_9399_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9400_cast_fp16 = mul(x = var_9398_cast_fp16, y = var_9399_to_fp16)[name = string("op_9400_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9402_cast_fp16 = add(x = var_9396_cast_fp16, y = var_9400_cast_fp16)[name = string("op_9402_cast_fp16")];
            fp32 var_9403_epsilon_0 = const()[name = string("op_9403_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9403_cast_fp16 = rsqrt(epsilon = var_9403_epsilon_0, x = var_9402_cast_fp16)[name = string("op_9403_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_309_cast_fp16 = mul(x = y_353_cast_fp16, y = var_9403_cast_fp16)[name = string("out_309_cast_fp16")];
            tensor<fp16, [512]> tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125369216)))];
            tensor<fp16, [1, 256, 512]> var_9405_cast_fp16 = mul(x = out_309_cast_fp16, y = tm_layers_21_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9405_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_43_cast_fp16 = add(x = input_261_cast_fp16, y = var_9405_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<fp16, [1]> tm_layers_21_layer_scalar_to_fp16 = const()[name = string("tm_layers_21_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.e8p-1])];
            tensor<fp16, [1, 256, 512]> x_353_cast_fp16 = mul(x = h_43_cast_fp16, y = tm_layers_21_layer_scalar_to_fp16)[name = string("x_353_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9409_cast_fp16 = abs(x = x_353_cast_fp16)[name = string("op_9409_cast_fp16")];
            bool var_9412 = const()[name = string("op_9412"), val = bool(true)];
            tensor<int32, [1]> var_9413_axes_0 = const()[name = string("op_9413_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9413_cast_fp16 = reduce_max(axes = var_9413_axes_0, keep_dims = var_9412, x = var_9409_cast_fp16)[name = string("op_9413_cast_fp16")];
            fp16 var_9414_to_fp16 = const()[name = string("op_9414_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_355_cast_fp16 = maximum(x = var_9413_cast_fp16, y = var_9414_to_fp16)[name = string("s_355_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_355_cast_fp16 = real_div(x = x_353_cast_fp16, y = s_355_cast_fp16)[name = string("y_355_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9417_cast_fp16 = mul(x = y_355_cast_fp16, y = y_355_cast_fp16)[name = string("op_9417_cast_fp16")];
            tensor<int32, [1]> var_9422_axes_0 = const()[name = string("op_9422_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9422_keep_dims_0 = const()[name = string("op_9422_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9422_cast_fp16 = reduce_mean(axes = var_9422_axes_0, keep_dims = var_9422_keep_dims_0, x = var_9417_cast_fp16)[name = string("op_9422_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9423_cast_fp16 = mul(x = s_355_cast_fp16, y = s_355_cast_fp16)[name = string("op_9423_cast_fp16")];
            fp32 var_9424_epsilon_0 = const()[name = string("op_9424_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9424_cast_fp16 = inverse(epsilon = var_9424_epsilon_0, x = var_9423_cast_fp16)[name = string("op_9424_cast_fp16")];
            fp16 var_9425_to_fp16 = const()[name = string("op_9425_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9426_cast_fp16 = mul(x = var_9424_cast_fp16, y = var_9425_to_fp16)[name = string("op_9426_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9428_cast_fp16 = add(x = var_9422_cast_fp16, y = var_9426_cast_fp16)[name = string("op_9428_cast_fp16")];
            fp32 var_9429_epsilon_0 = const()[name = string("op_9429_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9429_cast_fp16 = rsqrt(epsilon = var_9429_epsilon_0, x = var_9428_cast_fp16)[name = string("op_9429_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_311_cast_fp16 = mul(x = y_355_cast_fp16, y = var_9429_cast_fp16)[name = string("out_311_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125370304)))];
            tensor<fp16, [1, 256, 512]> input_265_cast_fp16 = mul(x = out_311_cast_fp16, y = tm_layers_22_input_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [1024, 512]> tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125371392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125895744))))[name = string("tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 1024]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = tm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<int32, [4]> var_9439 = const()[name = string("op_9439"), val = tensor<int32, [4]>([1, -1, 4, 256])];
            tensor<fp16, [1, 256, 4, 256]> var_9440_cast_fp16 = reshape(shape = var_9439, x = linear_199_cast_fp16)[name = string("op_9440_cast_fp16")];
            tensor<int32, [4]> x_355_perm_0 = const()[name = string("x_355_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125897856))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126160064))))[name = string("tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<int32, [4]> var_9451 = const()[name = string("op_9451"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_9452_cast_fp16 = reshape(shape = var_9451, x = linear_200_cast_fp16)[name = string("op_9452_cast_fp16")];
            tensor<int32, [4]> x_357_perm_0 = const()[name = string("x_357_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126161152))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126423360))))[name = string("tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<int32, [4]> var_9463 = const()[name = string("op_9463"), val = tensor<int32, [4]>([1, -1, 2, 256])];
            tensor<fp16, [1, 256, 2, 256]> var_9464_cast_fp16 = reshape(shape = var_9463, x = linear_201_cast_fp16)[name = string("op_9464_cast_fp16")];
            tensor<int32, [4]> x_359_perm_0 = const()[name = string("x_359_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 256]> x_355_cast_fp16 = transpose(perm = x_355_perm_0, x = var_9440_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 4, 256, 256]> var_9468_cast_fp16 = abs(x = x_355_cast_fp16)[name = string("op_9468_cast_fp16")];
            bool var_9471 = const()[name = string("op_9471"), val = bool(true)];
            tensor<int32, [1]> var_9472_axes_0 = const()[name = string("op_9472_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_9472_cast_fp16 = reduce_max(axes = var_9472_axes_0, keep_dims = var_9471, x = var_9468_cast_fp16)[name = string("op_9472_cast_fp16")];
            fp16 var_9473_to_fp16 = const()[name = string("op_9473_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_357_cast_fp16 = maximum(x = var_9472_cast_fp16, y = var_9473_to_fp16)[name = string("s_357_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> y_357_cast_fp16 = real_div(x = x_355_cast_fp16, y = s_357_cast_fp16)[name = string("y_357_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9476_cast_fp16 = mul(x = y_357_cast_fp16, y = y_357_cast_fp16)[name = string("op_9476_cast_fp16")];
            tensor<int32, [1]> var_9481_axes_0 = const()[name = string("op_9481_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9481_keep_dims_0 = const()[name = string("op_9481_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_9481_cast_fp16 = reduce_mean(axes = var_9481_axes_0, keep_dims = var_9481_keep_dims_0, x = var_9476_cast_fp16)[name = string("op_9481_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9482_cast_fp16 = mul(x = s_357_cast_fp16, y = s_357_cast_fp16)[name = string("op_9482_cast_fp16")];
            fp32 var_9483_epsilon_0 = const()[name = string("op_9483_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_9483_cast_fp16 = inverse(epsilon = var_9483_epsilon_0, x = var_9482_cast_fp16)[name = string("op_9483_cast_fp16")];
            fp16 var_9484_to_fp16 = const()[name = string("op_9484_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_9485_cast_fp16 = mul(x = var_9483_cast_fp16, y = var_9484_to_fp16)[name = string("op_9485_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9487_cast_fp16 = add(x = var_9481_cast_fp16, y = var_9485_cast_fp16)[name = string("op_9487_cast_fp16")];
            fp32 var_9488_epsilon_0 = const()[name = string("op_9488_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_9488_cast_fp16 = rsqrt(epsilon = var_9488_epsilon_0, x = var_9487_cast_fp16)[name = string("op_9488_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> out_313_cast_fp16 = mul(x = y_357_cast_fp16, y = var_9488_cast_fp16)[name = string("out_313_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_89_cast_fp16 = mul(x = out_313_cast_fp16, y = tm_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("q_89_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_357_cast_fp16 = transpose(perm = x_357_perm_0, x = var_9452_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 256, 256]> var_9491_cast_fp16 = abs(x = x_357_cast_fp16)[name = string("op_9491_cast_fp16")];
            bool var_9494 = const()[name = string("op_9494"), val = bool(true)];
            tensor<int32, [1]> var_9495_axes_0 = const()[name = string("op_9495_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_9495_cast_fp16 = reduce_max(axes = var_9495_axes_0, keep_dims = var_9494, x = var_9491_cast_fp16)[name = string("op_9495_cast_fp16")];
            fp16 var_9496_to_fp16 = const()[name = string("op_9496_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_359_cast_fp16 = maximum(x = var_9495_cast_fp16, y = var_9496_to_fp16)[name = string("s_359_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_359_cast_fp16 = real_div(x = x_357_cast_fp16, y = s_359_cast_fp16)[name = string("y_359_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9499_cast_fp16 = mul(x = y_359_cast_fp16, y = y_359_cast_fp16)[name = string("op_9499_cast_fp16")];
            tensor<int32, [1]> var_9504_axes_0 = const()[name = string("op_9504_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9504_keep_dims_0 = const()[name = string("op_9504_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_9504_cast_fp16 = reduce_mean(axes = var_9504_axes_0, keep_dims = var_9504_keep_dims_0, x = var_9499_cast_fp16)[name = string("op_9504_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9505_cast_fp16 = mul(x = s_359_cast_fp16, y = s_359_cast_fp16)[name = string("op_9505_cast_fp16")];
            fp32 var_9506_epsilon_0 = const()[name = string("op_9506_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_9506_cast_fp16 = inverse(epsilon = var_9506_epsilon_0, x = var_9505_cast_fp16)[name = string("op_9506_cast_fp16")];
            fp16 var_9507_to_fp16 = const()[name = string("op_9507_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_9508_cast_fp16 = mul(x = var_9506_cast_fp16, y = var_9507_to_fp16)[name = string("op_9508_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9510_cast_fp16 = add(x = var_9504_cast_fp16, y = var_9508_cast_fp16)[name = string("op_9510_cast_fp16")];
            fp32 var_9511_epsilon_0 = const()[name = string("op_9511_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_9511_cast_fp16 = rsqrt(epsilon = var_9511_epsilon_0, x = var_9510_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> out_315_cast_fp16 = mul(x = y_359_cast_fp16, y = var_9511_cast_fp16)[name = string("out_315_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_133_cast_fp16 = mul(x = out_315_cast_fp16, y = tm_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("k_133_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> x_359_cast_fp16 = transpose(perm = x_359_perm_0, x = var_9464_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 256, 256]> var_9514_cast_fp16 = abs(x = x_359_cast_fp16)[name = string("op_9514_cast_fp16")];
            bool var_9517 = const()[name = string("op_9517"), val = bool(true)];
            tensor<int32, [1]> var_9518_axes_0 = const()[name = string("op_9518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2, 256, 1]> var_9518_cast_fp16 = reduce_max(axes = var_9518_axes_0, keep_dims = var_9517, x = var_9514_cast_fp16)[name = string("op_9518_cast_fp16")];
            fp16 var_9519_to_fp16 = const()[name = string("op_9519_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2, 256, 1]> s_361_cast_fp16 = maximum(x = var_9518_cast_fp16, y = var_9519_to_fp16)[name = string("s_361_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> y_361_cast_fp16 = real_div(x = x_359_cast_fp16, y = s_361_cast_fp16)[name = string("y_361_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9522_cast_fp16 = mul(x = y_361_cast_fp16, y = y_361_cast_fp16)[name = string("op_9522_cast_fp16")];
            tensor<int32, [1]> var_9527_axes_0 = const()[name = string("op_9527_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9527_keep_dims_0 = const()[name = string("op_9527_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2, 256, 1]> var_9527_cast_fp16 = reduce_mean(axes = var_9527_axes_0, keep_dims = var_9527_keep_dims_0, x = var_9522_cast_fp16)[name = string("op_9527_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9528_cast_fp16 = mul(x = s_361_cast_fp16, y = s_361_cast_fp16)[name = string("op_9528_cast_fp16")];
            fp32 var_9529_epsilon_0 = const()[name = string("op_9529_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2, 256, 1]> var_9529_cast_fp16 = inverse(epsilon = var_9529_epsilon_0, x = var_9528_cast_fp16)[name = string("op_9529_cast_fp16")];
            fp16 var_9530_to_fp16 = const()[name = string("op_9530_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2, 256, 1]> var_9531_cast_fp16 = mul(x = var_9529_cast_fp16, y = var_9530_to_fp16)[name = string("op_9531_cast_fp16")];
            tensor<fp16, [1, 2, 256, 1]> var_9533_cast_fp16 = add(x = var_9527_cast_fp16, y = var_9531_cast_fp16)[name = string("op_9533_cast_fp16")];
            fp32 var_9534_epsilon_0 = const()[name = string("op_9534_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2, 256, 1]> var_9534_cast_fp16 = rsqrt(epsilon = var_9534_epsilon_0, x = var_9533_cast_fp16)[name = string("op_9534_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> v_89_cast_fp16 = mul(x = y_361_cast_fp16, y = var_9534_cast_fp16)[name = string("v_89_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9536_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9536_cast_fp16")];
            tensor<int32, [4]> var_9541_begin_0 = const()[name = string("op_9541_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9541_end_0 = const()[name = string("op_9541_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_9541_end_mask_0 = const()[name = string("op_9541_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 128]> var_9541_cast_fp16 = slice_by_index(begin = var_9541_begin_0, end = var_9541_end_0, end_mask = var_9541_end_mask_0, x = q_89_cast_fp16)[name = string("op_9541_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 128]> var_9542_cast_fp16 = mul(x = var_9541_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_9542_cast_fp16")];
            tensor<int32, [4]> var_9547_begin_0 = const()[name = string("op_9547_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9547_end_0 = const()[name = string("op_9547_end_0"), val = tensor<int32, [4]>([1, 4, 256, 128])];
            tensor<bool, [4]> var_9547_end_mask_0 = const()[name = string("op_9547_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 128]> var_9547_cast_fp16 = slice_by_index(begin = var_9547_begin_0, end = var_9547_end_0, end_mask = var_9547_end_mask_0, x = q_89_cast_fp16)[name = string("op_9547_cast_fp16")];
            int32 var_9549 = const()[name = string("op_9549"), val = int32(-1)];
            bool var_9550_interleave_0 = const()[name = string("op_9550_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> var_9550_cast_fp16 = concat(axis = var_9549, interleave = var_9550_interleave_0, values = (var_9542_cast_fp16, var_9547_cast_fp16))[name = string("op_9550_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9551_cast_fp16 = mul(x = var_9550_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9551_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> q_91_cast_fp16 = add(x = var_9536_cast_fp16, y = var_9551_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9554_cast_fp16 = mul(x = k_133_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9554_cast_fp16")];
            tensor<int32, [4]> var_9559_begin_0 = const()[name = string("op_9559_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 128])];
            tensor<int32, [4]> var_9559_end_0 = const()[name = string("op_9559_end_0"), val = tensor<int32, [4]>([1, 2, 256, 256])];
            tensor<bool, [4]> var_9559_end_mask_0 = const()[name = string("op_9559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 128]> var_9559_cast_fp16 = slice_by_index(begin = var_9559_begin_0, end = var_9559_end_0, end_mask = var_9559_end_mask_0, x = k_133_cast_fp16)[name = string("op_9559_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 128]> var_9560_cast_fp16 = mul(x = var_9559_cast_fp16, y = const_45_promoted_to_fp16)[name = string("op_9560_cast_fp16")];
            tensor<int32, [4]> var_9565_begin_0 = const()[name = string("op_9565_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9565_end_0 = const()[name = string("op_9565_end_0"), val = tensor<int32, [4]>([1, 2, 256, 128])];
            tensor<bool, [4]> var_9565_end_mask_0 = const()[name = string("op_9565_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 128]> var_9565_cast_fp16 = slice_by_index(begin = var_9565_begin_0, end = var_9565_end_0, end_mask = var_9565_end_mask_0, x = k_133_cast_fp16)[name = string("op_9565_cast_fp16")];
            int32 var_9567 = const()[name = string("op_9567"), val = int32(-1)];
            bool var_9568_interleave_0 = const()[name = string("op_9568_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 256]> var_9568_cast_fp16 = concat(axis = var_9567, interleave = var_9568_interleave_0, values = (var_9560_cast_fp16, var_9565_cast_fp16))[name = string("op_9568_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> var_9569_cast_fp16 = mul(x = var_9568_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9569_cast_fp16")];
            tensor<fp16, [1, 2, 256, 256]> k_135_cast_fp16 = add(x = var_9554_cast_fp16, y = var_9569_cast_fp16)[name = string("k_135_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_44_reps_0 = const()[name = string("tile_44_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = k_135_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 1, 256, 256]> tile_44_cast_fp16 = tile(reps = tile_44_reps_0, x = transpose_88_cast_fp16)[name = string("tile_44_cast_fp16")];
            tensor<int32, [5]> concat_88 = const()[name = string("concat_88"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_88_cast_fp16 = reshape(shape = concat_88, x = tile_44_cast_fp16)[name = string("reshape_88_cast_fp16")];
            tensor<int32, [5]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = reshape_88_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 1, 256, 256]> reshape_89_cast_fp16 = reshape(shape = concat_89, x = transpose_89_cast_fp16)[name = string("reshape_89_cast_fp16")];
            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_45_reps_0 = const()[name = string("tile_45_reps_0"), val = tensor<int32, [4]>([2, 1, 1, 1])];
            tensor<fp16, [2, 1, 256, 256]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = v_89_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 1, 256, 256]> tile_45_cast_fp16 = tile(reps = tile_45_reps_0, x = transpose_90_cast_fp16)[name = string("tile_45_cast_fp16")];
            tensor<int32, [5]> concat_90 = const()[name = string("concat_90"), val = tensor<int32, [5]>([2, 2, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> reshape_90_cast_fp16 = reshape(shape = concat_90, x = tile_45_cast_fp16)[name = string("reshape_90_cast_fp16")];
            tensor<int32, [5]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_91 = const()[name = string("concat_91"), val = tensor<int32, [4]>([-1, 1, 256, 256])];
            tensor<fp16, [2, 2, 1, 256, 256]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = reshape_90_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 1, 256, 256]> reshape_91_cast_fp16 = reshape(shape = concat_91, x = transpose_91_cast_fp16)[name = string("reshape_91_cast_fp16")];
            tensor<int32, [4]> v_91_perm_0 = const()[name = string("v_91_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9583_transpose_x_0 = const()[name = string("op_9583_transpose_x_0"), val = bool(false)];
            bool var_9583_transpose_y_0 = const()[name = string("op_9583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> transpose_118_cast_fp16 = transpose(perm = transpose_118_perm_0, x = reshape_89_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 4, 256, 256]> var_9583_cast_fp16 = matmul(transpose_x = var_9583_transpose_x_0, transpose_y = var_9583_transpose_y_0, x = q_91_cast_fp16, y = transpose_118_cast_fp16)[name = string("op_9583_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9585_cast_fp16 = add(x = var_9583_cast_fp16, y = attention_bias)[name = string("op_9585_cast_fp16")];
            int32 var_9586 = const()[name = string("op_9586"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_9588_cast_fp16 = softmax(axis = var_9586, x = var_9585_cast_fp16)[name = string("op_9588_cast_fp16")];
            bool var_9589_transpose_x_0 = const()[name = string("op_9589_transpose_x_0"), val = bool(false)];
            bool var_9589_transpose_y_0 = const()[name = string("op_9589_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 256]> v_91_cast_fp16 = transpose(perm = v_91_perm_0, x = reshape_91_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 4, 256, 256]> var_9589_cast_fp16 = matmul(transpose_x = var_9589_transpose_x_0, transpose_y = var_9589_transpose_y_0, x = var_9588_cast_fp16, y = v_91_cast_fp16)[name = string("op_9589_cast_fp16")];
            tensor<int32, [4]> var_9592_perm_0 = const()[name = string("op_9592_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9596 = const()[name = string("op_9596"), val = tensor<int32, [3]>([1, -1, 1024])];
            tensor<fp16, [1, 256, 4, 256]> var_9592_cast_fp16 = transpose(perm = var_9592_perm_0, x = var_9589_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 256, 1024]> input_267_cast_fp16 = reshape(shape = var_9596, x = var_9592_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<fp16, [512, 1024]> tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126424448))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126948800))))[name = string("tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9601_cast_fp16 = abs(x = linear_202_cast_fp16)[name = string("op_9601_cast_fp16")];
            bool var_9604 = const()[name = string("op_9604"), val = bool(true)];
            tensor<int32, [1]> var_9605_axes_0 = const()[name = string("op_9605_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9605_cast_fp16 = reduce_max(axes = var_9605_axes_0, keep_dims = var_9604, x = var_9601_cast_fp16)[name = string("op_9605_cast_fp16")];
            fp16 var_9606_to_fp16 = const()[name = string("op_9606_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_363_cast_fp16 = maximum(x = var_9605_cast_fp16, y = var_9606_to_fp16)[name = string("s_363_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_363_cast_fp16 = real_div(x = linear_202_cast_fp16, y = s_363_cast_fp16)[name = string("y_363_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9609_cast_fp16 = mul(x = y_363_cast_fp16, y = y_363_cast_fp16)[name = string("op_9609_cast_fp16")];
            tensor<int32, [1]> var_9614_axes_0 = const()[name = string("op_9614_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9614_keep_dims_0 = const()[name = string("op_9614_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9614_cast_fp16 = reduce_mean(axes = var_9614_axes_0, keep_dims = var_9614_keep_dims_0, x = var_9609_cast_fp16)[name = string("op_9614_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9615_cast_fp16 = mul(x = s_363_cast_fp16, y = s_363_cast_fp16)[name = string("op_9615_cast_fp16")];
            fp32 var_9616_epsilon_0 = const()[name = string("op_9616_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9616_cast_fp16 = inverse(epsilon = var_9616_epsilon_0, x = var_9615_cast_fp16)[name = string("op_9616_cast_fp16")];
            fp16 var_9617_to_fp16 = const()[name = string("op_9617_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9618_cast_fp16 = mul(x = var_9616_cast_fp16, y = var_9617_to_fp16)[name = string("op_9618_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9620_cast_fp16 = add(x = var_9614_cast_fp16, y = var_9618_cast_fp16)[name = string("op_9620_cast_fp16")];
            fp32 var_9621_epsilon_0 = const()[name = string("op_9621_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9621_cast_fp16 = rsqrt(epsilon = var_9621_epsilon_0, x = var_9620_cast_fp16)[name = string("op_9621_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_317_cast_fp16 = mul(x = y_363_cast_fp16, y = var_9621_cast_fp16)[name = string("out_317_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126949888)))];
            tensor<fp16, [1, 256, 512]> var_9623_cast_fp16 = mul(x = out_317_cast_fp16, y = tm_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("op_9623_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_363_cast_fp16 = add(x = x_353_cast_fp16, y = var_9623_cast_fp16)[name = string("x_363_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9626_cast_fp16 = abs(x = x_363_cast_fp16)[name = string("op_9626_cast_fp16")];
            bool var_9629 = const()[name = string("op_9629"), val = bool(true)];
            tensor<int32, [1]> var_9630_axes_0 = const()[name = string("op_9630_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9630_cast_fp16 = reduce_max(axes = var_9630_axes_0, keep_dims = var_9629, x = var_9626_cast_fp16)[name = string("op_9630_cast_fp16")];
            fp16 var_9631_to_fp16 = const()[name = string("op_9631_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_365_cast_fp16 = maximum(x = var_9630_cast_fp16, y = var_9631_to_fp16)[name = string("s_365_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_365_cast_fp16 = real_div(x = x_363_cast_fp16, y = s_365_cast_fp16)[name = string("y_365_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9634_cast_fp16 = mul(x = y_365_cast_fp16, y = y_365_cast_fp16)[name = string("op_9634_cast_fp16")];
            tensor<int32, [1]> var_9639_axes_0 = const()[name = string("op_9639_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9639_keep_dims_0 = const()[name = string("op_9639_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9639_cast_fp16 = reduce_mean(axes = var_9639_axes_0, keep_dims = var_9639_keep_dims_0, x = var_9634_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9640_cast_fp16 = mul(x = s_365_cast_fp16, y = s_365_cast_fp16)[name = string("op_9640_cast_fp16")];
            fp32 var_9641_epsilon_0 = const()[name = string("op_9641_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9641_cast_fp16 = inverse(epsilon = var_9641_epsilon_0, x = var_9640_cast_fp16)[name = string("op_9641_cast_fp16")];
            fp16 var_9642_to_fp16 = const()[name = string("op_9642_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9643_cast_fp16 = mul(x = var_9641_cast_fp16, y = var_9642_to_fp16)[name = string("op_9643_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9645_cast_fp16 = add(x = var_9639_cast_fp16, y = var_9643_cast_fp16)[name = string("op_9645_cast_fp16")];
            fp32 var_9646_epsilon_0 = const()[name = string("op_9646_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9646_cast_fp16 = rsqrt(epsilon = var_9646_epsilon_0, x = var_9645_cast_fp16)[name = string("op_9646_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_319_cast_fp16 = mul(x = y_365_cast_fp16, y = var_9646_cast_fp16)[name = string("out_319_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126950976)))];
            tensor<fp16, [1, 256, 512]> input_269_cast_fp16 = mul(x = out_319_cast_fp16, y = tm_layers_22_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126952064))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128000704))))[name = string("tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_203_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_203_cast_fp16")];
            string var_9653_mode_0 = const()[name = string("op_9653_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_9653_cast_fp16 = gelu(mode = var_9653_mode_0, x = linear_203_cast_fp16)[name = string("op_9653_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128004864))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129053504))))[name = string("tm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_204_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_271_cast_fp16 = mul(x = var_9653_cast_fp16, y = linear_204_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129057664))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130106304))))[name = string("tm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9661_cast_fp16 = abs(x = linear_205_cast_fp16)[name = string("op_9661_cast_fp16")];
            bool var_9664 = const()[name = string("op_9664"), val = bool(true)];
            tensor<int32, [1]> var_9665_axes_0 = const()[name = string("op_9665_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9665_cast_fp16 = reduce_max(axes = var_9665_axes_0, keep_dims = var_9664, x = var_9661_cast_fp16)[name = string("op_9665_cast_fp16")];
            fp16 var_9666_to_fp16 = const()[name = string("op_9666_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_367_cast_fp16 = maximum(x = var_9665_cast_fp16, y = var_9666_to_fp16)[name = string("s_367_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_367_cast_fp16 = real_div(x = linear_205_cast_fp16, y = s_367_cast_fp16)[name = string("y_367_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9669_cast_fp16 = mul(x = y_367_cast_fp16, y = y_367_cast_fp16)[name = string("op_9669_cast_fp16")];
            tensor<int32, [1]> var_9674_axes_0 = const()[name = string("op_9674_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9674_keep_dims_0 = const()[name = string("op_9674_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9674_cast_fp16 = reduce_mean(axes = var_9674_axes_0, keep_dims = var_9674_keep_dims_0, x = var_9669_cast_fp16)[name = string("op_9674_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9675_cast_fp16 = mul(x = s_367_cast_fp16, y = s_367_cast_fp16)[name = string("op_9675_cast_fp16")];
            fp32 var_9676_epsilon_0 = const()[name = string("op_9676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9676_cast_fp16 = inverse(epsilon = var_9676_epsilon_0, x = var_9675_cast_fp16)[name = string("op_9676_cast_fp16")];
            fp16 var_9677_to_fp16 = const()[name = string("op_9677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9678_cast_fp16 = mul(x = var_9676_cast_fp16, y = var_9677_to_fp16)[name = string("op_9678_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9680_cast_fp16 = add(x = var_9674_cast_fp16, y = var_9678_cast_fp16)[name = string("op_9680_cast_fp16")];
            fp32 var_9681_epsilon_0 = const()[name = string("op_9681_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9681_cast_fp16 = rsqrt(epsilon = var_9681_epsilon_0, x = var_9680_cast_fp16)[name = string("op_9681_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_321_cast_fp16 = mul(x = y_367_cast_fp16, y = var_9681_cast_fp16)[name = string("out_321_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_22_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130107392)))];
            tensor<fp16, [1, 256, 512]> var_9683_cast_fp16 = mul(x = out_321_cast_fp16, y = tm_layers_22_post_feedforward_layernorm_weight_to_fp16)[name = string("op_9683_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_273_cast_fp16 = add(x = x_363_cast_fp16, y = var_9683_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130108480))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130370688))))[name = string("tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_206_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_206_cast_fp16")];
            string var_9690_mode_0 = const()[name = string("op_9690_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_9690_cast_fp16 = gelu(mode = var_9690_mode_0, x = linear_206_cast_fp16)[name = string("op_9690_cast_fp16")];
            tensor<int32, [4]> var_9703_begin_0 = const()[name = string("op_9703_begin_0"), val = tensor<int32, [4]>([0, 0, 22, 0])];
            tensor<int32, [4]> var_9703_end_0 = const()[name = string("op_9703_end_0"), val = tensor<int32, [4]>([1, 256, 23, 512])];
            tensor<bool, [4]> var_9703_end_mask_0 = const()[name = string("op_9703_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_9703_squeeze_mask_0 = const()[name = string("op_9703_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_9703_cast_fp16 = slice_by_index(begin = var_9703_begin_0, end = var_9703_end_0, end_mask = var_9703_end_mask_0, squeeze_mask = var_9703_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_9703_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_275_cast_fp16 = mul(x = var_9690_cast_fp16, y = var_9703_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130371776))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130633984))))[name = string("tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_22_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9713_cast_fp16 = abs(x = linear_207_cast_fp16)[name = string("op_9713_cast_fp16")];
            bool var_9716 = const()[name = string("op_9716"), val = bool(true)];
            tensor<int32, [1]> var_9717_axes_0 = const()[name = string("op_9717_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9717_cast_fp16 = reduce_max(axes = var_9717_axes_0, keep_dims = var_9716, x = var_9713_cast_fp16)[name = string("op_9717_cast_fp16")];
            fp16 var_9718_to_fp16 = const()[name = string("op_9718_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_369_cast_fp16 = maximum(x = var_9717_cast_fp16, y = var_9718_to_fp16)[name = string("s_369_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_369_cast_fp16 = real_div(x = linear_207_cast_fp16, y = s_369_cast_fp16)[name = string("y_369_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9721_cast_fp16 = mul(x = y_369_cast_fp16, y = y_369_cast_fp16)[name = string("op_9721_cast_fp16")];
            tensor<int32, [1]> var_9726_axes_0 = const()[name = string("op_9726_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9726_keep_dims_0 = const()[name = string("op_9726_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9726_cast_fp16 = reduce_mean(axes = var_9726_axes_0, keep_dims = var_9726_keep_dims_0, x = var_9721_cast_fp16)[name = string("op_9726_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9727_cast_fp16 = mul(x = s_369_cast_fp16, y = s_369_cast_fp16)[name = string("op_9727_cast_fp16")];
            fp32 var_9728_epsilon_0 = const()[name = string("op_9728_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9728_cast_fp16 = inverse(epsilon = var_9728_epsilon_0, x = var_9727_cast_fp16)[name = string("op_9728_cast_fp16")];
            fp16 var_9729_to_fp16 = const()[name = string("op_9729_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9730_cast_fp16 = mul(x = var_9728_cast_fp16, y = var_9729_to_fp16)[name = string("op_9730_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9732_cast_fp16 = add(x = var_9726_cast_fp16, y = var_9730_cast_fp16)[name = string("op_9732_cast_fp16")];
            fp32 var_9733_epsilon_0 = const()[name = string("op_9733_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9733_cast_fp16 = rsqrt(epsilon = var_9733_epsilon_0, x = var_9732_cast_fp16)[name = string("op_9733_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_323_cast_fp16 = mul(x = y_369_cast_fp16, y = var_9733_cast_fp16)[name = string("out_323_cast_fp16")];
            tensor<fp16, [512]> tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130635072)))];
            tensor<fp16, [1, 256, 512]> var_9735_cast_fp16 = mul(x = out_323_cast_fp16, y = tm_layers_22_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_9735_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_45_cast_fp16 = add(x = input_273_cast_fp16, y = var_9735_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<fp16, [1]> tm_layers_22_layer_scalar_to_fp16 = const()[name = string("tm_layers_22_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.bep-1])];
            tensor<fp16, [1, 256, 512]> x_369_cast_fp16 = mul(x = h_45_cast_fp16, y = tm_layers_22_layer_scalar_to_fp16)[name = string("x_369_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9739_cast_fp16 = abs(x = x_369_cast_fp16)[name = string("op_9739_cast_fp16")];
            bool var_9742 = const()[name = string("op_9742"), val = bool(true)];
            tensor<int32, [1]> var_9743_axes_0 = const()[name = string("op_9743_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9743_cast_fp16 = reduce_max(axes = var_9743_axes_0, keep_dims = var_9742, x = var_9739_cast_fp16)[name = string("op_9743_cast_fp16")];
            fp16 var_9744_to_fp16 = const()[name = string("op_9744_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_371_cast_fp16 = maximum(x = var_9743_cast_fp16, y = var_9744_to_fp16)[name = string("s_371_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_371_cast_fp16 = real_div(x = x_369_cast_fp16, y = s_371_cast_fp16)[name = string("y_371_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9747_cast_fp16 = mul(x = y_371_cast_fp16, y = y_371_cast_fp16)[name = string("op_9747_cast_fp16")];
            tensor<int32, [1]> var_9752_axes_0 = const()[name = string("op_9752_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9752_keep_dims_0 = const()[name = string("op_9752_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9752_cast_fp16 = reduce_mean(axes = var_9752_axes_0, keep_dims = var_9752_keep_dims_0, x = var_9747_cast_fp16)[name = string("op_9752_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9753_cast_fp16 = mul(x = s_371_cast_fp16, y = s_371_cast_fp16)[name = string("op_9753_cast_fp16")];
            fp32 var_9754_epsilon_0 = const()[name = string("op_9754_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9754_cast_fp16 = inverse(epsilon = var_9754_epsilon_0, x = var_9753_cast_fp16)[name = string("op_9754_cast_fp16")];
            fp16 var_9755_to_fp16 = const()[name = string("op_9755_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9756_cast_fp16 = mul(x = var_9754_cast_fp16, y = var_9755_to_fp16)[name = string("op_9756_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9758_cast_fp16 = add(x = var_9752_cast_fp16, y = var_9756_cast_fp16)[name = string("op_9758_cast_fp16")];
            fp32 var_9759_epsilon_0 = const()[name = string("op_9759_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9759_cast_fp16 = rsqrt(epsilon = var_9759_epsilon_0, x = var_9758_cast_fp16)[name = string("op_9759_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_325_cast_fp16 = mul(x = y_371_cast_fp16, y = var_9759_cast_fp16)[name = string("out_325_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130636160)))];
            tensor<fp16, [1, 256, 512]> input_277_cast_fp16 = mul(x = out_325_cast_fp16, y = tm_layers_23_input_layernorm_weight_to_fp16)[name = string("input_277_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130637248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131685888))))[name = string("tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_208_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<int32, [4]> var_9769 = const()[name = string("op_9769"), val = tensor<int32, [4]>([1, -1, 4, 512])];
            tensor<fp16, [1, 256, 4, 512]> var_9770_cast_fp16 = reshape(shape = var_9769, x = linear_208_cast_fp16)[name = string("op_9770_cast_fp16")];
            tensor<int32, [4]> x_371_perm_0 = const()[name = string("x_371_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131690048))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131952256))))[name = string("tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<int32, [4]> var_9781 = const()[name = string("op_9781"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_9782_cast_fp16 = reshape(shape = var_9781, x = linear_209_cast_fp16)[name = string("op_9782_cast_fp16")];
            tensor<int32, [4]> x_373_perm_0 = const()[name = string("x_373_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131953344))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132215552))))[name = string("tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_277_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<int32, [4]> var_9793 = const()[name = string("op_9793"), val = tensor<int32, [4]>([1, -1, 1, 512])];
            tensor<fp16, [1, 256, 1, 512]> var_9794_cast_fp16 = reshape(shape = var_9793, x = linear_210_cast_fp16)[name = string("op_9794_cast_fp16")];
            tensor<int32, [4]> x_375_perm_0 = const()[name = string("x_375_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 4, 256, 512]> x_371_cast_fp16 = transpose(perm = x_371_perm_0, x = var_9770_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 4, 256, 512]> var_9798_cast_fp16 = abs(x = x_371_cast_fp16)[name = string("op_9798_cast_fp16")];
            bool var_9801 = const()[name = string("op_9801"), val = bool(true)];
            tensor<int32, [1]> var_9802_axes_0 = const()[name = string("op_9802_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 4, 256, 1]> var_9802_cast_fp16 = reduce_max(axes = var_9802_axes_0, keep_dims = var_9801, x = var_9798_cast_fp16)[name = string("op_9802_cast_fp16")];
            fp16 var_9803_to_fp16 = const()[name = string("op_9803_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 4, 256, 1]> s_373_cast_fp16 = maximum(x = var_9802_cast_fp16, y = var_9803_to_fp16)[name = string("s_373_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> y_373_cast_fp16 = real_div(x = x_371_cast_fp16, y = s_373_cast_fp16)[name = string("y_373_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_9806_cast_fp16 = mul(x = y_373_cast_fp16, y = y_373_cast_fp16)[name = string("op_9806_cast_fp16")];
            tensor<int32, [1]> var_9811_axes_0 = const()[name = string("op_9811_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9811_keep_dims_0 = const()[name = string("op_9811_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 4, 256, 1]> var_9811_cast_fp16 = reduce_mean(axes = var_9811_axes_0, keep_dims = var_9811_keep_dims_0, x = var_9806_cast_fp16)[name = string("op_9811_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9812_cast_fp16 = mul(x = s_373_cast_fp16, y = s_373_cast_fp16)[name = string("op_9812_cast_fp16")];
            fp32 var_9813_epsilon_0 = const()[name = string("op_9813_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 4, 256, 1]> var_9813_cast_fp16 = inverse(epsilon = var_9813_epsilon_0, x = var_9812_cast_fp16)[name = string("op_9813_cast_fp16")];
            fp16 var_9814_to_fp16 = const()[name = string("op_9814_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 4, 256, 1]> var_9815_cast_fp16 = mul(x = var_9813_cast_fp16, y = var_9814_to_fp16)[name = string("op_9815_cast_fp16")];
            tensor<fp16, [1, 4, 256, 1]> var_9817_cast_fp16 = add(x = var_9811_cast_fp16, y = var_9815_cast_fp16)[name = string("op_9817_cast_fp16")];
            fp32 var_9818_epsilon_0 = const()[name = string("op_9818_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 4, 256, 1]> var_9818_cast_fp16 = rsqrt(epsilon = var_9818_epsilon_0, x = var_9817_cast_fp16)[name = string("op_9818_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> out_327_cast_fp16 = mul(x = y_373_cast_fp16, y = var_9818_cast_fp16)[name = string("out_327_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132216640)))];
            tensor<fp16, [1, 4, 256, 512]> q_93_cast_fp16 = mul(x = out_327_cast_fp16, y = tm_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("q_93_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_373_cast_fp16 = transpose(perm = x_373_perm_0, x = var_9782_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1, 256, 512]> var_9821_cast_fp16 = abs(x = x_373_cast_fp16)[name = string("op_9821_cast_fp16")];
            bool var_9824 = const()[name = string("op_9824"), val = bool(true)];
            tensor<int32, [1]> var_9825_axes_0 = const()[name = string("op_9825_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_9825_cast_fp16 = reduce_max(axes = var_9825_axes_0, keep_dims = var_9824, x = var_9821_cast_fp16)[name = string("op_9825_cast_fp16")];
            fp16 var_9826_to_fp16 = const()[name = string("op_9826_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_375_cast_fp16 = maximum(x = var_9825_cast_fp16, y = var_9826_to_fp16)[name = string("s_375_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_375_cast_fp16 = real_div(x = x_373_cast_fp16, y = s_375_cast_fp16)[name = string("y_375_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_9829_cast_fp16 = mul(x = y_375_cast_fp16, y = y_375_cast_fp16)[name = string("op_9829_cast_fp16")];
            tensor<int32, [1]> var_9834_axes_0 = const()[name = string("op_9834_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9834_keep_dims_0 = const()[name = string("op_9834_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_9834_cast_fp16 = reduce_mean(axes = var_9834_axes_0, keep_dims = var_9834_keep_dims_0, x = var_9829_cast_fp16)[name = string("op_9834_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_9835_cast_fp16 = mul(x = s_375_cast_fp16, y = s_375_cast_fp16)[name = string("op_9835_cast_fp16")];
            fp32 var_9836_epsilon_0 = const()[name = string("op_9836_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_9836_cast_fp16 = inverse(epsilon = var_9836_epsilon_0, x = var_9835_cast_fp16)[name = string("op_9836_cast_fp16")];
            fp16 var_9837_to_fp16 = const()[name = string("op_9837_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_9838_cast_fp16 = mul(x = var_9836_cast_fp16, y = var_9837_to_fp16)[name = string("op_9838_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_9840_cast_fp16 = add(x = var_9834_cast_fp16, y = var_9838_cast_fp16)[name = string("op_9840_cast_fp16")];
            fp32 var_9841_epsilon_0 = const()[name = string("op_9841_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_9841_cast_fp16 = rsqrt(epsilon = var_9841_epsilon_0, x = var_9840_cast_fp16)[name = string("op_9841_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> out_329_cast_fp16 = mul(x = y_375_cast_fp16, y = var_9841_cast_fp16)[name = string("out_329_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("tm_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132217728)))];
            tensor<fp16, [1, 1, 256, 512]> k_139_cast_fp16 = mul(x = out_329_cast_fp16, y = tm_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("k_139_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> x_375_cast_fp16 = transpose(perm = x_375_perm_0, x = var_9794_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 1, 256, 512]> var_9844_cast_fp16 = abs(x = x_375_cast_fp16)[name = string("op_9844_cast_fp16")];
            bool var_9847 = const()[name = string("op_9847"), val = bool(true)];
            tensor<int32, [1]> var_9848_axes_0 = const()[name = string("op_9848_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 256, 1]> var_9848_cast_fp16 = reduce_max(axes = var_9848_axes_0, keep_dims = var_9847, x = var_9844_cast_fp16)[name = string("op_9848_cast_fp16")];
            fp16 var_9849_to_fp16 = const()[name = string("op_9849_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1, 256, 1]> s_377_cast_fp16 = maximum(x = var_9848_cast_fp16, y = var_9849_to_fp16)[name = string("s_377_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> y_377_cast_fp16 = real_div(x = x_375_cast_fp16, y = s_377_cast_fp16)[name = string("y_377_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_9852_cast_fp16 = mul(x = y_377_cast_fp16, y = y_377_cast_fp16)[name = string("op_9852_cast_fp16")];
            tensor<int32, [1]> var_9857_axes_0 = const()[name = string("op_9857_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9857_keep_dims_0 = const()[name = string("op_9857_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 256, 1]> var_9857_cast_fp16 = reduce_mean(axes = var_9857_axes_0, keep_dims = var_9857_keep_dims_0, x = var_9852_cast_fp16)[name = string("op_9857_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_9858_cast_fp16 = mul(x = s_377_cast_fp16, y = s_377_cast_fp16)[name = string("op_9858_cast_fp16")];
            fp32 var_9859_epsilon_0 = const()[name = string("op_9859_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 256, 1]> var_9859_cast_fp16 = inverse(epsilon = var_9859_epsilon_0, x = var_9858_cast_fp16)[name = string("op_9859_cast_fp16")];
            fp16 var_9860_to_fp16 = const()[name = string("op_9860_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 256, 1]> var_9861_cast_fp16 = mul(x = var_9859_cast_fp16, y = var_9860_to_fp16)[name = string("op_9861_cast_fp16")];
            tensor<fp16, [1, 1, 256, 1]> var_9863_cast_fp16 = add(x = var_9857_cast_fp16, y = var_9861_cast_fp16)[name = string("op_9863_cast_fp16")];
            fp32 var_9864_epsilon_0 = const()[name = string("op_9864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 256, 1]> var_9864_cast_fp16 = rsqrt(epsilon = var_9864_epsilon_0, x = var_9863_cast_fp16)[name = string("op_9864_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> v_93_cast_fp16 = mul(x = y_377_cast_fp16, y = var_9864_cast_fp16)[name = string("v_93_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_9866_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_cast_fp16)[name = string("op_9866_cast_fp16")];
            tensor<int32, [4]> var_9871_begin_0 = const()[name = string("op_9871_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9871_end_0 = const()[name = string("op_9871_end_0"), val = tensor<int32, [4]>([1, 4, 256, 512])];
            tensor<bool, [4]> var_9871_end_mask_0 = const()[name = string("op_9871_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 256, 256]> var_9871_cast_fp16 = slice_by_index(begin = var_9871_begin_0, end = var_9871_end_0, end_mask = var_9871_end_mask_0, x = q_93_cast_fp16)[name = string("op_9871_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 4, 256, 256]> var_9872_cast_fp16 = mul(x = var_9871_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_9872_cast_fp16")];
            tensor<int32, [4]> var_9877_begin_0 = const()[name = string("op_9877_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9877_end_0 = const()[name = string("op_9877_end_0"), val = tensor<int32, [4]>([1, 4, 256, 256])];
            tensor<bool, [4]> var_9877_end_mask_0 = const()[name = string("op_9877_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 256, 256]> var_9877_cast_fp16 = slice_by_index(begin = var_9877_begin_0, end = var_9877_end_0, end_mask = var_9877_end_mask_0, x = q_93_cast_fp16)[name = string("op_9877_cast_fp16")];
            int32 var_9879 = const()[name = string("op_9879"), val = int32(-1)];
            bool var_9880_interleave_0 = const()[name = string("op_9880_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> var_9880_cast_fp16 = concat(axis = var_9879, interleave = var_9880_interleave_0, values = (var_9872_cast_fp16, var_9877_cast_fp16))[name = string("op_9880_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> var_9881_cast_fp16 = mul(x = var_9880_cast_fp16, y = sin_cast_fp16)[name = string("op_9881_cast_fp16")];
            tensor<fp16, [1, 4, 256, 512]> q_cast_fp16 = add(x = var_9866_cast_fp16, y = var_9881_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_9884_cast_fp16 = mul(x = k_139_cast_fp16, y = cos_cast_fp16)[name = string("op_9884_cast_fp16")];
            tensor<int32, [4]> var_9889_begin_0 = const()[name = string("op_9889_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 256])];
            tensor<int32, [4]> var_9889_end_0 = const()[name = string("op_9889_end_0"), val = tensor<int32, [4]>([1, 1, 256, 512])];
            tensor<bool, [4]> var_9889_end_mask_0 = const()[name = string("op_9889_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 256, 256]> var_9889_cast_fp16 = slice_by_index(begin = var_9889_begin_0, end = var_9889_end_0, end_mask = var_9889_end_mask_0, x = k_139_cast_fp16)[name = string("op_9889_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1, 256, 256]> var_9890_cast_fp16 = mul(x = var_9889_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_9890_cast_fp16")];
            tensor<int32, [4]> var_9895_begin_0 = const()[name = string("op_9895_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9895_end_0 = const()[name = string("op_9895_end_0"), val = tensor<int32, [4]>([1, 1, 256, 256])];
            tensor<bool, [4]> var_9895_end_mask_0 = const()[name = string("op_9895_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 256, 256]> var_9895_cast_fp16 = slice_by_index(begin = var_9895_begin_0, end = var_9895_end_0, end_mask = var_9895_end_mask_0, x = k_139_cast_fp16)[name = string("op_9895_cast_fp16")];
            int32 var_9897 = const()[name = string("op_9897"), val = int32(-1)];
            bool var_9898_interleave_0 = const()[name = string("op_9898_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1, 256, 512]> var_9898_cast_fp16 = concat(axis = var_9897, interleave = var_9898_interleave_0, values = (var_9890_cast_fp16, var_9895_cast_fp16))[name = string("op_9898_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> var_9899_cast_fp16 = mul(x = var_9898_cast_fp16, y = sin_cast_fp16)[name = string("op_9899_cast_fp16")];
            tensor<fp16, [1, 1, 256, 512]> k_141_cast_fp16 = add(x = var_9884_cast_fp16, y = var_9899_cast_fp16)[name = string("k_141_cast_fp16")];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_46_reps_0 = const()[name = string("tile_46_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_92_cast_fp16 = transpose(perm = transpose_92_perm_0, x = k_141_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 1, 256, 512]> tile_46_cast_fp16 = tile(reps = tile_46_reps_0, x = transpose_92_cast_fp16)[name = string("tile_46_cast_fp16")];
            tensor<int32, [5]> concat_92 = const()[name = string("concat_92"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_92_cast_fp16 = reshape(shape = concat_92, x = tile_46_cast_fp16)[name = string("reshape_92_cast_fp16")];
            tensor<int32, [5]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_93 = const()[name = string("concat_93"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = reshape_92_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 1, 256, 512]> reshape_93_cast_fp16 = reshape(shape = concat_93, x = transpose_93_cast_fp16)[name = string("reshape_93_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([1, 0, -1, -2])];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            tensor<int32, [4]> tile_47_reps_0 = const()[name = string("tile_47_reps_0"), val = tensor<int32, [4]>([4, 1, 1, 1])];
            tensor<fp16, [1, 1, 256, 512]> transpose_94_cast_fp16 = transpose(perm = transpose_94_perm_0, x = v_93_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 1, 256, 512]> tile_47_cast_fp16 = tile(reps = tile_47_reps_0, x = transpose_94_cast_fp16)[name = string("tile_47_cast_fp16")];
            tensor<int32, [5]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [5]>([4, 1, 1, 256, 512])];
            tensor<fp16, [4, 1, 1, 256, 512]> reshape_94_cast_fp16 = reshape(shape = concat_94, x = tile_47_cast_fp16)[name = string("reshape_94_cast_fp16")];
            tensor<int32, [5]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [5]>([1, 0, 2, 3, 4])];
            tensor<int32, [4]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [4]>([-1, 1, 256, 512])];
            tensor<fp16, [1, 4, 1, 256, 512]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = reshape_94_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 1, 256, 512]> reshape_95_cast_fp16 = reshape(shape = concat_95, x = transpose_95_cast_fp16)[name = string("reshape_95_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([1, 0, -2, -1])];
            bool var_9913_transpose_x_0 = const()[name = string("op_9913_transpose_x_0"), val = bool(false)];
            bool var_9913_transpose_y_0 = const()[name = string("op_9913_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 512, 256]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = reshape_93_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 4, 256, 256]> var_9913_cast_fp16 = matmul(transpose_x = var_9913_transpose_x_0, transpose_y = var_9913_transpose_y_0, x = q_cast_fp16, y = transpose_119_cast_fp16)[name = string("op_9913_cast_fp16")];
            tensor<fp16, [1, 4, 256, 256]> var_9915_cast_fp16 = add(x = var_9913_cast_fp16, y = attention_bias)[name = string("op_9915_cast_fp16")];
            int32 var_9916 = const()[name = string("op_9916"), val = int32(-1)];
            tensor<fp16, [1, 4, 256, 256]> var_9918_cast_fp16 = softmax(axis = var_9916, x = var_9915_cast_fp16)[name = string("op_9918_cast_fp16")];
            bool var_9919_transpose_x_0 = const()[name = string("op_9919_transpose_x_0"), val = bool(false)];
            bool var_9919_transpose_y_0 = const()[name = string("op_9919_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 4, 256, 512]> v_cast_fp16 = transpose(perm = v_perm_0, x = reshape_95_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 4, 256, 512]> var_9919_cast_fp16 = matmul(transpose_x = var_9919_transpose_x_0, transpose_y = var_9919_transpose_y_0, x = var_9918_cast_fp16, y = v_cast_fp16)[name = string("op_9919_cast_fp16")];
            tensor<int32, [4]> var_9922_perm_0 = const()[name = string("op_9922_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9926 = const()[name = string("op_9926"), val = tensor<int32, [3]>([1, -1, 2048])];
            tensor<fp16, [1, 256, 4, 512]> var_9922_cast_fp16 = transpose(perm = var_9922_perm_0, x = var_9919_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 256, 2048]> input_279_cast_fp16 = reshape(shape = var_9926, x = var_9922_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132218816))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133267456))))[name = string("tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_211_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9931_cast_fp16 = abs(x = linear_211_cast_fp16)[name = string("op_9931_cast_fp16")];
            bool var_9934 = const()[name = string("op_9934"), val = bool(true)];
            tensor<int32, [1]> var_9935_axes_0 = const()[name = string("op_9935_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9935_cast_fp16 = reduce_max(axes = var_9935_axes_0, keep_dims = var_9934, x = var_9931_cast_fp16)[name = string("op_9935_cast_fp16")];
            fp16 var_9936_to_fp16 = const()[name = string("op_9936_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_379_cast_fp16 = maximum(x = var_9935_cast_fp16, y = var_9936_to_fp16)[name = string("s_379_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_379_cast_fp16 = real_div(x = linear_211_cast_fp16, y = s_379_cast_fp16)[name = string("y_379_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9939_cast_fp16 = mul(x = y_379_cast_fp16, y = y_379_cast_fp16)[name = string("op_9939_cast_fp16")];
            tensor<int32, [1]> var_9944_axes_0 = const()[name = string("op_9944_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9944_keep_dims_0 = const()[name = string("op_9944_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9944_cast_fp16 = reduce_mean(axes = var_9944_axes_0, keep_dims = var_9944_keep_dims_0, x = var_9939_cast_fp16)[name = string("op_9944_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9945_cast_fp16 = mul(x = s_379_cast_fp16, y = s_379_cast_fp16)[name = string("op_9945_cast_fp16")];
            fp32 var_9946_epsilon_0 = const()[name = string("op_9946_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9946_cast_fp16 = inverse(epsilon = var_9946_epsilon_0, x = var_9945_cast_fp16)[name = string("op_9946_cast_fp16")];
            fp16 var_9947_to_fp16 = const()[name = string("op_9947_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9948_cast_fp16 = mul(x = var_9946_cast_fp16, y = var_9947_to_fp16)[name = string("op_9948_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9950_cast_fp16 = add(x = var_9944_cast_fp16, y = var_9948_cast_fp16)[name = string("op_9950_cast_fp16")];
            fp32 var_9951_epsilon_0 = const()[name = string("op_9951_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9951_cast_fp16 = rsqrt(epsilon = var_9951_epsilon_0, x = var_9950_cast_fp16)[name = string("op_9951_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_331_cast_fp16 = mul(x = y_379_cast_fp16, y = var_9951_cast_fp16)[name = string("out_331_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133268544)))];
            tensor<fp16, [1, 256, 512]> var_9953_cast_fp16 = mul(x = out_331_cast_fp16, y = tm_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("op_9953_cast_fp16")];
            tensor<fp16, [1, 256, 512]> x_379_cast_fp16 = add(x = x_369_cast_fp16, y = var_9953_cast_fp16)[name = string("x_379_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9956_cast_fp16 = abs(x = x_379_cast_fp16)[name = string("op_9956_cast_fp16")];
            bool var_9959 = const()[name = string("op_9959"), val = bool(true)];
            tensor<int32, [1]> var_9960_axes_0 = const()[name = string("op_9960_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9960_cast_fp16 = reduce_max(axes = var_9960_axes_0, keep_dims = var_9959, x = var_9956_cast_fp16)[name = string("op_9960_cast_fp16")];
            fp16 var_9961_to_fp16 = const()[name = string("op_9961_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_381_cast_fp16 = maximum(x = var_9960_cast_fp16, y = var_9961_to_fp16)[name = string("s_381_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_381_cast_fp16 = real_div(x = x_379_cast_fp16, y = s_381_cast_fp16)[name = string("y_381_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9964_cast_fp16 = mul(x = y_381_cast_fp16, y = y_381_cast_fp16)[name = string("op_9964_cast_fp16")];
            tensor<int32, [1]> var_9969_axes_0 = const()[name = string("op_9969_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9969_keep_dims_0 = const()[name = string("op_9969_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9969_cast_fp16 = reduce_mean(axes = var_9969_axes_0, keep_dims = var_9969_keep_dims_0, x = var_9964_cast_fp16)[name = string("op_9969_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9970_cast_fp16 = mul(x = s_381_cast_fp16, y = s_381_cast_fp16)[name = string("op_9970_cast_fp16")];
            fp32 var_9971_epsilon_0 = const()[name = string("op_9971_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_9971_cast_fp16 = inverse(epsilon = var_9971_epsilon_0, x = var_9970_cast_fp16)[name = string("op_9971_cast_fp16")];
            fp16 var_9972_to_fp16 = const()[name = string("op_9972_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_9973_cast_fp16 = mul(x = var_9971_cast_fp16, y = var_9972_to_fp16)[name = string("op_9973_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_9975_cast_fp16 = add(x = var_9969_cast_fp16, y = var_9973_cast_fp16)[name = string("op_9975_cast_fp16")];
            fp32 var_9976_epsilon_0 = const()[name = string("op_9976_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_9976_cast_fp16 = rsqrt(epsilon = var_9976_epsilon_0, x = var_9975_cast_fp16)[name = string("op_9976_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_333_cast_fp16 = mul(x = y_381_cast_fp16, y = var_9976_cast_fp16)[name = string("out_333_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133269632)))];
            tensor<fp16, [1, 256, 512]> input_281_cast_fp16 = mul(x = out_333_cast_fp16, y = tm_layers_23_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133270720))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134319360))))[name = string("tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_212_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_212_cast_fp16")];
            string var_9983_mode_0 = const()[name = string("op_9983_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 2048]> var_9983_cast_fp16 = gelu(mode = var_9983_mode_0, x = linear_212_cast_fp16)[name = string("op_9983_cast_fp16")];
            tensor<fp16, [2048, 512]> tm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134323520))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135372160))))[name = string("tm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 2048]> linear_213_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = tm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = string("linear_213_cast_fp16")];
            tensor<fp16, [1, 256, 2048]> input_283_cast_fp16 = mul(x = var_9983_cast_fp16, y = linear_213_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [512, 2048]> tm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135376320))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136424960))))[name = string("tm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_283_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9991_cast_fp16 = abs(x = linear_214_cast_fp16)[name = string("op_9991_cast_fp16")];
            bool var_9994 = const()[name = string("op_9994"), val = bool(true)];
            tensor<int32, [1]> var_9995_axes_0 = const()[name = string("op_9995_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_9995_cast_fp16 = reduce_max(axes = var_9995_axes_0, keep_dims = var_9994, x = var_9991_cast_fp16)[name = string("op_9995_cast_fp16")];
            fp16 var_9996_to_fp16 = const()[name = string("op_9996_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_383_cast_fp16 = maximum(x = var_9995_cast_fp16, y = var_9996_to_fp16)[name = string("s_383_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_383_cast_fp16 = real_div(x = linear_214_cast_fp16, y = s_383_cast_fp16)[name = string("y_383_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_9999_cast_fp16 = mul(x = y_383_cast_fp16, y = y_383_cast_fp16)[name = string("op_9999_cast_fp16")];
            tensor<int32, [1]> var_10004_axes_0 = const()[name = string("op_10004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10004_keep_dims_0 = const()[name = string("op_10004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_10004_cast_fp16 = reduce_mean(axes = var_10004_axes_0, keep_dims = var_10004_keep_dims_0, x = var_9999_cast_fp16)[name = string("op_10004_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10005_cast_fp16 = mul(x = s_383_cast_fp16, y = s_383_cast_fp16)[name = string("op_10005_cast_fp16")];
            fp32 var_10006_epsilon_0 = const()[name = string("op_10006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_10006_cast_fp16 = inverse(epsilon = var_10006_epsilon_0, x = var_10005_cast_fp16)[name = string("op_10006_cast_fp16")];
            fp16 var_10007_to_fp16 = const()[name = string("op_10007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_10008_cast_fp16 = mul(x = var_10006_cast_fp16, y = var_10007_to_fp16)[name = string("op_10008_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10010_cast_fp16 = add(x = var_10004_cast_fp16, y = var_10008_cast_fp16)[name = string("op_10010_cast_fp16")];
            fp32 var_10011_epsilon_0 = const()[name = string("op_10011_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_10011_cast_fp16 = rsqrt(epsilon = var_10011_epsilon_0, x = var_10010_cast_fp16)[name = string("op_10011_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_335_cast_fp16 = mul(x = y_383_cast_fp16, y = var_10011_cast_fp16)[name = string("out_335_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("tm_layers_23_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136426048)))];
            tensor<fp16, [1, 256, 512]> var_10013_cast_fp16 = mul(x = out_335_cast_fp16, y = tm_layers_23_post_feedforward_layernorm_weight_to_fp16)[name = string("op_10013_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_285_cast_fp16 = add(x = x_379_cast_fp16, y = var_10013_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136427136))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136689344))))[name = string("tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_215_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_input_gate_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = string("linear_215_cast_fp16")];
            string var_10020_mode_0 = const()[name = string("op_10020_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 256, 512]> var_10020_cast_fp16 = gelu(mode = var_10020_mode_0, x = linear_215_cast_fp16)[name = string("op_10020_cast_fp16")];
            tensor<int32, [4]> var_10033_begin_0 = const()[name = string("op_10033_begin_0"), val = tensor<int32, [4]>([0, 0, 23, 0])];
            tensor<int32, [4]> var_10033_end_0 = const()[name = string("op_10033_end_0"), val = tensor<int32, [4]>([1, 256, 24, 512])];
            tensor<bool, [4]> var_10033_end_mask_0 = const()[name = string("op_10033_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<bool, [4]> var_10033_squeeze_mask_0 = const()[name = string("op_10033_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
            tensor<fp16, [1, 256, 512]> var_10033_cast_fp16 = slice_by_index(begin = var_10033_begin_0, end = var_10033_end_0, end_mask = var_10033_end_mask_0, squeeze_mask = var_10033_squeeze_mask_0, x = ple_1_cast_fp16)[name = string("op_10033_cast_fp16")];
            tensor<fp16, [1, 256, 512]> input_287_cast_fp16 = mul(x = var_10020_cast_fp16, y = var_10033_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [512, 512]> tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136690432))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136952640))))[name = string("tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 512]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = tm_layers_23_ple_block_per_layer_projection_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10043_cast_fp16 = abs(x = linear_216_cast_fp16)[name = string("op_10043_cast_fp16")];
            bool var_10046 = const()[name = string("op_10046"), val = bool(true)];
            tensor<int32, [1]> var_10047_axes_0 = const()[name = string("op_10047_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_10047_cast_fp16 = reduce_max(axes = var_10047_axes_0, keep_dims = var_10046, x = var_10043_cast_fp16)[name = string("op_10047_cast_fp16")];
            fp16 var_10048_to_fp16 = const()[name = string("op_10048_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_385_cast_fp16 = maximum(x = var_10047_cast_fp16, y = var_10048_to_fp16)[name = string("s_385_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_385_cast_fp16 = real_div(x = linear_216_cast_fp16, y = s_385_cast_fp16)[name = string("y_385_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10051_cast_fp16 = mul(x = y_385_cast_fp16, y = y_385_cast_fp16)[name = string("op_10051_cast_fp16")];
            tensor<int32, [1]> var_10056_axes_0 = const()[name = string("op_10056_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10056_keep_dims_0 = const()[name = string("op_10056_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_10056_cast_fp16 = reduce_mean(axes = var_10056_axes_0, keep_dims = var_10056_keep_dims_0, x = var_10051_cast_fp16)[name = string("op_10056_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10057_cast_fp16 = mul(x = s_385_cast_fp16, y = s_385_cast_fp16)[name = string("op_10057_cast_fp16")];
            fp32 var_10058_epsilon_0 = const()[name = string("op_10058_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_10058_cast_fp16 = inverse(epsilon = var_10058_epsilon_0, x = var_10057_cast_fp16)[name = string("op_10058_cast_fp16")];
            fp16 var_10059_to_fp16 = const()[name = string("op_10059_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_10060_cast_fp16 = mul(x = var_10058_cast_fp16, y = var_10059_to_fp16)[name = string("op_10060_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10062_cast_fp16 = add(x = var_10056_cast_fp16, y = var_10060_cast_fp16)[name = string("op_10062_cast_fp16")];
            fp32 var_10063_epsilon_0 = const()[name = string("op_10063_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_10063_cast_fp16 = rsqrt(epsilon = var_10063_epsilon_0, x = var_10062_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_337_cast_fp16 = mul(x = y_385_cast_fp16, y = var_10063_cast_fp16)[name = string("out_337_cast_fp16")];
            tensor<fp16, [512]> tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16 = const()[name = string("tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136953728)))];
            tensor<fp16, [1, 256, 512]> var_10065_cast_fp16 = mul(x = out_337_cast_fp16, y = tm_layers_23_ple_block_post_per_layer_input_norm_weight_to_fp16)[name = string("op_10065_cast_fp16")];
            tensor<fp16, [1, 256, 512]> h_47_cast_fp16 = add(x = input_285_cast_fp16, y = var_10065_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<fp16, [1]> tm_layers_23_layer_scalar_to_fp16 = const()[name = string("tm_layers_23_layer_scalar_to_fp16"), val = tensor<fp16, [1]>([0x1.d6p-4])];
            tensor<fp16, [1, 256, 512]> x_cast_fp16 = mul(x = h_47_cast_fp16, y = tm_layers_23_layer_scalar_to_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10069_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_10069_cast_fp16")];
            bool var_10072 = const()[name = string("op_10072"), val = bool(true)];
            tensor<int32, [1]> var_10073_axes_0 = const()[name = string("op_10073_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_10073_cast_fp16 = reduce_max(axes = var_10073_axes_0, keep_dims = var_10072, x = var_10069_cast_fp16)[name = string("op_10073_cast_fp16")];
            fp16 var_10074_to_fp16 = const()[name = string("op_10074_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 256, 1]> s_387_cast_fp16 = maximum(x = var_10073_cast_fp16, y = var_10074_to_fp16)[name = string("s_387_cast_fp16")];
            tensor<fp16, [1, 256, 512]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_387_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [1, 256, 512]> var_10077_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_10077_cast_fp16")];
            tensor<int32, [1]> var_10082_axes_0 = const()[name = string("op_10082_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10082_keep_dims_0 = const()[name = string("op_10082_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_10082_cast_fp16 = reduce_mean(axes = var_10082_axes_0, keep_dims = var_10082_keep_dims_0, x = var_10077_cast_fp16)[name = string("op_10082_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10083_cast_fp16 = mul(x = s_387_cast_fp16, y = s_387_cast_fp16)[name = string("op_10083_cast_fp16")];
            fp32 var_10084_epsilon_0 = const()[name = string("op_10084_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_10084_cast_fp16 = inverse(epsilon = var_10084_epsilon_0, x = var_10083_cast_fp16)[name = string("op_10084_cast_fp16")];
            fp16 var_10085_to_fp16 = const()[name = string("op_10085_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_10086_cast_fp16 = mul(x = var_10084_cast_fp16, y = var_10085_to_fp16)[name = string("op_10086_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_10088_cast_fp16 = add(x = var_10082_cast_fp16, y = var_10086_cast_fp16)[name = string("op_10088_cast_fp16")];
            fp32 var_10089_epsilon_0 = const()[name = string("op_10089_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_10089_cast_fp16 = rsqrt(epsilon = var_10089_epsilon_0, x = var_10088_cast_fp16)[name = string("op_10089_cast_fp16")];
            tensor<fp16, [1, 256, 512]> out_cast_fp16 = mul(x = y_cast_fp16, y = var_10089_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [512]> tm_norm_weight_to_fp16 = const()[name = string("tm_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136954816)))];
            tensor<fp16, [1, 256, 512]> h_cast_fp16 = mul(x = out_cast_fp16, y = tm_norm_weight_to_fp16)[name = string("h_cast_fp16")];
            tensor<int32, [3]> var_10094_begin_0 = const()[name = string("op_10094_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_10094_end_0 = const()[name = string("op_10094_end_0"), val = tensor<int32, [3]>([1, 256, 512])];
            tensor<bool, [3]> var_10094_end_mask_0 = const()[name = string("op_10094_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> var_10094_squeeze_mask_0 = const()[name = string("op_10094_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<fp16, [256, 512]> var_10094_cast_fp16 = slice_by_index(begin = var_10094_begin_0, end = var_10094_end_0, end_mask = var_10094_end_mask_0, squeeze_mask = var_10094_squeeze_mask_0, x = h_cast_fp16)[name = string("op_10094_cast_fp16")];
            bool input_transpose_x_0 = const()[name = string("input_transpose_x_0"), val = bool(false)];
            bool input_transpose_y_0 = const()[name = string("input_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 512]> input_cast_fp16 = matmul(transpose_x = input_transpose_x_0, transpose_y = input_transpose_y_0, x = pool, y = var_10094_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [768, 512]> tm_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136955904))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137349184))))[name = string("tm_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_217_bias_0_to_fp16 = const()[name = string("linear_217_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137350784)))];
            tensor<fp16, [8, 768]> linear_217_cast_fp16 = linear(bias = linear_217_bias_0_to_fp16, weight = tm_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<fp16, [8, 768]> var_10099_cast_fp16 = abs(x = linear_217_cast_fp16)[name = string("op_10099_cast_fp16")];
            bool var_10102 = const()[name = string("op_10102"), val = bool(true)];
            tensor<int32, [1]> var_10103_axes_0 = const()[name = string("op_10103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [8, 1]> var_10103_cast_fp16 = reduce_max(axes = var_10103_axes_0, keep_dims = var_10102, x = var_10099_cast_fp16)[name = string("op_10103_cast_fp16")];
            fp16 var_10104_to_fp16 = const()[name = string("op_10104_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [8, 1]> s_cast_fp16 = maximum(x = var_10103_cast_fp16, y = var_10104_to_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [8, 768]> pooled_cast_fp16 = real_div(x = linear_217_cast_fp16, y = s_cast_fp16)[name = string("pooled_cast_fp16")];
            tensor<fp16, [8, 768]> var_10107_cast_fp16 = mul(x = pooled_cast_fp16, y = pooled_cast_fp16)[name = string("op_10107_cast_fp16")];
            tensor<int32, [1]> var_10112_axes_0 = const()[name = string("op_10112_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10112_keep_dims_0 = const()[name = string("op_10112_keep_dims_0"), val = bool(true)];
            tensor<fp16, [8, 1]> var_10112_cast_fp16 = reduce_sum(axes = var_10112_axes_0, keep_dims = var_10112_keep_dims_0, x = var_10107_cast_fp16)[name = string("op_10112_cast_fp16")];
            fp16 var_10114_to_fp16 = const()[name = string("op_10114_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [8, 1]> var_10115_cast_fp16 = add(x = var_10112_cast_fp16, y = var_10114_to_fp16)[name = string("op_10115_cast_fp16")];
            fp32 var_10116_epsilon_0 = const()[name = string("op_10116_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [8, 1]> var_10116_cast_fp16 = rsqrt(epsilon = var_10116_epsilon_0, x = var_10115_cast_fp16)[name = string("op_10116_cast_fp16")];
            tensor<fp16, [8, 768]> var_10117_cast_fp16 = mul(x = pooled_cast_fp16, y = var_10116_cast_fp16)[name = string("op_10117_cast_fp16")];
            string var_10117_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_10117_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [8, 768]> embedding = cast(dtype = var_10117_cast_fp16_to_fp32_dtype_0, x = var_10117_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
}