program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.1"}})]
{
    func main<ios18>(tensor<fp16, [1, 188, 1, 1]> key_bias, tensor<fp16, [1, 128, 1, 1500]> mel, tensor<fp16, [1, 1, 1, 188]> pad_mask) {
            tensor<int32, [4]> var_70 = const()[name = string("op_70"), val = tensor<int32, [4]>([1, 1, 128, 1500])];
            tensor<fp16, [1, 1, 128, 1500]> input_1_cast_fp16 = reshape(shape = var_70, x = mel)[name = string("input_1_cast_fp16")];
            string input_3_pad_type_0 = const()[name = string("input_3_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> input_3_strides_0 = const()[name = string("input_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> input_3_dilations_0 = const()[name = string("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_3_groups_0 = const()[name = string("input_3_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1280))))[name = string("pre_encode_conv0_weight_to_fp16_quantized")];
            tensor<fp16, [256]> pre_encode_conv0_bias_to_fp16 = const()[name = string("pre_encode_conv0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1856)))];
            tensor<fp16, [1, 256, 64, 750]> input_3_cast_fp16 = conv(bias = pre_encode_conv0_bias_to_fp16, dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = pre_encode_conv0_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 256, 64, 750]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(256)];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2432))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3648))))[name = string("pre_encode_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [256]> pre_encode_conv2_bias_to_fp16 = const()[name = string("pre_encode_conv2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224)))];
            tensor<fp16, [1, 256, 32, 375]> input_7_cast_fp16 = conv(bias = pre_encode_conv2_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = pre_encode_conv2_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
            string input_9_pad_type_0 = const()[name = string("input_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_9_strides_0 = const()[name = string("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_9_dilations_0 = const()[name = string("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_9_groups_0 = const()[name = string("input_9_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4800))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37632))))[name = string("pre_encode_conv3_weight_to_fp16_quantized")];
            tensor<fp16, [256]> pre_encode_conv3_bias_to_fp16 = const()[name = string("pre_encode_conv3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38208)))];
            tensor<fp16, [1, 256, 32, 375]> input_9_cast_fp16 = conv(bias = pre_encode_conv3_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = pre_encode_conv3_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 256, 32, 375]> input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(256)];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38784))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40000))))[name = string("pre_encode_conv5_weight_to_fp16_quantized")];
            tensor<fp16, [256]> pre_encode_conv5_bias_to_fp16 = const()[name = string("pre_encode_conv5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40576)))];
            tensor<fp16, [1, 256, 16, 188]> input_13_cast_fp16 = conv(bias = pre_encode_conv5_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = pre_encode_conv5_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
            string input_15_pad_type_0 = const()[name = string("input_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_15_strides_0 = const()[name = string("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_15_dilations_0 = const()[name = string("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_15_groups_0 = const()[name = string("input_15_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41152))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73984))))[name = string("pre_encode_conv6_weight_to_fp16_quantized")];
            tensor<fp16, [256]> pre_encode_conv6_bias_to_fp16 = const()[name = string("pre_encode_conv6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74560)))];
            tensor<fp16, [1, 256, 16, 188]> input_15_cast_fp16 = conv(bias = pre_encode_conv6_bias_to_fp16, dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = pre_encode_conv6_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 256, 16, 188]> x_1_cast_fp16 = relu(x = input_15_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> var_118 = const()[name = string("op_118"), val = tensor<int32, [4]>([1, 4096, 1, 188])];
            tensor<fp16, [1, 4096, 1, 188]> input_17_cast_fp16 = reshape(shape = var_118, x = x_1_cast_fp16)[name = string("input_17_cast_fp16")];
            string x_3_pad_type_0 = const()[name = string("x_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_3_strides_0 = const()[name = string("x_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_3_pad_0 = const()[name = string("x_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_3_dilations_0 = const()[name = string("x_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_3_groups_0 = const()[name = string("x_3_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75136))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2172352))))[name = string("pre_encode_out_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> pre_encode_out_bias_to_fp16 = const()[name = string("pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2174464)))];
            tensor<fp16, [1, 1024, 1, 188]> x_3_cast_fp16 = conv(bias = pre_encode_out_bias_to_fp16, dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = pre_encode_out_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("x_3_cast_fp16")];
            int32 var_143 = const()[name = string("op_143"), val = int32(1)];
            tensor<int32, [1]> var_170_axes_0 = const()[name = string("op_170_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_0_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_0_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_170_cast_fp16 = layer_norm(axes = var_170_axes_0, epsilon = layers_0_norm_feed_forward1_eps_scaled_to_fp16, x = x_3_cast_fp16)[name = string("op_170_cast_fp16")];
            tensor<fp16, [1024]> input_19_mean_0_to_fp16 = const()[name = string("input_19_mean_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2176576)))];
            tensor<fp16, [1024]> input_19_variance_0_to_fp16 = const()[name = string("input_19_variance_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2178688)))];
            tensor<fp16, [1024]> input_19_gamma_0_to_fp16 = const()[name = string("input_19_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2180800)))];
            tensor<fp16, [1024]> input_19_beta_0_to_fp16 = const()[name = string("input_19_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2182912)))];
            fp16 input_19_epsilon_0_to_fp16 = const()[name = string("input_19_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_19_cast_fp16 = batch_norm(beta = input_19_beta_0_to_fp16, epsilon = input_19_epsilon_0_to_fp16, gamma = input_19_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_170_cast_fp16)[name = string("input_19_cast_fp16")];
            string input_21_pad_type_0 = const()[name = string("input_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_21_strides_0 = const()[name = string("input_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_21_pad_0 = const()[name = string("input_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_21_dilations_0 = const()[name = string("input_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_21_groups_0 = const()[name = string("input_21_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2185024))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4282240))))[name = string("layers_0_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_21_cast_fp16 = conv(dilations = input_21_dilations_0, groups = input_21_groups_0, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = input_21_strides_0, weight = layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_23_cast_fp16 = silu(x = input_21_cast_fp16)[name = string("input_23_cast_fp16")];
            string var_187_pad_type_0 = const()[name = string("op_187_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_187_strides_0 = const()[name = string("op_187_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_187_pad_0 = const()[name = string("op_187_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_187_dilations_0 = const()[name = string("op_187_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_187_groups_0 = const()[name = string("op_187_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_188_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4290496))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6387712))))[name = string("op_188_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_188_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_187_dilations_0, groups = var_187_groups_0, pad = var_187_pad_0, pad_type = var_187_pad_type_0, strides = var_187_strides_0, weight = op_188_weight_0_to_fp16_quantized, x = input_23_cast_fp16)[name = string("op_188_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = var_188_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<int32, [1]> var_204_axes_0 = const()[name = string("op_204_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_0_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_0_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_204_cast_fp16 = layer_norm(axes = var_204_axes_0, epsilon = layers_0_norm_self_att_eps_scaled_to_fp16, x = x_5_cast_fp16)[name = string("op_204_cast_fp16")];
            tensor<fp16, [1024]> x_7_gamma_0_to_fp16 = const()[name = string("x_7_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6389824)))];
            tensor<fp16, [1024]> x_7_beta_0_to_fp16 = const()[name = string("x_7_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6391936)))];
            fp16 x_7_epsilon_0_to_fp16 = const()[name = string("x_7_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_7_cast_fp16 = batch_norm(beta = x_7_beta_0_to_fp16, epsilon = x_7_epsilon_0_to_fp16, gamma = x_7_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_204_cast_fp16)[name = string("x_7_cast_fp16")];
            string q_1_pad_type_0 = const()[name = string("q_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_1_strides_0 = const()[name = string("q_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_1_pad_0 = const()[name = string("q_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_1_dilations_0 = const()[name = string("q_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_1_groups_0 = const()[name = string("q_1_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6394048))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6918400))))[name = string("layers_0_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_1_cast_fp16 = conv(dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = x_7_cast_fp16)[name = string("q_1_cast_fp16")];
            string k_1_pad_type_0 = const()[name = string("k_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_1_strides_0 = const()[name = string("k_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_1_pad_0 = const()[name = string("k_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_1_dilations_0 = const()[name = string("k_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_1_groups_0 = const()[name = string("k_1_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6920512))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7444864))))[name = string("layers_0_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = x_7_cast_fp16)[name = string("k_1_cast_fp16")];
            string v_1_pad_type_0 = const()[name = string("v_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_1_strides_0 = const()[name = string("v_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_1_pad_0 = const()[name = string("v_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_1_dilations_0 = const()[name = string("v_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_1_groups_0 = const()[name = string("v_1_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7446976))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7971328))))[name = string("layers_0_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_1_cast_fp16 = conv(dilations = v_1_dilations_0, groups = v_1_groups_0, pad = v_1_pad_0, pad_type = v_1_pad_type_0, strides = v_1_strides_0, weight = layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = x_7_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_1_to_fp16 = const()[name = string("bv_all_1_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7973440)))];
            tensor<fp16, [1, 1024, 1, 188]> var_237_cast_fp16 = add(x = q_1_cast_fp16, y = bv_all_1_to_fp16)[name = string("op_237_cast_fp16")];
            tensor<int32, [3]> var_238 = const()[name = string("op_238"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_1_cast_fp16 = reshape(shape = var_238, x = var_237_cast_fp16)[name = string("qb_1_cast_fp16")];
            bool bd_all_1_transpose_x_0 = const()[name = string("bd_all_1_transpose_x_0"), val = bool(false)];
            bool bd_all_1_transpose_y_0 = const()[name = string("bd_all_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_0_self_attn_pos_proj_to_fp16 = const()[name = string("layers_0_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7975552)))];
            tensor<fp16, [8, 375, 188]> bd_all_1_cast_fp16 = matmul(transpose_x = bd_all_1_transpose_x_0, transpose_y = bd_all_1_transpose_y_0, x = layers_0_self_attn_pos_proj_to_fp16, y = qb_1_cast_fp16)[name = string("bd_all_1_cast_fp16")];
            tensor<int32, [3]> x_9_perm_0 = const()[name = string("x_9_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_11_pad_0 = const()[name = string("x_11_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_11_mode_0 = const()[name = string("x_11_mode_0"), val = string("constant")];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = bd_all_1_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [8, 188, 376]> x_11_cast_fp16 = pad(constant_val = const_10_to_fp16, mode = x_11_mode_0, pad = x_11_pad_0, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [3]> var_245 = const()[name = string("op_245"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_13_cast_fp16 = reshape(shape = var_245, x = x_11_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<int32, [3]> var_248_begin_0 = const()[name = string("op_248_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_248_end_0 = const()[name = string("op_248_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_248_end_mask_0 = const()[name = string("op_248_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_248_cast_fp16 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = x_13_cast_fp16)[name = string("op_248_cast_fp16")];
            tensor<int32, [3]> var_249 = const()[name = string("op_249"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_15_cast_fp16 = reshape(shape = var_249, x = var_248_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [3]> bd_all_3_begin_0 = const()[name = string("bd_all_3_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_3_end_0 = const()[name = string("bd_all_3_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_3_end_mask_0 = const()[name = string("bd_all_3_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_3_cast_fp16 = slice_by_index(begin = bd_all_3_begin_0, end = bd_all_3_end_0, end_mask = bd_all_3_end_mask_0, x = x_15_cast_fp16)[name = string("bd_all_3_cast_fp16")];
            tensor<int32, [4]> var_254 = const()[name = string("op_254"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_255_cast_fp16 = reshape(shape = var_254, x = q_1_cast_fp16)[name = string("op_255_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_257_to_fp16 = const()[name = string("op_257_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8743616)))];
            tensor<fp16, [8, 128, 1, 188]> qh_1_cast_fp16 = add(x = var_255_cast_fp16, y = var_257_to_fp16)[name = string("qh_1_cast_fp16")];
            tensor<int32, [4]> var_259 = const()[name = string("op_259"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_260_cast_fp16 = reshape(shape = var_259, x = k_1_cast_fp16)[name = string("op_260_cast_fp16")];
            tensor<int32, [4]> var_261 = const()[name = string("op_261"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_263 = const()[name = string("op_263"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_264_cast_fp16 = reshape(shape = var_263, x = v_1_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_267_perm_0 = const()[name = string("op_267_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_1_axes_0 = const()[name = string("bd_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_267_cast_fp16 = transpose(perm = var_267_perm_0, x = bd_all_3_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [8, 188, 1, 188]> bd_1_cast_fp16 = expand_dims(axes = bd_1_axes_0, x = var_267_cast_fp16)[name = string("bd_1_cast_fp16")];
            string ac_1_equation_0 = const()[name = string("ac_1_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_262_cast_fp16 = transpose(perm = var_261, x = var_260_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [8, 188, 1, 188]> ac_1_cast_fp16 = einsum(equation = ac_1_equation_0, values = (var_262_cast_fp16, qh_1_cast_fp16))[name = string("ac_1_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_283_cast_fp16 = add(x = ac_1_cast_fp16, y = bd_1_cast_fp16)[name = string("op_283_cast_fp16")];
            fp16 var_284_to_fp16 = const()[name = string("op_284_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_1_cast_fp16 = mul(x = var_283_cast_fp16, y = var_284_to_fp16)[name = string("scores_1_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = key_bias)[name = string("scores_3_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_287_cast_fp16 = softmax(axis = var_143, x = scores_3_cast_fp16)[name = string("op_287_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = var_287_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [8, 188, 188]> reshape_0_cast_fp16 = reshape(shape = concat_4, x = transpose_0_cast_fp16)[name = string("reshape_0_cast_fp16")];
            tensor<int32, [3]> concat_5 = const()[name = string("concat_5"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = var_264_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [8, 188, 128]> reshape_1_cast_fp16 = reshape(shape = concat_5, x = transpose_48_cast_fp16)[name = string("reshape_1_cast_fp16")];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = reshape_0_cast_fp16, y = reshape_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<int32, [4]> concat_9 = const()[name = string("concat_9"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_2_cast_fp16 = reshape(shape = concat_9, x = matmul_0_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [4]> ctx_1_perm_0 = const()[name = string("ctx_1_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_290 = const()[name = string("op_290"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_1_cast_fp16 = transpose(perm = ctx_1_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 1024, 1, 188]> input_25_cast_fp16 = reshape(shape = var_290, x = ctx_1_cast_fp16)[name = string("input_25_cast_fp16")];
            string var_297_pad_type_0 = const()[name = string("op_297_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_297_strides_0 = const()[name = string("op_297_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_297_pad_0 = const()[name = string("op_297_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_297_dilations_0 = const()[name = string("op_297_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_297_groups_0 = const()[name = string("op_297_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8745728))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9270080))))[name = string("layers_0_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_297_cast_fp16 = conv(dilations = var_297_dilations_0, groups = var_297_groups_0, pad = var_297_pad_0, pad_type = var_297_pad_type_0, strides = var_297_strides_0, weight = layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_17_cast_fp16 = add(x = x_5_cast_fp16, y = var_297_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<int32, [1]> var_313_axes_0 = const()[name = string("op_313_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_0_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_0_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_313_cast_fp16 = layer_norm(axes = var_313_axes_0, epsilon = layers_0_norm_conv_eps_scaled_to_fp16, x = x_17_cast_fp16)[name = string("op_313_cast_fp16")];
            tensor<fp16, [1024]> input_27_gamma_0_to_fp16 = const()[name = string("input_27_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9272192)))];
            tensor<fp16, [1024]> input_27_beta_0_to_fp16 = const()[name = string("input_27_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9274304)))];
            fp16 input_27_epsilon_0_to_fp16 = const()[name = string("input_27_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_27_cast_fp16 = batch_norm(beta = input_27_beta_0_to_fp16, epsilon = input_27_epsilon_0_to_fp16, gamma = input_27_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_313_cast_fp16)[name = string("input_27_cast_fp16")];
            string input_29_pad_type_0 = const()[name = string("input_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_29_strides_0 = const()[name = string("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_29_pad_0 = const()[name = string("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_29_dilations_0 = const()[name = string("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_29_groups_0 = const()[name = string("input_29_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9276416))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10325056))))[name = string("layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_29_cast_fp16 = conv(dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
            int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)];
            int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_19_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_29_cast_fp16)[name = string("x_19_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_31_cast_fp16 = mul(x = x_19_cast_fp16, y = pad_mask)[name = string("input_31_cast_fp16")];
            string input_33_pad_type_0 = const()[name = string("input_33_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_33_pad_0 = const()[name = string("input_33_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_33_groups_0 = const()[name = string("input_33_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_33_strides_0 = const()[name = string("input_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_33_dilations_0 = const()[name = string("input_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_103_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10329216))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10333888))))[name = string("const_103_to_fp16_quantized")];
            tensor<fp16, [1024]> const_104_to_fp16 = const()[name = string("const_104_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10336000)))];
            tensor<fp16, [1, 1024, 1, 188]> input_35_cast_fp16 = conv(bias = const_104_to_fp16, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = const_103_to_fp16_quantized, x = input_31_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_37_cast_fp16 = silu(x = input_35_cast_fp16)[name = string("input_37_cast_fp16")];
            string var_345_pad_type_0 = const()[name = string("op_345_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_345_strides_0 = const()[name = string("op_345_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_345_pad_0 = const()[name = string("op_345_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_345_dilations_0 = const()[name = string("op_345_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_345_groups_0 = const()[name = string("op_345_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10338112))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10862464))))[name = string("layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_345_cast_fp16 = conv(dilations = var_345_dilations_0, groups = var_345_groups_0, pad = var_345_pad_0, pad_type = var_345_pad_type_0, strides = var_345_strides_0, weight = layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("op_345_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_21_cast_fp16 = add(x = x_17_cast_fp16, y = var_345_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [1]> var_361_axes_0 = const()[name = string("op_361_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_0_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_0_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_361_cast_fp16 = layer_norm(axes = var_361_axes_0, epsilon = layers_0_norm_feed_forward2_eps_scaled_to_fp16, x = x_21_cast_fp16)[name = string("op_361_cast_fp16")];
            tensor<fp16, [1024]> input_39_gamma_0_to_fp16 = const()[name = string("input_39_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10864576)))];
            tensor<fp16, [1024]> input_39_beta_0_to_fp16 = const()[name = string("input_39_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10866688)))];
            fp16 input_39_epsilon_0_to_fp16 = const()[name = string("input_39_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_39_cast_fp16 = batch_norm(beta = input_39_beta_0_to_fp16, epsilon = input_39_epsilon_0_to_fp16, gamma = input_39_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_361_cast_fp16)[name = string("input_39_cast_fp16")];
            string input_41_pad_type_0 = const()[name = string("input_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_41_strides_0 = const()[name = string("input_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_41_pad_0 = const()[name = string("input_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_41_dilations_0 = const()[name = string("input_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_41_groups_0 = const()[name = string("input_41_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10868800))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12966016))))[name = string("layers_0_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_41_cast_fp16 = conv(dilations = input_41_dilations_0, groups = input_41_groups_0, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = input_41_strides_0, weight = layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_43_cast_fp16 = silu(x = input_41_cast_fp16)[name = string("input_43_cast_fp16")];
            string var_378_pad_type_0 = const()[name = string("op_378_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_378_strides_0 = const()[name = string("op_378_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_378_pad_0 = const()[name = string("op_378_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_378_dilations_0 = const()[name = string("op_378_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_378_groups_0 = const()[name = string("op_378_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_379_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12974272))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15071488))))[name = string("op_379_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_379_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_378_dilations_0, groups = var_378_groups_0, pad = var_378_pad_0, pad_type = var_378_pad_type_0, strides = var_378_strides_0, weight = op_379_weight_0_to_fp16_quantized, x = input_43_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = var_379_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<int32, [1]> var_395_axes_0 = const()[name = string("op_395_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_0_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_0_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_395_cast_fp16 = layer_norm(axes = var_395_axes_0, epsilon = layers_0_norm_out_eps_scaled_to_fp16, x = x_23_cast_fp16)[name = string("op_395_cast_fp16")];
            tensor<fp16, [1024]> x_25_gamma_0_to_fp16 = const()[name = string("x_25_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15073600)))];
            tensor<fp16, [1024]> x_25_beta_0_to_fp16 = const()[name = string("x_25_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15075712)))];
            fp16 x_25_epsilon_0_to_fp16 = const()[name = string("x_25_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_25_cast_fp16 = batch_norm(beta = x_25_beta_0_to_fp16, epsilon = x_25_epsilon_0_to_fp16, gamma = x_25_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_395_cast_fp16)[name = string("x_25_cast_fp16")];
            int32 var_414 = const()[name = string("op_414"), val = int32(1)];
            tensor<int32, [1]> var_441_axes_0 = const()[name = string("op_441_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_1_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_1_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_441_cast_fp16 = layer_norm(axes = var_441_axes_0, epsilon = layers_1_norm_feed_forward1_eps_scaled_to_fp16, x = x_25_cast_fp16)[name = string("op_441_cast_fp16")];
            tensor<fp16, [1024]> input_45_gamma_0_to_fp16 = const()[name = string("input_45_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15077824)))];
            tensor<fp16, [1024]> input_45_beta_0_to_fp16 = const()[name = string("input_45_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15079936)))];
            fp16 input_45_epsilon_0_to_fp16 = const()[name = string("input_45_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_441_cast_fp16)[name = string("input_45_cast_fp16")];
            string input_47_pad_type_0 = const()[name = string("input_47_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_47_strides_0 = const()[name = string("input_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_47_pad_0 = const()[name = string("input_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_47_dilations_0 = const()[name = string("input_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_47_groups_0 = const()[name = string("input_47_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15082048))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17179264))))[name = string("layers_1_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_47_cast_fp16 = conv(dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_49_cast_fp16 = silu(x = input_47_cast_fp16)[name = string("input_49_cast_fp16")];
            string var_458_pad_type_0 = const()[name = string("op_458_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_458_strides_0 = const()[name = string("op_458_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_458_pad_0 = const()[name = string("op_458_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_458_dilations_0 = const()[name = string("op_458_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_458_groups_0 = const()[name = string("op_458_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_459_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17187520))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19284736))))[name = string("op_459_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_459_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_458_dilations_0, groups = var_458_groups_0, pad = var_458_pad_0, pad_type = var_458_pad_type_0, strides = var_458_strides_0, weight = op_459_weight_0_to_fp16_quantized, x = input_49_cast_fp16)[name = string("op_459_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = var_459_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<int32, [1]> var_475_axes_0 = const()[name = string("op_475_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_1_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_1_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_475_cast_fp16 = layer_norm(axes = var_475_axes_0, epsilon = layers_1_norm_self_att_eps_scaled_to_fp16, x = x_27_cast_fp16)[name = string("op_475_cast_fp16")];
            tensor<fp16, [1024]> x_29_gamma_0_to_fp16 = const()[name = string("x_29_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19286848)))];
            tensor<fp16, [1024]> x_29_beta_0_to_fp16 = const()[name = string("x_29_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19288960)))];
            fp16 x_29_epsilon_0_to_fp16 = const()[name = string("x_29_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_29_cast_fp16 = batch_norm(beta = x_29_beta_0_to_fp16, epsilon = x_29_epsilon_0_to_fp16, gamma = x_29_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_475_cast_fp16)[name = string("x_29_cast_fp16")];
            string q_3_pad_type_0 = const()[name = string("q_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_3_strides_0 = const()[name = string("q_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_3_pad_0 = const()[name = string("q_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_3_dilations_0 = const()[name = string("q_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_3_groups_0 = const()[name = string("q_3_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19291072))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19815424))))[name = string("layers_1_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_3_cast_fp16 = conv(dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = x_29_cast_fp16)[name = string("q_3_cast_fp16")];
            string k_3_pad_type_0 = const()[name = string("k_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_3_strides_0 = const()[name = string("k_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_3_pad_0 = const()[name = string("k_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_3_dilations_0 = const()[name = string("k_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_3_groups_0 = const()[name = string("k_3_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19817536))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20341888))))[name = string("layers_1_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = x_29_cast_fp16)[name = string("k_3_cast_fp16")];
            string v_3_pad_type_0 = const()[name = string("v_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_3_strides_0 = const()[name = string("v_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_3_pad_0 = const()[name = string("v_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_3_dilations_0 = const()[name = string("v_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_3_groups_0 = const()[name = string("v_3_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20344000))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20868352))))[name = string("layers_1_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_3_cast_fp16 = conv(dilations = v_3_dilations_0, groups = v_3_groups_0, pad = v_3_pad_0, pad_type = v_3_pad_type_0, strides = v_3_strides_0, weight = layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = x_29_cast_fp16)[name = string("v_3_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_3_to_fp16 = const()[name = string("bv_all_3_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20870464)))];
            tensor<fp16, [1, 1024, 1, 188]> var_508_cast_fp16 = add(x = q_3_cast_fp16, y = bv_all_3_to_fp16)[name = string("op_508_cast_fp16")];
            tensor<int32, [3]> var_509 = const()[name = string("op_509"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_3_cast_fp16 = reshape(shape = var_509, x = var_508_cast_fp16)[name = string("qb_3_cast_fp16")];
            bool bd_all_5_transpose_x_0 = const()[name = string("bd_all_5_transpose_x_0"), val = bool(false)];
            bool bd_all_5_transpose_y_0 = const()[name = string("bd_all_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_1_self_attn_pos_proj_to_fp16 = const()[name = string("layers_1_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20872576)))];
            tensor<fp16, [8, 375, 188]> bd_all_5_cast_fp16 = matmul(transpose_x = bd_all_5_transpose_x_0, transpose_y = bd_all_5_transpose_y_0, x = layers_1_self_attn_pos_proj_to_fp16, y = qb_3_cast_fp16)[name = string("bd_all_5_cast_fp16")];
            tensor<int32, [3]> x_31_perm_0 = const()[name = string("x_31_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_33_pad_0 = const()[name = string("x_33_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_33_mode_0 = const()[name = string("x_33_mode_0"), val = string("constant")];
            fp16 const_14_to_fp16 = const()[name = string("const_14_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_31_cast_fp16 = transpose(perm = x_31_perm_0, x = bd_all_5_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [8, 188, 376]> x_33_cast_fp16 = pad(constant_val = const_14_to_fp16, mode = x_33_mode_0, pad = x_33_pad_0, x = x_31_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [3]> var_516 = const()[name = string("op_516"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_35_cast_fp16 = reshape(shape = var_516, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<int32, [3]> var_519_begin_0 = const()[name = string("op_519_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_519_end_0 = const()[name = string("op_519_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_519_end_mask_0 = const()[name = string("op_519_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_519_cast_fp16 = slice_by_index(begin = var_519_begin_0, end = var_519_end_0, end_mask = var_519_end_mask_0, x = x_35_cast_fp16)[name = string("op_519_cast_fp16")];
            tensor<int32, [3]> var_520 = const()[name = string("op_520"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_37_cast_fp16 = reshape(shape = var_520, x = var_519_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<int32, [3]> bd_all_7_begin_0 = const()[name = string("bd_all_7_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_7_end_0 = const()[name = string("bd_all_7_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_7_end_mask_0 = const()[name = string("bd_all_7_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_7_cast_fp16 = slice_by_index(begin = bd_all_7_begin_0, end = bd_all_7_end_0, end_mask = bd_all_7_end_mask_0, x = x_37_cast_fp16)[name = string("bd_all_7_cast_fp16")];
            tensor<int32, [4]> var_525 = const()[name = string("op_525"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_526_cast_fp16 = reshape(shape = var_525, x = q_3_cast_fp16)[name = string("op_526_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_528_to_fp16 = const()[name = string("op_528_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21640640)))];
            tensor<fp16, [8, 128, 1, 188]> qh_3_cast_fp16 = add(x = var_526_cast_fp16, y = var_528_to_fp16)[name = string("qh_3_cast_fp16")];
            tensor<int32, [4]> var_530 = const()[name = string("op_530"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_531_cast_fp16 = reshape(shape = var_530, x = k_3_cast_fp16)[name = string("op_531_cast_fp16")];
            tensor<int32, [4]> var_532 = const()[name = string("op_532"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_534 = const()[name = string("op_534"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_535_cast_fp16 = reshape(shape = var_534, x = v_3_cast_fp16)[name = string("op_535_cast_fp16")];
            tensor<int32, [4]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_538_perm_0 = const()[name = string("op_538_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_3_axes_0 = const()[name = string("bd_3_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_538_cast_fp16 = transpose(perm = var_538_perm_0, x = bd_all_7_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [8, 188, 1, 188]> bd_3_cast_fp16 = expand_dims(axes = bd_3_axes_0, x = var_538_cast_fp16)[name = string("bd_3_cast_fp16")];
            string ac_3_equation_0 = const()[name = string("ac_3_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_533_cast_fp16 = transpose(perm = var_532, x = var_531_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [8, 188, 1, 188]> ac_3_cast_fp16 = einsum(equation = ac_3_equation_0, values = (var_533_cast_fp16, qh_3_cast_fp16))[name = string("ac_3_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_554_cast_fp16 = add(x = ac_3_cast_fp16, y = bd_3_cast_fp16)[name = string("op_554_cast_fp16")];
            fp16 var_555_to_fp16 = const()[name = string("op_555_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_5_cast_fp16 = mul(x = var_554_cast_fp16, y = var_555_to_fp16)[name = string("scores_5_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = key_bias)[name = string("scores_7_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_558_cast_fp16 = softmax(axis = var_414, x = scores_7_cast_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_14 = const()[name = string("concat_14"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = var_558_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [8, 188, 188]> reshape_3_cast_fp16 = reshape(shape = concat_14, x = transpose_2_cast_fp16)[name = string("reshape_3_cast_fp16")];
            tensor<int32, [3]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = var_535_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [8, 188, 128]> reshape_4_cast_fp16 = reshape(shape = concat_15, x = transpose_49_cast_fp16)[name = string("reshape_4_cast_fp16")];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = reshape_3_cast_fp16, y = reshape_4_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<int32, [4]> concat_19 = const()[name = string("concat_19"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_5_cast_fp16 = reshape(shape = concat_19, x = matmul_1_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> ctx_3_perm_0 = const()[name = string("ctx_3_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_561 = const()[name = string("op_561"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_3_cast_fp16 = transpose(perm = ctx_3_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 1024, 1, 188]> input_51_cast_fp16 = reshape(shape = var_561, x = ctx_3_cast_fp16)[name = string("input_51_cast_fp16")];
            string var_568_pad_type_0 = const()[name = string("op_568_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_568_strides_0 = const()[name = string("op_568_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_568_pad_0 = const()[name = string("op_568_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_568_dilations_0 = const()[name = string("op_568_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_568_groups_0 = const()[name = string("op_568_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21642752))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22167104))))[name = string("layers_1_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_568_cast_fp16 = conv(dilations = var_568_dilations_0, groups = var_568_groups_0, pad = var_568_pad_0, pad_type = var_568_pad_type_0, strides = var_568_strides_0, weight = layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("op_568_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_39_cast_fp16 = add(x = x_27_cast_fp16, y = var_568_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [1]> var_584_axes_0 = const()[name = string("op_584_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_1_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_1_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_584_cast_fp16 = layer_norm(axes = var_584_axes_0, epsilon = layers_1_norm_conv_eps_scaled_to_fp16, x = x_39_cast_fp16)[name = string("op_584_cast_fp16")];
            tensor<fp16, [1024]> input_53_gamma_0_to_fp16 = const()[name = string("input_53_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22169216)))];
            tensor<fp16, [1024]> input_53_beta_0_to_fp16 = const()[name = string("input_53_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22171328)))];
            fp16 input_53_epsilon_0_to_fp16 = const()[name = string("input_53_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_53_cast_fp16 = batch_norm(beta = input_53_beta_0_to_fp16, epsilon = input_53_epsilon_0_to_fp16, gamma = input_53_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_584_cast_fp16)[name = string("input_53_cast_fp16")];
            string input_55_pad_type_0 = const()[name = string("input_55_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_55_strides_0 = const()[name = string("input_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_55_pad_0 = const()[name = string("input_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_55_dilations_0 = const()[name = string("input_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_55_groups_0 = const()[name = string("input_55_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22173440))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23222080))))[name = string("layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_55_cast_fp16 = conv(dilations = input_55_dilations_0, groups = input_55_groups_0, pad = input_55_pad_0, pad_type = input_55_pad_type_0, strides = input_55_strides_0, weight = layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("input_55_cast_fp16")];
            int32 x_41_split_num_splits_0 = const()[name = string("x_41_split_num_splits_0"), val = int32(2)];
            int32 x_41_split_axis_0 = const()[name = string("x_41_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_41_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_41_split_cast_fp16_1 = split(axis = x_41_split_axis_0, num_splits = x_41_split_num_splits_0, x = input_55_cast_fp16)[name = string("x_41_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_41_split_1_sigmoid_cast_fp16 = sigmoid(x = x_41_split_cast_fp16_1)[name = string("x_41_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_41_cast_fp16 = mul(x = x_41_split_cast_fp16_0, y = x_41_split_1_sigmoid_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_57_cast_fp16 = mul(x = x_41_cast_fp16, y = pad_mask)[name = string("input_57_cast_fp16")];
            string input_59_pad_type_0 = const()[name = string("input_59_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_59_pad_0 = const()[name = string("input_59_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_59_groups_0 = const()[name = string("input_59_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_59_strides_0 = const()[name = string("input_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_59_dilations_0 = const()[name = string("input_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_105_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23226240))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23230912))))[name = string("const_105_to_fp16_quantized")];
            tensor<fp16, [1024]> const_106_to_fp16 = const()[name = string("const_106_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23233024)))];
            tensor<fp16, [1, 1024, 1, 188]> input_61_cast_fp16 = conv(bias = const_106_to_fp16, dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = const_105_to_fp16_quantized, x = input_57_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_63_cast_fp16 = silu(x = input_61_cast_fp16)[name = string("input_63_cast_fp16")];
            string var_616_pad_type_0 = const()[name = string("op_616_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_616_strides_0 = const()[name = string("op_616_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_616_pad_0 = const()[name = string("op_616_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_616_dilations_0 = const()[name = string("op_616_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_616_groups_0 = const()[name = string("op_616_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23235136))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23759488))))[name = string("layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_616_cast_fp16 = conv(dilations = var_616_dilations_0, groups = var_616_groups_0, pad = var_616_pad_0, pad_type = var_616_pad_type_0, strides = var_616_strides_0, weight = layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("op_616_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_43_cast_fp16 = add(x = x_39_cast_fp16, y = var_616_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<int32, [1]> var_632_axes_0 = const()[name = string("op_632_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_1_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_1_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_632_cast_fp16 = layer_norm(axes = var_632_axes_0, epsilon = layers_1_norm_feed_forward2_eps_scaled_to_fp16, x = x_43_cast_fp16)[name = string("op_632_cast_fp16")];
            tensor<fp16, [1024]> input_65_gamma_0_to_fp16 = const()[name = string("input_65_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23761600)))];
            tensor<fp16, [1024]> input_65_beta_0_to_fp16 = const()[name = string("input_65_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23763712)))];
            fp16 input_65_epsilon_0_to_fp16 = const()[name = string("input_65_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_65_cast_fp16 = batch_norm(beta = input_65_beta_0_to_fp16, epsilon = input_65_epsilon_0_to_fp16, gamma = input_65_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_632_cast_fp16)[name = string("input_65_cast_fp16")];
            string input_67_pad_type_0 = const()[name = string("input_67_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_67_strides_0 = const()[name = string("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_67_pad_0 = const()[name = string("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_67_dilations_0 = const()[name = string("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_67_groups_0 = const()[name = string("input_67_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23765824))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25863040))))[name = string("layers_1_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_67_cast_fp16 = conv(dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_69_cast_fp16 = silu(x = input_67_cast_fp16)[name = string("input_69_cast_fp16")];
            string var_649_pad_type_0 = const()[name = string("op_649_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_649_strides_0 = const()[name = string("op_649_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_649_pad_0 = const()[name = string("op_649_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_649_dilations_0 = const()[name = string("op_649_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_649_groups_0 = const()[name = string("op_649_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_650_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25871296))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27968512))))[name = string("op_650_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_650_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_649_dilations_0, groups = var_649_groups_0, pad = var_649_pad_0, pad_type = var_649_pad_type_0, strides = var_649_strides_0, weight = op_650_weight_0_to_fp16_quantized, x = input_69_cast_fp16)[name = string("op_650_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = var_650_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<int32, [1]> var_666_axes_0 = const()[name = string("op_666_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_1_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_1_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_666_cast_fp16 = layer_norm(axes = var_666_axes_0, epsilon = layers_1_norm_out_eps_scaled_to_fp16, x = x_45_cast_fp16)[name = string("op_666_cast_fp16")];
            tensor<fp16, [1024]> x_47_gamma_0_to_fp16 = const()[name = string("x_47_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27970624)))];
            tensor<fp16, [1024]> x_47_beta_0_to_fp16 = const()[name = string("x_47_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27972736)))];
            fp16 x_47_epsilon_0_to_fp16 = const()[name = string("x_47_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_47_cast_fp16 = batch_norm(beta = x_47_beta_0_to_fp16, epsilon = x_47_epsilon_0_to_fp16, gamma = x_47_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_666_cast_fp16)[name = string("x_47_cast_fp16")];
            int32 var_685 = const()[name = string("op_685"), val = int32(1)];
            tensor<int32, [1]> var_712_axes_0 = const()[name = string("op_712_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_2_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_2_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_712_cast_fp16 = layer_norm(axes = var_712_axes_0, epsilon = layers_2_norm_feed_forward1_eps_scaled_to_fp16, x = x_47_cast_fp16)[name = string("op_712_cast_fp16")];
            tensor<fp16, [1024]> input_71_gamma_0_to_fp16 = const()[name = string("input_71_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27974848)))];
            tensor<fp16, [1024]> input_71_beta_0_to_fp16 = const()[name = string("input_71_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27976960)))];
            fp16 input_71_epsilon_0_to_fp16 = const()[name = string("input_71_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_71_cast_fp16 = batch_norm(beta = input_71_beta_0_to_fp16, epsilon = input_71_epsilon_0_to_fp16, gamma = input_71_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_712_cast_fp16)[name = string("input_71_cast_fp16")];
            string input_73_pad_type_0 = const()[name = string("input_73_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_73_strides_0 = const()[name = string("input_73_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_73_pad_0 = const()[name = string("input_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_73_dilations_0 = const()[name = string("input_73_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_73_groups_0 = const()[name = string("input_73_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27979072))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30076288))))[name = string("layers_2_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_73_cast_fp16 = conv(dilations = input_73_dilations_0, groups = input_73_groups_0, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = input_73_strides_0, weight = layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_75_cast_fp16 = silu(x = input_73_cast_fp16)[name = string("input_75_cast_fp16")];
            string var_729_pad_type_0 = const()[name = string("op_729_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_729_strides_0 = const()[name = string("op_729_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_729_pad_0 = const()[name = string("op_729_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_729_dilations_0 = const()[name = string("op_729_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_729_groups_0 = const()[name = string("op_729_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_730_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30084544))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32181760))))[name = string("op_730_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_730_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_729_dilations_0, groups = var_729_groups_0, pad = var_729_pad_0, pad_type = var_729_pad_type_0, strides = var_729_strides_0, weight = op_730_weight_0_to_fp16_quantized, x = input_75_cast_fp16)[name = string("op_730_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = var_730_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [1]> var_746_axes_0 = const()[name = string("op_746_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_2_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_2_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_746_cast_fp16 = layer_norm(axes = var_746_axes_0, epsilon = layers_2_norm_self_att_eps_scaled_to_fp16, x = x_49_cast_fp16)[name = string("op_746_cast_fp16")];
            tensor<fp16, [1024]> x_51_gamma_0_to_fp16 = const()[name = string("x_51_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32183872)))];
            tensor<fp16, [1024]> x_51_beta_0_to_fp16 = const()[name = string("x_51_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32185984)))];
            fp16 x_51_epsilon_0_to_fp16 = const()[name = string("x_51_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_51_cast_fp16 = batch_norm(beta = x_51_beta_0_to_fp16, epsilon = x_51_epsilon_0_to_fp16, gamma = x_51_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_746_cast_fp16)[name = string("x_51_cast_fp16")];
            string q_5_pad_type_0 = const()[name = string("q_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_5_strides_0 = const()[name = string("q_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_5_pad_0 = const()[name = string("q_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_5_dilations_0 = const()[name = string("q_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_5_groups_0 = const()[name = string("q_5_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32188096))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32712448))))[name = string("layers_2_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_5_cast_fp16 = conv(dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = x_51_cast_fp16)[name = string("q_5_cast_fp16")];
            string k_5_pad_type_0 = const()[name = string("k_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_5_strides_0 = const()[name = string("k_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_5_pad_0 = const()[name = string("k_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_5_dilations_0 = const()[name = string("k_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_5_groups_0 = const()[name = string("k_5_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32714560))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33238912))))[name = string("layers_2_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = x_51_cast_fp16)[name = string("k_5_cast_fp16")];
            string v_5_pad_type_0 = const()[name = string("v_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_5_strides_0 = const()[name = string("v_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_5_pad_0 = const()[name = string("v_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_5_dilations_0 = const()[name = string("v_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_5_groups_0 = const()[name = string("v_5_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33241024))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33765376))))[name = string("layers_2_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_5_cast_fp16 = conv(dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = x_51_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_5_to_fp16 = const()[name = string("bv_all_5_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33767488)))];
            tensor<fp16, [1, 1024, 1, 188]> var_779_cast_fp16 = add(x = q_5_cast_fp16, y = bv_all_5_to_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [3]> var_780 = const()[name = string("op_780"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_5_cast_fp16 = reshape(shape = var_780, x = var_779_cast_fp16)[name = string("qb_5_cast_fp16")];
            bool bd_all_9_transpose_x_0 = const()[name = string("bd_all_9_transpose_x_0"), val = bool(false)];
            bool bd_all_9_transpose_y_0 = const()[name = string("bd_all_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_2_self_attn_pos_proj_to_fp16 = const()[name = string("layers_2_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33769600)))];
            tensor<fp16, [8, 375, 188]> bd_all_9_cast_fp16 = matmul(transpose_x = bd_all_9_transpose_x_0, transpose_y = bd_all_9_transpose_y_0, x = layers_2_self_attn_pos_proj_to_fp16, y = qb_5_cast_fp16)[name = string("bd_all_9_cast_fp16")];
            tensor<int32, [3]> x_53_perm_0 = const()[name = string("x_53_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_55_pad_0 = const()[name = string("x_55_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_55_mode_0 = const()[name = string("x_55_mode_0"), val = string("constant")];
            fp16 const_18_to_fp16 = const()[name = string("const_18_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_53_cast_fp16 = transpose(perm = x_53_perm_0, x = bd_all_9_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [8, 188, 376]> x_55_cast_fp16 = pad(constant_val = const_18_to_fp16, mode = x_55_mode_0, pad = x_55_pad_0, x = x_53_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<int32, [3]> var_787 = const()[name = string("op_787"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_57_cast_fp16 = reshape(shape = var_787, x = x_55_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<int32, [3]> var_790_begin_0 = const()[name = string("op_790_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_790_end_0 = const()[name = string("op_790_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_790_end_mask_0 = const()[name = string("op_790_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_790_cast_fp16 = slice_by_index(begin = var_790_begin_0, end = var_790_end_0, end_mask = var_790_end_mask_0, x = x_57_cast_fp16)[name = string("op_790_cast_fp16")];
            tensor<int32, [3]> var_791 = const()[name = string("op_791"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_59_cast_fp16 = reshape(shape = var_791, x = var_790_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<int32, [3]> bd_all_11_begin_0 = const()[name = string("bd_all_11_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_11_end_0 = const()[name = string("bd_all_11_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_11_end_mask_0 = const()[name = string("bd_all_11_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_11_cast_fp16 = slice_by_index(begin = bd_all_11_begin_0, end = bd_all_11_end_0, end_mask = bd_all_11_end_mask_0, x = x_59_cast_fp16)[name = string("bd_all_11_cast_fp16")];
            tensor<int32, [4]> var_796 = const()[name = string("op_796"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_797_cast_fp16 = reshape(shape = var_796, x = q_5_cast_fp16)[name = string("op_797_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_799_to_fp16 = const()[name = string("op_799_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34537664)))];
            tensor<fp16, [8, 128, 1, 188]> qh_5_cast_fp16 = add(x = var_797_cast_fp16, y = var_799_to_fp16)[name = string("qh_5_cast_fp16")];
            tensor<int32, [4]> var_801 = const()[name = string("op_801"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_802_cast_fp16 = reshape(shape = var_801, x = k_5_cast_fp16)[name = string("op_802_cast_fp16")];
            tensor<int32, [4]> var_803 = const()[name = string("op_803"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_805 = const()[name = string("op_805"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_806_cast_fp16 = reshape(shape = var_805, x = v_5_cast_fp16)[name = string("op_806_cast_fp16")];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_809_perm_0 = const()[name = string("op_809_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_5_axes_0 = const()[name = string("bd_5_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_809_cast_fp16 = transpose(perm = var_809_perm_0, x = bd_all_11_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [8, 188, 1, 188]> bd_5_cast_fp16 = expand_dims(axes = bd_5_axes_0, x = var_809_cast_fp16)[name = string("bd_5_cast_fp16")];
            string ac_5_equation_0 = const()[name = string("ac_5_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_804_cast_fp16 = transpose(perm = var_803, x = var_802_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [8, 188, 1, 188]> ac_5_cast_fp16 = einsum(equation = ac_5_equation_0, values = (var_804_cast_fp16, qh_5_cast_fp16))[name = string("ac_5_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_825_cast_fp16 = add(x = ac_5_cast_fp16, y = bd_5_cast_fp16)[name = string("op_825_cast_fp16")];
            fp16 var_826_to_fp16 = const()[name = string("op_826_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_9_cast_fp16 = mul(x = var_825_cast_fp16, y = var_826_to_fp16)[name = string("scores_9_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = key_bias)[name = string("scores_11_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_829_cast_fp16 = softmax(axis = var_685, x = scores_11_cast_fp16)[name = string("op_829_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_24 = const()[name = string("concat_24"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = var_829_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [8, 188, 188]> reshape_6_cast_fp16 = reshape(shape = concat_24, x = transpose_4_cast_fp16)[name = string("reshape_6_cast_fp16")];
            tensor<int32, [3]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = var_806_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [8, 188, 128]> reshape_7_cast_fp16 = reshape(shape = concat_25, x = transpose_50_cast_fp16)[name = string("reshape_7_cast_fp16")];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_6_cast_fp16, y = reshape_7_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<int32, [4]> concat_29 = const()[name = string("concat_29"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_8_cast_fp16 = reshape(shape = concat_29, x = matmul_2_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [4]> ctx_5_perm_0 = const()[name = string("ctx_5_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_832 = const()[name = string("op_832"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_5_cast_fp16 = transpose(perm = ctx_5_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 1024, 1, 188]> input_77_cast_fp16 = reshape(shape = var_832, x = ctx_5_cast_fp16)[name = string("input_77_cast_fp16")];
            string var_839_pad_type_0 = const()[name = string("op_839_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_839_strides_0 = const()[name = string("op_839_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_839_pad_0 = const()[name = string("op_839_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_839_dilations_0 = const()[name = string("op_839_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_839_groups_0 = const()[name = string("op_839_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34539776))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35064128))))[name = string("layers_2_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_839_cast_fp16 = conv(dilations = var_839_dilations_0, groups = var_839_groups_0, pad = var_839_pad_0, pad_type = var_839_pad_type_0, strides = var_839_strides_0, weight = layers_2_self_attn_linear_out_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("op_839_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_61_cast_fp16 = add(x = x_49_cast_fp16, y = var_839_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [1]> var_855_axes_0 = const()[name = string("op_855_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_2_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_2_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_855_cast_fp16 = layer_norm(axes = var_855_axes_0, epsilon = layers_2_norm_conv_eps_scaled_to_fp16, x = x_61_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<fp16, [1024]> input_79_gamma_0_to_fp16 = const()[name = string("input_79_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35066240)))];
            tensor<fp16, [1024]> input_79_beta_0_to_fp16 = const()[name = string("input_79_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35068352)))];
            fp16 input_79_epsilon_0_to_fp16 = const()[name = string("input_79_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_79_cast_fp16 = batch_norm(beta = input_79_beta_0_to_fp16, epsilon = input_79_epsilon_0_to_fp16, gamma = input_79_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_855_cast_fp16)[name = string("input_79_cast_fp16")];
            string input_81_pad_type_0 = const()[name = string("input_81_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_81_strides_0 = const()[name = string("input_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_81_pad_0 = const()[name = string("input_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_81_dilations_0 = const()[name = string("input_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_81_groups_0 = const()[name = string("input_81_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35070464))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36119104))))[name = string("layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_81_cast_fp16 = conv(dilations = input_81_dilations_0, groups = input_81_groups_0, pad = input_81_pad_0, pad_type = input_81_pad_type_0, strides = input_81_strides_0, weight = layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("input_81_cast_fp16")];
            int32 x_63_split_num_splits_0 = const()[name = string("x_63_split_num_splits_0"), val = int32(2)];
            int32 x_63_split_axis_0 = const()[name = string("x_63_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_63_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_63_split_cast_fp16_1 = split(axis = x_63_split_axis_0, num_splits = x_63_split_num_splits_0, x = input_81_cast_fp16)[name = string("x_63_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_63_split_1_sigmoid_cast_fp16 = sigmoid(x = x_63_split_cast_fp16_1)[name = string("x_63_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_63_cast_fp16 = mul(x = x_63_split_cast_fp16_0, y = x_63_split_1_sigmoid_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_83_cast_fp16 = mul(x = x_63_cast_fp16, y = pad_mask)[name = string("input_83_cast_fp16")];
            string input_85_pad_type_0 = const()[name = string("input_85_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_85_pad_0 = const()[name = string("input_85_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_85_groups_0 = const()[name = string("input_85_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_85_strides_0 = const()[name = string("input_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_85_dilations_0 = const()[name = string("input_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_107_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36123264))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36127936))))[name = string("const_107_to_fp16_quantized")];
            tensor<fp16, [1024]> const_108_to_fp16 = const()[name = string("const_108_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36130048)))];
            tensor<fp16, [1, 1024, 1, 188]> input_87_cast_fp16 = conv(bias = const_108_to_fp16, dilations = input_85_dilations_0, groups = input_85_groups_0, pad = input_85_pad_0, pad_type = input_85_pad_type_0, strides = input_85_strides_0, weight = const_107_to_fp16_quantized, x = input_83_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_89_cast_fp16 = silu(x = input_87_cast_fp16)[name = string("input_89_cast_fp16")];
            string var_887_pad_type_0 = const()[name = string("op_887_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_887_strides_0 = const()[name = string("op_887_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_887_pad_0 = const()[name = string("op_887_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_887_dilations_0 = const()[name = string("op_887_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_887_groups_0 = const()[name = string("op_887_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36132160))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36656512))))[name = string("layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_887_cast_fp16 = conv(dilations = var_887_dilations_0, groups = var_887_groups_0, pad = var_887_pad_0, pad_type = var_887_pad_type_0, strides = var_887_strides_0, weight = layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("op_887_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_65_cast_fp16 = add(x = x_61_cast_fp16, y = var_887_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<int32, [1]> var_903_axes_0 = const()[name = string("op_903_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_2_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_2_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_903_cast_fp16 = layer_norm(axes = var_903_axes_0, epsilon = layers_2_norm_feed_forward2_eps_scaled_to_fp16, x = x_65_cast_fp16)[name = string("op_903_cast_fp16")];
            tensor<fp16, [1024]> input_91_gamma_0_to_fp16 = const()[name = string("input_91_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36658624)))];
            tensor<fp16, [1024]> input_91_beta_0_to_fp16 = const()[name = string("input_91_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36660736)))];
            fp16 input_91_epsilon_0_to_fp16 = const()[name = string("input_91_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_91_cast_fp16 = batch_norm(beta = input_91_beta_0_to_fp16, epsilon = input_91_epsilon_0_to_fp16, gamma = input_91_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_903_cast_fp16)[name = string("input_91_cast_fp16")];
            string input_93_pad_type_0 = const()[name = string("input_93_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_93_strides_0 = const()[name = string("input_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_93_pad_0 = const()[name = string("input_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_93_dilations_0 = const()[name = string("input_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_93_groups_0 = const()[name = string("input_93_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_2_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36662848))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38760064))))[name = string("layers_2_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_93_cast_fp16 = conv(dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = layers_2_feed_forward2_linear1_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_95_cast_fp16 = silu(x = input_93_cast_fp16)[name = string("input_95_cast_fp16")];
            string var_920_pad_type_0 = const()[name = string("op_920_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_920_strides_0 = const()[name = string("op_920_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_920_pad_0 = const()[name = string("op_920_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_920_dilations_0 = const()[name = string("op_920_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_920_groups_0 = const()[name = string("op_920_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_921_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38768320))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40865536))))[name = string("op_921_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_921_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_920_dilations_0, groups = var_920_groups_0, pad = var_920_pad_0, pad_type = var_920_pad_type_0, strides = var_920_strides_0, weight = op_921_weight_0_to_fp16_quantized, x = input_95_cast_fp16)[name = string("op_921_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = var_921_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<int32, [1]> var_937_axes_0 = const()[name = string("op_937_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_2_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_2_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_937_cast_fp16 = layer_norm(axes = var_937_axes_0, epsilon = layers_2_norm_out_eps_scaled_to_fp16, x = x_67_cast_fp16)[name = string("op_937_cast_fp16")];
            tensor<fp16, [1024]> x_69_gamma_0_to_fp16 = const()[name = string("x_69_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40867648)))];
            tensor<fp16, [1024]> x_69_beta_0_to_fp16 = const()[name = string("x_69_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40869760)))];
            fp16 x_69_epsilon_0_to_fp16 = const()[name = string("x_69_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_69_cast_fp16 = batch_norm(beta = x_69_beta_0_to_fp16, epsilon = x_69_epsilon_0_to_fp16, gamma = x_69_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_937_cast_fp16)[name = string("x_69_cast_fp16")];
            int32 var_956 = const()[name = string("op_956"), val = int32(1)];
            tensor<int32, [1]> var_983_axes_0 = const()[name = string("op_983_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_3_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_3_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_983_cast_fp16 = layer_norm(axes = var_983_axes_0, epsilon = layers_3_norm_feed_forward1_eps_scaled_to_fp16, x = x_69_cast_fp16)[name = string("op_983_cast_fp16")];
            tensor<fp16, [1024]> input_97_gamma_0_to_fp16 = const()[name = string("input_97_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40871872)))];
            tensor<fp16, [1024]> input_97_beta_0_to_fp16 = const()[name = string("input_97_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40873984)))];
            fp16 input_97_epsilon_0_to_fp16 = const()[name = string("input_97_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_97_cast_fp16 = batch_norm(beta = input_97_beta_0_to_fp16, epsilon = input_97_epsilon_0_to_fp16, gamma = input_97_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_983_cast_fp16)[name = string("input_97_cast_fp16")];
            string input_99_pad_type_0 = const()[name = string("input_99_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_99_strides_0 = const()[name = string("input_99_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_99_pad_0 = const()[name = string("input_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_99_dilations_0 = const()[name = string("input_99_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_99_groups_0 = const()[name = string("input_99_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_3_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40876096))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42973312))))[name = string("layers_3_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_99_cast_fp16 = conv(dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = layers_3_feed_forward1_linear1_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_101_cast_fp16 = silu(x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
            string var_1000_pad_type_0 = const()[name = string("op_1000_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1000_strides_0 = const()[name = string("op_1000_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1000_pad_0 = const()[name = string("op_1000_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1000_dilations_0 = const()[name = string("op_1000_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1000_groups_0 = const()[name = string("op_1000_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_1001_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42981568))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45078784))))[name = string("op_1001_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1001_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_1000_dilations_0, groups = var_1000_groups_0, pad = var_1000_pad_0, pad_type = var_1000_pad_type_0, strides = var_1000_strides_0, weight = op_1001_weight_0_to_fp16_quantized, x = input_101_cast_fp16)[name = string("op_1001_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = var_1001_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [1]> var_1017_axes_0 = const()[name = string("op_1017_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_3_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_3_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1017_cast_fp16 = layer_norm(axes = var_1017_axes_0, epsilon = layers_3_norm_self_att_eps_scaled_to_fp16, x = x_71_cast_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1024]> x_73_gamma_0_to_fp16 = const()[name = string("x_73_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45080896)))];
            tensor<fp16, [1024]> x_73_beta_0_to_fp16 = const()[name = string("x_73_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45083008)))];
            fp16 x_73_epsilon_0_to_fp16 = const()[name = string("x_73_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_73_cast_fp16 = batch_norm(beta = x_73_beta_0_to_fp16, epsilon = x_73_epsilon_0_to_fp16, gamma = x_73_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1017_cast_fp16)[name = string("x_73_cast_fp16")];
            string q_7_pad_type_0 = const()[name = string("q_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_7_strides_0 = const()[name = string("q_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_7_pad_0 = const()[name = string("q_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_7_dilations_0 = const()[name = string("q_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_7_groups_0 = const()[name = string("q_7_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45085120))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45609472))))[name = string("layers_3_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_7_cast_fp16 = conv(dilations = q_7_dilations_0, groups = q_7_groups_0, pad = q_7_pad_0, pad_type = q_7_pad_type_0, strides = q_7_strides_0, weight = layers_3_self_attn_linear_q_weight_to_fp16_quantized, x = x_73_cast_fp16)[name = string("q_7_cast_fp16")];
            string k_7_pad_type_0 = const()[name = string("k_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_7_strides_0 = const()[name = string("k_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_7_pad_0 = const()[name = string("k_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_7_dilations_0 = const()[name = string("k_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_7_groups_0 = const()[name = string("k_7_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45611584))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46135936))))[name = string("layers_3_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_7_cast_fp16 = conv(dilations = k_7_dilations_0, groups = k_7_groups_0, pad = k_7_pad_0, pad_type = k_7_pad_type_0, strides = k_7_strides_0, weight = layers_3_self_attn_linear_k_weight_to_fp16_quantized, x = x_73_cast_fp16)[name = string("k_7_cast_fp16")];
            string v_7_pad_type_0 = const()[name = string("v_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_7_strides_0 = const()[name = string("v_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_7_pad_0 = const()[name = string("v_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_7_dilations_0 = const()[name = string("v_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_7_groups_0 = const()[name = string("v_7_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46138048))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46662400))))[name = string("layers_3_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_7_cast_fp16 = conv(dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = layers_3_self_attn_linear_v_weight_to_fp16_quantized, x = x_73_cast_fp16)[name = string("v_7_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_7_to_fp16 = const()[name = string("bv_all_7_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46664512)))];
            tensor<fp16, [1, 1024, 1, 188]> var_1050_cast_fp16 = add(x = q_7_cast_fp16, y = bv_all_7_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<int32, [3]> var_1051 = const()[name = string("op_1051"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_7_cast_fp16 = reshape(shape = var_1051, x = var_1050_cast_fp16)[name = string("qb_7_cast_fp16")];
            bool bd_all_13_transpose_x_0 = const()[name = string("bd_all_13_transpose_x_0"), val = bool(false)];
            bool bd_all_13_transpose_y_0 = const()[name = string("bd_all_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_3_self_attn_pos_proj_to_fp16 = const()[name = string("layers_3_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46666624)))];
            tensor<fp16, [8, 375, 188]> bd_all_13_cast_fp16 = matmul(transpose_x = bd_all_13_transpose_x_0, transpose_y = bd_all_13_transpose_y_0, x = layers_3_self_attn_pos_proj_to_fp16, y = qb_7_cast_fp16)[name = string("bd_all_13_cast_fp16")];
            tensor<int32, [3]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_77_mode_0 = const()[name = string("x_77_mode_0"), val = string("constant")];
            fp16 const_22_to_fp16 = const()[name = string("const_22_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = bd_all_13_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [8, 188, 376]> x_77_cast_fp16 = pad(constant_val = const_22_to_fp16, mode = x_77_mode_0, pad = x_77_pad_0, x = x_75_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<int32, [3]> var_1058 = const()[name = string("op_1058"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_79_cast_fp16 = reshape(shape = var_1058, x = x_77_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [3]> var_1061_begin_0 = const()[name = string("op_1061_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1061_end_0 = const()[name = string("op_1061_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_1061_end_mask_0 = const()[name = string("op_1061_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, x = x_79_cast_fp16)[name = string("op_1061_cast_fp16")];
            tensor<int32, [3]> var_1062 = const()[name = string("op_1062"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_81_cast_fp16 = reshape(shape = var_1062, x = var_1061_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [3]> bd_all_15_begin_0 = const()[name = string("bd_all_15_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_15_end_0 = const()[name = string("bd_all_15_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_15_end_mask_0 = const()[name = string("bd_all_15_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_15_cast_fp16 = slice_by_index(begin = bd_all_15_begin_0, end = bd_all_15_end_0, end_mask = bd_all_15_end_mask_0, x = x_81_cast_fp16)[name = string("bd_all_15_cast_fp16")];
            tensor<int32, [4]> var_1067 = const()[name = string("op_1067"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1068_cast_fp16 = reshape(shape = var_1067, x = q_7_cast_fp16)[name = string("op_1068_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_1070_to_fp16 = const()[name = string("op_1070_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47434688)))];
            tensor<fp16, [8, 128, 1, 188]> qh_7_cast_fp16 = add(x = var_1068_cast_fp16, y = var_1070_to_fp16)[name = string("qh_7_cast_fp16")];
            tensor<int32, [4]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1073_cast_fp16 = reshape(shape = var_1072, x = k_7_cast_fp16)[name = string("op_1073_cast_fp16")];
            tensor<int32, [4]> var_1074 = const()[name = string("op_1074"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1077_cast_fp16 = reshape(shape = var_1076, x = v_7_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1080_perm_0 = const()[name = string("op_1080_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_7_axes_0 = const()[name = string("bd_7_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_1080_cast_fp16 = transpose(perm = var_1080_perm_0, x = bd_all_15_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [8, 188, 1, 188]> bd_7_cast_fp16 = expand_dims(axes = bd_7_axes_0, x = var_1080_cast_fp16)[name = string("bd_7_cast_fp16")];
            string ac_7_equation_0 = const()[name = string("ac_7_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_1075_cast_fp16 = transpose(perm = var_1074, x = var_1073_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [8, 188, 1, 188]> ac_7_cast_fp16 = einsum(equation = ac_7_equation_0, values = (var_1075_cast_fp16, qh_7_cast_fp16))[name = string("ac_7_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_1096_cast_fp16 = add(x = ac_7_cast_fp16, y = bd_7_cast_fp16)[name = string("op_1096_cast_fp16")];
            fp16 var_1097_to_fp16 = const()[name = string("op_1097_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_13_cast_fp16 = mul(x = var_1096_cast_fp16, y = var_1097_to_fp16)[name = string("scores_13_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = key_bias)[name = string("scores_15_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_1100_cast_fp16 = softmax(axis = var_956, x = scores_15_cast_fp16)[name = string("op_1100_cast_fp16")];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_34 = const()[name = string("concat_34"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = var_1100_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [8, 188, 188]> reshape_9_cast_fp16 = reshape(shape = concat_34, x = transpose_6_cast_fp16)[name = string("reshape_9_cast_fp16")];
            tensor<int32, [3]> concat_35 = const()[name = string("concat_35"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = var_1077_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [8, 188, 128]> reshape_10_cast_fp16 = reshape(shape = concat_35, x = transpose_51_cast_fp16)[name = string("reshape_10_cast_fp16")];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = reshape_9_cast_fp16, y = reshape_10_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<int32, [4]> concat_39 = const()[name = string("concat_39"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_11_cast_fp16 = reshape(shape = concat_39, x = matmul_3_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> ctx_7_perm_0 = const()[name = string("ctx_7_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1103 = const()[name = string("op_1103"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_7_cast_fp16 = transpose(perm = ctx_7_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 1024, 1, 188]> input_103_cast_fp16 = reshape(shape = var_1103, x = ctx_7_cast_fp16)[name = string("input_103_cast_fp16")];
            string var_1110_pad_type_0 = const()[name = string("op_1110_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1110_strides_0 = const()[name = string("op_1110_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1110_pad_0 = const()[name = string("op_1110_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1110_dilations_0 = const()[name = string("op_1110_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1110_groups_0 = const()[name = string("op_1110_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47436800))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47961152))))[name = string("layers_3_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1110_cast_fp16 = conv(dilations = var_1110_dilations_0, groups = var_1110_groups_0, pad = var_1110_pad_0, pad_type = var_1110_pad_type_0, strides = var_1110_strides_0, weight = layers_3_self_attn_linear_out_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("op_1110_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_83_cast_fp16 = add(x = x_71_cast_fp16, y = var_1110_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<int32, [1]> var_1126_axes_0 = const()[name = string("op_1126_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_3_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_3_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1126_cast_fp16 = layer_norm(axes = var_1126_axes_0, epsilon = layers_3_norm_conv_eps_scaled_to_fp16, x = x_83_cast_fp16)[name = string("op_1126_cast_fp16")];
            tensor<fp16, [1024]> input_105_gamma_0_to_fp16 = const()[name = string("input_105_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47963264)))];
            tensor<fp16, [1024]> input_105_beta_0_to_fp16 = const()[name = string("input_105_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47965376)))];
            fp16 input_105_epsilon_0_to_fp16 = const()[name = string("input_105_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_105_cast_fp16 = batch_norm(beta = input_105_beta_0_to_fp16, epsilon = input_105_epsilon_0_to_fp16, gamma = input_105_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1126_cast_fp16)[name = string("input_105_cast_fp16")];
            string input_107_pad_type_0 = const()[name = string("input_107_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_107_strides_0 = const()[name = string("input_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_107_pad_0 = const()[name = string("input_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_107_dilations_0 = const()[name = string("input_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_107_groups_0 = const()[name = string("input_107_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47967488))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49016128))))[name = string("layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_107_cast_fp16 = conv(dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
            int32 x_85_split_num_splits_0 = const()[name = string("x_85_split_num_splits_0"), val = int32(2)];
            int32 x_85_split_axis_0 = const()[name = string("x_85_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_85_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_85_split_cast_fp16_1 = split(axis = x_85_split_axis_0, num_splits = x_85_split_num_splits_0, x = input_107_cast_fp16)[name = string("x_85_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_85_split_1_sigmoid_cast_fp16 = sigmoid(x = x_85_split_cast_fp16_1)[name = string("x_85_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_85_cast_fp16 = mul(x = x_85_split_cast_fp16_0, y = x_85_split_1_sigmoid_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_109_cast_fp16 = mul(x = x_85_cast_fp16, y = pad_mask)[name = string("input_109_cast_fp16")];
            string input_111_pad_type_0 = const()[name = string("input_111_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_111_pad_0 = const()[name = string("input_111_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_111_groups_0 = const()[name = string("input_111_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_111_strides_0 = const()[name = string("input_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_111_dilations_0 = const()[name = string("input_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_109_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49020288))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49024960))))[name = string("const_109_to_fp16_quantized")];
            tensor<fp16, [1024]> const_110_to_fp16 = const()[name = string("const_110_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49027072)))];
            tensor<fp16, [1, 1024, 1, 188]> input_113_cast_fp16 = conv(bias = const_110_to_fp16, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = const_109_to_fp16_quantized, x = input_109_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_115_cast_fp16 = silu(x = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
            string var_1158_pad_type_0 = const()[name = string("op_1158_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1158_strides_0 = const()[name = string("op_1158_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1158_pad_0 = const()[name = string("op_1158_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1158_dilations_0 = const()[name = string("op_1158_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1158_groups_0 = const()[name = string("op_1158_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49029184))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49553536))))[name = string("layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1158_cast_fp16 = conv(dilations = var_1158_dilations_0, groups = var_1158_groups_0, pad = var_1158_pad_0, pad_type = var_1158_pad_type_0, strides = var_1158_strides_0, weight = layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("op_1158_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_87_cast_fp16 = add(x = x_83_cast_fp16, y = var_1158_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [1]> var_1174_axes_0 = const()[name = string("op_1174_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_3_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_3_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1174_cast_fp16 = layer_norm(axes = var_1174_axes_0, epsilon = layers_3_norm_feed_forward2_eps_scaled_to_fp16, x = x_87_cast_fp16)[name = string("op_1174_cast_fp16")];
            tensor<fp16, [1024]> input_117_gamma_0_to_fp16 = const()[name = string("input_117_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49555648)))];
            tensor<fp16, [1024]> input_117_beta_0_to_fp16 = const()[name = string("input_117_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49557760)))];
            fp16 input_117_epsilon_0_to_fp16 = const()[name = string("input_117_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_117_cast_fp16 = batch_norm(beta = input_117_beta_0_to_fp16, epsilon = input_117_epsilon_0_to_fp16, gamma = input_117_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1174_cast_fp16)[name = string("input_117_cast_fp16")];
            string input_119_pad_type_0 = const()[name = string("input_119_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_119_strides_0 = const()[name = string("input_119_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_119_pad_0 = const()[name = string("input_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_119_dilations_0 = const()[name = string("input_119_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_119_groups_0 = const()[name = string("input_119_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_3_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49559872))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51657088))))[name = string("layers_3_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_119_cast_fp16 = conv(dilations = input_119_dilations_0, groups = input_119_groups_0, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = input_119_strides_0, weight = layers_3_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_121_cast_fp16 = silu(x = input_119_cast_fp16)[name = string("input_121_cast_fp16")];
            string var_1191_pad_type_0 = const()[name = string("op_1191_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1191_strides_0 = const()[name = string("op_1191_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1191_pad_0 = const()[name = string("op_1191_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1191_dilations_0 = const()[name = string("op_1191_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1191_groups_0 = const()[name = string("op_1191_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_1192_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51665344))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53762560))))[name = string("op_1192_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1192_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_1191_dilations_0, groups = var_1191_groups_0, pad = var_1191_pad_0, pad_type = var_1191_pad_type_0, strides = var_1191_strides_0, weight = op_1192_weight_0_to_fp16_quantized, x = input_121_cast_fp16)[name = string("op_1192_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = var_1192_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [1]> var_1208_axes_0 = const()[name = string("op_1208_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_3_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_3_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1208_cast_fp16 = layer_norm(axes = var_1208_axes_0, epsilon = layers_3_norm_out_eps_scaled_to_fp16, x = x_89_cast_fp16)[name = string("op_1208_cast_fp16")];
            tensor<fp16, [1024]> x_91_gamma_0_to_fp16 = const()[name = string("x_91_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53764672)))];
            tensor<fp16, [1024]> x_91_beta_0_to_fp16 = const()[name = string("x_91_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53766784)))];
            fp16 x_91_epsilon_0_to_fp16 = const()[name = string("x_91_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_91_cast_fp16 = batch_norm(beta = x_91_beta_0_to_fp16, epsilon = x_91_epsilon_0_to_fp16, gamma = x_91_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1208_cast_fp16)[name = string("x_91_cast_fp16")];
            int32 var_1227 = const()[name = string("op_1227"), val = int32(1)];
            tensor<int32, [1]> var_1254_axes_0 = const()[name = string("op_1254_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_4_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_4_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1254_cast_fp16 = layer_norm(axes = var_1254_axes_0, epsilon = layers_4_norm_feed_forward1_eps_scaled_to_fp16, x = x_91_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1024]> input_123_gamma_0_to_fp16 = const()[name = string("input_123_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53768896)))];
            tensor<fp16, [1024]> input_123_beta_0_to_fp16 = const()[name = string("input_123_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53771008)))];
            fp16 input_123_epsilon_0_to_fp16 = const()[name = string("input_123_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_123_cast_fp16 = batch_norm(beta = input_123_beta_0_to_fp16, epsilon = input_123_epsilon_0_to_fp16, gamma = input_123_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1254_cast_fp16)[name = string("input_123_cast_fp16")];
            string input_125_pad_type_0 = const()[name = string("input_125_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_125_strides_0 = const()[name = string("input_125_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_125_pad_0 = const()[name = string("input_125_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_125_dilations_0 = const()[name = string("input_125_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_125_groups_0 = const()[name = string("input_125_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_4_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53773120))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55870336))))[name = string("layers_4_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_125_cast_fp16 = conv(dilations = input_125_dilations_0, groups = input_125_groups_0, pad = input_125_pad_0, pad_type = input_125_pad_type_0, strides = input_125_strides_0, weight = layers_4_feed_forward1_linear1_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_127_cast_fp16 = silu(x = input_125_cast_fp16)[name = string("input_127_cast_fp16")];
            string var_1271_pad_type_0 = const()[name = string("op_1271_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1271_strides_0 = const()[name = string("op_1271_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1271_pad_0 = const()[name = string("op_1271_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1271_dilations_0 = const()[name = string("op_1271_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1271_groups_0 = const()[name = string("op_1271_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_1272_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55878592))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57975808))))[name = string("op_1272_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1272_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_1271_dilations_0, groups = var_1271_groups_0, pad = var_1271_pad_0, pad_type = var_1271_pad_type_0, strides = var_1271_strides_0, weight = op_1272_weight_0_to_fp16_quantized, x = input_127_cast_fp16)[name = string("op_1272_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = var_1272_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_4_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_4_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1288_cast_fp16 = layer_norm(axes = var_1288_axes_0, epsilon = layers_4_norm_self_att_eps_scaled_to_fp16, x = x_93_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<fp16, [1024]> x_95_gamma_0_to_fp16 = const()[name = string("x_95_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57977920)))];
            tensor<fp16, [1024]> x_95_beta_0_to_fp16 = const()[name = string("x_95_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57980032)))];
            fp16 x_95_epsilon_0_to_fp16 = const()[name = string("x_95_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_95_cast_fp16 = batch_norm(beta = x_95_beta_0_to_fp16, epsilon = x_95_epsilon_0_to_fp16, gamma = x_95_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1288_cast_fp16)[name = string("x_95_cast_fp16")];
            string q_9_pad_type_0 = const()[name = string("q_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_9_strides_0 = const()[name = string("q_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_9_pad_0 = const()[name = string("q_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_9_dilations_0 = const()[name = string("q_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_9_groups_0 = const()[name = string("q_9_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57982144))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58506496))))[name = string("layers_4_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_9_cast_fp16 = conv(dilations = q_9_dilations_0, groups = q_9_groups_0, pad = q_9_pad_0, pad_type = q_9_pad_type_0, strides = q_9_strides_0, weight = layers_4_self_attn_linear_q_weight_to_fp16_quantized, x = x_95_cast_fp16)[name = string("q_9_cast_fp16")];
            string k_9_pad_type_0 = const()[name = string("k_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_9_strides_0 = const()[name = string("k_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_9_pad_0 = const()[name = string("k_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_9_dilations_0 = const()[name = string("k_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_9_groups_0 = const()[name = string("k_9_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58508608))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59032960))))[name = string("layers_4_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_9_cast_fp16 = conv(dilations = k_9_dilations_0, groups = k_9_groups_0, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = k_9_strides_0, weight = layers_4_self_attn_linear_k_weight_to_fp16_quantized, x = x_95_cast_fp16)[name = string("k_9_cast_fp16")];
            string v_9_pad_type_0 = const()[name = string("v_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_9_strides_0 = const()[name = string("v_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_9_pad_0 = const()[name = string("v_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_9_dilations_0 = const()[name = string("v_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_9_groups_0 = const()[name = string("v_9_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59035072))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59559424))))[name = string("layers_4_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_9_cast_fp16 = conv(dilations = v_9_dilations_0, groups = v_9_groups_0, pad = v_9_pad_0, pad_type = v_9_pad_type_0, strides = v_9_strides_0, weight = layers_4_self_attn_linear_v_weight_to_fp16_quantized, x = x_95_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_9_to_fp16 = const()[name = string("bv_all_9_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59561536)))];
            tensor<fp16, [1, 1024, 1, 188]> var_1321_cast_fp16 = add(x = q_9_cast_fp16, y = bv_all_9_to_fp16)[name = string("op_1321_cast_fp16")];
            tensor<int32, [3]> var_1322 = const()[name = string("op_1322"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_9_cast_fp16 = reshape(shape = var_1322, x = var_1321_cast_fp16)[name = string("qb_9_cast_fp16")];
            bool bd_all_17_transpose_x_0 = const()[name = string("bd_all_17_transpose_x_0"), val = bool(false)];
            bool bd_all_17_transpose_y_0 = const()[name = string("bd_all_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_4_self_attn_pos_proj_to_fp16 = const()[name = string("layers_4_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59563648)))];
            tensor<fp16, [8, 375, 188]> bd_all_17_cast_fp16 = matmul(transpose_x = bd_all_17_transpose_x_0, transpose_y = bd_all_17_transpose_y_0, x = layers_4_self_attn_pos_proj_to_fp16, y = qb_9_cast_fp16)[name = string("bd_all_17_cast_fp16")];
            tensor<int32, [3]> x_97_perm_0 = const()[name = string("x_97_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_99_mode_0 = const()[name = string("x_99_mode_0"), val = string("constant")];
            fp16 const_26_to_fp16 = const()[name = string("const_26_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_97_cast_fp16 = transpose(perm = x_97_perm_0, x = bd_all_17_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [8, 188, 376]> x_99_cast_fp16 = pad(constant_val = const_26_to_fp16, mode = x_99_mode_0, pad = x_99_pad_0, x = x_97_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [3]> var_1329 = const()[name = string("op_1329"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_101_cast_fp16 = reshape(shape = var_1329, x = x_99_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [3]> var_1332_begin_0 = const()[name = string("op_1332_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1332_end_0 = const()[name = string("op_1332_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_1332_end_mask_0 = const()[name = string("op_1332_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_1332_cast_fp16 = slice_by_index(begin = var_1332_begin_0, end = var_1332_end_0, end_mask = var_1332_end_mask_0, x = x_101_cast_fp16)[name = string("op_1332_cast_fp16")];
            tensor<int32, [3]> var_1333 = const()[name = string("op_1333"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_103_cast_fp16 = reshape(shape = var_1333, x = var_1332_cast_fp16)[name = string("x_103_cast_fp16")];
            tensor<int32, [3]> bd_all_19_begin_0 = const()[name = string("bd_all_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_19_end_0 = const()[name = string("bd_all_19_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_19_end_mask_0 = const()[name = string("bd_all_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_19_cast_fp16 = slice_by_index(begin = bd_all_19_begin_0, end = bd_all_19_end_0, end_mask = bd_all_19_end_mask_0, x = x_103_cast_fp16)[name = string("bd_all_19_cast_fp16")];
            tensor<int32, [4]> var_1338 = const()[name = string("op_1338"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1339_cast_fp16 = reshape(shape = var_1338, x = q_9_cast_fp16)[name = string("op_1339_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_1341_to_fp16 = const()[name = string("op_1341_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60331712)))];
            tensor<fp16, [8, 128, 1, 188]> qh_9_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1341_to_fp16)[name = string("qh_9_cast_fp16")];
            tensor<int32, [4]> var_1343 = const()[name = string("op_1343"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1344_cast_fp16 = reshape(shape = var_1343, x = k_9_cast_fp16)[name = string("op_1344_cast_fp16")];
            tensor<int32, [4]> var_1345 = const()[name = string("op_1345"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_1347 = const()[name = string("op_1347"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1348_cast_fp16 = reshape(shape = var_1347, x = v_9_cast_fp16)[name = string("op_1348_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1351_perm_0 = const()[name = string("op_1351_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_9_axes_0 = const()[name = string("bd_9_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_1351_cast_fp16 = transpose(perm = var_1351_perm_0, x = bd_all_19_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [8, 188, 1, 188]> bd_9_cast_fp16 = expand_dims(axes = bd_9_axes_0, x = var_1351_cast_fp16)[name = string("bd_9_cast_fp16")];
            string ac_9_equation_0 = const()[name = string("ac_9_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_1346_cast_fp16 = transpose(perm = var_1345, x = var_1344_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [8, 188, 1, 188]> ac_9_cast_fp16 = einsum(equation = ac_9_equation_0, values = (var_1346_cast_fp16, qh_9_cast_fp16))[name = string("ac_9_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_1367_cast_fp16 = add(x = ac_9_cast_fp16, y = bd_9_cast_fp16)[name = string("op_1367_cast_fp16")];
            fp16 var_1368_to_fp16 = const()[name = string("op_1368_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_17_cast_fp16 = mul(x = var_1367_cast_fp16, y = var_1368_to_fp16)[name = string("scores_17_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = key_bias)[name = string("scores_19_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_1371_cast_fp16 = softmax(axis = var_1227, x = scores_19_cast_fp16)[name = string("op_1371_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_44 = const()[name = string("concat_44"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = var_1371_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [8, 188, 188]> reshape_12_cast_fp16 = reshape(shape = concat_44, x = transpose_8_cast_fp16)[name = string("reshape_12_cast_fp16")];
            tensor<int32, [3]> concat_45 = const()[name = string("concat_45"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = var_1348_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [8, 188, 128]> reshape_13_cast_fp16 = reshape(shape = concat_45, x = transpose_52_cast_fp16)[name = string("reshape_13_cast_fp16")];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = reshape_12_cast_fp16, y = reshape_13_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<int32, [4]> concat_49 = const()[name = string("concat_49"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_14_cast_fp16 = reshape(shape = concat_49, x = matmul_4_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [4]> ctx_9_perm_0 = const()[name = string("ctx_9_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1374 = const()[name = string("op_1374"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_9_cast_fp16 = transpose(perm = ctx_9_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 1024, 1, 188]> input_129_cast_fp16 = reshape(shape = var_1374, x = ctx_9_cast_fp16)[name = string("input_129_cast_fp16")];
            string var_1381_pad_type_0 = const()[name = string("op_1381_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1381_strides_0 = const()[name = string("op_1381_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1381_pad_0 = const()[name = string("op_1381_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1381_dilations_0 = const()[name = string("op_1381_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1381_groups_0 = const()[name = string("op_1381_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60333824))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60858176))))[name = string("layers_4_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1381_cast_fp16 = conv(dilations = var_1381_dilations_0, groups = var_1381_groups_0, pad = var_1381_pad_0, pad_type = var_1381_pad_type_0, strides = var_1381_strides_0, weight = layers_4_self_attn_linear_out_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("op_1381_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_105_cast_fp16 = add(x = x_93_cast_fp16, y = var_1381_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [1]> var_1397_axes_0 = const()[name = string("op_1397_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_4_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_4_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1397_cast_fp16 = layer_norm(axes = var_1397_axes_0, epsilon = layers_4_norm_conv_eps_scaled_to_fp16, x = x_105_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1024]> input_131_gamma_0_to_fp16 = const()[name = string("input_131_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60860288)))];
            tensor<fp16, [1024]> input_131_beta_0_to_fp16 = const()[name = string("input_131_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60862400)))];
            fp16 input_131_epsilon_0_to_fp16 = const()[name = string("input_131_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_131_cast_fp16 = batch_norm(beta = input_131_beta_0_to_fp16, epsilon = input_131_epsilon_0_to_fp16, gamma = input_131_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1397_cast_fp16)[name = string("input_131_cast_fp16")];
            string input_133_pad_type_0 = const()[name = string("input_133_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_133_strides_0 = const()[name = string("input_133_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_133_pad_0 = const()[name = string("input_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_133_dilations_0 = const()[name = string("input_133_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_133_groups_0 = const()[name = string("input_133_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60864512))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61913152))))[name = string("layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_133_cast_fp16 = conv(dilations = input_133_dilations_0, groups = input_133_groups_0, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = input_133_strides_0, weight = layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("input_133_cast_fp16")];
            int32 x_107_split_num_splits_0 = const()[name = string("x_107_split_num_splits_0"), val = int32(2)];
            int32 x_107_split_axis_0 = const()[name = string("x_107_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_107_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_107_split_cast_fp16_1 = split(axis = x_107_split_axis_0, num_splits = x_107_split_num_splits_0, x = input_133_cast_fp16)[name = string("x_107_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_107_split_1_sigmoid_cast_fp16 = sigmoid(x = x_107_split_cast_fp16_1)[name = string("x_107_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_107_cast_fp16 = mul(x = x_107_split_cast_fp16_0, y = x_107_split_1_sigmoid_cast_fp16)[name = string("x_107_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_135_cast_fp16 = mul(x = x_107_cast_fp16, y = pad_mask)[name = string("input_135_cast_fp16")];
            string input_137_pad_type_0 = const()[name = string("input_137_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_137_pad_0 = const()[name = string("input_137_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_137_groups_0 = const()[name = string("input_137_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_137_strides_0 = const()[name = string("input_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_137_dilations_0 = const()[name = string("input_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_111_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61917312))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61921984))))[name = string("const_111_to_fp16_quantized")];
            tensor<fp16, [1024]> const_112_to_fp16 = const()[name = string("const_112_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61924096)))];
            tensor<fp16, [1, 1024, 1, 188]> input_139_cast_fp16 = conv(bias = const_112_to_fp16, dilations = input_137_dilations_0, groups = input_137_groups_0, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = input_137_strides_0, weight = const_111_to_fp16_quantized, x = input_135_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_141_cast_fp16 = silu(x = input_139_cast_fp16)[name = string("input_141_cast_fp16")];
            string var_1429_pad_type_0 = const()[name = string("op_1429_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1429_strides_0 = const()[name = string("op_1429_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1429_pad_0 = const()[name = string("op_1429_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1429_dilations_0 = const()[name = string("op_1429_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1429_groups_0 = const()[name = string("op_1429_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61926208))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62450560))))[name = string("layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1429_cast_fp16 = conv(dilations = var_1429_dilations_0, groups = var_1429_groups_0, pad = var_1429_pad_0, pad_type = var_1429_pad_type_0, strides = var_1429_strides_0, weight = layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("op_1429_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_109_cast_fp16 = add(x = x_105_cast_fp16, y = var_1429_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [1]> var_1445_axes_0 = const()[name = string("op_1445_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_4_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_4_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1445_cast_fp16 = layer_norm(axes = var_1445_axes_0, epsilon = layers_4_norm_feed_forward2_eps_scaled_to_fp16, x = x_109_cast_fp16)[name = string("op_1445_cast_fp16")];
            tensor<fp16, [1024]> input_143_gamma_0_to_fp16 = const()[name = string("input_143_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62452672)))];
            tensor<fp16, [1024]> input_143_beta_0_to_fp16 = const()[name = string("input_143_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62454784)))];
            fp16 input_143_epsilon_0_to_fp16 = const()[name = string("input_143_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_143_cast_fp16 = batch_norm(beta = input_143_beta_0_to_fp16, epsilon = input_143_epsilon_0_to_fp16, gamma = input_143_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1445_cast_fp16)[name = string("input_143_cast_fp16")];
            string input_145_pad_type_0 = const()[name = string("input_145_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_145_strides_0 = const()[name = string("input_145_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_145_pad_0 = const()[name = string("input_145_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_145_dilations_0 = const()[name = string("input_145_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_145_groups_0 = const()[name = string("input_145_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_4_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62456896))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64554112))))[name = string("layers_4_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_145_cast_fp16 = conv(dilations = input_145_dilations_0, groups = input_145_groups_0, pad = input_145_pad_0, pad_type = input_145_pad_type_0, strides = input_145_strides_0, weight = layers_4_feed_forward2_linear1_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_147_cast_fp16 = silu(x = input_145_cast_fp16)[name = string("input_147_cast_fp16")];
            string var_1462_pad_type_0 = const()[name = string("op_1462_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1462_strides_0 = const()[name = string("op_1462_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1462_pad_0 = const()[name = string("op_1462_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1462_dilations_0 = const()[name = string("op_1462_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1462_groups_0 = const()[name = string("op_1462_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_1463_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64562368))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66659584))))[name = string("op_1463_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1463_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_1462_dilations_0, groups = var_1462_groups_0, pad = var_1462_pad_0, pad_type = var_1462_pad_type_0, strides = var_1462_strides_0, weight = op_1463_weight_0_to_fp16_quantized, x = input_147_cast_fp16)[name = string("op_1463_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_111_cast_fp16 = add(x = x_109_cast_fp16, y = var_1463_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<int32, [1]> var_1479_axes_0 = const()[name = string("op_1479_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_4_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_4_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1479_cast_fp16 = layer_norm(axes = var_1479_axes_0, epsilon = layers_4_norm_out_eps_scaled_to_fp16, x = x_111_cast_fp16)[name = string("op_1479_cast_fp16")];
            tensor<fp16, [1024]> x_113_gamma_0_to_fp16 = const()[name = string("x_113_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66661696)))];
            tensor<fp16, [1024]> x_113_beta_0_to_fp16 = const()[name = string("x_113_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66663808)))];
            fp16 x_113_epsilon_0_to_fp16 = const()[name = string("x_113_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_113_cast_fp16 = batch_norm(beta = x_113_beta_0_to_fp16, epsilon = x_113_epsilon_0_to_fp16, gamma = x_113_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1479_cast_fp16)[name = string("x_113_cast_fp16")];
            int32 var_1498 = const()[name = string("op_1498"), val = int32(1)];
            tensor<int32, [1]> var_1525_axes_0 = const()[name = string("op_1525_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_5_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_5_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1525_cast_fp16 = layer_norm(axes = var_1525_axes_0, epsilon = layers_5_norm_feed_forward1_eps_scaled_to_fp16, x = x_113_cast_fp16)[name = string("op_1525_cast_fp16")];
            tensor<fp16, [1024]> input_149_gamma_0_to_fp16 = const()[name = string("input_149_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66665920)))];
            tensor<fp16, [1024]> input_149_beta_0_to_fp16 = const()[name = string("input_149_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66668032)))];
            fp16 input_149_epsilon_0_to_fp16 = const()[name = string("input_149_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_149_cast_fp16 = batch_norm(beta = input_149_beta_0_to_fp16, epsilon = input_149_epsilon_0_to_fp16, gamma = input_149_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1525_cast_fp16)[name = string("input_149_cast_fp16")];
            string input_151_pad_type_0 = const()[name = string("input_151_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_151_strides_0 = const()[name = string("input_151_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_151_pad_0 = const()[name = string("input_151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_151_dilations_0 = const()[name = string("input_151_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_151_groups_0 = const()[name = string("input_151_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_5_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66670144))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68767360))))[name = string("layers_5_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_151_cast_fp16 = conv(dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = layers_5_feed_forward1_linear1_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_153_cast_fp16 = silu(x = input_151_cast_fp16)[name = string("input_153_cast_fp16")];
            string var_1542_pad_type_0 = const()[name = string("op_1542_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1542_strides_0 = const()[name = string("op_1542_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1542_pad_0 = const()[name = string("op_1542_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1542_dilations_0 = const()[name = string("op_1542_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1542_groups_0 = const()[name = string("op_1542_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_1543_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68775616))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70872832))))[name = string("op_1543_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1543_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_1542_dilations_0, groups = var_1542_groups_0, pad = var_1542_pad_0, pad_type = var_1542_pad_type_0, strides = var_1542_strides_0, weight = op_1543_weight_0_to_fp16_quantized, x = input_153_cast_fp16)[name = string("op_1543_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_115_cast_fp16 = add(x = x_113_cast_fp16, y = var_1543_cast_fp16)[name = string("x_115_cast_fp16")];
            tensor<int32, [1]> var_1559_axes_0 = const()[name = string("op_1559_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_5_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_5_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1559_cast_fp16 = layer_norm(axes = var_1559_axes_0, epsilon = layers_5_norm_self_att_eps_scaled_to_fp16, x = x_115_cast_fp16)[name = string("op_1559_cast_fp16")];
            tensor<fp16, [1024]> x_117_gamma_0_to_fp16 = const()[name = string("x_117_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70874944)))];
            tensor<fp16, [1024]> x_117_beta_0_to_fp16 = const()[name = string("x_117_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70877056)))];
            fp16 x_117_epsilon_0_to_fp16 = const()[name = string("x_117_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_117_cast_fp16 = batch_norm(beta = x_117_beta_0_to_fp16, epsilon = x_117_epsilon_0_to_fp16, gamma = x_117_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1559_cast_fp16)[name = string("x_117_cast_fp16")];
            string q_11_pad_type_0 = const()[name = string("q_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_11_strides_0 = const()[name = string("q_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_11_pad_0 = const()[name = string("q_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_11_dilations_0 = const()[name = string("q_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_11_groups_0 = const()[name = string("q_11_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70879168))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71403520))))[name = string("layers_5_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_11_cast_fp16 = conv(dilations = q_11_dilations_0, groups = q_11_groups_0, pad = q_11_pad_0, pad_type = q_11_pad_type_0, strides = q_11_strides_0, weight = layers_5_self_attn_linear_q_weight_to_fp16_quantized, x = x_117_cast_fp16)[name = string("q_11_cast_fp16")];
            string k_11_pad_type_0 = const()[name = string("k_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_11_strides_0 = const()[name = string("k_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_11_pad_0 = const()[name = string("k_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_11_dilations_0 = const()[name = string("k_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_11_groups_0 = const()[name = string("k_11_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71405632))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71929984))))[name = string("layers_5_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_11_cast_fp16 = conv(dilations = k_11_dilations_0, groups = k_11_groups_0, pad = k_11_pad_0, pad_type = k_11_pad_type_0, strides = k_11_strides_0, weight = layers_5_self_attn_linear_k_weight_to_fp16_quantized, x = x_117_cast_fp16)[name = string("k_11_cast_fp16")];
            string v_11_pad_type_0 = const()[name = string("v_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_11_strides_0 = const()[name = string("v_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_11_pad_0 = const()[name = string("v_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_11_dilations_0 = const()[name = string("v_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_11_groups_0 = const()[name = string("v_11_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71932096))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72456448))))[name = string("layers_5_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_11_cast_fp16 = conv(dilations = v_11_dilations_0, groups = v_11_groups_0, pad = v_11_pad_0, pad_type = v_11_pad_type_0, strides = v_11_strides_0, weight = layers_5_self_attn_linear_v_weight_to_fp16_quantized, x = x_117_cast_fp16)[name = string("v_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_11_to_fp16 = const()[name = string("bv_all_11_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72458560)))];
            tensor<fp16, [1, 1024, 1, 188]> var_1592_cast_fp16 = add(x = q_11_cast_fp16, y = bv_all_11_to_fp16)[name = string("op_1592_cast_fp16")];
            tensor<int32, [3]> var_1593 = const()[name = string("op_1593"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_11_cast_fp16 = reshape(shape = var_1593, x = var_1592_cast_fp16)[name = string("qb_11_cast_fp16")];
            bool bd_all_21_transpose_x_0 = const()[name = string("bd_all_21_transpose_x_0"), val = bool(false)];
            bool bd_all_21_transpose_y_0 = const()[name = string("bd_all_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_5_self_attn_pos_proj_to_fp16 = const()[name = string("layers_5_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72460672)))];
            tensor<fp16, [8, 375, 188]> bd_all_21_cast_fp16 = matmul(transpose_x = bd_all_21_transpose_x_0, transpose_y = bd_all_21_transpose_y_0, x = layers_5_self_attn_pos_proj_to_fp16, y = qb_11_cast_fp16)[name = string("bd_all_21_cast_fp16")];
            tensor<int32, [3]> x_119_perm_0 = const()[name = string("x_119_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_121_pad_0 = const()[name = string("x_121_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_121_mode_0 = const()[name = string("x_121_mode_0"), val = string("constant")];
            fp16 const_30_to_fp16 = const()[name = string("const_30_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_119_cast_fp16 = transpose(perm = x_119_perm_0, x = bd_all_21_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [8, 188, 376]> x_121_cast_fp16 = pad(constant_val = const_30_to_fp16, mode = x_121_mode_0, pad = x_121_pad_0, x = x_119_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [3]> var_1600 = const()[name = string("op_1600"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_123_cast_fp16 = reshape(shape = var_1600, x = x_121_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [3]> var_1603_begin_0 = const()[name = string("op_1603_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1603_end_0 = const()[name = string("op_1603_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_1603_end_mask_0 = const()[name = string("op_1603_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_1603_cast_fp16 = slice_by_index(begin = var_1603_begin_0, end = var_1603_end_0, end_mask = var_1603_end_mask_0, x = x_123_cast_fp16)[name = string("op_1603_cast_fp16")];
            tensor<int32, [3]> var_1604 = const()[name = string("op_1604"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_125_cast_fp16 = reshape(shape = var_1604, x = var_1603_cast_fp16)[name = string("x_125_cast_fp16")];
            tensor<int32, [3]> bd_all_23_begin_0 = const()[name = string("bd_all_23_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_23_end_0 = const()[name = string("bd_all_23_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_23_end_mask_0 = const()[name = string("bd_all_23_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_23_cast_fp16 = slice_by_index(begin = bd_all_23_begin_0, end = bd_all_23_end_0, end_mask = bd_all_23_end_mask_0, x = x_125_cast_fp16)[name = string("bd_all_23_cast_fp16")];
            tensor<int32, [4]> var_1609 = const()[name = string("op_1609"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1610_cast_fp16 = reshape(shape = var_1609, x = q_11_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_1612_to_fp16 = const()[name = string("op_1612_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73228736)))];
            tensor<fp16, [8, 128, 1, 188]> qh_11_cast_fp16 = add(x = var_1610_cast_fp16, y = var_1612_to_fp16)[name = string("qh_11_cast_fp16")];
            tensor<int32, [4]> var_1614 = const()[name = string("op_1614"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1615_cast_fp16 = reshape(shape = var_1614, x = k_11_cast_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [4]> var_1616 = const()[name = string("op_1616"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_1618 = const()[name = string("op_1618"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1619_cast_fp16 = reshape(shape = var_1618, x = v_11_cast_fp16)[name = string("op_1619_cast_fp16")];
            tensor<int32, [4]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1622_perm_0 = const()[name = string("op_1622_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_11_axes_0 = const()[name = string("bd_11_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_1622_cast_fp16 = transpose(perm = var_1622_perm_0, x = bd_all_23_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [8, 188, 1, 188]> bd_11_cast_fp16 = expand_dims(axes = bd_11_axes_0, x = var_1622_cast_fp16)[name = string("bd_11_cast_fp16")];
            string ac_11_equation_0 = const()[name = string("ac_11_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_1617_cast_fp16 = transpose(perm = var_1616, x = var_1615_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [8, 188, 1, 188]> ac_11_cast_fp16 = einsum(equation = ac_11_equation_0, values = (var_1617_cast_fp16, qh_11_cast_fp16))[name = string("ac_11_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_1638_cast_fp16 = add(x = ac_11_cast_fp16, y = bd_11_cast_fp16)[name = string("op_1638_cast_fp16")];
            fp16 var_1639_to_fp16 = const()[name = string("op_1639_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_21_cast_fp16 = mul(x = var_1638_cast_fp16, y = var_1639_to_fp16)[name = string("scores_21_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = key_bias)[name = string("scores_23_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_1642_cast_fp16 = softmax(axis = var_1498, x = scores_23_cast_fp16)[name = string("op_1642_cast_fp16")];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_54 = const()[name = string("concat_54"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = var_1642_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [8, 188, 188]> reshape_15_cast_fp16 = reshape(shape = concat_54, x = transpose_10_cast_fp16)[name = string("reshape_15_cast_fp16")];
            tensor<int32, [3]> concat_55 = const()[name = string("concat_55"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = var_1619_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [8, 188, 128]> reshape_16_cast_fp16 = reshape(shape = concat_55, x = transpose_53_cast_fp16)[name = string("reshape_16_cast_fp16")];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = reshape_15_cast_fp16, y = reshape_16_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<int32, [4]> concat_59 = const()[name = string("concat_59"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_17_cast_fp16 = reshape(shape = concat_59, x = matmul_5_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> ctx_11_perm_0 = const()[name = string("ctx_11_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1645 = const()[name = string("op_1645"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_11_cast_fp16 = transpose(perm = ctx_11_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 1024, 1, 188]> input_155_cast_fp16 = reshape(shape = var_1645, x = ctx_11_cast_fp16)[name = string("input_155_cast_fp16")];
            string var_1652_pad_type_0 = const()[name = string("op_1652_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1652_strides_0 = const()[name = string("op_1652_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1652_pad_0 = const()[name = string("op_1652_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1652_dilations_0 = const()[name = string("op_1652_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1652_groups_0 = const()[name = string("op_1652_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73230848))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73755200))))[name = string("layers_5_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1652_cast_fp16 = conv(dilations = var_1652_dilations_0, groups = var_1652_groups_0, pad = var_1652_pad_0, pad_type = var_1652_pad_type_0, strides = var_1652_strides_0, weight = layers_5_self_attn_linear_out_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("op_1652_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_127_cast_fp16 = add(x = x_115_cast_fp16, y = var_1652_cast_fp16)[name = string("x_127_cast_fp16")];
            tensor<int32, [1]> var_1668_axes_0 = const()[name = string("op_1668_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_5_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_5_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1668_cast_fp16 = layer_norm(axes = var_1668_axes_0, epsilon = layers_5_norm_conv_eps_scaled_to_fp16, x = x_127_cast_fp16)[name = string("op_1668_cast_fp16")];
            tensor<fp16, [1024]> input_157_gamma_0_to_fp16 = const()[name = string("input_157_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73757312)))];
            tensor<fp16, [1024]> input_157_beta_0_to_fp16 = const()[name = string("input_157_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73759424)))];
            fp16 input_157_epsilon_0_to_fp16 = const()[name = string("input_157_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_157_cast_fp16 = batch_norm(beta = input_157_beta_0_to_fp16, epsilon = input_157_epsilon_0_to_fp16, gamma = input_157_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1668_cast_fp16)[name = string("input_157_cast_fp16")];
            string input_159_pad_type_0 = const()[name = string("input_159_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_159_strides_0 = const()[name = string("input_159_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_159_pad_0 = const()[name = string("input_159_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_159_dilations_0 = const()[name = string("input_159_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_159_groups_0 = const()[name = string("input_159_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73761536))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74810176))))[name = string("layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_159_cast_fp16 = conv(dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("input_159_cast_fp16")];
            int32 x_129_split_num_splits_0 = const()[name = string("x_129_split_num_splits_0"), val = int32(2)];
            int32 x_129_split_axis_0 = const()[name = string("x_129_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_129_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_129_split_cast_fp16_1 = split(axis = x_129_split_axis_0, num_splits = x_129_split_num_splits_0, x = input_159_cast_fp16)[name = string("x_129_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_129_split_1_sigmoid_cast_fp16 = sigmoid(x = x_129_split_cast_fp16_1)[name = string("x_129_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_129_cast_fp16 = mul(x = x_129_split_cast_fp16_0, y = x_129_split_1_sigmoid_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_161_cast_fp16 = mul(x = x_129_cast_fp16, y = pad_mask)[name = string("input_161_cast_fp16")];
            string input_163_pad_type_0 = const()[name = string("input_163_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_163_pad_0 = const()[name = string("input_163_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_163_groups_0 = const()[name = string("input_163_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_163_strides_0 = const()[name = string("input_163_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_163_dilations_0 = const()[name = string("input_163_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_113_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74814336))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74819008))))[name = string("const_113_to_fp16_quantized")];
            tensor<fp16, [1024]> const_114_to_fp16 = const()[name = string("const_114_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74821120)))];
            tensor<fp16, [1, 1024, 1, 188]> input_165_cast_fp16 = conv(bias = const_114_to_fp16, dilations = input_163_dilations_0, groups = input_163_groups_0, pad = input_163_pad_0, pad_type = input_163_pad_type_0, strides = input_163_strides_0, weight = const_113_to_fp16_quantized, x = input_161_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_167_cast_fp16 = silu(x = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
            string var_1700_pad_type_0 = const()[name = string("op_1700_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1700_strides_0 = const()[name = string("op_1700_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1700_pad_0 = const()[name = string("op_1700_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1700_dilations_0 = const()[name = string("op_1700_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1700_groups_0 = const()[name = string("op_1700_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74823232))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75347584))))[name = string("layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1700_cast_fp16 = conv(dilations = var_1700_dilations_0, groups = var_1700_groups_0, pad = var_1700_pad_0, pad_type = var_1700_pad_type_0, strides = var_1700_strides_0, weight = layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("op_1700_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_131_cast_fp16 = add(x = x_127_cast_fp16, y = var_1700_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [1]> var_1716_axes_0 = const()[name = string("op_1716_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_5_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_5_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1716_cast_fp16 = layer_norm(axes = var_1716_axes_0, epsilon = layers_5_norm_feed_forward2_eps_scaled_to_fp16, x = x_131_cast_fp16)[name = string("op_1716_cast_fp16")];
            tensor<fp16, [1024]> input_169_gamma_0_to_fp16 = const()[name = string("input_169_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75349696)))];
            tensor<fp16, [1024]> input_169_beta_0_to_fp16 = const()[name = string("input_169_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75351808)))];
            fp16 input_169_epsilon_0_to_fp16 = const()[name = string("input_169_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_169_cast_fp16 = batch_norm(beta = input_169_beta_0_to_fp16, epsilon = input_169_epsilon_0_to_fp16, gamma = input_169_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1716_cast_fp16)[name = string("input_169_cast_fp16")];
            string input_171_pad_type_0 = const()[name = string("input_171_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_171_strides_0 = const()[name = string("input_171_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_171_pad_0 = const()[name = string("input_171_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_171_dilations_0 = const()[name = string("input_171_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_171_groups_0 = const()[name = string("input_171_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_5_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75353920))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77451136))))[name = string("layers_5_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_171_cast_fp16 = conv(dilations = input_171_dilations_0, groups = input_171_groups_0, pad = input_171_pad_0, pad_type = input_171_pad_type_0, strides = input_171_strides_0, weight = layers_5_feed_forward2_linear1_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_173_cast_fp16 = silu(x = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
            string var_1733_pad_type_0 = const()[name = string("op_1733_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1733_strides_0 = const()[name = string("op_1733_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1733_pad_0 = const()[name = string("op_1733_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1733_dilations_0 = const()[name = string("op_1733_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1733_groups_0 = const()[name = string("op_1733_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_1734_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77459392))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79556608))))[name = string("op_1734_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1734_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_1733_dilations_0, groups = var_1733_groups_0, pad = var_1733_pad_0, pad_type = var_1733_pad_type_0, strides = var_1733_strides_0, weight = op_1734_weight_0_to_fp16_quantized, x = input_173_cast_fp16)[name = string("op_1734_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_133_cast_fp16 = add(x = x_131_cast_fp16, y = var_1734_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<int32, [1]> var_1750_axes_0 = const()[name = string("op_1750_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_5_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_5_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1750_cast_fp16 = layer_norm(axes = var_1750_axes_0, epsilon = layers_5_norm_out_eps_scaled_to_fp16, x = x_133_cast_fp16)[name = string("op_1750_cast_fp16")];
            tensor<fp16, [1024]> x_135_gamma_0_to_fp16 = const()[name = string("x_135_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79558720)))];
            tensor<fp16, [1024]> x_135_beta_0_to_fp16 = const()[name = string("x_135_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79560832)))];
            fp16 x_135_epsilon_0_to_fp16 = const()[name = string("x_135_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_135_cast_fp16 = batch_norm(beta = x_135_beta_0_to_fp16, epsilon = x_135_epsilon_0_to_fp16, gamma = x_135_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1750_cast_fp16)[name = string("x_135_cast_fp16")];
            int32 var_1769 = const()[name = string("op_1769"), val = int32(1)];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_6_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_6_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1796_cast_fp16 = layer_norm(axes = var_1796_axes_0, epsilon = layers_6_norm_feed_forward1_eps_scaled_to_fp16, x = x_135_cast_fp16)[name = string("op_1796_cast_fp16")];
            tensor<fp16, [1024]> input_175_gamma_0_to_fp16 = const()[name = string("input_175_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79562944)))];
            tensor<fp16, [1024]> input_175_beta_0_to_fp16 = const()[name = string("input_175_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79565056)))];
            fp16 input_175_epsilon_0_to_fp16 = const()[name = string("input_175_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_175_cast_fp16 = batch_norm(beta = input_175_beta_0_to_fp16, epsilon = input_175_epsilon_0_to_fp16, gamma = input_175_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1796_cast_fp16)[name = string("input_175_cast_fp16")];
            string input_177_pad_type_0 = const()[name = string("input_177_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_177_strides_0 = const()[name = string("input_177_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_177_pad_0 = const()[name = string("input_177_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_177_dilations_0 = const()[name = string("input_177_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_177_groups_0 = const()[name = string("input_177_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_6_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79567168))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81664384))))[name = string("layers_6_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_177_cast_fp16 = conv(dilations = input_177_dilations_0, groups = input_177_groups_0, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = input_177_strides_0, weight = layers_6_feed_forward1_linear1_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_179_cast_fp16 = silu(x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
            string var_1813_pad_type_0 = const()[name = string("op_1813_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1813_strides_0 = const()[name = string("op_1813_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1813_pad_0 = const()[name = string("op_1813_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1813_dilations_0 = const()[name = string("op_1813_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1813_groups_0 = const()[name = string("op_1813_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_1814_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81672640))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83769856))))[name = string("op_1814_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1814_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_1813_dilations_0, groups = var_1813_groups_0, pad = var_1813_pad_0, pad_type = var_1813_pad_type_0, strides = var_1813_strides_0, weight = op_1814_weight_0_to_fp16_quantized, x = input_179_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_137_cast_fp16 = add(x = x_135_cast_fp16, y = var_1814_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<int32, [1]> var_1830_axes_0 = const()[name = string("op_1830_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_6_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_6_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1830_cast_fp16 = layer_norm(axes = var_1830_axes_0, epsilon = layers_6_norm_self_att_eps_scaled_to_fp16, x = x_137_cast_fp16)[name = string("op_1830_cast_fp16")];
            tensor<fp16, [1024]> x_139_gamma_0_to_fp16 = const()[name = string("x_139_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83771968)))];
            tensor<fp16, [1024]> x_139_beta_0_to_fp16 = const()[name = string("x_139_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83774080)))];
            fp16 x_139_epsilon_0_to_fp16 = const()[name = string("x_139_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_139_cast_fp16 = batch_norm(beta = x_139_beta_0_to_fp16, epsilon = x_139_epsilon_0_to_fp16, gamma = x_139_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1830_cast_fp16)[name = string("x_139_cast_fp16")];
            string q_13_pad_type_0 = const()[name = string("q_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_13_strides_0 = const()[name = string("q_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_13_pad_0 = const()[name = string("q_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_13_dilations_0 = const()[name = string("q_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_13_groups_0 = const()[name = string("q_13_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83776192))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84300544))))[name = string("layers_6_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_13_cast_fp16 = conv(dilations = q_13_dilations_0, groups = q_13_groups_0, pad = q_13_pad_0, pad_type = q_13_pad_type_0, strides = q_13_strides_0, weight = layers_6_self_attn_linear_q_weight_to_fp16_quantized, x = x_139_cast_fp16)[name = string("q_13_cast_fp16")];
            string k_13_pad_type_0 = const()[name = string("k_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_13_strides_0 = const()[name = string("k_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_13_pad_0 = const()[name = string("k_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_13_dilations_0 = const()[name = string("k_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_13_groups_0 = const()[name = string("k_13_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84302656))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84827008))))[name = string("layers_6_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_13_cast_fp16 = conv(dilations = k_13_dilations_0, groups = k_13_groups_0, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = k_13_strides_0, weight = layers_6_self_attn_linear_k_weight_to_fp16_quantized, x = x_139_cast_fp16)[name = string("k_13_cast_fp16")];
            string v_13_pad_type_0 = const()[name = string("v_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_13_strides_0 = const()[name = string("v_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_13_pad_0 = const()[name = string("v_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_13_dilations_0 = const()[name = string("v_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_13_groups_0 = const()[name = string("v_13_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84829120))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85353472))))[name = string("layers_6_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_13_cast_fp16 = conv(dilations = v_13_dilations_0, groups = v_13_groups_0, pad = v_13_pad_0, pad_type = v_13_pad_type_0, strides = v_13_strides_0, weight = layers_6_self_attn_linear_v_weight_to_fp16_quantized, x = x_139_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_13_to_fp16 = const()[name = string("bv_all_13_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85355584)))];
            tensor<fp16, [1, 1024, 1, 188]> var_1863_cast_fp16 = add(x = q_13_cast_fp16, y = bv_all_13_to_fp16)[name = string("op_1863_cast_fp16")];
            tensor<int32, [3]> var_1864 = const()[name = string("op_1864"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_13_cast_fp16 = reshape(shape = var_1864, x = var_1863_cast_fp16)[name = string("qb_13_cast_fp16")];
            bool bd_all_25_transpose_x_0 = const()[name = string("bd_all_25_transpose_x_0"), val = bool(false)];
            bool bd_all_25_transpose_y_0 = const()[name = string("bd_all_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_6_self_attn_pos_proj_to_fp16 = const()[name = string("layers_6_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85357696)))];
            tensor<fp16, [8, 375, 188]> bd_all_25_cast_fp16 = matmul(transpose_x = bd_all_25_transpose_x_0, transpose_y = bd_all_25_transpose_y_0, x = layers_6_self_attn_pos_proj_to_fp16, y = qb_13_cast_fp16)[name = string("bd_all_25_cast_fp16")];
            tensor<int32, [3]> x_141_perm_0 = const()[name = string("x_141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_143_pad_0 = const()[name = string("x_143_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_143_mode_0 = const()[name = string("x_143_mode_0"), val = string("constant")];
            fp16 const_34_to_fp16 = const()[name = string("const_34_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_141_cast_fp16 = transpose(perm = x_141_perm_0, x = bd_all_25_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [8, 188, 376]> x_143_cast_fp16 = pad(constant_val = const_34_to_fp16, mode = x_143_mode_0, pad = x_143_pad_0, x = x_141_cast_fp16)[name = string("x_143_cast_fp16")];
            tensor<int32, [3]> var_1871 = const()[name = string("op_1871"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_145_cast_fp16 = reshape(shape = var_1871, x = x_143_cast_fp16)[name = string("x_145_cast_fp16")];
            tensor<int32, [3]> var_1874_begin_0 = const()[name = string("op_1874_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_1874_end_0 = const()[name = string("op_1874_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_1874_end_mask_0 = const()[name = string("op_1874_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_1874_cast_fp16 = slice_by_index(begin = var_1874_begin_0, end = var_1874_end_0, end_mask = var_1874_end_mask_0, x = x_145_cast_fp16)[name = string("op_1874_cast_fp16")];
            tensor<int32, [3]> var_1875 = const()[name = string("op_1875"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_147_cast_fp16 = reshape(shape = var_1875, x = var_1874_cast_fp16)[name = string("x_147_cast_fp16")];
            tensor<int32, [3]> bd_all_27_begin_0 = const()[name = string("bd_all_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_27_end_0 = const()[name = string("bd_all_27_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_27_end_mask_0 = const()[name = string("bd_all_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_27_cast_fp16 = slice_by_index(begin = bd_all_27_begin_0, end = bd_all_27_end_0, end_mask = bd_all_27_end_mask_0, x = x_147_cast_fp16)[name = string("bd_all_27_cast_fp16")];
            tensor<int32, [4]> var_1880 = const()[name = string("op_1880"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1881_cast_fp16 = reshape(shape = var_1880, x = q_13_cast_fp16)[name = string("op_1881_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_1883_to_fp16 = const()[name = string("op_1883_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86125760)))];
            tensor<fp16, [8, 128, 1, 188]> qh_13_cast_fp16 = add(x = var_1881_cast_fp16, y = var_1883_to_fp16)[name = string("qh_13_cast_fp16")];
            tensor<int32, [4]> var_1885 = const()[name = string("op_1885"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1886_cast_fp16 = reshape(shape = var_1885, x = k_13_cast_fp16)[name = string("op_1886_cast_fp16")];
            tensor<int32, [4]> var_1887 = const()[name = string("op_1887"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_1889 = const()[name = string("op_1889"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_1890_cast_fp16 = reshape(shape = var_1889, x = v_13_cast_fp16)[name = string("op_1890_cast_fp16")];
            tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1893_perm_0 = const()[name = string("op_1893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_13_axes_0 = const()[name = string("bd_13_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_1893_cast_fp16 = transpose(perm = var_1893_perm_0, x = bd_all_27_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [8, 188, 1, 188]> bd_13_cast_fp16 = expand_dims(axes = bd_13_axes_0, x = var_1893_cast_fp16)[name = string("bd_13_cast_fp16")];
            string ac_13_equation_0 = const()[name = string("ac_13_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_1888_cast_fp16 = transpose(perm = var_1887, x = var_1886_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [8, 188, 1, 188]> ac_13_cast_fp16 = einsum(equation = ac_13_equation_0, values = (var_1888_cast_fp16, qh_13_cast_fp16))[name = string("ac_13_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_1909_cast_fp16 = add(x = ac_13_cast_fp16, y = bd_13_cast_fp16)[name = string("op_1909_cast_fp16")];
            fp16 var_1910_to_fp16 = const()[name = string("op_1910_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_25_cast_fp16 = mul(x = var_1909_cast_fp16, y = var_1910_to_fp16)[name = string("scores_25_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = key_bias)[name = string("scores_27_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_1913_cast_fp16 = softmax(axis = var_1769, x = scores_27_cast_fp16)[name = string("op_1913_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_64 = const()[name = string("concat_64"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = var_1913_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [8, 188, 188]> reshape_18_cast_fp16 = reshape(shape = concat_64, x = transpose_12_cast_fp16)[name = string("reshape_18_cast_fp16")];
            tensor<int32, [3]> concat_65 = const()[name = string("concat_65"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_54_cast_fp16 = transpose(perm = transpose_54_perm_0, x = var_1890_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [8, 188, 128]> reshape_19_cast_fp16 = reshape(shape = concat_65, x = transpose_54_cast_fp16)[name = string("reshape_19_cast_fp16")];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = reshape_18_cast_fp16, y = reshape_19_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_20_cast_fp16 = reshape(shape = concat_69, x = matmul_6_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [4]> ctx_13_perm_0 = const()[name = string("ctx_13_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1916 = const()[name = string("op_1916"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_13_cast_fp16 = transpose(perm = ctx_13_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 1024, 1, 188]> input_181_cast_fp16 = reshape(shape = var_1916, x = ctx_13_cast_fp16)[name = string("input_181_cast_fp16")];
            string var_1923_pad_type_0 = const()[name = string("op_1923_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1923_strides_0 = const()[name = string("op_1923_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1923_pad_0 = const()[name = string("op_1923_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1923_dilations_0 = const()[name = string("op_1923_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1923_groups_0 = const()[name = string("op_1923_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86127872))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86652224))))[name = string("layers_6_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1923_cast_fp16 = conv(dilations = var_1923_dilations_0, groups = var_1923_groups_0, pad = var_1923_pad_0, pad_type = var_1923_pad_type_0, strides = var_1923_strides_0, weight = layers_6_self_attn_linear_out_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_149_cast_fp16 = add(x = x_137_cast_fp16, y = var_1923_cast_fp16)[name = string("x_149_cast_fp16")];
            tensor<int32, [1]> var_1939_axes_0 = const()[name = string("op_1939_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_6_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_6_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1939_cast_fp16 = layer_norm(axes = var_1939_axes_0, epsilon = layers_6_norm_conv_eps_scaled_to_fp16, x = x_149_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<fp16, [1024]> input_183_gamma_0_to_fp16 = const()[name = string("input_183_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86654336)))];
            tensor<fp16, [1024]> input_183_beta_0_to_fp16 = const()[name = string("input_183_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86656448)))];
            fp16 input_183_epsilon_0_to_fp16 = const()[name = string("input_183_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_183_cast_fp16 = batch_norm(beta = input_183_beta_0_to_fp16, epsilon = input_183_epsilon_0_to_fp16, gamma = input_183_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1939_cast_fp16)[name = string("input_183_cast_fp16")];
            string input_185_pad_type_0 = const()[name = string("input_185_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_185_strides_0 = const()[name = string("input_185_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_185_pad_0 = const()[name = string("input_185_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_185_dilations_0 = const()[name = string("input_185_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_185_groups_0 = const()[name = string("input_185_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86658560))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87707200))))[name = string("layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_185_cast_fp16 = conv(dilations = input_185_dilations_0, groups = input_185_groups_0, pad = input_185_pad_0, pad_type = input_185_pad_type_0, strides = input_185_strides_0, weight = layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("input_185_cast_fp16")];
            int32 x_151_split_num_splits_0 = const()[name = string("x_151_split_num_splits_0"), val = int32(2)];
            int32 x_151_split_axis_0 = const()[name = string("x_151_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_151_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_151_split_cast_fp16_1 = split(axis = x_151_split_axis_0, num_splits = x_151_split_num_splits_0, x = input_185_cast_fp16)[name = string("x_151_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_151_split_1_sigmoid_cast_fp16 = sigmoid(x = x_151_split_cast_fp16_1)[name = string("x_151_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_151_cast_fp16 = mul(x = x_151_split_cast_fp16_0, y = x_151_split_1_sigmoid_cast_fp16)[name = string("x_151_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_187_cast_fp16 = mul(x = x_151_cast_fp16, y = pad_mask)[name = string("input_187_cast_fp16")];
            string input_189_pad_type_0 = const()[name = string("input_189_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_189_pad_0 = const()[name = string("input_189_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_189_groups_0 = const()[name = string("input_189_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_189_strides_0 = const()[name = string("input_189_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_189_dilations_0 = const()[name = string("input_189_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_115_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87711360))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87716032))))[name = string("const_115_to_fp16_quantized")];
            tensor<fp16, [1024]> const_116_to_fp16 = const()[name = string("const_116_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87718144)))];
            tensor<fp16, [1, 1024, 1, 188]> input_191_cast_fp16 = conv(bias = const_116_to_fp16, dilations = input_189_dilations_0, groups = input_189_groups_0, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = input_189_strides_0, weight = const_115_to_fp16_quantized, x = input_187_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_193_cast_fp16 = silu(x = input_191_cast_fp16)[name = string("input_193_cast_fp16")];
            string var_1971_pad_type_0 = const()[name = string("op_1971_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1971_strides_0 = const()[name = string("op_1971_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1971_pad_0 = const()[name = string("op_1971_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1971_dilations_0 = const()[name = string("op_1971_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1971_groups_0 = const()[name = string("op_1971_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87720256))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88244608))))[name = string("layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_1971_cast_fp16 = conv(dilations = var_1971_dilations_0, groups = var_1971_groups_0, pad = var_1971_pad_0, pad_type = var_1971_pad_type_0, strides = var_1971_strides_0, weight = layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_153_cast_fp16 = add(x = x_149_cast_fp16, y = var_1971_cast_fp16)[name = string("x_153_cast_fp16")];
            tensor<int32, [1]> var_1987_axes_0 = const()[name = string("op_1987_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_6_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_6_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_1987_cast_fp16 = layer_norm(axes = var_1987_axes_0, epsilon = layers_6_norm_feed_forward2_eps_scaled_to_fp16, x = x_153_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<fp16, [1024]> input_195_gamma_0_to_fp16 = const()[name = string("input_195_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88246720)))];
            tensor<fp16, [1024]> input_195_beta_0_to_fp16 = const()[name = string("input_195_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88248832)))];
            fp16 input_195_epsilon_0_to_fp16 = const()[name = string("input_195_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_1987_cast_fp16)[name = string("input_195_cast_fp16")];
            string input_197_pad_type_0 = const()[name = string("input_197_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_197_strides_0 = const()[name = string("input_197_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_197_pad_0 = const()[name = string("input_197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_197_dilations_0 = const()[name = string("input_197_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_197_groups_0 = const()[name = string("input_197_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_6_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88250944))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90348160))))[name = string("layers_6_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_197_cast_fp16 = conv(dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = layers_6_feed_forward2_linear1_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_199_cast_fp16 = silu(x = input_197_cast_fp16)[name = string("input_199_cast_fp16")];
            string var_2004_pad_type_0 = const()[name = string("op_2004_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2004_strides_0 = const()[name = string("op_2004_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2004_pad_0 = const()[name = string("op_2004_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2004_dilations_0 = const()[name = string("op_2004_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2004_groups_0 = const()[name = string("op_2004_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_2005_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90356416))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92453632))))[name = string("op_2005_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2005_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_2004_dilations_0, groups = var_2004_groups_0, pad = var_2004_pad_0, pad_type = var_2004_pad_type_0, strides = var_2004_strides_0, weight = op_2005_weight_0_to_fp16_quantized, x = input_199_cast_fp16)[name = string("op_2005_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_155_cast_fp16 = add(x = x_153_cast_fp16, y = var_2005_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<int32, [1]> var_2021_axes_0 = const()[name = string("op_2021_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_6_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_6_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2021_cast_fp16 = layer_norm(axes = var_2021_axes_0, epsilon = layers_6_norm_out_eps_scaled_to_fp16, x = x_155_cast_fp16)[name = string("op_2021_cast_fp16")];
            tensor<fp16, [1024]> x_157_gamma_0_to_fp16 = const()[name = string("x_157_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92455744)))];
            tensor<fp16, [1024]> x_157_beta_0_to_fp16 = const()[name = string("x_157_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92457856)))];
            fp16 x_157_epsilon_0_to_fp16 = const()[name = string("x_157_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_157_cast_fp16 = batch_norm(beta = x_157_beta_0_to_fp16, epsilon = x_157_epsilon_0_to_fp16, gamma = x_157_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2021_cast_fp16)[name = string("x_157_cast_fp16")];
            int32 var_2040 = const()[name = string("op_2040"), val = int32(1)];
            tensor<int32, [1]> var_2067_axes_0 = const()[name = string("op_2067_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_7_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_7_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2067_cast_fp16 = layer_norm(axes = var_2067_axes_0, epsilon = layers_7_norm_feed_forward1_eps_scaled_to_fp16, x = x_157_cast_fp16)[name = string("op_2067_cast_fp16")];
            tensor<fp16, [1024]> input_201_gamma_0_to_fp16 = const()[name = string("input_201_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92459968)))];
            tensor<fp16, [1024]> input_201_beta_0_to_fp16 = const()[name = string("input_201_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92462080)))];
            fp16 input_201_epsilon_0_to_fp16 = const()[name = string("input_201_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_201_cast_fp16 = batch_norm(beta = input_201_beta_0_to_fp16, epsilon = input_201_epsilon_0_to_fp16, gamma = input_201_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2067_cast_fp16)[name = string("input_201_cast_fp16")];
            string input_203_pad_type_0 = const()[name = string("input_203_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_203_strides_0 = const()[name = string("input_203_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_203_pad_0 = const()[name = string("input_203_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_203_dilations_0 = const()[name = string("input_203_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_203_groups_0 = const()[name = string("input_203_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_7_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92464192))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94561408))))[name = string("layers_7_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_203_cast_fp16 = conv(dilations = input_203_dilations_0, groups = input_203_groups_0, pad = input_203_pad_0, pad_type = input_203_pad_type_0, strides = input_203_strides_0, weight = layers_7_feed_forward1_linear1_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_205_cast_fp16 = silu(x = input_203_cast_fp16)[name = string("input_205_cast_fp16")];
            string var_2084_pad_type_0 = const()[name = string("op_2084_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2084_strides_0 = const()[name = string("op_2084_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2084_pad_0 = const()[name = string("op_2084_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2084_dilations_0 = const()[name = string("op_2084_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2084_groups_0 = const()[name = string("op_2084_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_2085_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94569664))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96666880))))[name = string("op_2085_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2085_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_2084_dilations_0, groups = var_2084_groups_0, pad = var_2084_pad_0, pad_type = var_2084_pad_type_0, strides = var_2084_strides_0, weight = op_2085_weight_0_to_fp16_quantized, x = input_205_cast_fp16)[name = string("op_2085_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_159_cast_fp16 = add(x = x_157_cast_fp16, y = var_2085_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<int32, [1]> var_2101_axes_0 = const()[name = string("op_2101_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_7_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_7_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2101_cast_fp16 = layer_norm(axes = var_2101_axes_0, epsilon = layers_7_norm_self_att_eps_scaled_to_fp16, x = x_159_cast_fp16)[name = string("op_2101_cast_fp16")];
            tensor<fp16, [1024]> x_161_gamma_0_to_fp16 = const()[name = string("x_161_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96668992)))];
            tensor<fp16, [1024]> x_161_beta_0_to_fp16 = const()[name = string("x_161_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96671104)))];
            fp16 x_161_epsilon_0_to_fp16 = const()[name = string("x_161_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_161_cast_fp16 = batch_norm(beta = x_161_beta_0_to_fp16, epsilon = x_161_epsilon_0_to_fp16, gamma = x_161_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2101_cast_fp16)[name = string("x_161_cast_fp16")];
            string q_15_pad_type_0 = const()[name = string("q_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_15_strides_0 = const()[name = string("q_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_15_pad_0 = const()[name = string("q_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_15_dilations_0 = const()[name = string("q_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_15_groups_0 = const()[name = string("q_15_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96673216))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97197568))))[name = string("layers_7_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_15_cast_fp16 = conv(dilations = q_15_dilations_0, groups = q_15_groups_0, pad = q_15_pad_0, pad_type = q_15_pad_type_0, strides = q_15_strides_0, weight = layers_7_self_attn_linear_q_weight_to_fp16_quantized, x = x_161_cast_fp16)[name = string("q_15_cast_fp16")];
            string k_15_pad_type_0 = const()[name = string("k_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_15_strides_0 = const()[name = string("k_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_15_pad_0 = const()[name = string("k_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_15_dilations_0 = const()[name = string("k_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_15_groups_0 = const()[name = string("k_15_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97199680))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97724032))))[name = string("layers_7_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_15_cast_fp16 = conv(dilations = k_15_dilations_0, groups = k_15_groups_0, pad = k_15_pad_0, pad_type = k_15_pad_type_0, strides = k_15_strides_0, weight = layers_7_self_attn_linear_k_weight_to_fp16_quantized, x = x_161_cast_fp16)[name = string("k_15_cast_fp16")];
            string v_15_pad_type_0 = const()[name = string("v_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_15_strides_0 = const()[name = string("v_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_15_pad_0 = const()[name = string("v_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_15_dilations_0 = const()[name = string("v_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_15_groups_0 = const()[name = string("v_15_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97726144))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98250496))))[name = string("layers_7_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_15_cast_fp16 = conv(dilations = v_15_dilations_0, groups = v_15_groups_0, pad = v_15_pad_0, pad_type = v_15_pad_type_0, strides = v_15_strides_0, weight = layers_7_self_attn_linear_v_weight_to_fp16_quantized, x = x_161_cast_fp16)[name = string("v_15_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_15_to_fp16 = const()[name = string("bv_all_15_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98252608)))];
            tensor<fp16, [1, 1024, 1, 188]> var_2134_cast_fp16 = add(x = q_15_cast_fp16, y = bv_all_15_to_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [3]> var_2135 = const()[name = string("op_2135"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_15_cast_fp16 = reshape(shape = var_2135, x = var_2134_cast_fp16)[name = string("qb_15_cast_fp16")];
            bool bd_all_29_transpose_x_0 = const()[name = string("bd_all_29_transpose_x_0"), val = bool(false)];
            bool bd_all_29_transpose_y_0 = const()[name = string("bd_all_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_7_self_attn_pos_proj_to_fp16 = const()[name = string("layers_7_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98254720)))];
            tensor<fp16, [8, 375, 188]> bd_all_29_cast_fp16 = matmul(transpose_x = bd_all_29_transpose_x_0, transpose_y = bd_all_29_transpose_y_0, x = layers_7_self_attn_pos_proj_to_fp16, y = qb_15_cast_fp16)[name = string("bd_all_29_cast_fp16")];
            tensor<int32, [3]> x_163_perm_0 = const()[name = string("x_163_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")];
            fp16 const_38_to_fp16 = const()[name = string("const_38_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_163_cast_fp16 = transpose(perm = x_163_perm_0, x = bd_all_29_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [8, 188, 376]> x_165_cast_fp16 = pad(constant_val = const_38_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")];
            tensor<int32, [3]> var_2142 = const()[name = string("op_2142"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_167_cast_fp16 = reshape(shape = var_2142, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")];
            tensor<int32, [3]> var_2145_begin_0 = const()[name = string("op_2145_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2145_end_0 = const()[name = string("op_2145_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_2145_end_mask_0 = const()[name = string("op_2145_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_2145_cast_fp16 = slice_by_index(begin = var_2145_begin_0, end = var_2145_end_0, end_mask = var_2145_end_mask_0, x = x_167_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<int32, [3]> var_2146 = const()[name = string("op_2146"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_169_cast_fp16 = reshape(shape = var_2146, x = var_2145_cast_fp16)[name = string("x_169_cast_fp16")];
            tensor<int32, [3]> bd_all_31_begin_0 = const()[name = string("bd_all_31_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_31_end_0 = const()[name = string("bd_all_31_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_31_end_mask_0 = const()[name = string("bd_all_31_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_31_cast_fp16 = slice_by_index(begin = bd_all_31_begin_0, end = bd_all_31_end_0, end_mask = bd_all_31_end_mask_0, x = x_169_cast_fp16)[name = string("bd_all_31_cast_fp16")];
            tensor<int32, [4]> var_2151 = const()[name = string("op_2151"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2152_cast_fp16 = reshape(shape = var_2151, x = q_15_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_2154_to_fp16 = const()[name = string("op_2154_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99022784)))];
            tensor<fp16, [8, 128, 1, 188]> qh_15_cast_fp16 = add(x = var_2152_cast_fp16, y = var_2154_to_fp16)[name = string("qh_15_cast_fp16")];
            tensor<int32, [4]> var_2156 = const()[name = string("op_2156"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2157_cast_fp16 = reshape(shape = var_2156, x = k_15_cast_fp16)[name = string("op_2157_cast_fp16")];
            tensor<int32, [4]> var_2158 = const()[name = string("op_2158"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_2160 = const()[name = string("op_2160"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2161_cast_fp16 = reshape(shape = var_2160, x = v_15_cast_fp16)[name = string("op_2161_cast_fp16")];
            tensor<int32, [4]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2164_perm_0 = const()[name = string("op_2164_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_15_axes_0 = const()[name = string("bd_15_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_2164_cast_fp16 = transpose(perm = var_2164_perm_0, x = bd_all_31_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [8, 188, 1, 188]> bd_15_cast_fp16 = expand_dims(axes = bd_15_axes_0, x = var_2164_cast_fp16)[name = string("bd_15_cast_fp16")];
            string ac_15_equation_0 = const()[name = string("ac_15_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_2159_cast_fp16 = transpose(perm = var_2158, x = var_2157_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [8, 188, 1, 188]> ac_15_cast_fp16 = einsum(equation = ac_15_equation_0, values = (var_2159_cast_fp16, qh_15_cast_fp16))[name = string("ac_15_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_2180_cast_fp16 = add(x = ac_15_cast_fp16, y = bd_15_cast_fp16)[name = string("op_2180_cast_fp16")];
            fp16 var_2181_to_fp16 = const()[name = string("op_2181_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_29_cast_fp16 = mul(x = var_2180_cast_fp16, y = var_2181_to_fp16)[name = string("scores_29_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = key_bias)[name = string("scores_31_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_2184_cast_fp16 = softmax(axis = var_2040, x = scores_31_cast_fp16)[name = string("op_2184_cast_fp16")];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_74 = const()[name = string("concat_74"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = var_2184_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [8, 188, 188]> reshape_21_cast_fp16 = reshape(shape = concat_74, x = transpose_14_cast_fp16)[name = string("reshape_21_cast_fp16")];
            tensor<int32, [3]> concat_75 = const()[name = string("concat_75"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = var_2161_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [8, 188, 128]> reshape_22_cast_fp16 = reshape(shape = concat_75, x = transpose_55_cast_fp16)[name = string("reshape_22_cast_fp16")];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = reshape_21_cast_fp16, y = reshape_22_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_23_cast_fp16 = reshape(shape = concat_79, x = matmul_7_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> ctx_15_perm_0 = const()[name = string("ctx_15_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2187 = const()[name = string("op_2187"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_15_cast_fp16 = transpose(perm = ctx_15_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 1024, 1, 188]> input_207_cast_fp16 = reshape(shape = var_2187, x = ctx_15_cast_fp16)[name = string("input_207_cast_fp16")];
            string var_2194_pad_type_0 = const()[name = string("op_2194_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2194_strides_0 = const()[name = string("op_2194_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2194_pad_0 = const()[name = string("op_2194_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2194_dilations_0 = const()[name = string("op_2194_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2194_groups_0 = const()[name = string("op_2194_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99024896))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99549248))))[name = string("layers_7_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2194_cast_fp16 = conv(dilations = var_2194_dilations_0, groups = var_2194_groups_0, pad = var_2194_pad_0, pad_type = var_2194_pad_type_0, strides = var_2194_strides_0, weight = layers_7_self_attn_linear_out_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_171_cast_fp16 = add(x = x_159_cast_fp16, y = var_2194_cast_fp16)[name = string("x_171_cast_fp16")];
            tensor<int32, [1]> var_2210_axes_0 = const()[name = string("op_2210_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_7_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_7_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2210_cast_fp16 = layer_norm(axes = var_2210_axes_0, epsilon = layers_7_norm_conv_eps_scaled_to_fp16, x = x_171_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<fp16, [1024]> input_209_gamma_0_to_fp16 = const()[name = string("input_209_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99551360)))];
            tensor<fp16, [1024]> input_209_beta_0_to_fp16 = const()[name = string("input_209_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99553472)))];
            fp16 input_209_epsilon_0_to_fp16 = const()[name = string("input_209_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_209_cast_fp16 = batch_norm(beta = input_209_beta_0_to_fp16, epsilon = input_209_epsilon_0_to_fp16, gamma = input_209_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2210_cast_fp16)[name = string("input_209_cast_fp16")];
            string input_211_pad_type_0 = const()[name = string("input_211_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_211_strides_0 = const()[name = string("input_211_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_211_pad_0 = const()[name = string("input_211_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_211_dilations_0 = const()[name = string("input_211_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_211_groups_0 = const()[name = string("input_211_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99555584))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100604224))))[name = string("layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_211_cast_fp16 = conv(dilations = input_211_dilations_0, groups = input_211_groups_0, pad = input_211_pad_0, pad_type = input_211_pad_type_0, strides = input_211_strides_0, weight = layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")];
            int32 x_173_split_num_splits_0 = const()[name = string("x_173_split_num_splits_0"), val = int32(2)];
            int32 x_173_split_axis_0 = const()[name = string("x_173_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_173_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_173_split_cast_fp16_1 = split(axis = x_173_split_axis_0, num_splits = x_173_split_num_splits_0, x = input_211_cast_fp16)[name = string("x_173_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_173_split_1_sigmoid_cast_fp16 = sigmoid(x = x_173_split_cast_fp16_1)[name = string("x_173_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_173_cast_fp16 = mul(x = x_173_split_cast_fp16_0, y = x_173_split_1_sigmoid_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_213_cast_fp16 = mul(x = x_173_cast_fp16, y = pad_mask)[name = string("input_213_cast_fp16")];
            string input_215_pad_type_0 = const()[name = string("input_215_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_215_pad_0 = const()[name = string("input_215_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_215_groups_0 = const()[name = string("input_215_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_215_strides_0 = const()[name = string("input_215_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_215_dilations_0 = const()[name = string("input_215_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_117_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100608384))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100613056))))[name = string("const_117_to_fp16_quantized")];
            tensor<fp16, [1024]> const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100615168)))];
            tensor<fp16, [1, 1024, 1, 188]> input_217_cast_fp16 = conv(bias = const_118_to_fp16, dilations = input_215_dilations_0, groups = input_215_groups_0, pad = input_215_pad_0, pad_type = input_215_pad_type_0, strides = input_215_strides_0, weight = const_117_to_fp16_quantized, x = input_213_cast_fp16)[name = string("input_217_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = string("input_219_cast_fp16")];
            string var_2242_pad_type_0 = const()[name = string("op_2242_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2242_strides_0 = const()[name = string("op_2242_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2242_pad_0 = const()[name = string("op_2242_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2242_dilations_0 = const()[name = string("op_2242_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2242_groups_0 = const()[name = string("op_2242_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100617280))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101141632))))[name = string("layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2242_cast_fp16 = conv(dilations = var_2242_dilations_0, groups = var_2242_groups_0, pad = var_2242_pad_0, pad_type = var_2242_pad_type_0, strides = var_2242_strides_0, weight = layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("op_2242_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_175_cast_fp16 = add(x = x_171_cast_fp16, y = var_2242_cast_fp16)[name = string("x_175_cast_fp16")];
            tensor<int32, [1]> var_2258_axes_0 = const()[name = string("op_2258_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_7_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_7_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2258_cast_fp16 = layer_norm(axes = var_2258_axes_0, epsilon = layers_7_norm_feed_forward2_eps_scaled_to_fp16, x = x_175_cast_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1024]> input_221_gamma_0_to_fp16 = const()[name = string("input_221_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101143744)))];
            tensor<fp16, [1024]> input_221_beta_0_to_fp16 = const()[name = string("input_221_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101145856)))];
            fp16 input_221_epsilon_0_to_fp16 = const()[name = string("input_221_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_221_cast_fp16 = batch_norm(beta = input_221_beta_0_to_fp16, epsilon = input_221_epsilon_0_to_fp16, gamma = input_221_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2258_cast_fp16)[name = string("input_221_cast_fp16")];
            string input_223_pad_type_0 = const()[name = string("input_223_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_223_strides_0 = const()[name = string("input_223_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_223_pad_0 = const()[name = string("input_223_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_223_dilations_0 = const()[name = string("input_223_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_223_groups_0 = const()[name = string("input_223_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_7_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101147968))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103245184))))[name = string("layers_7_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_223_cast_fp16 = conv(dilations = input_223_dilations_0, groups = input_223_groups_0, pad = input_223_pad_0, pad_type = input_223_pad_type_0, strides = input_223_strides_0, weight = layers_7_feed_forward2_linear1_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_225_cast_fp16 = silu(x = input_223_cast_fp16)[name = string("input_225_cast_fp16")];
            string var_2275_pad_type_0 = const()[name = string("op_2275_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2275_strides_0 = const()[name = string("op_2275_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2275_pad_0 = const()[name = string("op_2275_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2275_dilations_0 = const()[name = string("op_2275_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2275_groups_0 = const()[name = string("op_2275_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_2276_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103253440))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105350656))))[name = string("op_2276_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2276_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_2275_dilations_0, groups = var_2275_groups_0, pad = var_2275_pad_0, pad_type = var_2275_pad_type_0, strides = var_2275_strides_0, weight = op_2276_weight_0_to_fp16_quantized, x = input_225_cast_fp16)[name = string("op_2276_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_177_cast_fp16 = add(x = x_175_cast_fp16, y = var_2276_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [1]> var_2292_axes_0 = const()[name = string("op_2292_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_7_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_7_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2292_cast_fp16 = layer_norm(axes = var_2292_axes_0, epsilon = layers_7_norm_out_eps_scaled_to_fp16, x = x_177_cast_fp16)[name = string("op_2292_cast_fp16")];
            tensor<fp16, [1024]> x_179_gamma_0_to_fp16 = const()[name = string("x_179_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105352768)))];
            tensor<fp16, [1024]> x_179_beta_0_to_fp16 = const()[name = string("x_179_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105354880)))];
            fp16 x_179_epsilon_0_to_fp16 = const()[name = string("x_179_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_179_cast_fp16 = batch_norm(beta = x_179_beta_0_to_fp16, epsilon = x_179_epsilon_0_to_fp16, gamma = x_179_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2292_cast_fp16)[name = string("x_179_cast_fp16")];
            int32 var_2311 = const()[name = string("op_2311"), val = int32(1)];
            tensor<int32, [1]> var_2338_axes_0 = const()[name = string("op_2338_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_8_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_8_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2338_cast_fp16 = layer_norm(axes = var_2338_axes_0, epsilon = layers_8_norm_feed_forward1_eps_scaled_to_fp16, x = x_179_cast_fp16)[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1024]> input_227_gamma_0_to_fp16 = const()[name = string("input_227_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105356992)))];
            tensor<fp16, [1024]> input_227_beta_0_to_fp16 = const()[name = string("input_227_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105359104)))];
            fp16 input_227_epsilon_0_to_fp16 = const()[name = string("input_227_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_227_cast_fp16 = batch_norm(beta = input_227_beta_0_to_fp16, epsilon = input_227_epsilon_0_to_fp16, gamma = input_227_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2338_cast_fp16)[name = string("input_227_cast_fp16")];
            string input_229_pad_type_0 = const()[name = string("input_229_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_229_strides_0 = const()[name = string("input_229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_229_pad_0 = const()[name = string("input_229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_229_dilations_0 = const()[name = string("input_229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_229_groups_0 = const()[name = string("input_229_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_8_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105361216))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107458432))))[name = string("layers_8_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_229_cast_fp16 = conv(dilations = input_229_dilations_0, groups = input_229_groups_0, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = input_229_strides_0, weight = layers_8_feed_forward1_linear1_weight_to_fp16_quantized, x = input_227_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_231_cast_fp16 = silu(x = input_229_cast_fp16)[name = string("input_231_cast_fp16")];
            string var_2355_pad_type_0 = const()[name = string("op_2355_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2355_strides_0 = const()[name = string("op_2355_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2355_pad_0 = const()[name = string("op_2355_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2355_dilations_0 = const()[name = string("op_2355_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2355_groups_0 = const()[name = string("op_2355_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_2356_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107466688))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109563904))))[name = string("op_2356_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2356_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_2355_dilations_0, groups = var_2355_groups_0, pad = var_2355_pad_0, pad_type = var_2355_pad_type_0, strides = var_2355_strides_0, weight = op_2356_weight_0_to_fp16_quantized, x = input_231_cast_fp16)[name = string("op_2356_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_181_cast_fp16 = add(x = x_179_cast_fp16, y = var_2356_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<int32, [1]> var_2372_axes_0 = const()[name = string("op_2372_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_8_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_8_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2372_cast_fp16 = layer_norm(axes = var_2372_axes_0, epsilon = layers_8_norm_self_att_eps_scaled_to_fp16, x = x_181_cast_fp16)[name = string("op_2372_cast_fp16")];
            tensor<fp16, [1024]> x_183_gamma_0_to_fp16 = const()[name = string("x_183_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109566016)))];
            tensor<fp16, [1024]> x_183_beta_0_to_fp16 = const()[name = string("x_183_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109568128)))];
            fp16 x_183_epsilon_0_to_fp16 = const()[name = string("x_183_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_183_cast_fp16 = batch_norm(beta = x_183_beta_0_to_fp16, epsilon = x_183_epsilon_0_to_fp16, gamma = x_183_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2372_cast_fp16)[name = string("x_183_cast_fp16")];
            string q_17_pad_type_0 = const()[name = string("q_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_17_strides_0 = const()[name = string("q_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_17_pad_0 = const()[name = string("q_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_17_dilations_0 = const()[name = string("q_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_17_groups_0 = const()[name = string("q_17_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109570240))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110094592))))[name = string("layers_8_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_17_cast_fp16 = conv(dilations = q_17_dilations_0, groups = q_17_groups_0, pad = q_17_pad_0, pad_type = q_17_pad_type_0, strides = q_17_strides_0, weight = layers_8_self_attn_linear_q_weight_to_fp16_quantized, x = x_183_cast_fp16)[name = string("q_17_cast_fp16")];
            string k_17_pad_type_0 = const()[name = string("k_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_17_strides_0 = const()[name = string("k_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_17_pad_0 = const()[name = string("k_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_17_dilations_0 = const()[name = string("k_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_17_groups_0 = const()[name = string("k_17_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110096704))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110621056))))[name = string("layers_8_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_17_cast_fp16 = conv(dilations = k_17_dilations_0, groups = k_17_groups_0, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = k_17_strides_0, weight = layers_8_self_attn_linear_k_weight_to_fp16_quantized, x = x_183_cast_fp16)[name = string("k_17_cast_fp16")];
            string v_17_pad_type_0 = const()[name = string("v_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_17_strides_0 = const()[name = string("v_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_17_pad_0 = const()[name = string("v_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_17_dilations_0 = const()[name = string("v_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_17_groups_0 = const()[name = string("v_17_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110623168))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111147520))))[name = string("layers_8_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_17_cast_fp16 = conv(dilations = v_17_dilations_0, groups = v_17_groups_0, pad = v_17_pad_0, pad_type = v_17_pad_type_0, strides = v_17_strides_0, weight = layers_8_self_attn_linear_v_weight_to_fp16_quantized, x = x_183_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_17_to_fp16 = const()[name = string("bv_all_17_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111149632)))];
            tensor<fp16, [1, 1024, 1, 188]> var_2405_cast_fp16 = add(x = q_17_cast_fp16, y = bv_all_17_to_fp16)[name = string("op_2405_cast_fp16")];
            tensor<int32, [3]> var_2406 = const()[name = string("op_2406"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_17_cast_fp16 = reshape(shape = var_2406, x = var_2405_cast_fp16)[name = string("qb_17_cast_fp16")];
            bool bd_all_33_transpose_x_0 = const()[name = string("bd_all_33_transpose_x_0"), val = bool(false)];
            bool bd_all_33_transpose_y_0 = const()[name = string("bd_all_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_8_self_attn_pos_proj_to_fp16 = const()[name = string("layers_8_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111151744)))];
            tensor<fp16, [8, 375, 188]> bd_all_33_cast_fp16 = matmul(transpose_x = bd_all_33_transpose_x_0, transpose_y = bd_all_33_transpose_y_0, x = layers_8_self_attn_pos_proj_to_fp16, y = qb_17_cast_fp16)[name = string("bd_all_33_cast_fp16")];
            tensor<int32, [3]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_187_pad_0 = const()[name = string("x_187_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_187_mode_0 = const()[name = string("x_187_mode_0"), val = string("constant")];
            fp16 const_42_to_fp16 = const()[name = string("const_42_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = bd_all_33_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [8, 188, 376]> x_187_cast_fp16 = pad(constant_val = const_42_to_fp16, mode = x_187_mode_0, pad = x_187_pad_0, x = x_185_cast_fp16)[name = string("x_187_cast_fp16")];
            tensor<int32, [3]> var_2413 = const()[name = string("op_2413"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_189_cast_fp16 = reshape(shape = var_2413, x = x_187_cast_fp16)[name = string("x_189_cast_fp16")];
            tensor<int32, [3]> var_2416_begin_0 = const()[name = string("op_2416_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2416_end_0 = const()[name = string("op_2416_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_2416_end_mask_0 = const()[name = string("op_2416_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_2416_cast_fp16 = slice_by_index(begin = var_2416_begin_0, end = var_2416_end_0, end_mask = var_2416_end_mask_0, x = x_189_cast_fp16)[name = string("op_2416_cast_fp16")];
            tensor<int32, [3]> var_2417 = const()[name = string("op_2417"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_191_cast_fp16 = reshape(shape = var_2417, x = var_2416_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<int32, [3]> bd_all_35_begin_0 = const()[name = string("bd_all_35_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_35_end_0 = const()[name = string("bd_all_35_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_35_end_mask_0 = const()[name = string("bd_all_35_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_35_cast_fp16 = slice_by_index(begin = bd_all_35_begin_0, end = bd_all_35_end_0, end_mask = bd_all_35_end_mask_0, x = x_191_cast_fp16)[name = string("bd_all_35_cast_fp16")];
            tensor<int32, [4]> var_2422 = const()[name = string("op_2422"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2423_cast_fp16 = reshape(shape = var_2422, x = q_17_cast_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_2425_to_fp16 = const()[name = string("op_2425_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111919808)))];
            tensor<fp16, [8, 128, 1, 188]> qh_17_cast_fp16 = add(x = var_2423_cast_fp16, y = var_2425_to_fp16)[name = string("qh_17_cast_fp16")];
            tensor<int32, [4]> var_2427 = const()[name = string("op_2427"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2428_cast_fp16 = reshape(shape = var_2427, x = k_17_cast_fp16)[name = string("op_2428_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_2431 = const()[name = string("op_2431"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2432_cast_fp16 = reshape(shape = var_2431, x = v_17_cast_fp16)[name = string("op_2432_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2435_perm_0 = const()[name = string("op_2435_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_17_axes_0 = const()[name = string("bd_17_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_2435_cast_fp16 = transpose(perm = var_2435_perm_0, x = bd_all_35_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [8, 188, 1, 188]> bd_17_cast_fp16 = expand_dims(axes = bd_17_axes_0, x = var_2435_cast_fp16)[name = string("bd_17_cast_fp16")];
            string ac_17_equation_0 = const()[name = string("ac_17_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_2430_cast_fp16 = transpose(perm = var_2429, x = var_2428_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [8, 188, 1, 188]> ac_17_cast_fp16 = einsum(equation = ac_17_equation_0, values = (var_2430_cast_fp16, qh_17_cast_fp16))[name = string("ac_17_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_2451_cast_fp16 = add(x = ac_17_cast_fp16, y = bd_17_cast_fp16)[name = string("op_2451_cast_fp16")];
            fp16 var_2452_to_fp16 = const()[name = string("op_2452_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_33_cast_fp16 = mul(x = var_2451_cast_fp16, y = var_2452_to_fp16)[name = string("scores_33_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = key_bias)[name = string("scores_35_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_2455_cast_fp16 = softmax(axis = var_2311, x = scores_35_cast_fp16)[name = string("op_2455_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = var_2455_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [8, 188, 188]> reshape_24_cast_fp16 = reshape(shape = concat_84, x = transpose_16_cast_fp16)[name = string("reshape_24_cast_fp16")];
            tensor<int32, [3]> concat_85 = const()[name = string("concat_85"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = var_2432_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [8, 188, 128]> reshape_25_cast_fp16 = reshape(shape = concat_85, x = transpose_56_cast_fp16)[name = string("reshape_25_cast_fp16")];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = reshape_24_cast_fp16, y = reshape_25_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<int32, [4]> concat_89 = const()[name = string("concat_89"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_26_cast_fp16 = reshape(shape = concat_89, x = matmul_8_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [4]> ctx_17_perm_0 = const()[name = string("ctx_17_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2458 = const()[name = string("op_2458"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_17_cast_fp16 = transpose(perm = ctx_17_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 1024, 1, 188]> input_233_cast_fp16 = reshape(shape = var_2458, x = ctx_17_cast_fp16)[name = string("input_233_cast_fp16")];
            string var_2465_pad_type_0 = const()[name = string("op_2465_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2465_strides_0 = const()[name = string("op_2465_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2465_pad_0 = const()[name = string("op_2465_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2465_dilations_0 = const()[name = string("op_2465_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2465_groups_0 = const()[name = string("op_2465_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111921920))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112446272))))[name = string("layers_8_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2465_cast_fp16 = conv(dilations = var_2465_dilations_0, groups = var_2465_groups_0, pad = var_2465_pad_0, pad_type = var_2465_pad_type_0, strides = var_2465_strides_0, weight = layers_8_self_attn_linear_out_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("op_2465_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_193_cast_fp16 = add(x = x_181_cast_fp16, y = var_2465_cast_fp16)[name = string("x_193_cast_fp16")];
            tensor<int32, [1]> var_2481_axes_0 = const()[name = string("op_2481_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_8_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_8_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2481_cast_fp16 = layer_norm(axes = var_2481_axes_0, epsilon = layers_8_norm_conv_eps_scaled_to_fp16, x = x_193_cast_fp16)[name = string("op_2481_cast_fp16")];
            tensor<fp16, [1024]> input_235_gamma_0_to_fp16 = const()[name = string("input_235_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112448384)))];
            tensor<fp16, [1024]> input_235_beta_0_to_fp16 = const()[name = string("input_235_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112450496)))];
            fp16 input_235_epsilon_0_to_fp16 = const()[name = string("input_235_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2481_cast_fp16)[name = string("input_235_cast_fp16")];
            string input_237_pad_type_0 = const()[name = string("input_237_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_237_strides_0 = const()[name = string("input_237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_237_pad_0 = const()[name = string("input_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_237_dilations_0 = const()[name = string("input_237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_237_groups_0 = const()[name = string("input_237_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112452608))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113501248))))[name = string("layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_237_cast_fp16 = conv(dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("input_237_cast_fp16")];
            int32 x_195_split_num_splits_0 = const()[name = string("x_195_split_num_splits_0"), val = int32(2)];
            int32 x_195_split_axis_0 = const()[name = string("x_195_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_195_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_195_split_cast_fp16_1 = split(axis = x_195_split_axis_0, num_splits = x_195_split_num_splits_0, x = input_237_cast_fp16)[name = string("x_195_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_195_split_1_sigmoid_cast_fp16 = sigmoid(x = x_195_split_cast_fp16_1)[name = string("x_195_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_195_cast_fp16 = mul(x = x_195_split_cast_fp16_0, y = x_195_split_1_sigmoid_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_239_cast_fp16 = mul(x = x_195_cast_fp16, y = pad_mask)[name = string("input_239_cast_fp16")];
            string input_241_pad_type_0 = const()[name = string("input_241_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_241_pad_0 = const()[name = string("input_241_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_241_groups_0 = const()[name = string("input_241_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_241_strides_0 = const()[name = string("input_241_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_241_dilations_0 = const()[name = string("input_241_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_119_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113505408))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113510080))))[name = string("const_119_to_fp16_quantized")];
            tensor<fp16, [1024]> const_120_to_fp16 = const()[name = string("const_120_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113512192)))];
            tensor<fp16, [1, 1024, 1, 188]> input_243_cast_fp16 = conv(bias = const_120_to_fp16, dilations = input_241_dilations_0, groups = input_241_groups_0, pad = input_241_pad_0, pad_type = input_241_pad_type_0, strides = input_241_strides_0, weight = const_119_to_fp16_quantized, x = input_239_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_245_cast_fp16 = silu(x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
            string var_2513_pad_type_0 = const()[name = string("op_2513_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2513_strides_0 = const()[name = string("op_2513_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2513_pad_0 = const()[name = string("op_2513_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2513_dilations_0 = const()[name = string("op_2513_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2513_groups_0 = const()[name = string("op_2513_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113514304))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114038656))))[name = string("layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2513_cast_fp16 = conv(dilations = var_2513_dilations_0, groups = var_2513_groups_0, pad = var_2513_pad_0, pad_type = var_2513_pad_type_0, strides = var_2513_strides_0, weight = layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_197_cast_fp16 = add(x = x_193_cast_fp16, y = var_2513_cast_fp16)[name = string("x_197_cast_fp16")];
            tensor<int32, [1]> var_2529_axes_0 = const()[name = string("op_2529_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_8_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_8_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2529_cast_fp16 = layer_norm(axes = var_2529_axes_0, epsilon = layers_8_norm_feed_forward2_eps_scaled_to_fp16, x = x_197_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1024]> input_247_gamma_0_to_fp16 = const()[name = string("input_247_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114040768)))];
            tensor<fp16, [1024]> input_247_beta_0_to_fp16 = const()[name = string("input_247_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114042880)))];
            fp16 input_247_epsilon_0_to_fp16 = const()[name = string("input_247_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_247_cast_fp16 = batch_norm(beta = input_247_beta_0_to_fp16, epsilon = input_247_epsilon_0_to_fp16, gamma = input_247_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2529_cast_fp16)[name = string("input_247_cast_fp16")];
            string input_249_pad_type_0 = const()[name = string("input_249_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_249_strides_0 = const()[name = string("input_249_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_249_pad_0 = const()[name = string("input_249_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_249_dilations_0 = const()[name = string("input_249_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_249_groups_0 = const()[name = string("input_249_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_8_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114044992))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116142208))))[name = string("layers_8_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_249_cast_fp16 = conv(dilations = input_249_dilations_0, groups = input_249_groups_0, pad = input_249_pad_0, pad_type = input_249_pad_type_0, strides = input_249_strides_0, weight = layers_8_feed_forward2_linear1_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_251_cast_fp16 = silu(x = input_249_cast_fp16)[name = string("input_251_cast_fp16")];
            string var_2546_pad_type_0 = const()[name = string("op_2546_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2546_strides_0 = const()[name = string("op_2546_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2546_pad_0 = const()[name = string("op_2546_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2546_dilations_0 = const()[name = string("op_2546_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2546_groups_0 = const()[name = string("op_2546_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_2547_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116150464))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118247680))))[name = string("op_2547_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2547_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_2546_dilations_0, groups = var_2546_groups_0, pad = var_2546_pad_0, pad_type = var_2546_pad_type_0, strides = var_2546_strides_0, weight = op_2547_weight_0_to_fp16_quantized, x = input_251_cast_fp16)[name = string("op_2547_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_199_cast_fp16 = add(x = x_197_cast_fp16, y = var_2547_cast_fp16)[name = string("x_199_cast_fp16")];
            tensor<int32, [1]> var_2563_axes_0 = const()[name = string("op_2563_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_8_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_8_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2563_cast_fp16 = layer_norm(axes = var_2563_axes_0, epsilon = layers_8_norm_out_eps_scaled_to_fp16, x = x_199_cast_fp16)[name = string("op_2563_cast_fp16")];
            tensor<fp16, [1024]> x_201_gamma_0_to_fp16 = const()[name = string("x_201_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118249792)))];
            tensor<fp16, [1024]> x_201_beta_0_to_fp16 = const()[name = string("x_201_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118251904)))];
            fp16 x_201_epsilon_0_to_fp16 = const()[name = string("x_201_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_201_cast_fp16 = batch_norm(beta = x_201_beta_0_to_fp16, epsilon = x_201_epsilon_0_to_fp16, gamma = x_201_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2563_cast_fp16)[name = string("x_201_cast_fp16")];
            int32 var_2582 = const()[name = string("op_2582"), val = int32(1)];
            tensor<int32, [1]> var_2609_axes_0 = const()[name = string("op_2609_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_9_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_9_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2609_cast_fp16 = layer_norm(axes = var_2609_axes_0, epsilon = layers_9_norm_feed_forward1_eps_scaled_to_fp16, x = x_201_cast_fp16)[name = string("op_2609_cast_fp16")];
            tensor<fp16, [1024]> input_253_gamma_0_to_fp16 = const()[name = string("input_253_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118254016)))];
            tensor<fp16, [1024]> input_253_beta_0_to_fp16 = const()[name = string("input_253_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118256128)))];
            fp16 input_253_epsilon_0_to_fp16 = const()[name = string("input_253_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_253_cast_fp16 = batch_norm(beta = input_253_beta_0_to_fp16, epsilon = input_253_epsilon_0_to_fp16, gamma = input_253_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2609_cast_fp16)[name = string("input_253_cast_fp16")];
            string input_255_pad_type_0 = const()[name = string("input_255_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_255_strides_0 = const()[name = string("input_255_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_255_pad_0 = const()[name = string("input_255_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_255_dilations_0 = const()[name = string("input_255_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_255_groups_0 = const()[name = string("input_255_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_9_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118258240))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120355456))))[name = string("layers_9_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_255_cast_fp16 = conv(dilations = input_255_dilations_0, groups = input_255_groups_0, pad = input_255_pad_0, pad_type = input_255_pad_type_0, strides = input_255_strides_0, weight = layers_9_feed_forward1_linear1_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_257_cast_fp16 = silu(x = input_255_cast_fp16)[name = string("input_257_cast_fp16")];
            string var_2626_pad_type_0 = const()[name = string("op_2626_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2626_strides_0 = const()[name = string("op_2626_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2626_pad_0 = const()[name = string("op_2626_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2626_dilations_0 = const()[name = string("op_2626_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2626_groups_0 = const()[name = string("op_2626_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_2627_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120363712))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122460928))))[name = string("op_2627_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2627_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_2626_dilations_0, groups = var_2626_groups_0, pad = var_2626_pad_0, pad_type = var_2626_pad_type_0, strides = var_2626_strides_0, weight = op_2627_weight_0_to_fp16_quantized, x = input_257_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_203_cast_fp16 = add(x = x_201_cast_fp16, y = var_2627_cast_fp16)[name = string("x_203_cast_fp16")];
            tensor<int32, [1]> var_2643_axes_0 = const()[name = string("op_2643_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_9_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_9_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2643_cast_fp16 = layer_norm(axes = var_2643_axes_0, epsilon = layers_9_norm_self_att_eps_scaled_to_fp16, x = x_203_cast_fp16)[name = string("op_2643_cast_fp16")];
            tensor<fp16, [1024]> x_205_gamma_0_to_fp16 = const()[name = string("x_205_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122463040)))];
            tensor<fp16, [1024]> x_205_beta_0_to_fp16 = const()[name = string("x_205_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122465152)))];
            fp16 x_205_epsilon_0_to_fp16 = const()[name = string("x_205_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_205_cast_fp16 = batch_norm(beta = x_205_beta_0_to_fp16, epsilon = x_205_epsilon_0_to_fp16, gamma = x_205_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2643_cast_fp16)[name = string("x_205_cast_fp16")];
            string q_19_pad_type_0 = const()[name = string("q_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_19_strides_0 = const()[name = string("q_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_19_pad_0 = const()[name = string("q_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_19_dilations_0 = const()[name = string("q_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_19_groups_0 = const()[name = string("q_19_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122467264))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122991616))))[name = string("layers_9_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_19_cast_fp16 = conv(dilations = q_19_dilations_0, groups = q_19_groups_0, pad = q_19_pad_0, pad_type = q_19_pad_type_0, strides = q_19_strides_0, weight = layers_9_self_attn_linear_q_weight_to_fp16_quantized, x = x_205_cast_fp16)[name = string("q_19_cast_fp16")];
            string k_19_pad_type_0 = const()[name = string("k_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_19_strides_0 = const()[name = string("k_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_19_pad_0 = const()[name = string("k_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_19_dilations_0 = const()[name = string("k_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_19_groups_0 = const()[name = string("k_19_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122993728))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123518080))))[name = string("layers_9_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_19_cast_fp16 = conv(dilations = k_19_dilations_0, groups = k_19_groups_0, pad = k_19_pad_0, pad_type = k_19_pad_type_0, strides = k_19_strides_0, weight = layers_9_self_attn_linear_k_weight_to_fp16_quantized, x = x_205_cast_fp16)[name = string("k_19_cast_fp16")];
            string v_19_pad_type_0 = const()[name = string("v_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_19_strides_0 = const()[name = string("v_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_19_pad_0 = const()[name = string("v_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_19_dilations_0 = const()[name = string("v_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_19_groups_0 = const()[name = string("v_19_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123520192))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124044544))))[name = string("layers_9_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_19_cast_fp16 = conv(dilations = v_19_dilations_0, groups = v_19_groups_0, pad = v_19_pad_0, pad_type = v_19_pad_type_0, strides = v_19_strides_0, weight = layers_9_self_attn_linear_v_weight_to_fp16_quantized, x = x_205_cast_fp16)[name = string("v_19_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_19_to_fp16 = const()[name = string("bv_all_19_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124046656)))];
            tensor<fp16, [1, 1024, 1, 188]> var_2676_cast_fp16 = add(x = q_19_cast_fp16, y = bv_all_19_to_fp16)[name = string("op_2676_cast_fp16")];
            tensor<int32, [3]> var_2677 = const()[name = string("op_2677"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_19_cast_fp16 = reshape(shape = var_2677, x = var_2676_cast_fp16)[name = string("qb_19_cast_fp16")];
            bool bd_all_37_transpose_x_0 = const()[name = string("bd_all_37_transpose_x_0"), val = bool(false)];
            bool bd_all_37_transpose_y_0 = const()[name = string("bd_all_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_9_self_attn_pos_proj_to_fp16 = const()[name = string("layers_9_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124048768)))];
            tensor<fp16, [8, 375, 188]> bd_all_37_cast_fp16 = matmul(transpose_x = bd_all_37_transpose_x_0, transpose_y = bd_all_37_transpose_y_0, x = layers_9_self_attn_pos_proj_to_fp16, y = qb_19_cast_fp16)[name = string("bd_all_37_cast_fp16")];
            tensor<int32, [3]> x_207_perm_0 = const()[name = string("x_207_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_209_pad_0 = const()[name = string("x_209_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_209_mode_0 = const()[name = string("x_209_mode_0"), val = string("constant")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_207_cast_fp16 = transpose(perm = x_207_perm_0, x = bd_all_37_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [8, 188, 376]> x_209_cast_fp16 = pad(constant_val = const_46_to_fp16, mode = x_209_mode_0, pad = x_209_pad_0, x = x_207_cast_fp16)[name = string("x_209_cast_fp16")];
            tensor<int32, [3]> var_2684 = const()[name = string("op_2684"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_211_cast_fp16 = reshape(shape = var_2684, x = x_209_cast_fp16)[name = string("x_211_cast_fp16")];
            tensor<int32, [3]> var_2687_begin_0 = const()[name = string("op_2687_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2687_end_0 = const()[name = string("op_2687_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_2687_end_mask_0 = const()[name = string("op_2687_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_2687_cast_fp16 = slice_by_index(begin = var_2687_begin_0, end = var_2687_end_0, end_mask = var_2687_end_mask_0, x = x_211_cast_fp16)[name = string("op_2687_cast_fp16")];
            tensor<int32, [3]> var_2688 = const()[name = string("op_2688"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_213_cast_fp16 = reshape(shape = var_2688, x = var_2687_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [3]> bd_all_39_begin_0 = const()[name = string("bd_all_39_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_39_end_0 = const()[name = string("bd_all_39_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_39_end_mask_0 = const()[name = string("bd_all_39_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_39_cast_fp16 = slice_by_index(begin = bd_all_39_begin_0, end = bd_all_39_end_0, end_mask = bd_all_39_end_mask_0, x = x_213_cast_fp16)[name = string("bd_all_39_cast_fp16")];
            tensor<int32, [4]> var_2693 = const()[name = string("op_2693"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2694_cast_fp16 = reshape(shape = var_2693, x = q_19_cast_fp16)[name = string("op_2694_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_2696_to_fp16 = const()[name = string("op_2696_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124816832)))];
            tensor<fp16, [8, 128, 1, 188]> qh_19_cast_fp16 = add(x = var_2694_cast_fp16, y = var_2696_to_fp16)[name = string("qh_19_cast_fp16")];
            tensor<int32, [4]> var_2698 = const()[name = string("op_2698"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2699_cast_fp16 = reshape(shape = var_2698, x = k_19_cast_fp16)[name = string("op_2699_cast_fp16")];
            tensor<int32, [4]> var_2700 = const()[name = string("op_2700"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_2702 = const()[name = string("op_2702"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2703_cast_fp16 = reshape(shape = var_2702, x = v_19_cast_fp16)[name = string("op_2703_cast_fp16")];
            tensor<int32, [4]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2706_perm_0 = const()[name = string("op_2706_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_19_axes_0 = const()[name = string("bd_19_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_2706_cast_fp16 = transpose(perm = var_2706_perm_0, x = bd_all_39_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [8, 188, 1, 188]> bd_19_cast_fp16 = expand_dims(axes = bd_19_axes_0, x = var_2706_cast_fp16)[name = string("bd_19_cast_fp16")];
            string ac_19_equation_0 = const()[name = string("ac_19_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_2701_cast_fp16 = transpose(perm = var_2700, x = var_2699_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [8, 188, 1, 188]> ac_19_cast_fp16 = einsum(equation = ac_19_equation_0, values = (var_2701_cast_fp16, qh_19_cast_fp16))[name = string("ac_19_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_2722_cast_fp16 = add(x = ac_19_cast_fp16, y = bd_19_cast_fp16)[name = string("op_2722_cast_fp16")];
            fp16 var_2723_to_fp16 = const()[name = string("op_2723_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_37_cast_fp16 = mul(x = var_2722_cast_fp16, y = var_2723_to_fp16)[name = string("scores_37_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = key_bias)[name = string("scores_39_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_2726_cast_fp16 = softmax(axis = var_2582, x = scores_39_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = var_2726_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [8, 188, 188]> reshape_27_cast_fp16 = reshape(shape = concat_94, x = transpose_18_cast_fp16)[name = string("reshape_27_cast_fp16")];
            tensor<int32, [3]> concat_95 = const()[name = string("concat_95"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = var_2703_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [8, 188, 128]> reshape_28_cast_fp16 = reshape(shape = concat_95, x = transpose_57_cast_fp16)[name = string("reshape_28_cast_fp16")];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = reshape_27_cast_fp16, y = reshape_28_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<int32, [4]> concat_99 = const()[name = string("concat_99"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_29_cast_fp16 = reshape(shape = concat_99, x = matmul_9_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> ctx_19_perm_0 = const()[name = string("ctx_19_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2729 = const()[name = string("op_2729"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_19_cast_fp16 = transpose(perm = ctx_19_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 1024, 1, 188]> input_259_cast_fp16 = reshape(shape = var_2729, x = ctx_19_cast_fp16)[name = string("input_259_cast_fp16")];
            string var_2736_pad_type_0 = const()[name = string("op_2736_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2736_strides_0 = const()[name = string("op_2736_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2736_pad_0 = const()[name = string("op_2736_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2736_dilations_0 = const()[name = string("op_2736_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2736_groups_0 = const()[name = string("op_2736_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124818944))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125343296))))[name = string("layers_9_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2736_cast_fp16 = conv(dilations = var_2736_dilations_0, groups = var_2736_groups_0, pad = var_2736_pad_0, pad_type = var_2736_pad_type_0, strides = var_2736_strides_0, weight = layers_9_self_attn_linear_out_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("op_2736_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_215_cast_fp16 = add(x = x_203_cast_fp16, y = var_2736_cast_fp16)[name = string("x_215_cast_fp16")];
            tensor<int32, [1]> var_2752_axes_0 = const()[name = string("op_2752_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_9_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_9_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2752_cast_fp16 = layer_norm(axes = var_2752_axes_0, epsilon = layers_9_norm_conv_eps_scaled_to_fp16, x = x_215_cast_fp16)[name = string("op_2752_cast_fp16")];
            tensor<fp16, [1024]> input_261_gamma_0_to_fp16 = const()[name = string("input_261_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125345408)))];
            tensor<fp16, [1024]> input_261_beta_0_to_fp16 = const()[name = string("input_261_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125347520)))];
            fp16 input_261_epsilon_0_to_fp16 = const()[name = string("input_261_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_261_cast_fp16 = batch_norm(beta = input_261_beta_0_to_fp16, epsilon = input_261_epsilon_0_to_fp16, gamma = input_261_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2752_cast_fp16)[name = string("input_261_cast_fp16")];
            string input_263_pad_type_0 = const()[name = string("input_263_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_263_strides_0 = const()[name = string("input_263_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_263_pad_0 = const()[name = string("input_263_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_263_dilations_0 = const()[name = string("input_263_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_263_groups_0 = const()[name = string("input_263_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125349632))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126398272))))[name = string("layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_263_cast_fp16 = conv(dilations = input_263_dilations_0, groups = input_263_groups_0, pad = input_263_pad_0, pad_type = input_263_pad_type_0, strides = input_263_strides_0, weight = layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("input_263_cast_fp16")];
            int32 x_217_split_num_splits_0 = const()[name = string("x_217_split_num_splits_0"), val = int32(2)];
            int32 x_217_split_axis_0 = const()[name = string("x_217_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_217_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_217_split_cast_fp16_1 = split(axis = x_217_split_axis_0, num_splits = x_217_split_num_splits_0, x = input_263_cast_fp16)[name = string("x_217_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_217_split_1_sigmoid_cast_fp16 = sigmoid(x = x_217_split_cast_fp16_1)[name = string("x_217_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_217_cast_fp16 = mul(x = x_217_split_cast_fp16_0, y = x_217_split_1_sigmoid_cast_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_265_cast_fp16 = mul(x = x_217_cast_fp16, y = pad_mask)[name = string("input_265_cast_fp16")];
            string input_267_pad_type_0 = const()[name = string("input_267_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_267_pad_0 = const()[name = string("input_267_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_267_groups_0 = const()[name = string("input_267_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_267_strides_0 = const()[name = string("input_267_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_267_dilations_0 = const()[name = string("input_267_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_121_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126402432))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126407104))))[name = string("const_121_to_fp16_quantized")];
            tensor<fp16, [1024]> const_122_to_fp16 = const()[name = string("const_122_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126409216)))];
            tensor<fp16, [1, 1024, 1, 188]> input_269_cast_fp16 = conv(bias = const_122_to_fp16, dilations = input_267_dilations_0, groups = input_267_groups_0, pad = input_267_pad_0, pad_type = input_267_pad_type_0, strides = input_267_strides_0, weight = const_121_to_fp16_quantized, x = input_265_cast_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_271_cast_fp16 = silu(x = input_269_cast_fp16)[name = string("input_271_cast_fp16")];
            string var_2784_pad_type_0 = const()[name = string("op_2784_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2784_strides_0 = const()[name = string("op_2784_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2784_pad_0 = const()[name = string("op_2784_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2784_dilations_0 = const()[name = string("op_2784_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2784_groups_0 = const()[name = string("op_2784_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126411328))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126935680))))[name = string("layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2784_cast_fp16 = conv(dilations = var_2784_dilations_0, groups = var_2784_groups_0, pad = var_2784_pad_0, pad_type = var_2784_pad_type_0, strides = var_2784_strides_0, weight = layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("op_2784_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_219_cast_fp16 = add(x = x_215_cast_fp16, y = var_2784_cast_fp16)[name = string("x_219_cast_fp16")];
            tensor<int32, [1]> var_2800_axes_0 = const()[name = string("op_2800_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_9_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_9_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2800_cast_fp16 = layer_norm(axes = var_2800_axes_0, epsilon = layers_9_norm_feed_forward2_eps_scaled_to_fp16, x = x_219_cast_fp16)[name = string("op_2800_cast_fp16")];
            tensor<fp16, [1024]> input_273_gamma_0_to_fp16 = const()[name = string("input_273_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126937792)))];
            tensor<fp16, [1024]> input_273_beta_0_to_fp16 = const()[name = string("input_273_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126939904)))];
            fp16 input_273_epsilon_0_to_fp16 = const()[name = string("input_273_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_273_cast_fp16 = batch_norm(beta = input_273_beta_0_to_fp16, epsilon = input_273_epsilon_0_to_fp16, gamma = input_273_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2800_cast_fp16)[name = string("input_273_cast_fp16")];
            string input_275_pad_type_0 = const()[name = string("input_275_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_275_strides_0 = const()[name = string("input_275_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_275_pad_0 = const()[name = string("input_275_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_275_dilations_0 = const()[name = string("input_275_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_275_groups_0 = const()[name = string("input_275_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_9_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126942016))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129039232))))[name = string("layers_9_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_275_cast_fp16 = conv(dilations = input_275_dilations_0, groups = input_275_groups_0, pad = input_275_pad_0, pad_type = input_275_pad_type_0, strides = input_275_strides_0, weight = layers_9_feed_forward2_linear1_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_277_cast_fp16 = silu(x = input_275_cast_fp16)[name = string("input_277_cast_fp16")];
            string var_2817_pad_type_0 = const()[name = string("op_2817_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2817_strides_0 = const()[name = string("op_2817_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2817_pad_0 = const()[name = string("op_2817_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2817_dilations_0 = const()[name = string("op_2817_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2817_groups_0 = const()[name = string("op_2817_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_2818_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129047488))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131144704))))[name = string("op_2818_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2818_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_2817_dilations_0, groups = var_2817_groups_0, pad = var_2817_pad_0, pad_type = var_2817_pad_type_0, strides = var_2817_strides_0, weight = op_2818_weight_0_to_fp16_quantized, x = input_277_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_221_cast_fp16 = add(x = x_219_cast_fp16, y = var_2818_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<int32, [1]> var_2834_axes_0 = const()[name = string("op_2834_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_9_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_9_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2834_cast_fp16 = layer_norm(axes = var_2834_axes_0, epsilon = layers_9_norm_out_eps_scaled_to_fp16, x = x_221_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [1024]> x_223_gamma_0_to_fp16 = const()[name = string("x_223_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131146816)))];
            tensor<fp16, [1024]> x_223_beta_0_to_fp16 = const()[name = string("x_223_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131148928)))];
            fp16 x_223_epsilon_0_to_fp16 = const()[name = string("x_223_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_223_cast_fp16 = batch_norm(beta = x_223_beta_0_to_fp16, epsilon = x_223_epsilon_0_to_fp16, gamma = x_223_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2834_cast_fp16)[name = string("x_223_cast_fp16")];
            int32 var_2853 = const()[name = string("op_2853"), val = int32(1)];
            tensor<int32, [1]> var_2880_axes_0 = const()[name = string("op_2880_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_10_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_10_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2880_cast_fp16 = layer_norm(axes = var_2880_axes_0, epsilon = layers_10_norm_feed_forward1_eps_scaled_to_fp16, x = x_223_cast_fp16)[name = string("op_2880_cast_fp16")];
            tensor<fp16, [1024]> input_279_gamma_0_to_fp16 = const()[name = string("input_279_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131151040)))];
            tensor<fp16, [1024]> input_279_beta_0_to_fp16 = const()[name = string("input_279_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131153152)))];
            fp16 input_279_epsilon_0_to_fp16 = const()[name = string("input_279_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_279_cast_fp16 = batch_norm(beta = input_279_beta_0_to_fp16, epsilon = input_279_epsilon_0_to_fp16, gamma = input_279_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2880_cast_fp16)[name = string("input_279_cast_fp16")];
            string input_281_pad_type_0 = const()[name = string("input_281_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_281_strides_0 = const()[name = string("input_281_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_281_pad_0 = const()[name = string("input_281_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_281_dilations_0 = const()[name = string("input_281_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_281_groups_0 = const()[name = string("input_281_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_10_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131155264))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133252480))))[name = string("layers_10_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_281_cast_fp16 = conv(dilations = input_281_dilations_0, groups = input_281_groups_0, pad = input_281_pad_0, pad_type = input_281_pad_type_0, strides = input_281_strides_0, weight = layers_10_feed_forward1_linear1_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_283_cast_fp16 = silu(x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
            string var_2897_pad_type_0 = const()[name = string("op_2897_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2897_strides_0 = const()[name = string("op_2897_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2897_pad_0 = const()[name = string("op_2897_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2897_dilations_0 = const()[name = string("op_2897_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2897_groups_0 = const()[name = string("op_2897_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_2898_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133260736))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135357952))))[name = string("op_2898_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_2898_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_2897_dilations_0, groups = var_2897_groups_0, pad = var_2897_pad_0, pad_type = var_2897_pad_type_0, strides = var_2897_strides_0, weight = op_2898_weight_0_to_fp16_quantized, x = input_283_cast_fp16)[name = string("op_2898_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_225_cast_fp16 = add(x = x_223_cast_fp16, y = var_2898_cast_fp16)[name = string("x_225_cast_fp16")];
            tensor<int32, [1]> var_2914_axes_0 = const()[name = string("op_2914_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_10_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_10_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_2914_cast_fp16 = layer_norm(axes = var_2914_axes_0, epsilon = layers_10_norm_self_att_eps_scaled_to_fp16, x = x_225_cast_fp16)[name = string("op_2914_cast_fp16")];
            tensor<fp16, [1024]> x_227_gamma_0_to_fp16 = const()[name = string("x_227_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135360064)))];
            tensor<fp16, [1024]> x_227_beta_0_to_fp16 = const()[name = string("x_227_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135362176)))];
            fp16 x_227_epsilon_0_to_fp16 = const()[name = string("x_227_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_227_cast_fp16 = batch_norm(beta = x_227_beta_0_to_fp16, epsilon = x_227_epsilon_0_to_fp16, gamma = x_227_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_2914_cast_fp16)[name = string("x_227_cast_fp16")];
            string q_21_pad_type_0 = const()[name = string("q_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_21_strides_0 = const()[name = string("q_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_21_pad_0 = const()[name = string("q_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_21_dilations_0 = const()[name = string("q_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_21_groups_0 = const()[name = string("q_21_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135364288))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135888640))))[name = string("layers_10_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_21_cast_fp16 = conv(dilations = q_21_dilations_0, groups = q_21_groups_0, pad = q_21_pad_0, pad_type = q_21_pad_type_0, strides = q_21_strides_0, weight = layers_10_self_attn_linear_q_weight_to_fp16_quantized, x = x_227_cast_fp16)[name = string("q_21_cast_fp16")];
            string k_21_pad_type_0 = const()[name = string("k_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_21_strides_0 = const()[name = string("k_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_21_pad_0 = const()[name = string("k_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_21_dilations_0 = const()[name = string("k_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_21_groups_0 = const()[name = string("k_21_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135890752))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136415104))))[name = string("layers_10_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_21_cast_fp16 = conv(dilations = k_21_dilations_0, groups = k_21_groups_0, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = k_21_strides_0, weight = layers_10_self_attn_linear_k_weight_to_fp16_quantized, x = x_227_cast_fp16)[name = string("k_21_cast_fp16")];
            string v_21_pad_type_0 = const()[name = string("v_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_21_strides_0 = const()[name = string("v_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_21_pad_0 = const()[name = string("v_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_21_dilations_0 = const()[name = string("v_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_21_groups_0 = const()[name = string("v_21_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136417216))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136941568))))[name = string("layers_10_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_21_cast_fp16 = conv(dilations = v_21_dilations_0, groups = v_21_groups_0, pad = v_21_pad_0, pad_type = v_21_pad_type_0, strides = v_21_strides_0, weight = layers_10_self_attn_linear_v_weight_to_fp16_quantized, x = x_227_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_21_to_fp16 = const()[name = string("bv_all_21_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136943680)))];
            tensor<fp16, [1, 1024, 1, 188]> var_2947_cast_fp16 = add(x = q_21_cast_fp16, y = bv_all_21_to_fp16)[name = string("op_2947_cast_fp16")];
            tensor<int32, [3]> var_2948 = const()[name = string("op_2948"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_21_cast_fp16 = reshape(shape = var_2948, x = var_2947_cast_fp16)[name = string("qb_21_cast_fp16")];
            bool bd_all_41_transpose_x_0 = const()[name = string("bd_all_41_transpose_x_0"), val = bool(false)];
            bool bd_all_41_transpose_y_0 = const()[name = string("bd_all_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_10_self_attn_pos_proj_to_fp16 = const()[name = string("layers_10_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136945792)))];
            tensor<fp16, [8, 375, 188]> bd_all_41_cast_fp16 = matmul(transpose_x = bd_all_41_transpose_x_0, transpose_y = bd_all_41_transpose_y_0, x = layers_10_self_attn_pos_proj_to_fp16, y = qb_21_cast_fp16)[name = string("bd_all_41_cast_fp16")];
            tensor<int32, [3]> x_229_perm_0 = const()[name = string("x_229_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_231_pad_0 = const()[name = string("x_231_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_231_mode_0 = const()[name = string("x_231_mode_0"), val = string("constant")];
            fp16 const_50_to_fp16 = const()[name = string("const_50_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_229_cast_fp16 = transpose(perm = x_229_perm_0, x = bd_all_41_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [8, 188, 376]> x_231_cast_fp16 = pad(constant_val = const_50_to_fp16, mode = x_231_mode_0, pad = x_231_pad_0, x = x_229_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [3]> var_2955 = const()[name = string("op_2955"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_233_cast_fp16 = reshape(shape = var_2955, x = x_231_cast_fp16)[name = string("x_233_cast_fp16")];
            tensor<int32, [3]> var_2958_begin_0 = const()[name = string("op_2958_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_2958_end_0 = const()[name = string("op_2958_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_2958_end_mask_0 = const()[name = string("op_2958_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_2958_cast_fp16 = slice_by_index(begin = var_2958_begin_0, end = var_2958_end_0, end_mask = var_2958_end_mask_0, x = x_233_cast_fp16)[name = string("op_2958_cast_fp16")];
            tensor<int32, [3]> var_2959 = const()[name = string("op_2959"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_235_cast_fp16 = reshape(shape = var_2959, x = var_2958_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<int32, [3]> bd_all_43_begin_0 = const()[name = string("bd_all_43_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_43_end_0 = const()[name = string("bd_all_43_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_43_end_mask_0 = const()[name = string("bd_all_43_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_43_cast_fp16 = slice_by_index(begin = bd_all_43_begin_0, end = bd_all_43_end_0, end_mask = bd_all_43_end_mask_0, x = x_235_cast_fp16)[name = string("bd_all_43_cast_fp16")];
            tensor<int32, [4]> var_2964 = const()[name = string("op_2964"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2965_cast_fp16 = reshape(shape = var_2964, x = q_21_cast_fp16)[name = string("op_2965_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_2967_to_fp16 = const()[name = string("op_2967_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137713856)))];
            tensor<fp16, [8, 128, 1, 188]> qh_21_cast_fp16 = add(x = var_2965_cast_fp16, y = var_2967_to_fp16)[name = string("qh_21_cast_fp16")];
            tensor<int32, [4]> var_2969 = const()[name = string("op_2969"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2970_cast_fp16 = reshape(shape = var_2969, x = k_21_cast_fp16)[name = string("op_2970_cast_fp16")];
            tensor<int32, [4]> var_2971 = const()[name = string("op_2971"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_2973 = const()[name = string("op_2973"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_2974_cast_fp16 = reshape(shape = var_2973, x = v_21_cast_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2977_perm_0 = const()[name = string("op_2977_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_21_axes_0 = const()[name = string("bd_21_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_2977_cast_fp16 = transpose(perm = var_2977_perm_0, x = bd_all_43_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [8, 188, 1, 188]> bd_21_cast_fp16 = expand_dims(axes = bd_21_axes_0, x = var_2977_cast_fp16)[name = string("bd_21_cast_fp16")];
            string ac_21_equation_0 = const()[name = string("ac_21_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_2972_cast_fp16 = transpose(perm = var_2971, x = var_2970_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [8, 188, 1, 188]> ac_21_cast_fp16 = einsum(equation = ac_21_equation_0, values = (var_2972_cast_fp16, qh_21_cast_fp16))[name = string("ac_21_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_2993_cast_fp16 = add(x = ac_21_cast_fp16, y = bd_21_cast_fp16)[name = string("op_2993_cast_fp16")];
            fp16 var_2994_to_fp16 = const()[name = string("op_2994_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_41_cast_fp16 = mul(x = var_2993_cast_fp16, y = var_2994_to_fp16)[name = string("scores_41_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = key_bias)[name = string("scores_43_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_2997_cast_fp16 = softmax(axis = var_2853, x = scores_43_cast_fp16)[name = string("op_2997_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_104 = const()[name = string("concat_104"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = var_2997_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [8, 188, 188]> reshape_30_cast_fp16 = reshape(shape = concat_104, x = transpose_20_cast_fp16)[name = string("reshape_30_cast_fp16")];
            tensor<int32, [3]> concat_105 = const()[name = string("concat_105"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = var_2974_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [8, 188, 128]> reshape_31_cast_fp16 = reshape(shape = concat_105, x = transpose_58_cast_fp16)[name = string("reshape_31_cast_fp16")];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = reshape_30_cast_fp16, y = reshape_31_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<int32, [4]> concat_109 = const()[name = string("concat_109"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_32_cast_fp16 = reshape(shape = concat_109, x = matmul_10_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [4]> ctx_21_perm_0 = const()[name = string("ctx_21_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3000 = const()[name = string("op_3000"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_21_cast_fp16 = transpose(perm = ctx_21_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 1024, 1, 188]> input_285_cast_fp16 = reshape(shape = var_3000, x = ctx_21_cast_fp16)[name = string("input_285_cast_fp16")];
            string var_3007_pad_type_0 = const()[name = string("op_3007_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3007_strides_0 = const()[name = string("op_3007_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3007_pad_0 = const()[name = string("op_3007_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3007_dilations_0 = const()[name = string("op_3007_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3007_groups_0 = const()[name = string("op_3007_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137715968))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138240320))))[name = string("layers_10_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3007_cast_fp16 = conv(dilations = var_3007_dilations_0, groups = var_3007_groups_0, pad = var_3007_pad_0, pad_type = var_3007_pad_type_0, strides = var_3007_strides_0, weight = layers_10_self_attn_linear_out_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = string("op_3007_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_237_cast_fp16 = add(x = x_225_cast_fp16, y = var_3007_cast_fp16)[name = string("x_237_cast_fp16")];
            tensor<int32, [1]> var_3023_axes_0 = const()[name = string("op_3023_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_10_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_10_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3023_cast_fp16 = layer_norm(axes = var_3023_axes_0, epsilon = layers_10_norm_conv_eps_scaled_to_fp16, x = x_237_cast_fp16)[name = string("op_3023_cast_fp16")];
            tensor<fp16, [1024]> input_287_gamma_0_to_fp16 = const()[name = string("input_287_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138242432)))];
            tensor<fp16, [1024]> input_287_beta_0_to_fp16 = const()[name = string("input_287_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138244544)))];
            fp16 input_287_epsilon_0_to_fp16 = const()[name = string("input_287_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_287_cast_fp16 = batch_norm(beta = input_287_beta_0_to_fp16, epsilon = input_287_epsilon_0_to_fp16, gamma = input_287_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3023_cast_fp16)[name = string("input_287_cast_fp16")];
            string input_289_pad_type_0 = const()[name = string("input_289_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_289_strides_0 = const()[name = string("input_289_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_289_pad_0 = const()[name = string("input_289_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_289_dilations_0 = const()[name = string("input_289_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_289_groups_0 = const()[name = string("input_289_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138246656))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139295296))))[name = string("layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_289_cast_fp16 = conv(dilations = input_289_dilations_0, groups = input_289_groups_0, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = input_289_strides_0, weight = layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("input_289_cast_fp16")];
            int32 x_239_split_num_splits_0 = const()[name = string("x_239_split_num_splits_0"), val = int32(2)];
            int32 x_239_split_axis_0 = const()[name = string("x_239_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_239_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_239_split_cast_fp16_1 = split(axis = x_239_split_axis_0, num_splits = x_239_split_num_splits_0, x = input_289_cast_fp16)[name = string("x_239_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_239_split_1_sigmoid_cast_fp16 = sigmoid(x = x_239_split_cast_fp16_1)[name = string("x_239_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_239_cast_fp16 = mul(x = x_239_split_cast_fp16_0, y = x_239_split_1_sigmoid_cast_fp16)[name = string("x_239_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_291_cast_fp16 = mul(x = x_239_cast_fp16, y = pad_mask)[name = string("input_291_cast_fp16")];
            string input_293_pad_type_0 = const()[name = string("input_293_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_293_pad_0 = const()[name = string("input_293_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_293_groups_0 = const()[name = string("input_293_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_293_strides_0 = const()[name = string("input_293_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_293_dilations_0 = const()[name = string("input_293_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_123_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139299456))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139304128))))[name = string("const_123_to_fp16_quantized")];
            tensor<fp16, [1024]> const_124_to_fp16 = const()[name = string("const_124_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139306240)))];
            tensor<fp16, [1, 1024, 1, 188]> input_295_cast_fp16 = conv(bias = const_124_to_fp16, dilations = input_293_dilations_0, groups = input_293_groups_0, pad = input_293_pad_0, pad_type = input_293_pad_type_0, strides = input_293_strides_0, weight = const_123_to_fp16_quantized, x = input_291_cast_fp16)[name = string("input_295_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_297_cast_fp16 = silu(x = input_295_cast_fp16)[name = string("input_297_cast_fp16")];
            string var_3055_pad_type_0 = const()[name = string("op_3055_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3055_strides_0 = const()[name = string("op_3055_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3055_pad_0 = const()[name = string("op_3055_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3055_dilations_0 = const()[name = string("op_3055_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3055_groups_0 = const()[name = string("op_3055_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139308352))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139832704))))[name = string("layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3055_cast_fp16 = conv(dilations = var_3055_dilations_0, groups = var_3055_groups_0, pad = var_3055_pad_0, pad_type = var_3055_pad_type_0, strides = var_3055_strides_0, weight = layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_297_cast_fp16)[name = string("op_3055_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_241_cast_fp16 = add(x = x_237_cast_fp16, y = var_3055_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<int32, [1]> var_3071_axes_0 = const()[name = string("op_3071_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_10_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_10_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3071_cast_fp16 = layer_norm(axes = var_3071_axes_0, epsilon = layers_10_norm_feed_forward2_eps_scaled_to_fp16, x = x_241_cast_fp16)[name = string("op_3071_cast_fp16")];
            tensor<fp16, [1024]> input_299_gamma_0_to_fp16 = const()[name = string("input_299_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139834816)))];
            tensor<fp16, [1024]> input_299_beta_0_to_fp16 = const()[name = string("input_299_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139836928)))];
            fp16 input_299_epsilon_0_to_fp16 = const()[name = string("input_299_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_299_cast_fp16 = batch_norm(beta = input_299_beta_0_to_fp16, epsilon = input_299_epsilon_0_to_fp16, gamma = input_299_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3071_cast_fp16)[name = string("input_299_cast_fp16")];
            string input_301_pad_type_0 = const()[name = string("input_301_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_301_strides_0 = const()[name = string("input_301_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_301_pad_0 = const()[name = string("input_301_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_301_dilations_0 = const()[name = string("input_301_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_301_groups_0 = const()[name = string("input_301_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_10_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139839040))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141936256))))[name = string("layers_10_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_301_cast_fp16 = conv(dilations = input_301_dilations_0, groups = input_301_groups_0, pad = input_301_pad_0, pad_type = input_301_pad_type_0, strides = input_301_strides_0, weight = layers_10_feed_forward2_linear1_weight_to_fp16_quantized, x = input_299_cast_fp16)[name = string("input_301_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_303_cast_fp16 = silu(x = input_301_cast_fp16)[name = string("input_303_cast_fp16")];
            string var_3088_pad_type_0 = const()[name = string("op_3088_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3088_strides_0 = const()[name = string("op_3088_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3088_pad_0 = const()[name = string("op_3088_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3088_dilations_0 = const()[name = string("op_3088_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3088_groups_0 = const()[name = string("op_3088_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_3089_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141944512))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144041728))))[name = string("op_3089_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3089_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_3088_dilations_0, groups = var_3088_groups_0, pad = var_3088_pad_0, pad_type = var_3088_pad_type_0, strides = var_3088_strides_0, weight = op_3089_weight_0_to_fp16_quantized, x = input_303_cast_fp16)[name = string("op_3089_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_243_cast_fp16 = add(x = x_241_cast_fp16, y = var_3089_cast_fp16)[name = string("x_243_cast_fp16")];
            tensor<int32, [1]> var_3105_axes_0 = const()[name = string("op_3105_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_10_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_10_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3105_cast_fp16 = layer_norm(axes = var_3105_axes_0, epsilon = layers_10_norm_out_eps_scaled_to_fp16, x = x_243_cast_fp16)[name = string("op_3105_cast_fp16")];
            tensor<fp16, [1024]> x_245_gamma_0_to_fp16 = const()[name = string("x_245_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144043840)))];
            tensor<fp16, [1024]> x_245_beta_0_to_fp16 = const()[name = string("x_245_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144045952)))];
            fp16 x_245_epsilon_0_to_fp16 = const()[name = string("x_245_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_245_cast_fp16 = batch_norm(beta = x_245_beta_0_to_fp16, epsilon = x_245_epsilon_0_to_fp16, gamma = x_245_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3105_cast_fp16)[name = string("x_245_cast_fp16")];
            int32 var_3124 = const()[name = string("op_3124"), val = int32(1)];
            tensor<int32, [1]> var_3151_axes_0 = const()[name = string("op_3151_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_11_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_11_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3151_cast_fp16 = layer_norm(axes = var_3151_axes_0, epsilon = layers_11_norm_feed_forward1_eps_scaled_to_fp16, x = x_245_cast_fp16)[name = string("op_3151_cast_fp16")];
            tensor<fp16, [1024]> input_305_gamma_0_to_fp16 = const()[name = string("input_305_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144048064)))];
            tensor<fp16, [1024]> input_305_beta_0_to_fp16 = const()[name = string("input_305_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144050176)))];
            fp16 input_305_epsilon_0_to_fp16 = const()[name = string("input_305_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_305_cast_fp16 = batch_norm(beta = input_305_beta_0_to_fp16, epsilon = input_305_epsilon_0_to_fp16, gamma = input_305_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3151_cast_fp16)[name = string("input_305_cast_fp16")];
            string input_307_pad_type_0 = const()[name = string("input_307_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_307_strides_0 = const()[name = string("input_307_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_307_pad_0 = const()[name = string("input_307_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_307_dilations_0 = const()[name = string("input_307_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_307_groups_0 = const()[name = string("input_307_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_11_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144052288))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146149504))))[name = string("layers_11_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_307_cast_fp16 = conv(dilations = input_307_dilations_0, groups = input_307_groups_0, pad = input_307_pad_0, pad_type = input_307_pad_type_0, strides = input_307_strides_0, weight = layers_11_feed_forward1_linear1_weight_to_fp16_quantized, x = input_305_cast_fp16)[name = string("input_307_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_309_cast_fp16 = silu(x = input_307_cast_fp16)[name = string("input_309_cast_fp16")];
            string var_3168_pad_type_0 = const()[name = string("op_3168_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3168_strides_0 = const()[name = string("op_3168_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3168_pad_0 = const()[name = string("op_3168_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3168_dilations_0 = const()[name = string("op_3168_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3168_groups_0 = const()[name = string("op_3168_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_3169_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146157760))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148254976))))[name = string("op_3169_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3169_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_3168_dilations_0, groups = var_3168_groups_0, pad = var_3168_pad_0, pad_type = var_3168_pad_type_0, strides = var_3168_strides_0, weight = op_3169_weight_0_to_fp16_quantized, x = input_309_cast_fp16)[name = string("op_3169_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_247_cast_fp16 = add(x = x_245_cast_fp16, y = var_3169_cast_fp16)[name = string("x_247_cast_fp16")];
            tensor<int32, [1]> var_3185_axes_0 = const()[name = string("op_3185_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_11_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_11_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3185_cast_fp16 = layer_norm(axes = var_3185_axes_0, epsilon = layers_11_norm_self_att_eps_scaled_to_fp16, x = x_247_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<fp16, [1024]> x_249_gamma_0_to_fp16 = const()[name = string("x_249_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148257088)))];
            tensor<fp16, [1024]> x_249_beta_0_to_fp16 = const()[name = string("x_249_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148259200)))];
            fp16 x_249_epsilon_0_to_fp16 = const()[name = string("x_249_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_249_cast_fp16 = batch_norm(beta = x_249_beta_0_to_fp16, epsilon = x_249_epsilon_0_to_fp16, gamma = x_249_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3185_cast_fp16)[name = string("x_249_cast_fp16")];
            string q_23_pad_type_0 = const()[name = string("q_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_23_strides_0 = const()[name = string("q_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_23_pad_0 = const()[name = string("q_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_23_dilations_0 = const()[name = string("q_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_23_groups_0 = const()[name = string("q_23_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148261312))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148785664))))[name = string("layers_11_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_23_cast_fp16 = conv(dilations = q_23_dilations_0, groups = q_23_groups_0, pad = q_23_pad_0, pad_type = q_23_pad_type_0, strides = q_23_strides_0, weight = layers_11_self_attn_linear_q_weight_to_fp16_quantized, x = x_249_cast_fp16)[name = string("q_23_cast_fp16")];
            string k_23_pad_type_0 = const()[name = string("k_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_23_strides_0 = const()[name = string("k_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_23_pad_0 = const()[name = string("k_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_23_dilations_0 = const()[name = string("k_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_23_groups_0 = const()[name = string("k_23_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148787776))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149312128))))[name = string("layers_11_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_23_cast_fp16 = conv(dilations = k_23_dilations_0, groups = k_23_groups_0, pad = k_23_pad_0, pad_type = k_23_pad_type_0, strides = k_23_strides_0, weight = layers_11_self_attn_linear_k_weight_to_fp16_quantized, x = x_249_cast_fp16)[name = string("k_23_cast_fp16")];
            string v_23_pad_type_0 = const()[name = string("v_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_23_strides_0 = const()[name = string("v_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_23_pad_0 = const()[name = string("v_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_23_dilations_0 = const()[name = string("v_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_23_groups_0 = const()[name = string("v_23_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149314240))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149838592))))[name = string("layers_11_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_23_cast_fp16 = conv(dilations = v_23_dilations_0, groups = v_23_groups_0, pad = v_23_pad_0, pad_type = v_23_pad_type_0, strides = v_23_strides_0, weight = layers_11_self_attn_linear_v_weight_to_fp16_quantized, x = x_249_cast_fp16)[name = string("v_23_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_23_to_fp16 = const()[name = string("bv_all_23_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149840704)))];
            tensor<fp16, [1, 1024, 1, 188]> var_3218_cast_fp16 = add(x = q_23_cast_fp16, y = bv_all_23_to_fp16)[name = string("op_3218_cast_fp16")];
            tensor<int32, [3]> var_3219 = const()[name = string("op_3219"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_23_cast_fp16 = reshape(shape = var_3219, x = var_3218_cast_fp16)[name = string("qb_23_cast_fp16")];
            bool bd_all_45_transpose_x_0 = const()[name = string("bd_all_45_transpose_x_0"), val = bool(false)];
            bool bd_all_45_transpose_y_0 = const()[name = string("bd_all_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_11_self_attn_pos_proj_to_fp16 = const()[name = string("layers_11_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149842816)))];
            tensor<fp16, [8, 375, 188]> bd_all_45_cast_fp16 = matmul(transpose_x = bd_all_45_transpose_x_0, transpose_y = bd_all_45_transpose_y_0, x = layers_11_self_attn_pos_proj_to_fp16, y = qb_23_cast_fp16)[name = string("bd_all_45_cast_fp16")];
            tensor<int32, [3]> x_251_perm_0 = const()[name = string("x_251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_253_pad_0 = const()[name = string("x_253_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_253_mode_0 = const()[name = string("x_253_mode_0"), val = string("constant")];
            fp16 const_54_to_fp16 = const()[name = string("const_54_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_251_cast_fp16 = transpose(perm = x_251_perm_0, x = bd_all_45_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [8, 188, 376]> x_253_cast_fp16 = pad(constant_val = const_54_to_fp16, mode = x_253_mode_0, pad = x_253_pad_0, x = x_251_cast_fp16)[name = string("x_253_cast_fp16")];
            tensor<int32, [3]> var_3226 = const()[name = string("op_3226"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_255_cast_fp16 = reshape(shape = var_3226, x = x_253_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<int32, [3]> var_3229_begin_0 = const()[name = string("op_3229_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3229_end_0 = const()[name = string("op_3229_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_3229_end_mask_0 = const()[name = string("op_3229_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_3229_cast_fp16 = slice_by_index(begin = var_3229_begin_0, end = var_3229_end_0, end_mask = var_3229_end_mask_0, x = x_255_cast_fp16)[name = string("op_3229_cast_fp16")];
            tensor<int32, [3]> var_3230 = const()[name = string("op_3230"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_257_cast_fp16 = reshape(shape = var_3230, x = var_3229_cast_fp16)[name = string("x_257_cast_fp16")];
            tensor<int32, [3]> bd_all_47_begin_0 = const()[name = string("bd_all_47_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_47_end_0 = const()[name = string("bd_all_47_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_47_end_mask_0 = const()[name = string("bd_all_47_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_47_cast_fp16 = slice_by_index(begin = bd_all_47_begin_0, end = bd_all_47_end_0, end_mask = bd_all_47_end_mask_0, x = x_257_cast_fp16)[name = string("bd_all_47_cast_fp16")];
            tensor<int32, [4]> var_3235 = const()[name = string("op_3235"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3236_cast_fp16 = reshape(shape = var_3235, x = q_23_cast_fp16)[name = string("op_3236_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_3238_to_fp16 = const()[name = string("op_3238_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150610880)))];
            tensor<fp16, [8, 128, 1, 188]> qh_23_cast_fp16 = add(x = var_3236_cast_fp16, y = var_3238_to_fp16)[name = string("qh_23_cast_fp16")];
            tensor<int32, [4]> var_3240 = const()[name = string("op_3240"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3241_cast_fp16 = reshape(shape = var_3240, x = k_23_cast_fp16)[name = string("op_3241_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_3244 = const()[name = string("op_3244"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3245_cast_fp16 = reshape(shape = var_3244, x = v_23_cast_fp16)[name = string("op_3245_cast_fp16")];
            tensor<int32, [4]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3248_perm_0 = const()[name = string("op_3248_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_23_axes_0 = const()[name = string("bd_23_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_3248_cast_fp16 = transpose(perm = var_3248_perm_0, x = bd_all_47_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [8, 188, 1, 188]> bd_23_cast_fp16 = expand_dims(axes = bd_23_axes_0, x = var_3248_cast_fp16)[name = string("bd_23_cast_fp16")];
            string ac_23_equation_0 = const()[name = string("ac_23_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_3243_cast_fp16 = transpose(perm = var_3242, x = var_3241_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [8, 188, 1, 188]> ac_23_cast_fp16 = einsum(equation = ac_23_equation_0, values = (var_3243_cast_fp16, qh_23_cast_fp16))[name = string("ac_23_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_3264_cast_fp16 = add(x = ac_23_cast_fp16, y = bd_23_cast_fp16)[name = string("op_3264_cast_fp16")];
            fp16 var_3265_to_fp16 = const()[name = string("op_3265_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_45_cast_fp16 = mul(x = var_3264_cast_fp16, y = var_3265_to_fp16)[name = string("scores_45_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = key_bias)[name = string("scores_47_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_3268_cast_fp16 = softmax(axis = var_3124, x = scores_47_cast_fp16)[name = string("op_3268_cast_fp16")];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_114 = const()[name = string("concat_114"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = var_3268_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [8, 188, 188]> reshape_33_cast_fp16 = reshape(shape = concat_114, x = transpose_22_cast_fp16)[name = string("reshape_33_cast_fp16")];
            tensor<int32, [3]> concat_115 = const()[name = string("concat_115"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = var_3245_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [8, 188, 128]> reshape_34_cast_fp16 = reshape(shape = concat_115, x = transpose_59_cast_fp16)[name = string("reshape_34_cast_fp16")];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = reshape_33_cast_fp16, y = reshape_34_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<int32, [4]> concat_119 = const()[name = string("concat_119"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_35_cast_fp16 = reshape(shape = concat_119, x = matmul_11_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> ctx_23_perm_0 = const()[name = string("ctx_23_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3271 = const()[name = string("op_3271"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_23_cast_fp16 = transpose(perm = ctx_23_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 1024, 1, 188]> input_311_cast_fp16 = reshape(shape = var_3271, x = ctx_23_cast_fp16)[name = string("input_311_cast_fp16")];
            string var_3278_pad_type_0 = const()[name = string("op_3278_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3278_strides_0 = const()[name = string("op_3278_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3278_pad_0 = const()[name = string("op_3278_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3278_dilations_0 = const()[name = string("op_3278_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3278_groups_0 = const()[name = string("op_3278_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150612992))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151137344))))[name = string("layers_11_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3278_cast_fp16 = conv(dilations = var_3278_dilations_0, groups = var_3278_groups_0, pad = var_3278_pad_0, pad_type = var_3278_pad_type_0, strides = var_3278_strides_0, weight = layers_11_self_attn_linear_out_weight_to_fp16_quantized, x = input_311_cast_fp16)[name = string("op_3278_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_259_cast_fp16 = add(x = x_247_cast_fp16, y = var_3278_cast_fp16)[name = string("x_259_cast_fp16")];
            tensor<int32, [1]> var_3294_axes_0 = const()[name = string("op_3294_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_11_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_11_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3294_cast_fp16 = layer_norm(axes = var_3294_axes_0, epsilon = layers_11_norm_conv_eps_scaled_to_fp16, x = x_259_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1024]> input_313_gamma_0_to_fp16 = const()[name = string("input_313_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151139456)))];
            tensor<fp16, [1024]> input_313_beta_0_to_fp16 = const()[name = string("input_313_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151141568)))];
            fp16 input_313_epsilon_0_to_fp16 = const()[name = string("input_313_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_313_cast_fp16 = batch_norm(beta = input_313_beta_0_to_fp16, epsilon = input_313_epsilon_0_to_fp16, gamma = input_313_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3294_cast_fp16)[name = string("input_313_cast_fp16")];
            string input_315_pad_type_0 = const()[name = string("input_315_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_315_strides_0 = const()[name = string("input_315_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_315_pad_0 = const()[name = string("input_315_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_315_dilations_0 = const()[name = string("input_315_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_315_groups_0 = const()[name = string("input_315_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151143680))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152192320))))[name = string("layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_315_cast_fp16 = conv(dilations = input_315_dilations_0, groups = input_315_groups_0, pad = input_315_pad_0, pad_type = input_315_pad_type_0, strides = input_315_strides_0, weight = layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_313_cast_fp16)[name = string("input_315_cast_fp16")];
            int32 x_261_split_num_splits_0 = const()[name = string("x_261_split_num_splits_0"), val = int32(2)];
            int32 x_261_split_axis_0 = const()[name = string("x_261_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_261_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_261_split_cast_fp16_1 = split(axis = x_261_split_axis_0, num_splits = x_261_split_num_splits_0, x = input_315_cast_fp16)[name = string("x_261_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_261_split_1_sigmoid_cast_fp16 = sigmoid(x = x_261_split_cast_fp16_1)[name = string("x_261_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_261_cast_fp16 = mul(x = x_261_split_cast_fp16_0, y = x_261_split_1_sigmoid_cast_fp16)[name = string("x_261_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_317_cast_fp16 = mul(x = x_261_cast_fp16, y = pad_mask)[name = string("input_317_cast_fp16")];
            string input_319_pad_type_0 = const()[name = string("input_319_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_319_pad_0 = const()[name = string("input_319_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_319_groups_0 = const()[name = string("input_319_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_319_strides_0 = const()[name = string("input_319_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_319_dilations_0 = const()[name = string("input_319_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_125_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152196480))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201152))))[name = string("const_125_to_fp16_quantized")];
            tensor<fp16, [1024]> const_126_to_fp16 = const()[name = string("const_126_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152203264)))];
            tensor<fp16, [1, 1024, 1, 188]> input_321_cast_fp16 = conv(bias = const_126_to_fp16, dilations = input_319_dilations_0, groups = input_319_groups_0, pad = input_319_pad_0, pad_type = input_319_pad_type_0, strides = input_319_strides_0, weight = const_125_to_fp16_quantized, x = input_317_cast_fp16)[name = string("input_321_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_323_cast_fp16 = silu(x = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
            string var_3326_pad_type_0 = const()[name = string("op_3326_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3326_strides_0 = const()[name = string("op_3326_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3326_pad_0 = const()[name = string("op_3326_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3326_dilations_0 = const()[name = string("op_3326_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3326_groups_0 = const()[name = string("op_3326_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152205376))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152729728))))[name = string("layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3326_cast_fp16 = conv(dilations = var_3326_dilations_0, groups = var_3326_groups_0, pad = var_3326_pad_0, pad_type = var_3326_pad_type_0, strides = var_3326_strides_0, weight = layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_323_cast_fp16)[name = string("op_3326_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_263_cast_fp16 = add(x = x_259_cast_fp16, y = var_3326_cast_fp16)[name = string("x_263_cast_fp16")];
            tensor<int32, [1]> var_3342_axes_0 = const()[name = string("op_3342_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_11_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_11_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3342_cast_fp16 = layer_norm(axes = var_3342_axes_0, epsilon = layers_11_norm_feed_forward2_eps_scaled_to_fp16, x = x_263_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<fp16, [1024]> input_325_gamma_0_to_fp16 = const()[name = string("input_325_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152731840)))];
            tensor<fp16, [1024]> input_325_beta_0_to_fp16 = const()[name = string("input_325_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152733952)))];
            fp16 input_325_epsilon_0_to_fp16 = const()[name = string("input_325_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_325_cast_fp16 = batch_norm(beta = input_325_beta_0_to_fp16, epsilon = input_325_epsilon_0_to_fp16, gamma = input_325_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3342_cast_fp16)[name = string("input_325_cast_fp16")];
            string input_327_pad_type_0 = const()[name = string("input_327_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_327_strides_0 = const()[name = string("input_327_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_327_pad_0 = const()[name = string("input_327_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_327_dilations_0 = const()[name = string("input_327_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_327_groups_0 = const()[name = string("input_327_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_11_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152736064))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154833280))))[name = string("layers_11_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_327_cast_fp16 = conv(dilations = input_327_dilations_0, groups = input_327_groups_0, pad = input_327_pad_0, pad_type = input_327_pad_type_0, strides = input_327_strides_0, weight = layers_11_feed_forward2_linear1_weight_to_fp16_quantized, x = input_325_cast_fp16)[name = string("input_327_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_329_cast_fp16 = silu(x = input_327_cast_fp16)[name = string("input_329_cast_fp16")];
            string var_3359_pad_type_0 = const()[name = string("op_3359_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3359_strides_0 = const()[name = string("op_3359_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3359_pad_0 = const()[name = string("op_3359_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3359_dilations_0 = const()[name = string("op_3359_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3359_groups_0 = const()[name = string("op_3359_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_3360_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154841536))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156938752))))[name = string("op_3360_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3360_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_3359_dilations_0, groups = var_3359_groups_0, pad = var_3359_pad_0, pad_type = var_3359_pad_type_0, strides = var_3359_strides_0, weight = op_3360_weight_0_to_fp16_quantized, x = input_329_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_265_cast_fp16 = add(x = x_263_cast_fp16, y = var_3360_cast_fp16)[name = string("x_265_cast_fp16")];
            tensor<int32, [1]> var_3376_axes_0 = const()[name = string("op_3376_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_11_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_11_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3376_cast_fp16 = layer_norm(axes = var_3376_axes_0, epsilon = layers_11_norm_out_eps_scaled_to_fp16, x = x_265_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1024]> x_267_gamma_0_to_fp16 = const()[name = string("x_267_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156940864)))];
            tensor<fp16, [1024]> x_267_beta_0_to_fp16 = const()[name = string("x_267_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156942976)))];
            fp16 x_267_epsilon_0_to_fp16 = const()[name = string("x_267_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_267_cast_fp16 = batch_norm(beta = x_267_beta_0_to_fp16, epsilon = x_267_epsilon_0_to_fp16, gamma = x_267_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3376_cast_fp16)[name = string("x_267_cast_fp16")];
            int32 var_3395 = const()[name = string("op_3395"), val = int32(1)];
            tensor<int32, [1]> var_3422_axes_0 = const()[name = string("op_3422_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_12_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_12_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3422_cast_fp16 = layer_norm(axes = var_3422_axes_0, epsilon = layers_12_norm_feed_forward1_eps_scaled_to_fp16, x = x_267_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<fp16, [1024]> input_331_gamma_0_to_fp16 = const()[name = string("input_331_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156945088)))];
            tensor<fp16, [1024]> input_331_beta_0_to_fp16 = const()[name = string("input_331_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156947200)))];
            fp16 input_331_epsilon_0_to_fp16 = const()[name = string("input_331_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_331_cast_fp16 = batch_norm(beta = input_331_beta_0_to_fp16, epsilon = input_331_epsilon_0_to_fp16, gamma = input_331_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3422_cast_fp16)[name = string("input_331_cast_fp16")];
            string input_333_pad_type_0 = const()[name = string("input_333_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_333_strides_0 = const()[name = string("input_333_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_333_pad_0 = const()[name = string("input_333_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_333_dilations_0 = const()[name = string("input_333_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_333_groups_0 = const()[name = string("input_333_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_12_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156949312))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159046528))))[name = string("layers_12_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_333_cast_fp16 = conv(dilations = input_333_dilations_0, groups = input_333_groups_0, pad = input_333_pad_0, pad_type = input_333_pad_type_0, strides = input_333_strides_0, weight = layers_12_feed_forward1_linear1_weight_to_fp16_quantized, x = input_331_cast_fp16)[name = string("input_333_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_335_cast_fp16 = silu(x = input_333_cast_fp16)[name = string("input_335_cast_fp16")];
            string var_3439_pad_type_0 = const()[name = string("op_3439_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3439_strides_0 = const()[name = string("op_3439_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3439_pad_0 = const()[name = string("op_3439_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3439_dilations_0 = const()[name = string("op_3439_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3439_groups_0 = const()[name = string("op_3439_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_3440_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159054784))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161152000))))[name = string("op_3440_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3440_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_3439_dilations_0, groups = var_3439_groups_0, pad = var_3439_pad_0, pad_type = var_3439_pad_type_0, strides = var_3439_strides_0, weight = op_3440_weight_0_to_fp16_quantized, x = input_335_cast_fp16)[name = string("op_3440_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_269_cast_fp16 = add(x = x_267_cast_fp16, y = var_3440_cast_fp16)[name = string("x_269_cast_fp16")];
            tensor<int32, [1]> var_3456_axes_0 = const()[name = string("op_3456_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_12_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_12_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3456_cast_fp16 = layer_norm(axes = var_3456_axes_0, epsilon = layers_12_norm_self_att_eps_scaled_to_fp16, x = x_269_cast_fp16)[name = string("op_3456_cast_fp16")];
            tensor<fp16, [1024]> x_271_gamma_0_to_fp16 = const()[name = string("x_271_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161154112)))];
            tensor<fp16, [1024]> x_271_beta_0_to_fp16 = const()[name = string("x_271_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161156224)))];
            fp16 x_271_epsilon_0_to_fp16 = const()[name = string("x_271_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_271_cast_fp16 = batch_norm(beta = x_271_beta_0_to_fp16, epsilon = x_271_epsilon_0_to_fp16, gamma = x_271_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3456_cast_fp16)[name = string("x_271_cast_fp16")];
            string q_25_pad_type_0 = const()[name = string("q_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_25_strides_0 = const()[name = string("q_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_25_pad_0 = const()[name = string("q_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_25_dilations_0 = const()[name = string("q_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_25_groups_0 = const()[name = string("q_25_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161158336))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161682688))))[name = string("layers_12_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_25_cast_fp16 = conv(dilations = q_25_dilations_0, groups = q_25_groups_0, pad = q_25_pad_0, pad_type = q_25_pad_type_0, strides = q_25_strides_0, weight = layers_12_self_attn_linear_q_weight_to_fp16_quantized, x = x_271_cast_fp16)[name = string("q_25_cast_fp16")];
            string k_25_pad_type_0 = const()[name = string("k_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_25_strides_0 = const()[name = string("k_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_25_pad_0 = const()[name = string("k_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_25_dilations_0 = const()[name = string("k_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_25_groups_0 = const()[name = string("k_25_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161684800))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162209152))))[name = string("layers_12_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_25_cast_fp16 = conv(dilations = k_25_dilations_0, groups = k_25_groups_0, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = k_25_strides_0, weight = layers_12_self_attn_linear_k_weight_to_fp16_quantized, x = x_271_cast_fp16)[name = string("k_25_cast_fp16")];
            string v_25_pad_type_0 = const()[name = string("v_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_25_strides_0 = const()[name = string("v_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_25_pad_0 = const()[name = string("v_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_25_dilations_0 = const()[name = string("v_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_25_groups_0 = const()[name = string("v_25_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162211264))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162735616))))[name = string("layers_12_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_25_cast_fp16 = conv(dilations = v_25_dilations_0, groups = v_25_groups_0, pad = v_25_pad_0, pad_type = v_25_pad_type_0, strides = v_25_strides_0, weight = layers_12_self_attn_linear_v_weight_to_fp16_quantized, x = x_271_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_25_to_fp16 = const()[name = string("bv_all_25_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162737728)))];
            tensor<fp16, [1, 1024, 1, 188]> var_3489_cast_fp16 = add(x = q_25_cast_fp16, y = bv_all_25_to_fp16)[name = string("op_3489_cast_fp16")];
            tensor<int32, [3]> var_3490 = const()[name = string("op_3490"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_25_cast_fp16 = reshape(shape = var_3490, x = var_3489_cast_fp16)[name = string("qb_25_cast_fp16")];
            bool bd_all_49_transpose_x_0 = const()[name = string("bd_all_49_transpose_x_0"), val = bool(false)];
            bool bd_all_49_transpose_y_0 = const()[name = string("bd_all_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_12_self_attn_pos_proj_to_fp16 = const()[name = string("layers_12_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162739840)))];
            tensor<fp16, [8, 375, 188]> bd_all_49_cast_fp16 = matmul(transpose_x = bd_all_49_transpose_x_0, transpose_y = bd_all_49_transpose_y_0, x = layers_12_self_attn_pos_proj_to_fp16, y = qb_25_cast_fp16)[name = string("bd_all_49_cast_fp16")];
            tensor<int32, [3]> x_273_perm_0 = const()[name = string("x_273_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_275_pad_0 = const()[name = string("x_275_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_275_mode_0 = const()[name = string("x_275_mode_0"), val = string("constant")];
            fp16 const_58_to_fp16 = const()[name = string("const_58_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_273_cast_fp16 = transpose(perm = x_273_perm_0, x = bd_all_49_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [8, 188, 376]> x_275_cast_fp16 = pad(constant_val = const_58_to_fp16, mode = x_275_mode_0, pad = x_275_pad_0, x = x_273_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<int32, [3]> var_3497 = const()[name = string("op_3497"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_277_cast_fp16 = reshape(shape = var_3497, x = x_275_cast_fp16)[name = string("x_277_cast_fp16")];
            tensor<int32, [3]> var_3500_begin_0 = const()[name = string("op_3500_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3500_end_0 = const()[name = string("op_3500_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_3500_end_mask_0 = const()[name = string("op_3500_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_3500_cast_fp16 = slice_by_index(begin = var_3500_begin_0, end = var_3500_end_0, end_mask = var_3500_end_mask_0, x = x_277_cast_fp16)[name = string("op_3500_cast_fp16")];
            tensor<int32, [3]> var_3501 = const()[name = string("op_3501"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_279_cast_fp16 = reshape(shape = var_3501, x = var_3500_cast_fp16)[name = string("x_279_cast_fp16")];
            tensor<int32, [3]> bd_all_51_begin_0 = const()[name = string("bd_all_51_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_51_end_0 = const()[name = string("bd_all_51_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_51_end_mask_0 = const()[name = string("bd_all_51_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_51_cast_fp16 = slice_by_index(begin = bd_all_51_begin_0, end = bd_all_51_end_0, end_mask = bd_all_51_end_mask_0, x = x_279_cast_fp16)[name = string("bd_all_51_cast_fp16")];
            tensor<int32, [4]> var_3506 = const()[name = string("op_3506"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3507_cast_fp16 = reshape(shape = var_3506, x = q_25_cast_fp16)[name = string("op_3507_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_3509_to_fp16 = const()[name = string("op_3509_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163507904)))];
            tensor<fp16, [8, 128, 1, 188]> qh_25_cast_fp16 = add(x = var_3507_cast_fp16, y = var_3509_to_fp16)[name = string("qh_25_cast_fp16")];
            tensor<int32, [4]> var_3511 = const()[name = string("op_3511"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3512_cast_fp16 = reshape(shape = var_3511, x = k_25_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<int32, [4]> var_3513 = const()[name = string("op_3513"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_3515 = const()[name = string("op_3515"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3516_cast_fp16 = reshape(shape = var_3515, x = v_25_cast_fp16)[name = string("op_3516_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3519_perm_0 = const()[name = string("op_3519_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_25_axes_0 = const()[name = string("bd_25_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_3519_cast_fp16 = transpose(perm = var_3519_perm_0, x = bd_all_51_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [8, 188, 1, 188]> bd_25_cast_fp16 = expand_dims(axes = bd_25_axes_0, x = var_3519_cast_fp16)[name = string("bd_25_cast_fp16")];
            string ac_25_equation_0 = const()[name = string("ac_25_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_3514_cast_fp16 = transpose(perm = var_3513, x = var_3512_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [8, 188, 1, 188]> ac_25_cast_fp16 = einsum(equation = ac_25_equation_0, values = (var_3514_cast_fp16, qh_25_cast_fp16))[name = string("ac_25_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_3535_cast_fp16 = add(x = ac_25_cast_fp16, y = bd_25_cast_fp16)[name = string("op_3535_cast_fp16")];
            fp16 var_3536_to_fp16 = const()[name = string("op_3536_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_49_cast_fp16 = mul(x = var_3535_cast_fp16, y = var_3536_to_fp16)[name = string("scores_49_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = key_bias)[name = string("scores_51_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_3539_cast_fp16 = softmax(axis = var_3395, x = scores_51_cast_fp16)[name = string("op_3539_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_124 = const()[name = string("concat_124"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = var_3539_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [8, 188, 188]> reshape_36_cast_fp16 = reshape(shape = concat_124, x = transpose_24_cast_fp16)[name = string("reshape_36_cast_fp16")];
            tensor<int32, [3]> concat_125 = const()[name = string("concat_125"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = var_3516_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [8, 188, 128]> reshape_37_cast_fp16 = reshape(shape = concat_125, x = transpose_60_cast_fp16)[name = string("reshape_37_cast_fp16")];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = reshape_36_cast_fp16, y = reshape_37_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<int32, [4]> concat_129 = const()[name = string("concat_129"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_38_cast_fp16 = reshape(shape = concat_129, x = matmul_12_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [4]> ctx_25_perm_0 = const()[name = string("ctx_25_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3542 = const()[name = string("op_3542"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_25_cast_fp16 = transpose(perm = ctx_25_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 1024, 1, 188]> input_337_cast_fp16 = reshape(shape = var_3542, x = ctx_25_cast_fp16)[name = string("input_337_cast_fp16")];
            string var_3549_pad_type_0 = const()[name = string("op_3549_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3549_strides_0 = const()[name = string("op_3549_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3549_pad_0 = const()[name = string("op_3549_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3549_dilations_0 = const()[name = string("op_3549_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3549_groups_0 = const()[name = string("op_3549_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163510016))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164034368))))[name = string("layers_12_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3549_cast_fp16 = conv(dilations = var_3549_dilations_0, groups = var_3549_groups_0, pad = var_3549_pad_0, pad_type = var_3549_pad_type_0, strides = var_3549_strides_0, weight = layers_12_self_attn_linear_out_weight_to_fp16_quantized, x = input_337_cast_fp16)[name = string("op_3549_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_281_cast_fp16 = add(x = x_269_cast_fp16, y = var_3549_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<int32, [1]> var_3565_axes_0 = const()[name = string("op_3565_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_12_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_12_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3565_cast_fp16 = layer_norm(axes = var_3565_axes_0, epsilon = layers_12_norm_conv_eps_scaled_to_fp16, x = x_281_cast_fp16)[name = string("op_3565_cast_fp16")];
            tensor<fp16, [1024]> input_339_gamma_0_to_fp16 = const()[name = string("input_339_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164036480)))];
            tensor<fp16, [1024]> input_339_beta_0_to_fp16 = const()[name = string("input_339_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164038592)))];
            fp16 input_339_epsilon_0_to_fp16 = const()[name = string("input_339_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_339_cast_fp16 = batch_norm(beta = input_339_beta_0_to_fp16, epsilon = input_339_epsilon_0_to_fp16, gamma = input_339_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3565_cast_fp16)[name = string("input_339_cast_fp16")];
            string input_341_pad_type_0 = const()[name = string("input_341_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_341_strides_0 = const()[name = string("input_341_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_341_pad_0 = const()[name = string("input_341_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_341_dilations_0 = const()[name = string("input_341_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_341_groups_0 = const()[name = string("input_341_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164040704))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165089344))))[name = string("layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_341_cast_fp16 = conv(dilations = input_341_dilations_0, groups = input_341_groups_0, pad = input_341_pad_0, pad_type = input_341_pad_type_0, strides = input_341_strides_0, weight = layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_339_cast_fp16)[name = string("input_341_cast_fp16")];
            int32 x_283_split_num_splits_0 = const()[name = string("x_283_split_num_splits_0"), val = int32(2)];
            int32 x_283_split_axis_0 = const()[name = string("x_283_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_283_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_283_split_cast_fp16_1 = split(axis = x_283_split_axis_0, num_splits = x_283_split_num_splits_0, x = input_341_cast_fp16)[name = string("x_283_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_283_split_1_sigmoid_cast_fp16 = sigmoid(x = x_283_split_cast_fp16_1)[name = string("x_283_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_283_cast_fp16 = mul(x = x_283_split_cast_fp16_0, y = x_283_split_1_sigmoid_cast_fp16)[name = string("x_283_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_343_cast_fp16 = mul(x = x_283_cast_fp16, y = pad_mask)[name = string("input_343_cast_fp16")];
            string input_345_pad_type_0 = const()[name = string("input_345_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_345_pad_0 = const()[name = string("input_345_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_345_groups_0 = const()[name = string("input_345_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_345_strides_0 = const()[name = string("input_345_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_345_dilations_0 = const()[name = string("input_345_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_127_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165093504))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165098176))))[name = string("const_127_to_fp16_quantized")];
            tensor<fp16, [1024]> const_128_to_fp16 = const()[name = string("const_128_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165100288)))];
            tensor<fp16, [1, 1024, 1, 188]> input_347_cast_fp16 = conv(bias = const_128_to_fp16, dilations = input_345_dilations_0, groups = input_345_groups_0, pad = input_345_pad_0, pad_type = input_345_pad_type_0, strides = input_345_strides_0, weight = const_127_to_fp16_quantized, x = input_343_cast_fp16)[name = string("input_347_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_349_cast_fp16 = silu(x = input_347_cast_fp16)[name = string("input_349_cast_fp16")];
            string var_3597_pad_type_0 = const()[name = string("op_3597_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3597_strides_0 = const()[name = string("op_3597_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3597_pad_0 = const()[name = string("op_3597_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3597_dilations_0 = const()[name = string("op_3597_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3597_groups_0 = const()[name = string("op_3597_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165102400))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165626752))))[name = string("layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3597_cast_fp16 = conv(dilations = var_3597_dilations_0, groups = var_3597_groups_0, pad = var_3597_pad_0, pad_type = var_3597_pad_type_0, strides = var_3597_strides_0, weight = layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_349_cast_fp16)[name = string("op_3597_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_285_cast_fp16 = add(x = x_281_cast_fp16, y = var_3597_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [1]> var_3613_axes_0 = const()[name = string("op_3613_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_12_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_12_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3613_cast_fp16 = layer_norm(axes = var_3613_axes_0, epsilon = layers_12_norm_feed_forward2_eps_scaled_to_fp16, x = x_285_cast_fp16)[name = string("op_3613_cast_fp16")];
            tensor<fp16, [1024]> input_351_gamma_0_to_fp16 = const()[name = string("input_351_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165628864)))];
            tensor<fp16, [1024]> input_351_beta_0_to_fp16 = const()[name = string("input_351_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165630976)))];
            fp16 input_351_epsilon_0_to_fp16 = const()[name = string("input_351_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_351_cast_fp16 = batch_norm(beta = input_351_beta_0_to_fp16, epsilon = input_351_epsilon_0_to_fp16, gamma = input_351_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3613_cast_fp16)[name = string("input_351_cast_fp16")];
            string input_353_pad_type_0 = const()[name = string("input_353_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_353_strides_0 = const()[name = string("input_353_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_353_pad_0 = const()[name = string("input_353_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_353_dilations_0 = const()[name = string("input_353_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_353_groups_0 = const()[name = string("input_353_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_12_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165633088))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167730304))))[name = string("layers_12_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_353_cast_fp16 = conv(dilations = input_353_dilations_0, groups = input_353_groups_0, pad = input_353_pad_0, pad_type = input_353_pad_type_0, strides = input_353_strides_0, weight = layers_12_feed_forward2_linear1_weight_to_fp16_quantized, x = input_351_cast_fp16)[name = string("input_353_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = string("input_355_cast_fp16")];
            string var_3630_pad_type_0 = const()[name = string("op_3630_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3630_strides_0 = const()[name = string("op_3630_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3630_pad_0 = const()[name = string("op_3630_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3630_dilations_0 = const()[name = string("op_3630_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3630_groups_0 = const()[name = string("op_3630_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_3631_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167738560))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169835776))))[name = string("op_3631_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3631_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_3630_dilations_0, groups = var_3630_groups_0, pad = var_3630_pad_0, pad_type = var_3630_pad_type_0, strides = var_3630_strides_0, weight = op_3631_weight_0_to_fp16_quantized, x = input_355_cast_fp16)[name = string("op_3631_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_287_cast_fp16 = add(x = x_285_cast_fp16, y = var_3631_cast_fp16)[name = string("x_287_cast_fp16")];
            tensor<int32, [1]> var_3647_axes_0 = const()[name = string("op_3647_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_12_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_12_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3647_cast_fp16 = layer_norm(axes = var_3647_axes_0, epsilon = layers_12_norm_out_eps_scaled_to_fp16, x = x_287_cast_fp16)[name = string("op_3647_cast_fp16")];
            tensor<fp16, [1024]> x_289_gamma_0_to_fp16 = const()[name = string("x_289_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169837888)))];
            tensor<fp16, [1024]> x_289_beta_0_to_fp16 = const()[name = string("x_289_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169840000)))];
            fp16 x_289_epsilon_0_to_fp16 = const()[name = string("x_289_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_289_cast_fp16 = batch_norm(beta = x_289_beta_0_to_fp16, epsilon = x_289_epsilon_0_to_fp16, gamma = x_289_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3647_cast_fp16)[name = string("x_289_cast_fp16")];
            int32 var_3666 = const()[name = string("op_3666"), val = int32(1)];
            tensor<int32, [1]> var_3693_axes_0 = const()[name = string("op_3693_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_13_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_13_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3693_cast_fp16 = layer_norm(axes = var_3693_axes_0, epsilon = layers_13_norm_feed_forward1_eps_scaled_to_fp16, x = x_289_cast_fp16)[name = string("op_3693_cast_fp16")];
            tensor<fp16, [1024]> input_357_gamma_0_to_fp16 = const()[name = string("input_357_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169842112)))];
            tensor<fp16, [1024]> input_357_beta_0_to_fp16 = const()[name = string("input_357_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169844224)))];
            fp16 input_357_epsilon_0_to_fp16 = const()[name = string("input_357_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_357_cast_fp16 = batch_norm(beta = input_357_beta_0_to_fp16, epsilon = input_357_epsilon_0_to_fp16, gamma = input_357_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3693_cast_fp16)[name = string("input_357_cast_fp16")];
            string input_359_pad_type_0 = const()[name = string("input_359_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_359_strides_0 = const()[name = string("input_359_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_359_pad_0 = const()[name = string("input_359_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_359_dilations_0 = const()[name = string("input_359_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_359_groups_0 = const()[name = string("input_359_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_13_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169846336))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171943552))))[name = string("layers_13_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_359_cast_fp16 = conv(dilations = input_359_dilations_0, groups = input_359_groups_0, pad = input_359_pad_0, pad_type = input_359_pad_type_0, strides = input_359_strides_0, weight = layers_13_feed_forward1_linear1_weight_to_fp16_quantized, x = input_357_cast_fp16)[name = string("input_359_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_361_cast_fp16 = silu(x = input_359_cast_fp16)[name = string("input_361_cast_fp16")];
            string var_3710_pad_type_0 = const()[name = string("op_3710_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3710_strides_0 = const()[name = string("op_3710_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3710_pad_0 = const()[name = string("op_3710_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3710_dilations_0 = const()[name = string("op_3710_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3710_groups_0 = const()[name = string("op_3710_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_3711_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171951808))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174049024))))[name = string("op_3711_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3711_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_3710_dilations_0, groups = var_3710_groups_0, pad = var_3710_pad_0, pad_type = var_3710_pad_type_0, strides = var_3710_strides_0, weight = op_3711_weight_0_to_fp16_quantized, x = input_361_cast_fp16)[name = string("op_3711_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_291_cast_fp16 = add(x = x_289_cast_fp16, y = var_3711_cast_fp16)[name = string("x_291_cast_fp16")];
            tensor<int32, [1]> var_3727_axes_0 = const()[name = string("op_3727_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_13_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_13_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3727_cast_fp16 = layer_norm(axes = var_3727_axes_0, epsilon = layers_13_norm_self_att_eps_scaled_to_fp16, x = x_291_cast_fp16)[name = string("op_3727_cast_fp16")];
            tensor<fp16, [1024]> x_293_gamma_0_to_fp16 = const()[name = string("x_293_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174051136)))];
            tensor<fp16, [1024]> x_293_beta_0_to_fp16 = const()[name = string("x_293_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174053248)))];
            fp16 x_293_epsilon_0_to_fp16 = const()[name = string("x_293_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_293_cast_fp16 = batch_norm(beta = x_293_beta_0_to_fp16, epsilon = x_293_epsilon_0_to_fp16, gamma = x_293_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3727_cast_fp16)[name = string("x_293_cast_fp16")];
            string q_27_pad_type_0 = const()[name = string("q_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_27_strides_0 = const()[name = string("q_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_27_pad_0 = const()[name = string("q_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_27_dilations_0 = const()[name = string("q_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_27_groups_0 = const()[name = string("q_27_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174055360))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174579712))))[name = string("layers_13_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_27_cast_fp16 = conv(dilations = q_27_dilations_0, groups = q_27_groups_0, pad = q_27_pad_0, pad_type = q_27_pad_type_0, strides = q_27_strides_0, weight = layers_13_self_attn_linear_q_weight_to_fp16_quantized, x = x_293_cast_fp16)[name = string("q_27_cast_fp16")];
            string k_27_pad_type_0 = const()[name = string("k_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_27_strides_0 = const()[name = string("k_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_27_pad_0 = const()[name = string("k_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_27_dilations_0 = const()[name = string("k_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_27_groups_0 = const()[name = string("k_27_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174581824))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175106176))))[name = string("layers_13_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_27_cast_fp16 = conv(dilations = k_27_dilations_0, groups = k_27_groups_0, pad = k_27_pad_0, pad_type = k_27_pad_type_0, strides = k_27_strides_0, weight = layers_13_self_attn_linear_k_weight_to_fp16_quantized, x = x_293_cast_fp16)[name = string("k_27_cast_fp16")];
            string v_27_pad_type_0 = const()[name = string("v_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_27_strides_0 = const()[name = string("v_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_27_pad_0 = const()[name = string("v_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_27_dilations_0 = const()[name = string("v_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_27_groups_0 = const()[name = string("v_27_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175108288))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175632640))))[name = string("layers_13_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_27_cast_fp16 = conv(dilations = v_27_dilations_0, groups = v_27_groups_0, pad = v_27_pad_0, pad_type = v_27_pad_type_0, strides = v_27_strides_0, weight = layers_13_self_attn_linear_v_weight_to_fp16_quantized, x = x_293_cast_fp16)[name = string("v_27_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_27_to_fp16 = const()[name = string("bv_all_27_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175634752)))];
            tensor<fp16, [1, 1024, 1, 188]> var_3760_cast_fp16 = add(x = q_27_cast_fp16, y = bv_all_27_to_fp16)[name = string("op_3760_cast_fp16")];
            tensor<int32, [3]> var_3761 = const()[name = string("op_3761"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_27_cast_fp16 = reshape(shape = var_3761, x = var_3760_cast_fp16)[name = string("qb_27_cast_fp16")];
            bool bd_all_53_transpose_x_0 = const()[name = string("bd_all_53_transpose_x_0"), val = bool(false)];
            bool bd_all_53_transpose_y_0 = const()[name = string("bd_all_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_13_self_attn_pos_proj_to_fp16 = const()[name = string("layers_13_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175636864)))];
            tensor<fp16, [8, 375, 188]> bd_all_53_cast_fp16 = matmul(transpose_x = bd_all_53_transpose_x_0, transpose_y = bd_all_53_transpose_y_0, x = layers_13_self_attn_pos_proj_to_fp16, y = qb_27_cast_fp16)[name = string("bd_all_53_cast_fp16")];
            tensor<int32, [3]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_297_pad_0 = const()[name = string("x_297_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_297_mode_0 = const()[name = string("x_297_mode_0"), val = string("constant")];
            fp16 const_62_to_fp16 = const()[name = string("const_62_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = bd_all_53_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [8, 188, 376]> x_297_cast_fp16 = pad(constant_val = const_62_to_fp16, mode = x_297_mode_0, pad = x_297_pad_0, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")];
            tensor<int32, [3]> var_3768 = const()[name = string("op_3768"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_299_cast_fp16 = reshape(shape = var_3768, x = x_297_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<int32, [3]> var_3771_begin_0 = const()[name = string("op_3771_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_3771_end_0 = const()[name = string("op_3771_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_3771_end_mask_0 = const()[name = string("op_3771_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_3771_cast_fp16 = slice_by_index(begin = var_3771_begin_0, end = var_3771_end_0, end_mask = var_3771_end_mask_0, x = x_299_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [3]> var_3772 = const()[name = string("op_3772"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_301_cast_fp16 = reshape(shape = var_3772, x = var_3771_cast_fp16)[name = string("x_301_cast_fp16")];
            tensor<int32, [3]> bd_all_55_begin_0 = const()[name = string("bd_all_55_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_55_end_0 = const()[name = string("bd_all_55_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_55_end_mask_0 = const()[name = string("bd_all_55_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_55_cast_fp16 = slice_by_index(begin = bd_all_55_begin_0, end = bd_all_55_end_0, end_mask = bd_all_55_end_mask_0, x = x_301_cast_fp16)[name = string("bd_all_55_cast_fp16")];
            tensor<int32, [4]> var_3777 = const()[name = string("op_3777"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3778_cast_fp16 = reshape(shape = var_3777, x = q_27_cast_fp16)[name = string("op_3778_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_3780_to_fp16 = const()[name = string("op_3780_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176404928)))];
            tensor<fp16, [8, 128, 1, 188]> qh_27_cast_fp16 = add(x = var_3778_cast_fp16, y = var_3780_to_fp16)[name = string("qh_27_cast_fp16")];
            tensor<int32, [4]> var_3782 = const()[name = string("op_3782"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3783_cast_fp16 = reshape(shape = var_3782, x = k_27_cast_fp16)[name = string("op_3783_cast_fp16")];
            tensor<int32, [4]> var_3784 = const()[name = string("op_3784"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_3786 = const()[name = string("op_3786"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_3787_cast_fp16 = reshape(shape = var_3786, x = v_27_cast_fp16)[name = string("op_3787_cast_fp16")];
            tensor<int32, [4]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3790_perm_0 = const()[name = string("op_3790_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_27_axes_0 = const()[name = string("bd_27_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_3790_cast_fp16 = transpose(perm = var_3790_perm_0, x = bd_all_55_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [8, 188, 1, 188]> bd_27_cast_fp16 = expand_dims(axes = bd_27_axes_0, x = var_3790_cast_fp16)[name = string("bd_27_cast_fp16")];
            string ac_27_equation_0 = const()[name = string("ac_27_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_3785_cast_fp16 = transpose(perm = var_3784, x = var_3783_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [8, 188, 1, 188]> ac_27_cast_fp16 = einsum(equation = ac_27_equation_0, values = (var_3785_cast_fp16, qh_27_cast_fp16))[name = string("ac_27_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_3806_cast_fp16 = add(x = ac_27_cast_fp16, y = bd_27_cast_fp16)[name = string("op_3806_cast_fp16")];
            fp16 var_3807_to_fp16 = const()[name = string("op_3807_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_53_cast_fp16 = mul(x = var_3806_cast_fp16, y = var_3807_to_fp16)[name = string("scores_53_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = key_bias)[name = string("scores_55_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_3810_cast_fp16 = softmax(axis = var_3666, x = scores_55_cast_fp16)[name = string("op_3810_cast_fp16")];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_134 = const()[name = string("concat_134"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = var_3810_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [8, 188, 188]> reshape_39_cast_fp16 = reshape(shape = concat_134, x = transpose_26_cast_fp16)[name = string("reshape_39_cast_fp16")];
            tensor<int32, [3]> concat_135 = const()[name = string("concat_135"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = var_3787_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [8, 188, 128]> reshape_40_cast_fp16 = reshape(shape = concat_135, x = transpose_61_cast_fp16)[name = string("reshape_40_cast_fp16")];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = reshape_39_cast_fp16, y = reshape_40_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<int32, [4]> concat_139 = const()[name = string("concat_139"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_41_cast_fp16 = reshape(shape = concat_139, x = matmul_13_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> ctx_27_perm_0 = const()[name = string("ctx_27_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3813 = const()[name = string("op_3813"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_27_cast_fp16 = transpose(perm = ctx_27_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 1024, 1, 188]> input_363_cast_fp16 = reshape(shape = var_3813, x = ctx_27_cast_fp16)[name = string("input_363_cast_fp16")];
            string var_3820_pad_type_0 = const()[name = string("op_3820_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3820_strides_0 = const()[name = string("op_3820_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3820_pad_0 = const()[name = string("op_3820_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3820_dilations_0 = const()[name = string("op_3820_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3820_groups_0 = const()[name = string("op_3820_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176407040))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176931392))))[name = string("layers_13_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3820_cast_fp16 = conv(dilations = var_3820_dilations_0, groups = var_3820_groups_0, pad = var_3820_pad_0, pad_type = var_3820_pad_type_0, strides = var_3820_strides_0, weight = layers_13_self_attn_linear_out_weight_to_fp16_quantized, x = input_363_cast_fp16)[name = string("op_3820_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_303_cast_fp16 = add(x = x_291_cast_fp16, y = var_3820_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<int32, [1]> var_3836_axes_0 = const()[name = string("op_3836_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_13_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_13_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3836_cast_fp16 = layer_norm(axes = var_3836_axes_0, epsilon = layers_13_norm_conv_eps_scaled_to_fp16, x = x_303_cast_fp16)[name = string("op_3836_cast_fp16")];
            tensor<fp16, [1024]> input_365_gamma_0_to_fp16 = const()[name = string("input_365_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176933504)))];
            tensor<fp16, [1024]> input_365_beta_0_to_fp16 = const()[name = string("input_365_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176935616)))];
            fp16 input_365_epsilon_0_to_fp16 = const()[name = string("input_365_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_365_cast_fp16 = batch_norm(beta = input_365_beta_0_to_fp16, epsilon = input_365_epsilon_0_to_fp16, gamma = input_365_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3836_cast_fp16)[name = string("input_365_cast_fp16")];
            string input_367_pad_type_0 = const()[name = string("input_367_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_367_strides_0 = const()[name = string("input_367_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_367_pad_0 = const()[name = string("input_367_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_367_dilations_0 = const()[name = string("input_367_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_367_groups_0 = const()[name = string("input_367_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176937728))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177986368))))[name = string("layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_367_cast_fp16 = conv(dilations = input_367_dilations_0, groups = input_367_groups_0, pad = input_367_pad_0, pad_type = input_367_pad_type_0, strides = input_367_strides_0, weight = layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_365_cast_fp16)[name = string("input_367_cast_fp16")];
            int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)];
            int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_305_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_367_cast_fp16)[name = string("x_305_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_369_cast_fp16 = mul(x = x_305_cast_fp16, y = pad_mask)[name = string("input_369_cast_fp16")];
            string input_371_pad_type_0 = const()[name = string("input_371_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_371_pad_0 = const()[name = string("input_371_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_371_groups_0 = const()[name = string("input_371_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_371_strides_0 = const()[name = string("input_371_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_371_dilations_0 = const()[name = string("input_371_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_129_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177990528))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177995200))))[name = string("const_129_to_fp16_quantized")];
            tensor<fp16, [1024]> const_130_to_fp16 = const()[name = string("const_130_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177997312)))];
            tensor<fp16, [1, 1024, 1, 188]> input_373_cast_fp16 = conv(bias = const_130_to_fp16, dilations = input_371_dilations_0, groups = input_371_groups_0, pad = input_371_pad_0, pad_type = input_371_pad_type_0, strides = input_371_strides_0, weight = const_129_to_fp16_quantized, x = input_369_cast_fp16)[name = string("input_373_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_375_cast_fp16 = silu(x = input_373_cast_fp16)[name = string("input_375_cast_fp16")];
            string var_3868_pad_type_0 = const()[name = string("op_3868_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3868_strides_0 = const()[name = string("op_3868_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3868_pad_0 = const()[name = string("op_3868_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3868_dilations_0 = const()[name = string("op_3868_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3868_groups_0 = const()[name = string("op_3868_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177999424))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178523776))))[name = string("layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3868_cast_fp16 = conv(dilations = var_3868_dilations_0, groups = var_3868_groups_0, pad = var_3868_pad_0, pad_type = var_3868_pad_type_0, strides = var_3868_strides_0, weight = layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_375_cast_fp16)[name = string("op_3868_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_307_cast_fp16 = add(x = x_303_cast_fp16, y = var_3868_cast_fp16)[name = string("x_307_cast_fp16")];
            tensor<int32, [1]> var_3884_axes_0 = const()[name = string("op_3884_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_13_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_13_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3884_cast_fp16 = layer_norm(axes = var_3884_axes_0, epsilon = layers_13_norm_feed_forward2_eps_scaled_to_fp16, x = x_307_cast_fp16)[name = string("op_3884_cast_fp16")];
            tensor<fp16, [1024]> input_377_gamma_0_to_fp16 = const()[name = string("input_377_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178525888)))];
            tensor<fp16, [1024]> input_377_beta_0_to_fp16 = const()[name = string("input_377_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178528000)))];
            fp16 input_377_epsilon_0_to_fp16 = const()[name = string("input_377_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_377_cast_fp16 = batch_norm(beta = input_377_beta_0_to_fp16, epsilon = input_377_epsilon_0_to_fp16, gamma = input_377_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3884_cast_fp16)[name = string("input_377_cast_fp16")];
            string input_379_pad_type_0 = const()[name = string("input_379_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_379_strides_0 = const()[name = string("input_379_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_379_pad_0 = const()[name = string("input_379_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_379_dilations_0 = const()[name = string("input_379_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_379_groups_0 = const()[name = string("input_379_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_13_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178530112))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180627328))))[name = string("layers_13_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_379_cast_fp16 = conv(dilations = input_379_dilations_0, groups = input_379_groups_0, pad = input_379_pad_0, pad_type = input_379_pad_type_0, strides = input_379_strides_0, weight = layers_13_feed_forward2_linear1_weight_to_fp16_quantized, x = input_377_cast_fp16)[name = string("input_379_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_381_cast_fp16 = silu(x = input_379_cast_fp16)[name = string("input_381_cast_fp16")];
            string var_3901_pad_type_0 = const()[name = string("op_3901_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3901_strides_0 = const()[name = string("op_3901_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3901_pad_0 = const()[name = string("op_3901_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3901_dilations_0 = const()[name = string("op_3901_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3901_groups_0 = const()[name = string("op_3901_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_3902_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180635584))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182732800))))[name = string("op_3902_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3902_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_3901_dilations_0, groups = var_3901_groups_0, pad = var_3901_pad_0, pad_type = var_3901_pad_type_0, strides = var_3901_strides_0, weight = op_3902_weight_0_to_fp16_quantized, x = input_381_cast_fp16)[name = string("op_3902_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_309_cast_fp16 = add(x = x_307_cast_fp16, y = var_3902_cast_fp16)[name = string("x_309_cast_fp16")];
            tensor<int32, [1]> var_3918_axes_0 = const()[name = string("op_3918_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_13_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_13_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3918_cast_fp16 = layer_norm(axes = var_3918_axes_0, epsilon = layers_13_norm_out_eps_scaled_to_fp16, x = x_309_cast_fp16)[name = string("op_3918_cast_fp16")];
            tensor<fp16, [1024]> x_311_gamma_0_to_fp16 = const()[name = string("x_311_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182734912)))];
            tensor<fp16, [1024]> x_311_beta_0_to_fp16 = const()[name = string("x_311_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182737024)))];
            fp16 x_311_epsilon_0_to_fp16 = const()[name = string("x_311_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_311_cast_fp16 = batch_norm(beta = x_311_beta_0_to_fp16, epsilon = x_311_epsilon_0_to_fp16, gamma = x_311_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3918_cast_fp16)[name = string("x_311_cast_fp16")];
            int32 var_3937 = const()[name = string("op_3937"), val = int32(1)];
            tensor<int32, [1]> var_3964_axes_0 = const()[name = string("op_3964_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_14_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_14_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3964_cast_fp16 = layer_norm(axes = var_3964_axes_0, epsilon = layers_14_norm_feed_forward1_eps_scaled_to_fp16, x = x_311_cast_fp16)[name = string("op_3964_cast_fp16")];
            tensor<fp16, [1024]> input_383_gamma_0_to_fp16 = const()[name = string("input_383_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182739136)))];
            tensor<fp16, [1024]> input_383_beta_0_to_fp16 = const()[name = string("input_383_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182741248)))];
            fp16 input_383_epsilon_0_to_fp16 = const()[name = string("input_383_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_383_cast_fp16 = batch_norm(beta = input_383_beta_0_to_fp16, epsilon = input_383_epsilon_0_to_fp16, gamma = input_383_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3964_cast_fp16)[name = string("input_383_cast_fp16")];
            string input_385_pad_type_0 = const()[name = string("input_385_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_385_strides_0 = const()[name = string("input_385_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_385_pad_0 = const()[name = string("input_385_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_385_dilations_0 = const()[name = string("input_385_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_385_groups_0 = const()[name = string("input_385_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_14_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182743360))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184840576))))[name = string("layers_14_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_385_cast_fp16 = conv(dilations = input_385_dilations_0, groups = input_385_groups_0, pad = input_385_pad_0, pad_type = input_385_pad_type_0, strides = input_385_strides_0, weight = layers_14_feed_forward1_linear1_weight_to_fp16_quantized, x = input_383_cast_fp16)[name = string("input_385_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_387_cast_fp16 = silu(x = input_385_cast_fp16)[name = string("input_387_cast_fp16")];
            string var_3981_pad_type_0 = const()[name = string("op_3981_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3981_strides_0 = const()[name = string("op_3981_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3981_pad_0 = const()[name = string("op_3981_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3981_dilations_0 = const()[name = string("op_3981_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3981_groups_0 = const()[name = string("op_3981_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_3982_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184848832))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186946048))))[name = string("op_3982_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_3982_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_3981_dilations_0, groups = var_3981_groups_0, pad = var_3981_pad_0, pad_type = var_3981_pad_type_0, strides = var_3981_strides_0, weight = op_3982_weight_0_to_fp16_quantized, x = input_387_cast_fp16)[name = string("op_3982_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_313_cast_fp16 = add(x = x_311_cast_fp16, y = var_3982_cast_fp16)[name = string("x_313_cast_fp16")];
            tensor<int32, [1]> var_3998_axes_0 = const()[name = string("op_3998_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_14_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_14_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_3998_cast_fp16 = layer_norm(axes = var_3998_axes_0, epsilon = layers_14_norm_self_att_eps_scaled_to_fp16, x = x_313_cast_fp16)[name = string("op_3998_cast_fp16")];
            tensor<fp16, [1024]> x_315_gamma_0_to_fp16 = const()[name = string("x_315_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186948160)))];
            tensor<fp16, [1024]> x_315_beta_0_to_fp16 = const()[name = string("x_315_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186950272)))];
            fp16 x_315_epsilon_0_to_fp16 = const()[name = string("x_315_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_315_cast_fp16 = batch_norm(beta = x_315_beta_0_to_fp16, epsilon = x_315_epsilon_0_to_fp16, gamma = x_315_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_3998_cast_fp16)[name = string("x_315_cast_fp16")];
            string q_29_pad_type_0 = const()[name = string("q_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_29_strides_0 = const()[name = string("q_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_29_pad_0 = const()[name = string("q_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_29_dilations_0 = const()[name = string("q_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_29_groups_0 = const()[name = string("q_29_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186952384))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187476736))))[name = string("layers_14_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_29_cast_fp16 = conv(dilations = q_29_dilations_0, groups = q_29_groups_0, pad = q_29_pad_0, pad_type = q_29_pad_type_0, strides = q_29_strides_0, weight = layers_14_self_attn_linear_q_weight_to_fp16_quantized, x = x_315_cast_fp16)[name = string("q_29_cast_fp16")];
            string k_29_pad_type_0 = const()[name = string("k_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_29_strides_0 = const()[name = string("k_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_29_pad_0 = const()[name = string("k_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_29_dilations_0 = const()[name = string("k_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_29_groups_0 = const()[name = string("k_29_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187478848))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188003200))))[name = string("layers_14_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_29_cast_fp16 = conv(dilations = k_29_dilations_0, groups = k_29_groups_0, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = k_29_strides_0, weight = layers_14_self_attn_linear_k_weight_to_fp16_quantized, x = x_315_cast_fp16)[name = string("k_29_cast_fp16")];
            string v_29_pad_type_0 = const()[name = string("v_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_29_strides_0 = const()[name = string("v_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_29_pad_0 = const()[name = string("v_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_29_dilations_0 = const()[name = string("v_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_29_groups_0 = const()[name = string("v_29_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188005312))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188529664))))[name = string("layers_14_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_29_cast_fp16 = conv(dilations = v_29_dilations_0, groups = v_29_groups_0, pad = v_29_pad_0, pad_type = v_29_pad_type_0, strides = v_29_strides_0, weight = layers_14_self_attn_linear_v_weight_to_fp16_quantized, x = x_315_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_29_to_fp16 = const()[name = string("bv_all_29_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188531776)))];
            tensor<fp16, [1, 1024, 1, 188]> var_4031_cast_fp16 = add(x = q_29_cast_fp16, y = bv_all_29_to_fp16)[name = string("op_4031_cast_fp16")];
            tensor<int32, [3]> var_4032 = const()[name = string("op_4032"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_29_cast_fp16 = reshape(shape = var_4032, x = var_4031_cast_fp16)[name = string("qb_29_cast_fp16")];
            bool bd_all_57_transpose_x_0 = const()[name = string("bd_all_57_transpose_x_0"), val = bool(false)];
            bool bd_all_57_transpose_y_0 = const()[name = string("bd_all_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_14_self_attn_pos_proj_to_fp16 = const()[name = string("layers_14_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188533888)))];
            tensor<fp16, [8, 375, 188]> bd_all_57_cast_fp16 = matmul(transpose_x = bd_all_57_transpose_x_0, transpose_y = bd_all_57_transpose_y_0, x = layers_14_self_attn_pos_proj_to_fp16, y = qb_29_cast_fp16)[name = string("bd_all_57_cast_fp16")];
            tensor<int32, [3]> x_317_perm_0 = const()[name = string("x_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_319_pad_0 = const()[name = string("x_319_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_319_mode_0 = const()[name = string("x_319_mode_0"), val = string("constant")];
            fp16 const_66_to_fp16 = const()[name = string("const_66_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_317_cast_fp16 = transpose(perm = x_317_perm_0, x = bd_all_57_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [8, 188, 376]> x_319_cast_fp16 = pad(constant_val = const_66_to_fp16, mode = x_319_mode_0, pad = x_319_pad_0, x = x_317_cast_fp16)[name = string("x_319_cast_fp16")];
            tensor<int32, [3]> var_4039 = const()[name = string("op_4039"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_321_cast_fp16 = reshape(shape = var_4039, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [3]> var_4042_begin_0 = const()[name = string("op_4042_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_4042_end_0 = const()[name = string("op_4042_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_4042_end_mask_0 = const()[name = string("op_4042_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_4042_cast_fp16 = slice_by_index(begin = var_4042_begin_0, end = var_4042_end_0, end_mask = var_4042_end_mask_0, x = x_321_cast_fp16)[name = string("op_4042_cast_fp16")];
            tensor<int32, [3]> var_4043 = const()[name = string("op_4043"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_323_cast_fp16 = reshape(shape = var_4043, x = var_4042_cast_fp16)[name = string("x_323_cast_fp16")];
            tensor<int32, [3]> bd_all_59_begin_0 = const()[name = string("bd_all_59_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_59_end_0 = const()[name = string("bd_all_59_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_59_end_mask_0 = const()[name = string("bd_all_59_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_59_cast_fp16 = slice_by_index(begin = bd_all_59_begin_0, end = bd_all_59_end_0, end_mask = bd_all_59_end_mask_0, x = x_323_cast_fp16)[name = string("bd_all_59_cast_fp16")];
            tensor<int32, [4]> var_4048 = const()[name = string("op_4048"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4049_cast_fp16 = reshape(shape = var_4048, x = q_29_cast_fp16)[name = string("op_4049_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_4051_to_fp16 = const()[name = string("op_4051_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189301952)))];
            tensor<fp16, [8, 128, 1, 188]> qh_29_cast_fp16 = add(x = var_4049_cast_fp16, y = var_4051_to_fp16)[name = string("qh_29_cast_fp16")];
            tensor<int32, [4]> var_4053 = const()[name = string("op_4053"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4054_cast_fp16 = reshape(shape = var_4053, x = k_29_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [4]> var_4055 = const()[name = string("op_4055"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_4057 = const()[name = string("op_4057"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4058_cast_fp16 = reshape(shape = var_4057, x = v_29_cast_fp16)[name = string("op_4058_cast_fp16")];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4061_perm_0 = const()[name = string("op_4061_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_29_axes_0 = const()[name = string("bd_29_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_4061_cast_fp16 = transpose(perm = var_4061_perm_0, x = bd_all_59_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [8, 188, 1, 188]> bd_29_cast_fp16 = expand_dims(axes = bd_29_axes_0, x = var_4061_cast_fp16)[name = string("bd_29_cast_fp16")];
            string ac_29_equation_0 = const()[name = string("ac_29_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_4056_cast_fp16 = transpose(perm = var_4055, x = var_4054_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [8, 188, 1, 188]> ac_29_cast_fp16 = einsum(equation = ac_29_equation_0, values = (var_4056_cast_fp16, qh_29_cast_fp16))[name = string("ac_29_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_4077_cast_fp16 = add(x = ac_29_cast_fp16, y = bd_29_cast_fp16)[name = string("op_4077_cast_fp16")];
            fp16 var_4078_to_fp16 = const()[name = string("op_4078_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_57_cast_fp16 = mul(x = var_4077_cast_fp16, y = var_4078_to_fp16)[name = string("scores_57_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = key_bias)[name = string("scores_59_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_4081_cast_fp16 = softmax(axis = var_3937, x = scores_59_cast_fp16)[name = string("op_4081_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_144 = const()[name = string("concat_144"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = var_4081_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [8, 188, 188]> reshape_42_cast_fp16 = reshape(shape = concat_144, x = transpose_28_cast_fp16)[name = string("reshape_42_cast_fp16")];
            tensor<int32, [3]> concat_145 = const()[name = string("concat_145"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = var_4058_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [8, 188, 128]> reshape_43_cast_fp16 = reshape(shape = concat_145, x = transpose_62_cast_fp16)[name = string("reshape_43_cast_fp16")];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = reshape_42_cast_fp16, y = reshape_43_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<int32, [4]> concat_149 = const()[name = string("concat_149"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_44_cast_fp16 = reshape(shape = concat_149, x = matmul_14_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [4]> ctx_29_perm_0 = const()[name = string("ctx_29_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4084 = const()[name = string("op_4084"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_29_cast_fp16 = transpose(perm = ctx_29_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 1024, 1, 188]> input_389_cast_fp16 = reshape(shape = var_4084, x = ctx_29_cast_fp16)[name = string("input_389_cast_fp16")];
            string var_4091_pad_type_0 = const()[name = string("op_4091_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4091_strides_0 = const()[name = string("op_4091_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4091_pad_0 = const()[name = string("op_4091_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4091_dilations_0 = const()[name = string("op_4091_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4091_groups_0 = const()[name = string("op_4091_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189304064))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189828416))))[name = string("layers_14_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4091_cast_fp16 = conv(dilations = var_4091_dilations_0, groups = var_4091_groups_0, pad = var_4091_pad_0, pad_type = var_4091_pad_type_0, strides = var_4091_strides_0, weight = layers_14_self_attn_linear_out_weight_to_fp16_quantized, x = input_389_cast_fp16)[name = string("op_4091_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_325_cast_fp16 = add(x = x_313_cast_fp16, y = var_4091_cast_fp16)[name = string("x_325_cast_fp16")];
            tensor<int32, [1]> var_4107_axes_0 = const()[name = string("op_4107_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_14_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_14_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4107_cast_fp16 = layer_norm(axes = var_4107_axes_0, epsilon = layers_14_norm_conv_eps_scaled_to_fp16, x = x_325_cast_fp16)[name = string("op_4107_cast_fp16")];
            tensor<fp16, [1024]> input_391_gamma_0_to_fp16 = const()[name = string("input_391_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189830528)))];
            tensor<fp16, [1024]> input_391_beta_0_to_fp16 = const()[name = string("input_391_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189832640)))];
            fp16 input_391_epsilon_0_to_fp16 = const()[name = string("input_391_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_391_cast_fp16 = batch_norm(beta = input_391_beta_0_to_fp16, epsilon = input_391_epsilon_0_to_fp16, gamma = input_391_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4107_cast_fp16)[name = string("input_391_cast_fp16")];
            string input_393_pad_type_0 = const()[name = string("input_393_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_393_strides_0 = const()[name = string("input_393_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_393_pad_0 = const()[name = string("input_393_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_393_dilations_0 = const()[name = string("input_393_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_393_groups_0 = const()[name = string("input_393_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189834752))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190883392))))[name = string("layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_393_cast_fp16 = conv(dilations = input_393_dilations_0, groups = input_393_groups_0, pad = input_393_pad_0, pad_type = input_393_pad_type_0, strides = input_393_strides_0, weight = layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_391_cast_fp16)[name = string("input_393_cast_fp16")];
            int32 x_327_split_num_splits_0 = const()[name = string("x_327_split_num_splits_0"), val = int32(2)];
            int32 x_327_split_axis_0 = const()[name = string("x_327_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_327_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_327_split_cast_fp16_1 = split(axis = x_327_split_axis_0, num_splits = x_327_split_num_splits_0, x = input_393_cast_fp16)[name = string("x_327_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_327_split_1_sigmoid_cast_fp16 = sigmoid(x = x_327_split_cast_fp16_1)[name = string("x_327_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_327_cast_fp16 = mul(x = x_327_split_cast_fp16_0, y = x_327_split_1_sigmoid_cast_fp16)[name = string("x_327_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_395_cast_fp16 = mul(x = x_327_cast_fp16, y = pad_mask)[name = string("input_395_cast_fp16")];
            string input_397_pad_type_0 = const()[name = string("input_397_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_397_pad_0 = const()[name = string("input_397_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_397_groups_0 = const()[name = string("input_397_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_397_strides_0 = const()[name = string("input_397_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_397_dilations_0 = const()[name = string("input_397_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_131_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190887552))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190892224))))[name = string("const_131_to_fp16_quantized")];
            tensor<fp16, [1024]> const_132_to_fp16 = const()[name = string("const_132_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190894336)))];
            tensor<fp16, [1, 1024, 1, 188]> input_399_cast_fp16 = conv(bias = const_132_to_fp16, dilations = input_397_dilations_0, groups = input_397_groups_0, pad = input_397_pad_0, pad_type = input_397_pad_type_0, strides = input_397_strides_0, weight = const_131_to_fp16_quantized, x = input_395_cast_fp16)[name = string("input_399_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_401_cast_fp16 = silu(x = input_399_cast_fp16)[name = string("input_401_cast_fp16")];
            string var_4139_pad_type_0 = const()[name = string("op_4139_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4139_strides_0 = const()[name = string("op_4139_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4139_pad_0 = const()[name = string("op_4139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4139_dilations_0 = const()[name = string("op_4139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4139_groups_0 = const()[name = string("op_4139_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190896448))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191420800))))[name = string("layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4139_cast_fp16 = conv(dilations = var_4139_dilations_0, groups = var_4139_groups_0, pad = var_4139_pad_0, pad_type = var_4139_pad_type_0, strides = var_4139_strides_0, weight = layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_401_cast_fp16)[name = string("op_4139_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_329_cast_fp16 = add(x = x_325_cast_fp16, y = var_4139_cast_fp16)[name = string("x_329_cast_fp16")];
            tensor<int32, [1]> var_4155_axes_0 = const()[name = string("op_4155_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_14_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_14_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4155_cast_fp16 = layer_norm(axes = var_4155_axes_0, epsilon = layers_14_norm_feed_forward2_eps_scaled_to_fp16, x = x_329_cast_fp16)[name = string("op_4155_cast_fp16")];
            tensor<fp16, [1024]> input_403_gamma_0_to_fp16 = const()[name = string("input_403_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191422912)))];
            tensor<fp16, [1024]> input_403_beta_0_to_fp16 = const()[name = string("input_403_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191425024)))];
            fp16 input_403_epsilon_0_to_fp16 = const()[name = string("input_403_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_403_cast_fp16 = batch_norm(beta = input_403_beta_0_to_fp16, epsilon = input_403_epsilon_0_to_fp16, gamma = input_403_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4155_cast_fp16)[name = string("input_403_cast_fp16")];
            string input_405_pad_type_0 = const()[name = string("input_405_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_405_strides_0 = const()[name = string("input_405_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_405_pad_0 = const()[name = string("input_405_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_405_dilations_0 = const()[name = string("input_405_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_405_groups_0 = const()[name = string("input_405_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_14_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191427136))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193524352))))[name = string("layers_14_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_405_cast_fp16 = conv(dilations = input_405_dilations_0, groups = input_405_groups_0, pad = input_405_pad_0, pad_type = input_405_pad_type_0, strides = input_405_strides_0, weight = layers_14_feed_forward2_linear1_weight_to_fp16_quantized, x = input_403_cast_fp16)[name = string("input_405_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_407_cast_fp16 = silu(x = input_405_cast_fp16)[name = string("input_407_cast_fp16")];
            string var_4172_pad_type_0 = const()[name = string("op_4172_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4172_strides_0 = const()[name = string("op_4172_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4172_pad_0 = const()[name = string("op_4172_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4172_dilations_0 = const()[name = string("op_4172_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4172_groups_0 = const()[name = string("op_4172_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_4173_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193532608))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195629824))))[name = string("op_4173_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4173_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_4172_dilations_0, groups = var_4172_groups_0, pad = var_4172_pad_0, pad_type = var_4172_pad_type_0, strides = var_4172_strides_0, weight = op_4173_weight_0_to_fp16_quantized, x = input_407_cast_fp16)[name = string("op_4173_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_331_cast_fp16 = add(x = x_329_cast_fp16, y = var_4173_cast_fp16)[name = string("x_331_cast_fp16")];
            tensor<int32, [1]> var_4189_axes_0 = const()[name = string("op_4189_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_14_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_14_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4189_cast_fp16 = layer_norm(axes = var_4189_axes_0, epsilon = layers_14_norm_out_eps_scaled_to_fp16, x = x_331_cast_fp16)[name = string("op_4189_cast_fp16")];
            tensor<fp16, [1024]> x_333_gamma_0_to_fp16 = const()[name = string("x_333_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195631936)))];
            tensor<fp16, [1024]> x_333_beta_0_to_fp16 = const()[name = string("x_333_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195634048)))];
            fp16 x_333_epsilon_0_to_fp16 = const()[name = string("x_333_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_333_cast_fp16 = batch_norm(beta = x_333_beta_0_to_fp16, epsilon = x_333_epsilon_0_to_fp16, gamma = x_333_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4189_cast_fp16)[name = string("x_333_cast_fp16")];
            int32 var_4208 = const()[name = string("op_4208"), val = int32(1)];
            tensor<int32, [1]> var_4235_axes_0 = const()[name = string("op_4235_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_15_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_15_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4235_cast_fp16 = layer_norm(axes = var_4235_axes_0, epsilon = layers_15_norm_feed_forward1_eps_scaled_to_fp16, x = x_333_cast_fp16)[name = string("op_4235_cast_fp16")];
            tensor<fp16, [1024]> input_409_gamma_0_to_fp16 = const()[name = string("input_409_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195636160)))];
            tensor<fp16, [1024]> input_409_beta_0_to_fp16 = const()[name = string("input_409_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195638272)))];
            fp16 input_409_epsilon_0_to_fp16 = const()[name = string("input_409_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_409_cast_fp16 = batch_norm(beta = input_409_beta_0_to_fp16, epsilon = input_409_epsilon_0_to_fp16, gamma = input_409_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4235_cast_fp16)[name = string("input_409_cast_fp16")];
            string input_411_pad_type_0 = const()[name = string("input_411_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_411_strides_0 = const()[name = string("input_411_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_411_pad_0 = const()[name = string("input_411_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_411_dilations_0 = const()[name = string("input_411_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_411_groups_0 = const()[name = string("input_411_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_15_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195640384))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197737600))))[name = string("layers_15_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_411_cast_fp16 = conv(dilations = input_411_dilations_0, groups = input_411_groups_0, pad = input_411_pad_0, pad_type = input_411_pad_type_0, strides = input_411_strides_0, weight = layers_15_feed_forward1_linear1_weight_to_fp16_quantized, x = input_409_cast_fp16)[name = string("input_411_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_413_cast_fp16 = silu(x = input_411_cast_fp16)[name = string("input_413_cast_fp16")];
            string var_4252_pad_type_0 = const()[name = string("op_4252_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4252_strides_0 = const()[name = string("op_4252_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4252_pad_0 = const()[name = string("op_4252_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4252_dilations_0 = const()[name = string("op_4252_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4252_groups_0 = const()[name = string("op_4252_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_4253_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197745856))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199843072))))[name = string("op_4253_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4253_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_4252_dilations_0, groups = var_4252_groups_0, pad = var_4252_pad_0, pad_type = var_4252_pad_type_0, strides = var_4252_strides_0, weight = op_4253_weight_0_to_fp16_quantized, x = input_413_cast_fp16)[name = string("op_4253_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_335_cast_fp16 = add(x = x_333_cast_fp16, y = var_4253_cast_fp16)[name = string("x_335_cast_fp16")];
            tensor<int32, [1]> var_4269_axes_0 = const()[name = string("op_4269_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_15_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_15_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4269_cast_fp16 = layer_norm(axes = var_4269_axes_0, epsilon = layers_15_norm_self_att_eps_scaled_to_fp16, x = x_335_cast_fp16)[name = string("op_4269_cast_fp16")];
            tensor<fp16, [1024]> x_337_gamma_0_to_fp16 = const()[name = string("x_337_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199845184)))];
            tensor<fp16, [1024]> x_337_beta_0_to_fp16 = const()[name = string("x_337_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199847296)))];
            fp16 x_337_epsilon_0_to_fp16 = const()[name = string("x_337_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_337_cast_fp16 = batch_norm(beta = x_337_beta_0_to_fp16, epsilon = x_337_epsilon_0_to_fp16, gamma = x_337_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4269_cast_fp16)[name = string("x_337_cast_fp16")];
            string q_31_pad_type_0 = const()[name = string("q_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_31_strides_0 = const()[name = string("q_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_31_pad_0 = const()[name = string("q_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_31_dilations_0 = const()[name = string("q_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_31_groups_0 = const()[name = string("q_31_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199849408))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200373760))))[name = string("layers_15_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_31_cast_fp16 = conv(dilations = q_31_dilations_0, groups = q_31_groups_0, pad = q_31_pad_0, pad_type = q_31_pad_type_0, strides = q_31_strides_0, weight = layers_15_self_attn_linear_q_weight_to_fp16_quantized, x = x_337_cast_fp16)[name = string("q_31_cast_fp16")];
            string k_31_pad_type_0 = const()[name = string("k_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_31_strides_0 = const()[name = string("k_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_31_pad_0 = const()[name = string("k_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_31_dilations_0 = const()[name = string("k_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_31_groups_0 = const()[name = string("k_31_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200375872))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200900224))))[name = string("layers_15_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_31_cast_fp16 = conv(dilations = k_31_dilations_0, groups = k_31_groups_0, pad = k_31_pad_0, pad_type = k_31_pad_type_0, strides = k_31_strides_0, weight = layers_15_self_attn_linear_k_weight_to_fp16_quantized, x = x_337_cast_fp16)[name = string("k_31_cast_fp16")];
            string v_31_pad_type_0 = const()[name = string("v_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_31_strides_0 = const()[name = string("v_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_31_pad_0 = const()[name = string("v_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_31_dilations_0 = const()[name = string("v_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_31_groups_0 = const()[name = string("v_31_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200902336))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201426688))))[name = string("layers_15_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_31_cast_fp16 = conv(dilations = v_31_dilations_0, groups = v_31_groups_0, pad = v_31_pad_0, pad_type = v_31_pad_type_0, strides = v_31_strides_0, weight = layers_15_self_attn_linear_v_weight_to_fp16_quantized, x = x_337_cast_fp16)[name = string("v_31_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_31_to_fp16 = const()[name = string("bv_all_31_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201428800)))];
            tensor<fp16, [1, 1024, 1, 188]> var_4302_cast_fp16 = add(x = q_31_cast_fp16, y = bv_all_31_to_fp16)[name = string("op_4302_cast_fp16")];
            tensor<int32, [3]> var_4303 = const()[name = string("op_4303"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_31_cast_fp16 = reshape(shape = var_4303, x = var_4302_cast_fp16)[name = string("qb_31_cast_fp16")];
            bool bd_all_61_transpose_x_0 = const()[name = string("bd_all_61_transpose_x_0"), val = bool(false)];
            bool bd_all_61_transpose_y_0 = const()[name = string("bd_all_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_15_self_attn_pos_proj_to_fp16 = const()[name = string("layers_15_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201430912)))];
            tensor<fp16, [8, 375, 188]> bd_all_61_cast_fp16 = matmul(transpose_x = bd_all_61_transpose_x_0, transpose_y = bd_all_61_transpose_y_0, x = layers_15_self_attn_pos_proj_to_fp16, y = qb_31_cast_fp16)[name = string("bd_all_61_cast_fp16")];
            tensor<int32, [3]> x_339_perm_0 = const()[name = string("x_339_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_341_pad_0 = const()[name = string("x_341_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_341_mode_0 = const()[name = string("x_341_mode_0"), val = string("constant")];
            fp16 const_70_to_fp16 = const()[name = string("const_70_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_339_cast_fp16 = transpose(perm = x_339_perm_0, x = bd_all_61_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [8, 188, 376]> x_341_cast_fp16 = pad(constant_val = const_70_to_fp16, mode = x_341_mode_0, pad = x_341_pad_0, x = x_339_cast_fp16)[name = string("x_341_cast_fp16")];
            tensor<int32, [3]> var_4310 = const()[name = string("op_4310"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_343_cast_fp16 = reshape(shape = var_4310, x = x_341_cast_fp16)[name = string("x_343_cast_fp16")];
            tensor<int32, [3]> var_4313_begin_0 = const()[name = string("op_4313_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_4313_end_0 = const()[name = string("op_4313_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_4313_end_mask_0 = const()[name = string("op_4313_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_4313_cast_fp16 = slice_by_index(begin = var_4313_begin_0, end = var_4313_end_0, end_mask = var_4313_end_mask_0, x = x_343_cast_fp16)[name = string("op_4313_cast_fp16")];
            tensor<int32, [3]> var_4314 = const()[name = string("op_4314"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_345_cast_fp16 = reshape(shape = var_4314, x = var_4313_cast_fp16)[name = string("x_345_cast_fp16")];
            tensor<int32, [3]> bd_all_63_begin_0 = const()[name = string("bd_all_63_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_63_end_0 = const()[name = string("bd_all_63_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_63_end_mask_0 = const()[name = string("bd_all_63_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_63_cast_fp16 = slice_by_index(begin = bd_all_63_begin_0, end = bd_all_63_end_0, end_mask = bd_all_63_end_mask_0, x = x_345_cast_fp16)[name = string("bd_all_63_cast_fp16")];
            tensor<int32, [4]> var_4319 = const()[name = string("op_4319"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4320_cast_fp16 = reshape(shape = var_4319, x = q_31_cast_fp16)[name = string("op_4320_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_4322_to_fp16 = const()[name = string("op_4322_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202198976)))];
            tensor<fp16, [8, 128, 1, 188]> qh_31_cast_fp16 = add(x = var_4320_cast_fp16, y = var_4322_to_fp16)[name = string("qh_31_cast_fp16")];
            tensor<int32, [4]> var_4324 = const()[name = string("op_4324"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4325_cast_fp16 = reshape(shape = var_4324, x = k_31_cast_fp16)[name = string("op_4325_cast_fp16")];
            tensor<int32, [4]> var_4326 = const()[name = string("op_4326"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_4328 = const()[name = string("op_4328"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4329_cast_fp16 = reshape(shape = var_4328, x = v_31_cast_fp16)[name = string("op_4329_cast_fp16")];
            tensor<int32, [4]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4332_perm_0 = const()[name = string("op_4332_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_31_axes_0 = const()[name = string("bd_31_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_4332_cast_fp16 = transpose(perm = var_4332_perm_0, x = bd_all_63_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [8, 188, 1, 188]> bd_31_cast_fp16 = expand_dims(axes = bd_31_axes_0, x = var_4332_cast_fp16)[name = string("bd_31_cast_fp16")];
            string ac_31_equation_0 = const()[name = string("ac_31_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_4327_cast_fp16 = transpose(perm = var_4326, x = var_4325_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [8, 188, 1, 188]> ac_31_cast_fp16 = einsum(equation = ac_31_equation_0, values = (var_4327_cast_fp16, qh_31_cast_fp16))[name = string("ac_31_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_4348_cast_fp16 = add(x = ac_31_cast_fp16, y = bd_31_cast_fp16)[name = string("op_4348_cast_fp16")];
            fp16 var_4349_to_fp16 = const()[name = string("op_4349_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_61_cast_fp16 = mul(x = var_4348_cast_fp16, y = var_4349_to_fp16)[name = string("scores_61_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = key_bias)[name = string("scores_63_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_4352_cast_fp16 = softmax(axis = var_4208, x = scores_63_cast_fp16)[name = string("op_4352_cast_fp16")];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_154 = const()[name = string("concat_154"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_30_cast_fp16 = transpose(perm = transpose_30_perm_0, x = var_4352_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [8, 188, 188]> reshape_45_cast_fp16 = reshape(shape = concat_154, x = transpose_30_cast_fp16)[name = string("reshape_45_cast_fp16")];
            tensor<int32, [3]> concat_155 = const()[name = string("concat_155"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = var_4329_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [8, 188, 128]> reshape_46_cast_fp16 = reshape(shape = concat_155, x = transpose_63_cast_fp16)[name = string("reshape_46_cast_fp16")];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = reshape_45_cast_fp16, y = reshape_46_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<int32, [4]> concat_159 = const()[name = string("concat_159"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_47_cast_fp16 = reshape(shape = concat_159, x = matmul_15_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> ctx_31_perm_0 = const()[name = string("ctx_31_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4355 = const()[name = string("op_4355"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_31_cast_fp16 = transpose(perm = ctx_31_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 1024, 1, 188]> input_415_cast_fp16 = reshape(shape = var_4355, x = ctx_31_cast_fp16)[name = string("input_415_cast_fp16")];
            string var_4362_pad_type_0 = const()[name = string("op_4362_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4362_strides_0 = const()[name = string("op_4362_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4362_pad_0 = const()[name = string("op_4362_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4362_dilations_0 = const()[name = string("op_4362_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4362_groups_0 = const()[name = string("op_4362_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202201088))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202725440))))[name = string("layers_15_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4362_cast_fp16 = conv(dilations = var_4362_dilations_0, groups = var_4362_groups_0, pad = var_4362_pad_0, pad_type = var_4362_pad_type_0, strides = var_4362_strides_0, weight = layers_15_self_attn_linear_out_weight_to_fp16_quantized, x = input_415_cast_fp16)[name = string("op_4362_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_347_cast_fp16 = add(x = x_335_cast_fp16, y = var_4362_cast_fp16)[name = string("x_347_cast_fp16")];
            tensor<int32, [1]> var_4378_axes_0 = const()[name = string("op_4378_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_15_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_15_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4378_cast_fp16 = layer_norm(axes = var_4378_axes_0, epsilon = layers_15_norm_conv_eps_scaled_to_fp16, x = x_347_cast_fp16)[name = string("op_4378_cast_fp16")];
            tensor<fp16, [1024]> input_417_gamma_0_to_fp16 = const()[name = string("input_417_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202727552)))];
            tensor<fp16, [1024]> input_417_beta_0_to_fp16 = const()[name = string("input_417_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202729664)))];
            fp16 input_417_epsilon_0_to_fp16 = const()[name = string("input_417_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_417_cast_fp16 = batch_norm(beta = input_417_beta_0_to_fp16, epsilon = input_417_epsilon_0_to_fp16, gamma = input_417_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4378_cast_fp16)[name = string("input_417_cast_fp16")];
            string input_419_pad_type_0 = const()[name = string("input_419_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_419_strides_0 = const()[name = string("input_419_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_419_pad_0 = const()[name = string("input_419_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_419_dilations_0 = const()[name = string("input_419_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_419_groups_0 = const()[name = string("input_419_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202731776))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203780416))))[name = string("layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_419_cast_fp16 = conv(dilations = input_419_dilations_0, groups = input_419_groups_0, pad = input_419_pad_0, pad_type = input_419_pad_type_0, strides = input_419_strides_0, weight = layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_417_cast_fp16)[name = string("input_419_cast_fp16")];
            int32 x_349_split_num_splits_0 = const()[name = string("x_349_split_num_splits_0"), val = int32(2)];
            int32 x_349_split_axis_0 = const()[name = string("x_349_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_349_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_349_split_cast_fp16_1 = split(axis = x_349_split_axis_0, num_splits = x_349_split_num_splits_0, x = input_419_cast_fp16)[name = string("x_349_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_349_split_1_sigmoid_cast_fp16 = sigmoid(x = x_349_split_cast_fp16_1)[name = string("x_349_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_349_cast_fp16 = mul(x = x_349_split_cast_fp16_0, y = x_349_split_1_sigmoid_cast_fp16)[name = string("x_349_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_421_cast_fp16 = mul(x = x_349_cast_fp16, y = pad_mask)[name = string("input_421_cast_fp16")];
            string input_423_pad_type_0 = const()[name = string("input_423_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_423_pad_0 = const()[name = string("input_423_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_423_groups_0 = const()[name = string("input_423_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_423_strides_0 = const()[name = string("input_423_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_423_dilations_0 = const()[name = string("input_423_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_133_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203784576))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203789248))))[name = string("const_133_to_fp16_quantized")];
            tensor<fp16, [1024]> const_134_to_fp16 = const()[name = string("const_134_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203791360)))];
            tensor<fp16, [1, 1024, 1, 188]> input_425_cast_fp16 = conv(bias = const_134_to_fp16, dilations = input_423_dilations_0, groups = input_423_groups_0, pad = input_423_pad_0, pad_type = input_423_pad_type_0, strides = input_423_strides_0, weight = const_133_to_fp16_quantized, x = input_421_cast_fp16)[name = string("input_425_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_427_cast_fp16 = silu(x = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
            string var_4410_pad_type_0 = const()[name = string("op_4410_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4410_strides_0 = const()[name = string("op_4410_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4410_pad_0 = const()[name = string("op_4410_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4410_dilations_0 = const()[name = string("op_4410_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4410_groups_0 = const()[name = string("op_4410_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203793472))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204317824))))[name = string("layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4410_cast_fp16 = conv(dilations = var_4410_dilations_0, groups = var_4410_groups_0, pad = var_4410_pad_0, pad_type = var_4410_pad_type_0, strides = var_4410_strides_0, weight = layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_427_cast_fp16)[name = string("op_4410_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_351_cast_fp16 = add(x = x_347_cast_fp16, y = var_4410_cast_fp16)[name = string("x_351_cast_fp16")];
            tensor<int32, [1]> var_4426_axes_0 = const()[name = string("op_4426_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_15_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_15_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4426_cast_fp16 = layer_norm(axes = var_4426_axes_0, epsilon = layers_15_norm_feed_forward2_eps_scaled_to_fp16, x = x_351_cast_fp16)[name = string("op_4426_cast_fp16")];
            tensor<fp16, [1024]> input_429_gamma_0_to_fp16 = const()[name = string("input_429_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204319936)))];
            tensor<fp16, [1024]> input_429_beta_0_to_fp16 = const()[name = string("input_429_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204322048)))];
            fp16 input_429_epsilon_0_to_fp16 = const()[name = string("input_429_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_429_cast_fp16 = batch_norm(beta = input_429_beta_0_to_fp16, epsilon = input_429_epsilon_0_to_fp16, gamma = input_429_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4426_cast_fp16)[name = string("input_429_cast_fp16")];
            string input_431_pad_type_0 = const()[name = string("input_431_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_431_strides_0 = const()[name = string("input_431_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_431_pad_0 = const()[name = string("input_431_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_431_dilations_0 = const()[name = string("input_431_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_431_groups_0 = const()[name = string("input_431_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_15_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204324160))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206421376))))[name = string("layers_15_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_431_cast_fp16 = conv(dilations = input_431_dilations_0, groups = input_431_groups_0, pad = input_431_pad_0, pad_type = input_431_pad_type_0, strides = input_431_strides_0, weight = layers_15_feed_forward2_linear1_weight_to_fp16_quantized, x = input_429_cast_fp16)[name = string("input_431_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_433_cast_fp16 = silu(x = input_431_cast_fp16)[name = string("input_433_cast_fp16")];
            string var_4443_pad_type_0 = const()[name = string("op_4443_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4443_strides_0 = const()[name = string("op_4443_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4443_pad_0 = const()[name = string("op_4443_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4443_dilations_0 = const()[name = string("op_4443_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4443_groups_0 = const()[name = string("op_4443_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_4444_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206429632))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208526848))))[name = string("op_4444_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4444_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_4443_dilations_0, groups = var_4443_groups_0, pad = var_4443_pad_0, pad_type = var_4443_pad_type_0, strides = var_4443_strides_0, weight = op_4444_weight_0_to_fp16_quantized, x = input_433_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_353_cast_fp16 = add(x = x_351_cast_fp16, y = var_4444_cast_fp16)[name = string("x_353_cast_fp16")];
            tensor<int32, [1]> var_4460_axes_0 = const()[name = string("op_4460_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_15_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_15_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4460_cast_fp16 = layer_norm(axes = var_4460_axes_0, epsilon = layers_15_norm_out_eps_scaled_to_fp16, x = x_353_cast_fp16)[name = string("op_4460_cast_fp16")];
            tensor<fp16, [1024]> x_355_gamma_0_to_fp16 = const()[name = string("x_355_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208528960)))];
            tensor<fp16, [1024]> x_355_beta_0_to_fp16 = const()[name = string("x_355_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208531072)))];
            fp16 x_355_epsilon_0_to_fp16 = const()[name = string("x_355_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_355_cast_fp16 = batch_norm(beta = x_355_beta_0_to_fp16, epsilon = x_355_epsilon_0_to_fp16, gamma = x_355_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4460_cast_fp16)[name = string("x_355_cast_fp16")];
            int32 var_4479 = const()[name = string("op_4479"), val = int32(1)];
            tensor<int32, [1]> var_4506_axes_0 = const()[name = string("op_4506_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_16_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_16_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4506_cast_fp16 = layer_norm(axes = var_4506_axes_0, epsilon = layers_16_norm_feed_forward1_eps_scaled_to_fp16, x = x_355_cast_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1024]> input_435_gamma_0_to_fp16 = const()[name = string("input_435_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208533184)))];
            tensor<fp16, [1024]> input_435_beta_0_to_fp16 = const()[name = string("input_435_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208535296)))];
            fp16 input_435_epsilon_0_to_fp16 = const()[name = string("input_435_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_435_cast_fp16 = batch_norm(beta = input_435_beta_0_to_fp16, epsilon = input_435_epsilon_0_to_fp16, gamma = input_435_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4506_cast_fp16)[name = string("input_435_cast_fp16")];
            string input_437_pad_type_0 = const()[name = string("input_437_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_437_strides_0 = const()[name = string("input_437_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_437_pad_0 = const()[name = string("input_437_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_437_dilations_0 = const()[name = string("input_437_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_437_groups_0 = const()[name = string("input_437_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_16_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208537408))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210634624))))[name = string("layers_16_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_437_cast_fp16 = conv(dilations = input_437_dilations_0, groups = input_437_groups_0, pad = input_437_pad_0, pad_type = input_437_pad_type_0, strides = input_437_strides_0, weight = layers_16_feed_forward1_linear1_weight_to_fp16_quantized, x = input_435_cast_fp16)[name = string("input_437_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_439_cast_fp16 = silu(x = input_437_cast_fp16)[name = string("input_439_cast_fp16")];
            string var_4523_pad_type_0 = const()[name = string("op_4523_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4523_strides_0 = const()[name = string("op_4523_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4523_pad_0 = const()[name = string("op_4523_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4523_dilations_0 = const()[name = string("op_4523_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4523_groups_0 = const()[name = string("op_4523_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_4524_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210642880))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212740096))))[name = string("op_4524_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4524_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_4523_dilations_0, groups = var_4523_groups_0, pad = var_4523_pad_0, pad_type = var_4523_pad_type_0, strides = var_4523_strides_0, weight = op_4524_weight_0_to_fp16_quantized, x = input_439_cast_fp16)[name = string("op_4524_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_357_cast_fp16 = add(x = x_355_cast_fp16, y = var_4524_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [1]> var_4540_axes_0 = const()[name = string("op_4540_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_16_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_16_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4540_cast_fp16 = layer_norm(axes = var_4540_axes_0, epsilon = layers_16_norm_self_att_eps_scaled_to_fp16, x = x_357_cast_fp16)[name = string("op_4540_cast_fp16")];
            tensor<fp16, [1024]> x_359_gamma_0_to_fp16 = const()[name = string("x_359_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212742208)))];
            tensor<fp16, [1024]> x_359_beta_0_to_fp16 = const()[name = string("x_359_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212744320)))];
            fp16 x_359_epsilon_0_to_fp16 = const()[name = string("x_359_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_359_cast_fp16 = batch_norm(beta = x_359_beta_0_to_fp16, epsilon = x_359_epsilon_0_to_fp16, gamma = x_359_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4540_cast_fp16)[name = string("x_359_cast_fp16")];
            string q_33_pad_type_0 = const()[name = string("q_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_33_strides_0 = const()[name = string("q_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_33_pad_0 = const()[name = string("q_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_33_dilations_0 = const()[name = string("q_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_33_groups_0 = const()[name = string("q_33_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212746432))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213270784))))[name = string("layers_16_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_33_cast_fp16 = conv(dilations = q_33_dilations_0, groups = q_33_groups_0, pad = q_33_pad_0, pad_type = q_33_pad_type_0, strides = q_33_strides_0, weight = layers_16_self_attn_linear_q_weight_to_fp16_quantized, x = x_359_cast_fp16)[name = string("q_33_cast_fp16")];
            string k_33_pad_type_0 = const()[name = string("k_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_33_strides_0 = const()[name = string("k_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_33_pad_0 = const()[name = string("k_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_33_dilations_0 = const()[name = string("k_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_33_groups_0 = const()[name = string("k_33_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213272896))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213797248))))[name = string("layers_16_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_33_cast_fp16 = conv(dilations = k_33_dilations_0, groups = k_33_groups_0, pad = k_33_pad_0, pad_type = k_33_pad_type_0, strides = k_33_strides_0, weight = layers_16_self_attn_linear_k_weight_to_fp16_quantized, x = x_359_cast_fp16)[name = string("k_33_cast_fp16")];
            string v_33_pad_type_0 = const()[name = string("v_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_33_strides_0 = const()[name = string("v_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_33_pad_0 = const()[name = string("v_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_33_dilations_0 = const()[name = string("v_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_33_groups_0 = const()[name = string("v_33_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213799360))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214323712))))[name = string("layers_16_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_33_cast_fp16 = conv(dilations = v_33_dilations_0, groups = v_33_groups_0, pad = v_33_pad_0, pad_type = v_33_pad_type_0, strides = v_33_strides_0, weight = layers_16_self_attn_linear_v_weight_to_fp16_quantized, x = x_359_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_33_to_fp16 = const()[name = string("bv_all_33_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214325824)))];
            tensor<fp16, [1, 1024, 1, 188]> var_4573_cast_fp16 = add(x = q_33_cast_fp16, y = bv_all_33_to_fp16)[name = string("op_4573_cast_fp16")];
            tensor<int32, [3]> var_4574 = const()[name = string("op_4574"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_33_cast_fp16 = reshape(shape = var_4574, x = var_4573_cast_fp16)[name = string("qb_33_cast_fp16")];
            bool bd_all_65_transpose_x_0 = const()[name = string("bd_all_65_transpose_x_0"), val = bool(false)];
            bool bd_all_65_transpose_y_0 = const()[name = string("bd_all_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_16_self_attn_pos_proj_to_fp16 = const()[name = string("layers_16_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214327936)))];
            tensor<fp16, [8, 375, 188]> bd_all_65_cast_fp16 = matmul(transpose_x = bd_all_65_transpose_x_0, transpose_y = bd_all_65_transpose_y_0, x = layers_16_self_attn_pos_proj_to_fp16, y = qb_33_cast_fp16)[name = string("bd_all_65_cast_fp16")];
            tensor<int32, [3]> x_361_perm_0 = const()[name = string("x_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_363_mode_0 = const()[name = string("x_363_mode_0"), val = string("constant")];
            fp16 const_74_to_fp16 = const()[name = string("const_74_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_361_cast_fp16 = transpose(perm = x_361_perm_0, x = bd_all_65_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [8, 188, 376]> x_363_cast_fp16 = pad(constant_val = const_74_to_fp16, mode = x_363_mode_0, pad = x_363_pad_0, x = x_361_cast_fp16)[name = string("x_363_cast_fp16")];
            tensor<int32, [3]> var_4581 = const()[name = string("op_4581"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_365_cast_fp16 = reshape(shape = var_4581, x = x_363_cast_fp16)[name = string("x_365_cast_fp16")];
            tensor<int32, [3]> var_4584_begin_0 = const()[name = string("op_4584_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_4584_end_0 = const()[name = string("op_4584_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_4584_end_mask_0 = const()[name = string("op_4584_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_4584_cast_fp16 = slice_by_index(begin = var_4584_begin_0, end = var_4584_end_0, end_mask = var_4584_end_mask_0, x = x_365_cast_fp16)[name = string("op_4584_cast_fp16")];
            tensor<int32, [3]> var_4585 = const()[name = string("op_4585"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_367_cast_fp16 = reshape(shape = var_4585, x = var_4584_cast_fp16)[name = string("x_367_cast_fp16")];
            tensor<int32, [3]> bd_all_67_begin_0 = const()[name = string("bd_all_67_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_67_end_0 = const()[name = string("bd_all_67_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_67_end_mask_0 = const()[name = string("bd_all_67_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_67_cast_fp16 = slice_by_index(begin = bd_all_67_begin_0, end = bd_all_67_end_0, end_mask = bd_all_67_end_mask_0, x = x_367_cast_fp16)[name = string("bd_all_67_cast_fp16")];
            tensor<int32, [4]> var_4590 = const()[name = string("op_4590"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4591_cast_fp16 = reshape(shape = var_4590, x = q_33_cast_fp16)[name = string("op_4591_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_4593_to_fp16 = const()[name = string("op_4593_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215096000)))];
            tensor<fp16, [8, 128, 1, 188]> qh_33_cast_fp16 = add(x = var_4591_cast_fp16, y = var_4593_to_fp16)[name = string("qh_33_cast_fp16")];
            tensor<int32, [4]> var_4595 = const()[name = string("op_4595"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4596_cast_fp16 = reshape(shape = var_4595, x = k_33_cast_fp16)[name = string("op_4596_cast_fp16")];
            tensor<int32, [4]> var_4597 = const()[name = string("op_4597"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_4599 = const()[name = string("op_4599"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4600_cast_fp16 = reshape(shape = var_4599, x = v_33_cast_fp16)[name = string("op_4600_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4603_perm_0 = const()[name = string("op_4603_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_33_axes_0 = const()[name = string("bd_33_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_4603_cast_fp16 = transpose(perm = var_4603_perm_0, x = bd_all_67_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [8, 188, 1, 188]> bd_33_cast_fp16 = expand_dims(axes = bd_33_axes_0, x = var_4603_cast_fp16)[name = string("bd_33_cast_fp16")];
            string ac_33_equation_0 = const()[name = string("ac_33_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_4598_cast_fp16 = transpose(perm = var_4597, x = var_4596_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [8, 188, 1, 188]> ac_33_cast_fp16 = einsum(equation = ac_33_equation_0, values = (var_4598_cast_fp16, qh_33_cast_fp16))[name = string("ac_33_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_4619_cast_fp16 = add(x = ac_33_cast_fp16, y = bd_33_cast_fp16)[name = string("op_4619_cast_fp16")];
            fp16 var_4620_to_fp16 = const()[name = string("op_4620_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_65_cast_fp16 = mul(x = var_4619_cast_fp16, y = var_4620_to_fp16)[name = string("scores_65_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = key_bias)[name = string("scores_67_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_4623_cast_fp16 = softmax(axis = var_4479, x = scores_67_cast_fp16)[name = string("op_4623_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_164 = const()[name = string("concat_164"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = var_4623_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [8, 188, 188]> reshape_48_cast_fp16 = reshape(shape = concat_164, x = transpose_32_cast_fp16)[name = string("reshape_48_cast_fp16")];
            tensor<int32, [3]> concat_165 = const()[name = string("concat_165"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = var_4600_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [8, 188, 128]> reshape_49_cast_fp16 = reshape(shape = concat_165, x = transpose_64_cast_fp16)[name = string("reshape_49_cast_fp16")];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = reshape_48_cast_fp16, y = reshape_49_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<int32, [4]> concat_169 = const()[name = string("concat_169"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_50_cast_fp16 = reshape(shape = concat_169, x = matmul_16_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [4]> ctx_33_perm_0 = const()[name = string("ctx_33_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4626 = const()[name = string("op_4626"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_33_cast_fp16 = transpose(perm = ctx_33_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 1024, 1, 188]> input_441_cast_fp16 = reshape(shape = var_4626, x = ctx_33_cast_fp16)[name = string("input_441_cast_fp16")];
            string var_4633_pad_type_0 = const()[name = string("op_4633_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4633_strides_0 = const()[name = string("op_4633_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4633_pad_0 = const()[name = string("op_4633_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4633_dilations_0 = const()[name = string("op_4633_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4633_groups_0 = const()[name = string("op_4633_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215098112))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215622464))))[name = string("layers_16_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4633_cast_fp16 = conv(dilations = var_4633_dilations_0, groups = var_4633_groups_0, pad = var_4633_pad_0, pad_type = var_4633_pad_type_0, strides = var_4633_strides_0, weight = layers_16_self_attn_linear_out_weight_to_fp16_quantized, x = input_441_cast_fp16)[name = string("op_4633_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_369_cast_fp16 = add(x = x_357_cast_fp16, y = var_4633_cast_fp16)[name = string("x_369_cast_fp16")];
            tensor<int32, [1]> var_4649_axes_0 = const()[name = string("op_4649_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_16_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_16_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4649_cast_fp16 = layer_norm(axes = var_4649_axes_0, epsilon = layers_16_norm_conv_eps_scaled_to_fp16, x = x_369_cast_fp16)[name = string("op_4649_cast_fp16")];
            tensor<fp16, [1024]> input_443_gamma_0_to_fp16 = const()[name = string("input_443_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215624576)))];
            tensor<fp16, [1024]> input_443_beta_0_to_fp16 = const()[name = string("input_443_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215626688)))];
            fp16 input_443_epsilon_0_to_fp16 = const()[name = string("input_443_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_443_cast_fp16 = batch_norm(beta = input_443_beta_0_to_fp16, epsilon = input_443_epsilon_0_to_fp16, gamma = input_443_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4649_cast_fp16)[name = string("input_443_cast_fp16")];
            string input_445_pad_type_0 = const()[name = string("input_445_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_445_strides_0 = const()[name = string("input_445_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_445_pad_0 = const()[name = string("input_445_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_445_dilations_0 = const()[name = string("input_445_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_445_groups_0 = const()[name = string("input_445_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215628800))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216677440))))[name = string("layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_445_cast_fp16 = conv(dilations = input_445_dilations_0, groups = input_445_groups_0, pad = input_445_pad_0, pad_type = input_445_pad_type_0, strides = input_445_strides_0, weight = layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_443_cast_fp16)[name = string("input_445_cast_fp16")];
            int32 x_371_split_num_splits_0 = const()[name = string("x_371_split_num_splits_0"), val = int32(2)];
            int32 x_371_split_axis_0 = const()[name = string("x_371_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_371_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_371_split_cast_fp16_1 = split(axis = x_371_split_axis_0, num_splits = x_371_split_num_splits_0, x = input_445_cast_fp16)[name = string("x_371_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_371_split_1_sigmoid_cast_fp16 = sigmoid(x = x_371_split_cast_fp16_1)[name = string("x_371_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_371_cast_fp16 = mul(x = x_371_split_cast_fp16_0, y = x_371_split_1_sigmoid_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_447_cast_fp16 = mul(x = x_371_cast_fp16, y = pad_mask)[name = string("input_447_cast_fp16")];
            string input_449_pad_type_0 = const()[name = string("input_449_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_449_pad_0 = const()[name = string("input_449_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_449_groups_0 = const()[name = string("input_449_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_449_strides_0 = const()[name = string("input_449_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_449_dilations_0 = const()[name = string("input_449_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_135_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216681600))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216686272))))[name = string("const_135_to_fp16_quantized")];
            tensor<fp16, [1024]> const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216688384)))];
            tensor<fp16, [1, 1024, 1, 188]> input_451_cast_fp16 = conv(bias = const_136_to_fp16, dilations = input_449_dilations_0, groups = input_449_groups_0, pad = input_449_pad_0, pad_type = input_449_pad_type_0, strides = input_449_strides_0, weight = const_135_to_fp16_quantized, x = input_447_cast_fp16)[name = string("input_451_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_453_cast_fp16 = silu(x = input_451_cast_fp16)[name = string("input_453_cast_fp16")];
            string var_4681_pad_type_0 = const()[name = string("op_4681_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4681_strides_0 = const()[name = string("op_4681_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4681_pad_0 = const()[name = string("op_4681_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4681_dilations_0 = const()[name = string("op_4681_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4681_groups_0 = const()[name = string("op_4681_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216690496))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217214848))))[name = string("layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4681_cast_fp16 = conv(dilations = var_4681_dilations_0, groups = var_4681_groups_0, pad = var_4681_pad_0, pad_type = var_4681_pad_type_0, strides = var_4681_strides_0, weight = layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_453_cast_fp16)[name = string("op_4681_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_373_cast_fp16 = add(x = x_369_cast_fp16, y = var_4681_cast_fp16)[name = string("x_373_cast_fp16")];
            tensor<int32, [1]> var_4697_axes_0 = const()[name = string("op_4697_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_16_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_16_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4697_cast_fp16 = layer_norm(axes = var_4697_axes_0, epsilon = layers_16_norm_feed_forward2_eps_scaled_to_fp16, x = x_373_cast_fp16)[name = string("op_4697_cast_fp16")];
            tensor<fp16, [1024]> input_455_gamma_0_to_fp16 = const()[name = string("input_455_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217216960)))];
            tensor<fp16, [1024]> input_455_beta_0_to_fp16 = const()[name = string("input_455_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217219072)))];
            fp16 input_455_epsilon_0_to_fp16 = const()[name = string("input_455_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_455_cast_fp16 = batch_norm(beta = input_455_beta_0_to_fp16, epsilon = input_455_epsilon_0_to_fp16, gamma = input_455_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4697_cast_fp16)[name = string("input_455_cast_fp16")];
            string input_457_pad_type_0 = const()[name = string("input_457_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_457_strides_0 = const()[name = string("input_457_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_457_pad_0 = const()[name = string("input_457_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_457_dilations_0 = const()[name = string("input_457_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_457_groups_0 = const()[name = string("input_457_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_16_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217221184))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219318400))))[name = string("layers_16_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_457_cast_fp16 = conv(dilations = input_457_dilations_0, groups = input_457_groups_0, pad = input_457_pad_0, pad_type = input_457_pad_type_0, strides = input_457_strides_0, weight = layers_16_feed_forward2_linear1_weight_to_fp16_quantized, x = input_455_cast_fp16)[name = string("input_457_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_459_cast_fp16 = silu(x = input_457_cast_fp16)[name = string("input_459_cast_fp16")];
            string var_4714_pad_type_0 = const()[name = string("op_4714_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4714_strides_0 = const()[name = string("op_4714_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4714_pad_0 = const()[name = string("op_4714_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4714_dilations_0 = const()[name = string("op_4714_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4714_groups_0 = const()[name = string("op_4714_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_4715_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219326656))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221423872))))[name = string("op_4715_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4715_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_4714_dilations_0, groups = var_4714_groups_0, pad = var_4714_pad_0, pad_type = var_4714_pad_type_0, strides = var_4714_strides_0, weight = op_4715_weight_0_to_fp16_quantized, x = input_459_cast_fp16)[name = string("op_4715_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_375_cast_fp16 = add(x = x_373_cast_fp16, y = var_4715_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [1]> var_4731_axes_0 = const()[name = string("op_4731_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_16_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_16_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4731_cast_fp16 = layer_norm(axes = var_4731_axes_0, epsilon = layers_16_norm_out_eps_scaled_to_fp16, x = x_375_cast_fp16)[name = string("op_4731_cast_fp16")];
            tensor<fp16, [1024]> x_377_gamma_0_to_fp16 = const()[name = string("x_377_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221425984)))];
            tensor<fp16, [1024]> x_377_beta_0_to_fp16 = const()[name = string("x_377_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221428096)))];
            fp16 x_377_epsilon_0_to_fp16 = const()[name = string("x_377_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_377_cast_fp16 = batch_norm(beta = x_377_beta_0_to_fp16, epsilon = x_377_epsilon_0_to_fp16, gamma = x_377_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4731_cast_fp16)[name = string("x_377_cast_fp16")];
            int32 var_4750 = const()[name = string("op_4750"), val = int32(1)];
            tensor<int32, [1]> var_4777_axes_0 = const()[name = string("op_4777_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_17_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_17_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4777_cast_fp16 = layer_norm(axes = var_4777_axes_0, epsilon = layers_17_norm_feed_forward1_eps_scaled_to_fp16, x = x_377_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<fp16, [1024]> input_461_gamma_0_to_fp16 = const()[name = string("input_461_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221430208)))];
            tensor<fp16, [1024]> input_461_beta_0_to_fp16 = const()[name = string("input_461_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221432320)))];
            fp16 input_461_epsilon_0_to_fp16 = const()[name = string("input_461_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_461_cast_fp16 = batch_norm(beta = input_461_beta_0_to_fp16, epsilon = input_461_epsilon_0_to_fp16, gamma = input_461_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4777_cast_fp16)[name = string("input_461_cast_fp16")];
            string input_463_pad_type_0 = const()[name = string("input_463_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_463_strides_0 = const()[name = string("input_463_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_463_pad_0 = const()[name = string("input_463_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_463_dilations_0 = const()[name = string("input_463_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_463_groups_0 = const()[name = string("input_463_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_17_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221434432))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223531648))))[name = string("layers_17_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_463_cast_fp16 = conv(dilations = input_463_dilations_0, groups = input_463_groups_0, pad = input_463_pad_0, pad_type = input_463_pad_type_0, strides = input_463_strides_0, weight = layers_17_feed_forward1_linear1_weight_to_fp16_quantized, x = input_461_cast_fp16)[name = string("input_463_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_465_cast_fp16 = silu(x = input_463_cast_fp16)[name = string("input_465_cast_fp16")];
            string var_4794_pad_type_0 = const()[name = string("op_4794_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4794_strides_0 = const()[name = string("op_4794_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4794_pad_0 = const()[name = string("op_4794_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4794_dilations_0 = const()[name = string("op_4794_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4794_groups_0 = const()[name = string("op_4794_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_4795_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223539904))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225637120))))[name = string("op_4795_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4795_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_4794_dilations_0, groups = var_4794_groups_0, pad = var_4794_pad_0, pad_type = var_4794_pad_type_0, strides = var_4794_strides_0, weight = op_4795_weight_0_to_fp16_quantized, x = input_465_cast_fp16)[name = string("op_4795_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_379_cast_fp16 = add(x = x_377_cast_fp16, y = var_4795_cast_fp16)[name = string("x_379_cast_fp16")];
            tensor<int32, [1]> var_4811_axes_0 = const()[name = string("op_4811_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_17_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_17_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4811_cast_fp16 = layer_norm(axes = var_4811_axes_0, epsilon = layers_17_norm_self_att_eps_scaled_to_fp16, x = x_379_cast_fp16)[name = string("op_4811_cast_fp16")];
            tensor<fp16, [1024]> x_381_gamma_0_to_fp16 = const()[name = string("x_381_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225639232)))];
            tensor<fp16, [1024]> x_381_beta_0_to_fp16 = const()[name = string("x_381_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225641344)))];
            fp16 x_381_epsilon_0_to_fp16 = const()[name = string("x_381_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_381_cast_fp16 = batch_norm(beta = x_381_beta_0_to_fp16, epsilon = x_381_epsilon_0_to_fp16, gamma = x_381_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4811_cast_fp16)[name = string("x_381_cast_fp16")];
            string q_35_pad_type_0 = const()[name = string("q_35_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_35_strides_0 = const()[name = string("q_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_35_pad_0 = const()[name = string("q_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_35_dilations_0 = const()[name = string("q_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_35_groups_0 = const()[name = string("q_35_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225643456))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226167808))))[name = string("layers_17_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_35_cast_fp16 = conv(dilations = q_35_dilations_0, groups = q_35_groups_0, pad = q_35_pad_0, pad_type = q_35_pad_type_0, strides = q_35_strides_0, weight = layers_17_self_attn_linear_q_weight_to_fp16_quantized, x = x_381_cast_fp16)[name = string("q_35_cast_fp16")];
            string k_35_pad_type_0 = const()[name = string("k_35_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_35_strides_0 = const()[name = string("k_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_35_pad_0 = const()[name = string("k_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_35_dilations_0 = const()[name = string("k_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_35_groups_0 = const()[name = string("k_35_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226169920))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226694272))))[name = string("layers_17_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_35_cast_fp16 = conv(dilations = k_35_dilations_0, groups = k_35_groups_0, pad = k_35_pad_0, pad_type = k_35_pad_type_0, strides = k_35_strides_0, weight = layers_17_self_attn_linear_k_weight_to_fp16_quantized, x = x_381_cast_fp16)[name = string("k_35_cast_fp16")];
            string v_35_pad_type_0 = const()[name = string("v_35_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_35_strides_0 = const()[name = string("v_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_35_pad_0 = const()[name = string("v_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_35_dilations_0 = const()[name = string("v_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_35_groups_0 = const()[name = string("v_35_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226696384))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227220736))))[name = string("layers_17_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_35_cast_fp16 = conv(dilations = v_35_dilations_0, groups = v_35_groups_0, pad = v_35_pad_0, pad_type = v_35_pad_type_0, strides = v_35_strides_0, weight = layers_17_self_attn_linear_v_weight_to_fp16_quantized, x = x_381_cast_fp16)[name = string("v_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_35_to_fp16 = const()[name = string("bv_all_35_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227222848)))];
            tensor<fp16, [1, 1024, 1, 188]> var_4844_cast_fp16 = add(x = q_35_cast_fp16, y = bv_all_35_to_fp16)[name = string("op_4844_cast_fp16")];
            tensor<int32, [3]> var_4845 = const()[name = string("op_4845"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_35_cast_fp16 = reshape(shape = var_4845, x = var_4844_cast_fp16)[name = string("qb_35_cast_fp16")];
            bool bd_all_69_transpose_x_0 = const()[name = string("bd_all_69_transpose_x_0"), val = bool(false)];
            bool bd_all_69_transpose_y_0 = const()[name = string("bd_all_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_17_self_attn_pos_proj_to_fp16 = const()[name = string("layers_17_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227224960)))];
            tensor<fp16, [8, 375, 188]> bd_all_69_cast_fp16 = matmul(transpose_x = bd_all_69_transpose_x_0, transpose_y = bd_all_69_transpose_y_0, x = layers_17_self_attn_pos_proj_to_fp16, y = qb_35_cast_fp16)[name = string("bd_all_69_cast_fp16")];
            tensor<int32, [3]> x_383_perm_0 = const()[name = string("x_383_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_385_mode_0 = const()[name = string("x_385_mode_0"), val = string("constant")];
            fp16 const_78_to_fp16 = const()[name = string("const_78_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_383_cast_fp16 = transpose(perm = x_383_perm_0, x = bd_all_69_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [8, 188, 376]> x_385_cast_fp16 = pad(constant_val = const_78_to_fp16, mode = x_385_mode_0, pad = x_385_pad_0, x = x_383_cast_fp16)[name = string("x_385_cast_fp16")];
            tensor<int32, [3]> var_4852 = const()[name = string("op_4852"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_387_cast_fp16 = reshape(shape = var_4852, x = x_385_cast_fp16)[name = string("x_387_cast_fp16")];
            tensor<int32, [3]> var_4855_begin_0 = const()[name = string("op_4855_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_4855_end_0 = const()[name = string("op_4855_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_4855_end_mask_0 = const()[name = string("op_4855_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_4855_cast_fp16 = slice_by_index(begin = var_4855_begin_0, end = var_4855_end_0, end_mask = var_4855_end_mask_0, x = x_387_cast_fp16)[name = string("op_4855_cast_fp16")];
            tensor<int32, [3]> var_4856 = const()[name = string("op_4856"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_389_cast_fp16 = reshape(shape = var_4856, x = var_4855_cast_fp16)[name = string("x_389_cast_fp16")];
            tensor<int32, [3]> bd_all_71_begin_0 = const()[name = string("bd_all_71_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_71_end_0 = const()[name = string("bd_all_71_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_71_end_mask_0 = const()[name = string("bd_all_71_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_71_cast_fp16 = slice_by_index(begin = bd_all_71_begin_0, end = bd_all_71_end_0, end_mask = bd_all_71_end_mask_0, x = x_389_cast_fp16)[name = string("bd_all_71_cast_fp16")];
            tensor<int32, [4]> var_4861 = const()[name = string("op_4861"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4862_cast_fp16 = reshape(shape = var_4861, x = q_35_cast_fp16)[name = string("op_4862_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_4864_to_fp16 = const()[name = string("op_4864_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227993024)))];
            tensor<fp16, [8, 128, 1, 188]> qh_35_cast_fp16 = add(x = var_4862_cast_fp16, y = var_4864_to_fp16)[name = string("qh_35_cast_fp16")];
            tensor<int32, [4]> var_4866 = const()[name = string("op_4866"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4867_cast_fp16 = reshape(shape = var_4866, x = k_35_cast_fp16)[name = string("op_4867_cast_fp16")];
            tensor<int32, [4]> var_4868 = const()[name = string("op_4868"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_4870 = const()[name = string("op_4870"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_4871_cast_fp16 = reshape(shape = var_4870, x = v_35_cast_fp16)[name = string("op_4871_cast_fp16")];
            tensor<int32, [4]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4874_perm_0 = const()[name = string("op_4874_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_35_axes_0 = const()[name = string("bd_35_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_4874_cast_fp16 = transpose(perm = var_4874_perm_0, x = bd_all_71_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [8, 188, 1, 188]> bd_35_cast_fp16 = expand_dims(axes = bd_35_axes_0, x = var_4874_cast_fp16)[name = string("bd_35_cast_fp16")];
            string ac_35_equation_0 = const()[name = string("ac_35_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_4869_cast_fp16 = transpose(perm = var_4868, x = var_4867_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [8, 188, 1, 188]> ac_35_cast_fp16 = einsum(equation = ac_35_equation_0, values = (var_4869_cast_fp16, qh_35_cast_fp16))[name = string("ac_35_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_4890_cast_fp16 = add(x = ac_35_cast_fp16, y = bd_35_cast_fp16)[name = string("op_4890_cast_fp16")];
            fp16 var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_69_cast_fp16 = mul(x = var_4890_cast_fp16, y = var_4891_to_fp16)[name = string("scores_69_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = key_bias)[name = string("scores_71_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_4894_cast_fp16 = softmax(axis = var_4750, x = scores_71_cast_fp16)[name = string("op_4894_cast_fp16")];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_174 = const()[name = string("concat_174"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = var_4894_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [8, 188, 188]> reshape_51_cast_fp16 = reshape(shape = concat_174, x = transpose_34_cast_fp16)[name = string("reshape_51_cast_fp16")];
            tensor<int32, [3]> concat_175 = const()[name = string("concat_175"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = var_4871_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [8, 188, 128]> reshape_52_cast_fp16 = reshape(shape = concat_175, x = transpose_65_cast_fp16)[name = string("reshape_52_cast_fp16")];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = reshape_51_cast_fp16, y = reshape_52_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<int32, [4]> concat_179 = const()[name = string("concat_179"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_53_cast_fp16 = reshape(shape = concat_179, x = matmul_17_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> ctx_35_perm_0 = const()[name = string("ctx_35_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4897 = const()[name = string("op_4897"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_35_cast_fp16 = transpose(perm = ctx_35_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 1024, 1, 188]> input_467_cast_fp16 = reshape(shape = var_4897, x = ctx_35_cast_fp16)[name = string("input_467_cast_fp16")];
            string var_4904_pad_type_0 = const()[name = string("op_4904_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4904_strides_0 = const()[name = string("op_4904_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4904_pad_0 = const()[name = string("op_4904_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4904_dilations_0 = const()[name = string("op_4904_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4904_groups_0 = const()[name = string("op_4904_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227995136))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228519488))))[name = string("layers_17_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4904_cast_fp16 = conv(dilations = var_4904_dilations_0, groups = var_4904_groups_0, pad = var_4904_pad_0, pad_type = var_4904_pad_type_0, strides = var_4904_strides_0, weight = layers_17_self_attn_linear_out_weight_to_fp16_quantized, x = input_467_cast_fp16)[name = string("op_4904_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_391_cast_fp16 = add(x = x_379_cast_fp16, y = var_4904_cast_fp16)[name = string("x_391_cast_fp16")];
            tensor<int32, [1]> var_4920_axes_0 = const()[name = string("op_4920_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_17_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_17_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4920_cast_fp16 = layer_norm(axes = var_4920_axes_0, epsilon = layers_17_norm_conv_eps_scaled_to_fp16, x = x_391_cast_fp16)[name = string("op_4920_cast_fp16")];
            tensor<fp16, [1024]> input_469_gamma_0_to_fp16 = const()[name = string("input_469_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228521600)))];
            tensor<fp16, [1024]> input_469_beta_0_to_fp16 = const()[name = string("input_469_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228523712)))];
            fp16 input_469_epsilon_0_to_fp16 = const()[name = string("input_469_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_469_cast_fp16 = batch_norm(beta = input_469_beta_0_to_fp16, epsilon = input_469_epsilon_0_to_fp16, gamma = input_469_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4920_cast_fp16)[name = string("input_469_cast_fp16")];
            string input_471_pad_type_0 = const()[name = string("input_471_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_471_strides_0 = const()[name = string("input_471_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_471_pad_0 = const()[name = string("input_471_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_471_dilations_0 = const()[name = string("input_471_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_471_groups_0 = const()[name = string("input_471_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228525824))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229574464))))[name = string("layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_471_cast_fp16 = conv(dilations = input_471_dilations_0, groups = input_471_groups_0, pad = input_471_pad_0, pad_type = input_471_pad_type_0, strides = input_471_strides_0, weight = layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_469_cast_fp16)[name = string("input_471_cast_fp16")];
            int32 x_393_split_num_splits_0 = const()[name = string("x_393_split_num_splits_0"), val = int32(2)];
            int32 x_393_split_axis_0 = const()[name = string("x_393_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_393_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_393_split_cast_fp16_1 = split(axis = x_393_split_axis_0, num_splits = x_393_split_num_splits_0, x = input_471_cast_fp16)[name = string("x_393_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_393_split_1_sigmoid_cast_fp16 = sigmoid(x = x_393_split_cast_fp16_1)[name = string("x_393_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_393_cast_fp16 = mul(x = x_393_split_cast_fp16_0, y = x_393_split_1_sigmoid_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_473_cast_fp16 = mul(x = x_393_cast_fp16, y = pad_mask)[name = string("input_473_cast_fp16")];
            string input_475_pad_type_0 = const()[name = string("input_475_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_475_pad_0 = const()[name = string("input_475_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_475_groups_0 = const()[name = string("input_475_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_475_strides_0 = const()[name = string("input_475_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_475_dilations_0 = const()[name = string("input_475_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_137_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229578624))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229583296))))[name = string("const_137_to_fp16_quantized")];
            tensor<fp16, [1024]> const_138_to_fp16 = const()[name = string("const_138_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229585408)))];
            tensor<fp16, [1, 1024, 1, 188]> input_477_cast_fp16 = conv(bias = const_138_to_fp16, dilations = input_475_dilations_0, groups = input_475_groups_0, pad = input_475_pad_0, pad_type = input_475_pad_type_0, strides = input_475_strides_0, weight = const_137_to_fp16_quantized, x = input_473_cast_fp16)[name = string("input_477_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_479_cast_fp16 = silu(x = input_477_cast_fp16)[name = string("input_479_cast_fp16")];
            string var_4952_pad_type_0 = const()[name = string("op_4952_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4952_strides_0 = const()[name = string("op_4952_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4952_pad_0 = const()[name = string("op_4952_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4952_dilations_0 = const()[name = string("op_4952_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4952_groups_0 = const()[name = string("op_4952_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229587520))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230111872))))[name = string("layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4952_cast_fp16 = conv(dilations = var_4952_dilations_0, groups = var_4952_groups_0, pad = var_4952_pad_0, pad_type = var_4952_pad_type_0, strides = var_4952_strides_0, weight = layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_479_cast_fp16)[name = string("op_4952_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_395_cast_fp16 = add(x = x_391_cast_fp16, y = var_4952_cast_fp16)[name = string("x_395_cast_fp16")];
            tensor<int32, [1]> var_4968_axes_0 = const()[name = string("op_4968_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_17_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_17_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_4968_cast_fp16 = layer_norm(axes = var_4968_axes_0, epsilon = layers_17_norm_feed_forward2_eps_scaled_to_fp16, x = x_395_cast_fp16)[name = string("op_4968_cast_fp16")];
            tensor<fp16, [1024]> input_481_gamma_0_to_fp16 = const()[name = string("input_481_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230113984)))];
            tensor<fp16, [1024]> input_481_beta_0_to_fp16 = const()[name = string("input_481_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230116096)))];
            fp16 input_481_epsilon_0_to_fp16 = const()[name = string("input_481_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_481_cast_fp16 = batch_norm(beta = input_481_beta_0_to_fp16, epsilon = input_481_epsilon_0_to_fp16, gamma = input_481_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_4968_cast_fp16)[name = string("input_481_cast_fp16")];
            string input_483_pad_type_0 = const()[name = string("input_483_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_483_strides_0 = const()[name = string("input_483_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_483_pad_0 = const()[name = string("input_483_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_483_dilations_0 = const()[name = string("input_483_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_483_groups_0 = const()[name = string("input_483_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_17_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230118208))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232215424))))[name = string("layers_17_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_483_cast_fp16 = conv(dilations = input_483_dilations_0, groups = input_483_groups_0, pad = input_483_pad_0, pad_type = input_483_pad_type_0, strides = input_483_strides_0, weight = layers_17_feed_forward2_linear1_weight_to_fp16_quantized, x = input_481_cast_fp16)[name = string("input_483_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_485_cast_fp16 = silu(x = input_483_cast_fp16)[name = string("input_485_cast_fp16")];
            string var_4985_pad_type_0 = const()[name = string("op_4985_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4985_strides_0 = const()[name = string("op_4985_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4985_pad_0 = const()[name = string("op_4985_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4985_dilations_0 = const()[name = string("op_4985_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4985_groups_0 = const()[name = string("op_4985_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_4986_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232223680))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234320896))))[name = string("op_4986_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_4986_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_4985_dilations_0, groups = var_4985_groups_0, pad = var_4985_pad_0, pad_type = var_4985_pad_type_0, strides = var_4985_strides_0, weight = op_4986_weight_0_to_fp16_quantized, x = input_485_cast_fp16)[name = string("op_4986_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_397_cast_fp16 = add(x = x_395_cast_fp16, y = var_4986_cast_fp16)[name = string("x_397_cast_fp16")];
            tensor<int32, [1]> var_5002_axes_0 = const()[name = string("op_5002_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_17_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_17_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5002_cast_fp16 = layer_norm(axes = var_5002_axes_0, epsilon = layers_17_norm_out_eps_scaled_to_fp16, x = x_397_cast_fp16)[name = string("op_5002_cast_fp16")];
            tensor<fp16, [1024]> x_399_gamma_0_to_fp16 = const()[name = string("x_399_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234323008)))];
            tensor<fp16, [1024]> x_399_beta_0_to_fp16 = const()[name = string("x_399_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234325120)))];
            fp16 x_399_epsilon_0_to_fp16 = const()[name = string("x_399_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_399_cast_fp16 = batch_norm(beta = x_399_beta_0_to_fp16, epsilon = x_399_epsilon_0_to_fp16, gamma = x_399_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5002_cast_fp16)[name = string("x_399_cast_fp16")];
            int32 var_5021 = const()[name = string("op_5021"), val = int32(1)];
            tensor<int32, [1]> var_5048_axes_0 = const()[name = string("op_5048_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_18_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_18_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5048_cast_fp16 = layer_norm(axes = var_5048_axes_0, epsilon = layers_18_norm_feed_forward1_eps_scaled_to_fp16, x = x_399_cast_fp16)[name = string("op_5048_cast_fp16")];
            tensor<fp16, [1024]> input_487_gamma_0_to_fp16 = const()[name = string("input_487_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234327232)))];
            tensor<fp16, [1024]> input_487_beta_0_to_fp16 = const()[name = string("input_487_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234329344)))];
            fp16 input_487_epsilon_0_to_fp16 = const()[name = string("input_487_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_487_cast_fp16 = batch_norm(beta = input_487_beta_0_to_fp16, epsilon = input_487_epsilon_0_to_fp16, gamma = input_487_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5048_cast_fp16)[name = string("input_487_cast_fp16")];
            string input_489_pad_type_0 = const()[name = string("input_489_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_489_strides_0 = const()[name = string("input_489_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_489_pad_0 = const()[name = string("input_489_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_489_dilations_0 = const()[name = string("input_489_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_489_groups_0 = const()[name = string("input_489_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_18_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234331456))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236428672))))[name = string("layers_18_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_489_cast_fp16 = conv(dilations = input_489_dilations_0, groups = input_489_groups_0, pad = input_489_pad_0, pad_type = input_489_pad_type_0, strides = input_489_strides_0, weight = layers_18_feed_forward1_linear1_weight_to_fp16_quantized, x = input_487_cast_fp16)[name = string("input_489_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = string("input_491_cast_fp16")];
            string var_5065_pad_type_0 = const()[name = string("op_5065_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5065_strides_0 = const()[name = string("op_5065_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5065_pad_0 = const()[name = string("op_5065_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5065_dilations_0 = const()[name = string("op_5065_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5065_groups_0 = const()[name = string("op_5065_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_5066_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236436928))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238534144))))[name = string("op_5066_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5066_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_5065_dilations_0, groups = var_5065_groups_0, pad = var_5065_pad_0, pad_type = var_5065_pad_type_0, strides = var_5065_strides_0, weight = op_5066_weight_0_to_fp16_quantized, x = input_491_cast_fp16)[name = string("op_5066_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_401_cast_fp16 = add(x = x_399_cast_fp16, y = var_5066_cast_fp16)[name = string("x_401_cast_fp16")];
            tensor<int32, [1]> var_5082_axes_0 = const()[name = string("op_5082_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_18_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_18_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5082_cast_fp16 = layer_norm(axes = var_5082_axes_0, epsilon = layers_18_norm_self_att_eps_scaled_to_fp16, x = x_401_cast_fp16)[name = string("op_5082_cast_fp16")];
            tensor<fp16, [1024]> x_403_gamma_0_to_fp16 = const()[name = string("x_403_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238536256)))];
            tensor<fp16, [1024]> x_403_beta_0_to_fp16 = const()[name = string("x_403_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238538368)))];
            fp16 x_403_epsilon_0_to_fp16 = const()[name = string("x_403_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_403_cast_fp16 = batch_norm(beta = x_403_beta_0_to_fp16, epsilon = x_403_epsilon_0_to_fp16, gamma = x_403_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5082_cast_fp16)[name = string("x_403_cast_fp16")];
            string q_37_pad_type_0 = const()[name = string("q_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_37_strides_0 = const()[name = string("q_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_37_pad_0 = const()[name = string("q_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_37_dilations_0 = const()[name = string("q_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_37_groups_0 = const()[name = string("q_37_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238540480))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239064832))))[name = string("layers_18_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_37_cast_fp16 = conv(dilations = q_37_dilations_0, groups = q_37_groups_0, pad = q_37_pad_0, pad_type = q_37_pad_type_0, strides = q_37_strides_0, weight = layers_18_self_attn_linear_q_weight_to_fp16_quantized, x = x_403_cast_fp16)[name = string("q_37_cast_fp16")];
            string k_37_pad_type_0 = const()[name = string("k_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_37_strides_0 = const()[name = string("k_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_37_pad_0 = const()[name = string("k_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_37_dilations_0 = const()[name = string("k_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_37_groups_0 = const()[name = string("k_37_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239066944))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239591296))))[name = string("layers_18_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_37_cast_fp16 = conv(dilations = k_37_dilations_0, groups = k_37_groups_0, pad = k_37_pad_0, pad_type = k_37_pad_type_0, strides = k_37_strides_0, weight = layers_18_self_attn_linear_k_weight_to_fp16_quantized, x = x_403_cast_fp16)[name = string("k_37_cast_fp16")];
            string v_37_pad_type_0 = const()[name = string("v_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_37_strides_0 = const()[name = string("v_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_37_pad_0 = const()[name = string("v_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_37_dilations_0 = const()[name = string("v_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_37_groups_0 = const()[name = string("v_37_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239593408))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240117760))))[name = string("layers_18_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_37_cast_fp16 = conv(dilations = v_37_dilations_0, groups = v_37_groups_0, pad = v_37_pad_0, pad_type = v_37_pad_type_0, strides = v_37_strides_0, weight = layers_18_self_attn_linear_v_weight_to_fp16_quantized, x = x_403_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_37_to_fp16 = const()[name = string("bv_all_37_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240119872)))];
            tensor<fp16, [1, 1024, 1, 188]> var_5115_cast_fp16 = add(x = q_37_cast_fp16, y = bv_all_37_to_fp16)[name = string("op_5115_cast_fp16")];
            tensor<int32, [3]> var_5116 = const()[name = string("op_5116"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_37_cast_fp16 = reshape(shape = var_5116, x = var_5115_cast_fp16)[name = string("qb_37_cast_fp16")];
            bool bd_all_73_transpose_x_0 = const()[name = string("bd_all_73_transpose_x_0"), val = bool(false)];
            bool bd_all_73_transpose_y_0 = const()[name = string("bd_all_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_18_self_attn_pos_proj_to_fp16 = const()[name = string("layers_18_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240121984)))];
            tensor<fp16, [8, 375, 188]> bd_all_73_cast_fp16 = matmul(transpose_x = bd_all_73_transpose_x_0, transpose_y = bd_all_73_transpose_y_0, x = layers_18_self_attn_pos_proj_to_fp16, y = qb_37_cast_fp16)[name = string("bd_all_73_cast_fp16")];
            tensor<int32, [3]> x_405_perm_0 = const()[name = string("x_405_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_407_pad_0 = const()[name = string("x_407_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_407_mode_0 = const()[name = string("x_407_mode_0"), val = string("constant")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_405_cast_fp16 = transpose(perm = x_405_perm_0, x = bd_all_73_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [8, 188, 376]> x_407_cast_fp16 = pad(constant_val = const_82_to_fp16, mode = x_407_mode_0, pad = x_407_pad_0, x = x_405_cast_fp16)[name = string("x_407_cast_fp16")];
            tensor<int32, [3]> var_5123 = const()[name = string("op_5123"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_409_cast_fp16 = reshape(shape = var_5123, x = x_407_cast_fp16)[name = string("x_409_cast_fp16")];
            tensor<int32, [3]> var_5126_begin_0 = const()[name = string("op_5126_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_5126_end_0 = const()[name = string("op_5126_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_5126_end_mask_0 = const()[name = string("op_5126_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_5126_cast_fp16 = slice_by_index(begin = var_5126_begin_0, end = var_5126_end_0, end_mask = var_5126_end_mask_0, x = x_409_cast_fp16)[name = string("op_5126_cast_fp16")];
            tensor<int32, [3]> var_5127 = const()[name = string("op_5127"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_411_cast_fp16 = reshape(shape = var_5127, x = var_5126_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<int32, [3]> bd_all_75_begin_0 = const()[name = string("bd_all_75_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_75_end_0 = const()[name = string("bd_all_75_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_75_end_mask_0 = const()[name = string("bd_all_75_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_75_cast_fp16 = slice_by_index(begin = bd_all_75_begin_0, end = bd_all_75_end_0, end_mask = bd_all_75_end_mask_0, x = x_411_cast_fp16)[name = string("bd_all_75_cast_fp16")];
            tensor<int32, [4]> var_5132 = const()[name = string("op_5132"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5133_cast_fp16 = reshape(shape = var_5132, x = q_37_cast_fp16)[name = string("op_5133_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_5135_to_fp16 = const()[name = string("op_5135_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240890048)))];
            tensor<fp16, [8, 128, 1, 188]> qh_37_cast_fp16 = add(x = var_5133_cast_fp16, y = var_5135_to_fp16)[name = string("qh_37_cast_fp16")];
            tensor<int32, [4]> var_5137 = const()[name = string("op_5137"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5138_cast_fp16 = reshape(shape = var_5137, x = k_37_cast_fp16)[name = string("op_5138_cast_fp16")];
            tensor<int32, [4]> var_5139 = const()[name = string("op_5139"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_5141 = const()[name = string("op_5141"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5142_cast_fp16 = reshape(shape = var_5141, x = v_37_cast_fp16)[name = string("op_5142_cast_fp16")];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5145_perm_0 = const()[name = string("op_5145_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_37_axes_0 = const()[name = string("bd_37_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_5145_cast_fp16 = transpose(perm = var_5145_perm_0, x = bd_all_75_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [8, 188, 1, 188]> bd_37_cast_fp16 = expand_dims(axes = bd_37_axes_0, x = var_5145_cast_fp16)[name = string("bd_37_cast_fp16")];
            string ac_37_equation_0 = const()[name = string("ac_37_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_5140_cast_fp16 = transpose(perm = var_5139, x = var_5138_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [8, 188, 1, 188]> ac_37_cast_fp16 = einsum(equation = ac_37_equation_0, values = (var_5140_cast_fp16, qh_37_cast_fp16))[name = string("ac_37_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_5161_cast_fp16 = add(x = ac_37_cast_fp16, y = bd_37_cast_fp16)[name = string("op_5161_cast_fp16")];
            fp16 var_5162_to_fp16 = const()[name = string("op_5162_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_73_cast_fp16 = mul(x = var_5161_cast_fp16, y = var_5162_to_fp16)[name = string("scores_73_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = key_bias)[name = string("scores_75_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_5165_cast_fp16 = softmax(axis = var_5021, x = scores_75_cast_fp16)[name = string("op_5165_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_184 = const()[name = string("concat_184"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = var_5165_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [8, 188, 188]> reshape_54_cast_fp16 = reshape(shape = concat_184, x = transpose_36_cast_fp16)[name = string("reshape_54_cast_fp16")];
            tensor<int32, [3]> concat_185 = const()[name = string("concat_185"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = var_5142_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [8, 188, 128]> reshape_55_cast_fp16 = reshape(shape = concat_185, x = transpose_66_cast_fp16)[name = string("reshape_55_cast_fp16")];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = reshape_54_cast_fp16, y = reshape_55_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<int32, [4]> concat_189 = const()[name = string("concat_189"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_56_cast_fp16 = reshape(shape = concat_189, x = matmul_18_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [4]> ctx_37_perm_0 = const()[name = string("ctx_37_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_5168 = const()[name = string("op_5168"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_37_cast_fp16 = transpose(perm = ctx_37_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 1024, 1, 188]> input_493_cast_fp16 = reshape(shape = var_5168, x = ctx_37_cast_fp16)[name = string("input_493_cast_fp16")];
            string var_5175_pad_type_0 = const()[name = string("op_5175_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5175_strides_0 = const()[name = string("op_5175_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5175_pad_0 = const()[name = string("op_5175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5175_dilations_0 = const()[name = string("op_5175_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5175_groups_0 = const()[name = string("op_5175_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240892160))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241416512))))[name = string("layers_18_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5175_cast_fp16 = conv(dilations = var_5175_dilations_0, groups = var_5175_groups_0, pad = var_5175_pad_0, pad_type = var_5175_pad_type_0, strides = var_5175_strides_0, weight = layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_493_cast_fp16)[name = string("op_5175_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_413_cast_fp16 = add(x = x_401_cast_fp16, y = var_5175_cast_fp16)[name = string("x_413_cast_fp16")];
            tensor<int32, [1]> var_5191_axes_0 = const()[name = string("op_5191_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_18_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_18_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5191_cast_fp16 = layer_norm(axes = var_5191_axes_0, epsilon = layers_18_norm_conv_eps_scaled_to_fp16, x = x_413_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<fp16, [1024]> input_495_gamma_0_to_fp16 = const()[name = string("input_495_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241418624)))];
            tensor<fp16, [1024]> input_495_beta_0_to_fp16 = const()[name = string("input_495_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241420736)))];
            fp16 input_495_epsilon_0_to_fp16 = const()[name = string("input_495_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_495_cast_fp16 = batch_norm(beta = input_495_beta_0_to_fp16, epsilon = input_495_epsilon_0_to_fp16, gamma = input_495_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5191_cast_fp16)[name = string("input_495_cast_fp16")];
            string input_497_pad_type_0 = const()[name = string("input_497_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_497_strides_0 = const()[name = string("input_497_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_497_pad_0 = const()[name = string("input_497_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_497_dilations_0 = const()[name = string("input_497_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_497_groups_0 = const()[name = string("input_497_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241422848))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242471488))))[name = string("layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_497_cast_fp16 = conv(dilations = input_497_dilations_0, groups = input_497_groups_0, pad = input_497_pad_0, pad_type = input_497_pad_type_0, strides = input_497_strides_0, weight = layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_495_cast_fp16)[name = string("input_497_cast_fp16")];
            int32 x_415_split_num_splits_0 = const()[name = string("x_415_split_num_splits_0"), val = int32(2)];
            int32 x_415_split_axis_0 = const()[name = string("x_415_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_415_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_415_split_cast_fp16_1 = split(axis = x_415_split_axis_0, num_splits = x_415_split_num_splits_0, x = input_497_cast_fp16)[name = string("x_415_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_415_split_1_sigmoid_cast_fp16 = sigmoid(x = x_415_split_cast_fp16_1)[name = string("x_415_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_415_cast_fp16 = mul(x = x_415_split_cast_fp16_0, y = x_415_split_1_sigmoid_cast_fp16)[name = string("x_415_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_499_cast_fp16 = mul(x = x_415_cast_fp16, y = pad_mask)[name = string("input_499_cast_fp16")];
            string input_501_pad_type_0 = const()[name = string("input_501_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_501_pad_0 = const()[name = string("input_501_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_501_groups_0 = const()[name = string("input_501_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_501_strides_0 = const()[name = string("input_501_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_501_dilations_0 = const()[name = string("input_501_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_139_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242475648))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242480320))))[name = string("const_139_to_fp16_quantized")];
            tensor<fp16, [1024]> const_140_to_fp16 = const()[name = string("const_140_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242482432)))];
            tensor<fp16, [1, 1024, 1, 188]> input_503_cast_fp16 = conv(bias = const_140_to_fp16, dilations = input_501_dilations_0, groups = input_501_groups_0, pad = input_501_pad_0, pad_type = input_501_pad_type_0, strides = input_501_strides_0, weight = const_139_to_fp16_quantized, x = input_499_cast_fp16)[name = string("input_503_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_505_cast_fp16 = silu(x = input_503_cast_fp16)[name = string("input_505_cast_fp16")];
            string var_5223_pad_type_0 = const()[name = string("op_5223_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5223_strides_0 = const()[name = string("op_5223_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5223_pad_0 = const()[name = string("op_5223_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5223_dilations_0 = const()[name = string("op_5223_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5223_groups_0 = const()[name = string("op_5223_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242484544))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243008896))))[name = string("layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5223_cast_fp16 = conv(dilations = var_5223_dilations_0, groups = var_5223_groups_0, pad = var_5223_pad_0, pad_type = var_5223_pad_type_0, strides = var_5223_strides_0, weight = layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_505_cast_fp16)[name = string("op_5223_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_417_cast_fp16 = add(x = x_413_cast_fp16, y = var_5223_cast_fp16)[name = string("x_417_cast_fp16")];
            tensor<int32, [1]> var_5239_axes_0 = const()[name = string("op_5239_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_18_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_18_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5239_cast_fp16 = layer_norm(axes = var_5239_axes_0, epsilon = layers_18_norm_feed_forward2_eps_scaled_to_fp16, x = x_417_cast_fp16)[name = string("op_5239_cast_fp16")];
            tensor<fp16, [1024]> input_507_gamma_0_to_fp16 = const()[name = string("input_507_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243011008)))];
            tensor<fp16, [1024]> input_507_beta_0_to_fp16 = const()[name = string("input_507_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243013120)))];
            fp16 input_507_epsilon_0_to_fp16 = const()[name = string("input_507_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_507_cast_fp16 = batch_norm(beta = input_507_beta_0_to_fp16, epsilon = input_507_epsilon_0_to_fp16, gamma = input_507_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5239_cast_fp16)[name = string("input_507_cast_fp16")];
            string input_509_pad_type_0 = const()[name = string("input_509_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_509_strides_0 = const()[name = string("input_509_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_509_pad_0 = const()[name = string("input_509_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_509_dilations_0 = const()[name = string("input_509_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_509_groups_0 = const()[name = string("input_509_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243015232))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245112448))))[name = string("layers_18_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_509_cast_fp16 = conv(dilations = input_509_dilations_0, groups = input_509_groups_0, pad = input_509_pad_0, pad_type = input_509_pad_type_0, strides = input_509_strides_0, weight = layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_507_cast_fp16)[name = string("input_509_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_511_cast_fp16 = silu(x = input_509_cast_fp16)[name = string("input_511_cast_fp16")];
            string var_5256_pad_type_0 = const()[name = string("op_5256_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5256_strides_0 = const()[name = string("op_5256_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5256_pad_0 = const()[name = string("op_5256_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5256_dilations_0 = const()[name = string("op_5256_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5256_groups_0 = const()[name = string("op_5256_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_5257_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245120704))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247217920))))[name = string("op_5257_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5257_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_5256_dilations_0, groups = var_5256_groups_0, pad = var_5256_pad_0, pad_type = var_5256_pad_type_0, strides = var_5256_strides_0, weight = op_5257_weight_0_to_fp16_quantized, x = input_511_cast_fp16)[name = string("op_5257_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_419_cast_fp16 = add(x = x_417_cast_fp16, y = var_5257_cast_fp16)[name = string("x_419_cast_fp16")];
            tensor<int32, [1]> var_5273_axes_0 = const()[name = string("op_5273_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_18_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_18_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5273_cast_fp16 = layer_norm(axes = var_5273_axes_0, epsilon = layers_18_norm_out_eps_scaled_to_fp16, x = x_419_cast_fp16)[name = string("op_5273_cast_fp16")];
            tensor<fp16, [1024]> x_421_gamma_0_to_fp16 = const()[name = string("x_421_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247220032)))];
            tensor<fp16, [1024]> x_421_beta_0_to_fp16 = const()[name = string("x_421_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247222144)))];
            fp16 x_421_epsilon_0_to_fp16 = const()[name = string("x_421_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_421_cast_fp16 = batch_norm(beta = x_421_beta_0_to_fp16, epsilon = x_421_epsilon_0_to_fp16, gamma = x_421_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5273_cast_fp16)[name = string("x_421_cast_fp16")];
            int32 var_5292 = const()[name = string("op_5292"), val = int32(1)];
            tensor<int32, [1]> var_5319_axes_0 = const()[name = string("op_5319_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_19_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_19_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5319_cast_fp16 = layer_norm(axes = var_5319_axes_0, epsilon = layers_19_norm_feed_forward1_eps_scaled_to_fp16, x = x_421_cast_fp16)[name = string("op_5319_cast_fp16")];
            tensor<fp16, [1024]> input_513_gamma_0_to_fp16 = const()[name = string("input_513_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247224256)))];
            tensor<fp16, [1024]> input_513_beta_0_to_fp16 = const()[name = string("input_513_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247226368)))];
            fp16 input_513_epsilon_0_to_fp16 = const()[name = string("input_513_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_513_cast_fp16 = batch_norm(beta = input_513_beta_0_to_fp16, epsilon = input_513_epsilon_0_to_fp16, gamma = input_513_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5319_cast_fp16)[name = string("input_513_cast_fp16")];
            string input_515_pad_type_0 = const()[name = string("input_515_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_515_strides_0 = const()[name = string("input_515_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_515_pad_0 = const()[name = string("input_515_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_515_dilations_0 = const()[name = string("input_515_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_515_groups_0 = const()[name = string("input_515_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247228480))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249325696))))[name = string("layers_19_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_515_cast_fp16 = conv(dilations = input_515_dilations_0, groups = input_515_groups_0, pad = input_515_pad_0, pad_type = input_515_pad_type_0, strides = input_515_strides_0, weight = layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_513_cast_fp16)[name = string("input_515_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_517_cast_fp16 = silu(x = input_515_cast_fp16)[name = string("input_517_cast_fp16")];
            string var_5336_pad_type_0 = const()[name = string("op_5336_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5336_strides_0 = const()[name = string("op_5336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5336_pad_0 = const()[name = string("op_5336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5336_dilations_0 = const()[name = string("op_5336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5336_groups_0 = const()[name = string("op_5336_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_5337_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249333952))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251431168))))[name = string("op_5337_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5337_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_5336_dilations_0, groups = var_5336_groups_0, pad = var_5336_pad_0, pad_type = var_5336_pad_type_0, strides = var_5336_strides_0, weight = op_5337_weight_0_to_fp16_quantized, x = input_517_cast_fp16)[name = string("op_5337_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_423_cast_fp16 = add(x = x_421_cast_fp16, y = var_5337_cast_fp16)[name = string("x_423_cast_fp16")];
            tensor<int32, [1]> var_5353_axes_0 = const()[name = string("op_5353_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_19_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_19_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5353_cast_fp16 = layer_norm(axes = var_5353_axes_0, epsilon = layers_19_norm_self_att_eps_scaled_to_fp16, x = x_423_cast_fp16)[name = string("op_5353_cast_fp16")];
            tensor<fp16, [1024]> x_425_gamma_0_to_fp16 = const()[name = string("x_425_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251433280)))];
            tensor<fp16, [1024]> x_425_beta_0_to_fp16 = const()[name = string("x_425_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251435392)))];
            fp16 x_425_epsilon_0_to_fp16 = const()[name = string("x_425_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_425_cast_fp16 = batch_norm(beta = x_425_beta_0_to_fp16, epsilon = x_425_epsilon_0_to_fp16, gamma = x_425_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5353_cast_fp16)[name = string("x_425_cast_fp16")];
            string q_39_pad_type_0 = const()[name = string("q_39_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_39_strides_0 = const()[name = string("q_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_39_pad_0 = const()[name = string("q_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_39_dilations_0 = const()[name = string("q_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_39_groups_0 = const()[name = string("q_39_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251437504))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251961856))))[name = string("layers_19_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_39_cast_fp16 = conv(dilations = q_39_dilations_0, groups = q_39_groups_0, pad = q_39_pad_0, pad_type = q_39_pad_type_0, strides = q_39_strides_0, weight = layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = x_425_cast_fp16)[name = string("q_39_cast_fp16")];
            string k_39_pad_type_0 = const()[name = string("k_39_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_39_strides_0 = const()[name = string("k_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_39_pad_0 = const()[name = string("k_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_39_dilations_0 = const()[name = string("k_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_39_groups_0 = const()[name = string("k_39_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251963968))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252488320))))[name = string("layers_19_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_39_cast_fp16 = conv(dilations = k_39_dilations_0, groups = k_39_groups_0, pad = k_39_pad_0, pad_type = k_39_pad_type_0, strides = k_39_strides_0, weight = layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = x_425_cast_fp16)[name = string("k_39_cast_fp16")];
            string v_39_pad_type_0 = const()[name = string("v_39_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_39_strides_0 = const()[name = string("v_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_39_pad_0 = const()[name = string("v_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_39_dilations_0 = const()[name = string("v_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_39_groups_0 = const()[name = string("v_39_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252490432))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253014784))))[name = string("layers_19_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_39_cast_fp16 = conv(dilations = v_39_dilations_0, groups = v_39_groups_0, pad = v_39_pad_0, pad_type = v_39_pad_type_0, strides = v_39_strides_0, weight = layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = x_425_cast_fp16)[name = string("v_39_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_39_to_fp16 = const()[name = string("bv_all_39_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253016896)))];
            tensor<fp16, [1, 1024, 1, 188]> var_5386_cast_fp16 = add(x = q_39_cast_fp16, y = bv_all_39_to_fp16)[name = string("op_5386_cast_fp16")];
            tensor<int32, [3]> var_5387 = const()[name = string("op_5387"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_39_cast_fp16 = reshape(shape = var_5387, x = var_5386_cast_fp16)[name = string("qb_39_cast_fp16")];
            bool bd_all_77_transpose_x_0 = const()[name = string("bd_all_77_transpose_x_0"), val = bool(false)];
            bool bd_all_77_transpose_y_0 = const()[name = string("bd_all_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_19_self_attn_pos_proj_to_fp16 = const()[name = string("layers_19_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253019008)))];
            tensor<fp16, [8, 375, 188]> bd_all_77_cast_fp16 = matmul(transpose_x = bd_all_77_transpose_x_0, transpose_y = bd_all_77_transpose_y_0, x = layers_19_self_attn_pos_proj_to_fp16, y = qb_39_cast_fp16)[name = string("bd_all_77_cast_fp16")];
            tensor<int32, [3]> x_427_perm_0 = const()[name = string("x_427_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_429_pad_0 = const()[name = string("x_429_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_429_mode_0 = const()[name = string("x_429_mode_0"), val = string("constant")];
            fp16 const_86_to_fp16 = const()[name = string("const_86_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_427_cast_fp16 = transpose(perm = x_427_perm_0, x = bd_all_77_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [8, 188, 376]> x_429_cast_fp16 = pad(constant_val = const_86_to_fp16, mode = x_429_mode_0, pad = x_429_pad_0, x = x_427_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [3]> var_5394 = const()[name = string("op_5394"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_431_cast_fp16 = reshape(shape = var_5394, x = x_429_cast_fp16)[name = string("x_431_cast_fp16")];
            tensor<int32, [3]> var_5397_begin_0 = const()[name = string("op_5397_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_5397_end_0 = const()[name = string("op_5397_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_5397_end_mask_0 = const()[name = string("op_5397_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_5397_cast_fp16 = slice_by_index(begin = var_5397_begin_0, end = var_5397_end_0, end_mask = var_5397_end_mask_0, x = x_431_cast_fp16)[name = string("op_5397_cast_fp16")];
            tensor<int32, [3]> var_5398 = const()[name = string("op_5398"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_433_cast_fp16 = reshape(shape = var_5398, x = var_5397_cast_fp16)[name = string("x_433_cast_fp16")];
            tensor<int32, [3]> bd_all_79_begin_0 = const()[name = string("bd_all_79_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_79_end_0 = const()[name = string("bd_all_79_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_79_end_mask_0 = const()[name = string("bd_all_79_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_79_cast_fp16 = slice_by_index(begin = bd_all_79_begin_0, end = bd_all_79_end_0, end_mask = bd_all_79_end_mask_0, x = x_433_cast_fp16)[name = string("bd_all_79_cast_fp16")];
            tensor<int32, [4]> var_5403 = const()[name = string("op_5403"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5404_cast_fp16 = reshape(shape = var_5403, x = q_39_cast_fp16)[name = string("op_5404_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_5406_to_fp16 = const()[name = string("op_5406_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253787072)))];
            tensor<fp16, [8, 128, 1, 188]> qh_39_cast_fp16 = add(x = var_5404_cast_fp16, y = var_5406_to_fp16)[name = string("qh_39_cast_fp16")];
            tensor<int32, [4]> var_5408 = const()[name = string("op_5408"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5409_cast_fp16 = reshape(shape = var_5408, x = k_39_cast_fp16)[name = string("op_5409_cast_fp16")];
            tensor<int32, [4]> var_5410 = const()[name = string("op_5410"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_5412 = const()[name = string("op_5412"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5413_cast_fp16 = reshape(shape = var_5412, x = v_39_cast_fp16)[name = string("op_5413_cast_fp16")];
            tensor<int32, [4]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5416_perm_0 = const()[name = string("op_5416_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_39_axes_0 = const()[name = string("bd_39_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_5416_cast_fp16 = transpose(perm = var_5416_perm_0, x = bd_all_79_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [8, 188, 1, 188]> bd_39_cast_fp16 = expand_dims(axes = bd_39_axes_0, x = var_5416_cast_fp16)[name = string("bd_39_cast_fp16")];
            string ac_39_equation_0 = const()[name = string("ac_39_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_5411_cast_fp16 = transpose(perm = var_5410, x = var_5409_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [8, 188, 1, 188]> ac_39_cast_fp16 = einsum(equation = ac_39_equation_0, values = (var_5411_cast_fp16, qh_39_cast_fp16))[name = string("ac_39_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_5432_cast_fp16 = add(x = ac_39_cast_fp16, y = bd_39_cast_fp16)[name = string("op_5432_cast_fp16")];
            fp16 var_5433_to_fp16 = const()[name = string("op_5433_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_77_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = string("scores_77_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = key_bias)[name = string("scores_79_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_5436_cast_fp16 = softmax(axis = var_5292, x = scores_79_cast_fp16)[name = string("op_5436_cast_fp16")];
            tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_194 = const()[name = string("concat_194"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_38_cast_fp16 = transpose(perm = transpose_38_perm_0, x = var_5436_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [8, 188, 188]> reshape_57_cast_fp16 = reshape(shape = concat_194, x = transpose_38_cast_fp16)[name = string("reshape_57_cast_fp16")];
            tensor<int32, [3]> concat_195 = const()[name = string("concat_195"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = var_5413_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [8, 188, 128]> reshape_58_cast_fp16 = reshape(shape = concat_195, x = transpose_67_cast_fp16)[name = string("reshape_58_cast_fp16")];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = reshape_57_cast_fp16, y = reshape_58_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<int32, [4]> concat_199 = const()[name = string("concat_199"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_59_cast_fp16 = reshape(shape = concat_199, x = matmul_19_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> ctx_39_perm_0 = const()[name = string("ctx_39_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_5439 = const()[name = string("op_5439"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_39_cast_fp16 = transpose(perm = ctx_39_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 1024, 1, 188]> input_519_cast_fp16 = reshape(shape = var_5439, x = ctx_39_cast_fp16)[name = string("input_519_cast_fp16")];
            string var_5446_pad_type_0 = const()[name = string("op_5446_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5446_strides_0 = const()[name = string("op_5446_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5446_pad_0 = const()[name = string("op_5446_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5446_dilations_0 = const()[name = string("op_5446_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5446_groups_0 = const()[name = string("op_5446_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253789184))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254313536))))[name = string("layers_19_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5446_cast_fp16 = conv(dilations = var_5446_dilations_0, groups = var_5446_groups_0, pad = var_5446_pad_0, pad_type = var_5446_pad_type_0, strides = var_5446_strides_0, weight = layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_519_cast_fp16)[name = string("op_5446_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_435_cast_fp16 = add(x = x_423_cast_fp16, y = var_5446_cast_fp16)[name = string("x_435_cast_fp16")];
            tensor<int32, [1]> var_5462_axes_0 = const()[name = string("op_5462_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_19_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_19_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5462_cast_fp16 = layer_norm(axes = var_5462_axes_0, epsilon = layers_19_norm_conv_eps_scaled_to_fp16, x = x_435_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1024]> input_521_gamma_0_to_fp16 = const()[name = string("input_521_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254315648)))];
            tensor<fp16, [1024]> input_521_beta_0_to_fp16 = const()[name = string("input_521_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254317760)))];
            fp16 input_521_epsilon_0_to_fp16 = const()[name = string("input_521_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_521_cast_fp16 = batch_norm(beta = input_521_beta_0_to_fp16, epsilon = input_521_epsilon_0_to_fp16, gamma = input_521_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5462_cast_fp16)[name = string("input_521_cast_fp16")];
            string input_523_pad_type_0 = const()[name = string("input_523_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_523_strides_0 = const()[name = string("input_523_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_523_pad_0 = const()[name = string("input_523_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_523_dilations_0 = const()[name = string("input_523_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_523_groups_0 = const()[name = string("input_523_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254319872))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255368512))))[name = string("layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_523_cast_fp16 = conv(dilations = input_523_dilations_0, groups = input_523_groups_0, pad = input_523_pad_0, pad_type = input_523_pad_type_0, strides = input_523_strides_0, weight = layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_521_cast_fp16)[name = string("input_523_cast_fp16")];
            int32 x_437_split_num_splits_0 = const()[name = string("x_437_split_num_splits_0"), val = int32(2)];
            int32 x_437_split_axis_0 = const()[name = string("x_437_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_437_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_437_split_cast_fp16_1 = split(axis = x_437_split_axis_0, num_splits = x_437_split_num_splits_0, x = input_523_cast_fp16)[name = string("x_437_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_437_split_1_sigmoid_cast_fp16 = sigmoid(x = x_437_split_cast_fp16_1)[name = string("x_437_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_437_cast_fp16 = mul(x = x_437_split_cast_fp16_0, y = x_437_split_1_sigmoid_cast_fp16)[name = string("x_437_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_525_cast_fp16 = mul(x = x_437_cast_fp16, y = pad_mask)[name = string("input_525_cast_fp16")];
            string input_527_pad_type_0 = const()[name = string("input_527_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_527_pad_0 = const()[name = string("input_527_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_527_groups_0 = const()[name = string("input_527_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_527_strides_0 = const()[name = string("input_527_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_527_dilations_0 = const()[name = string("input_527_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_141_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255372672))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255377344))))[name = string("const_141_to_fp16_quantized")];
            tensor<fp16, [1024]> const_142_to_fp16 = const()[name = string("const_142_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255379456)))];
            tensor<fp16, [1, 1024, 1, 188]> input_529_cast_fp16 = conv(bias = const_142_to_fp16, dilations = input_527_dilations_0, groups = input_527_groups_0, pad = input_527_pad_0, pad_type = input_527_pad_type_0, strides = input_527_strides_0, weight = const_141_to_fp16_quantized, x = input_525_cast_fp16)[name = string("input_529_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_531_cast_fp16 = silu(x = input_529_cast_fp16)[name = string("input_531_cast_fp16")];
            string var_5494_pad_type_0 = const()[name = string("op_5494_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5494_strides_0 = const()[name = string("op_5494_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5494_pad_0 = const()[name = string("op_5494_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5494_dilations_0 = const()[name = string("op_5494_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5494_groups_0 = const()[name = string("op_5494_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255381568))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255905920))))[name = string("layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5494_cast_fp16 = conv(dilations = var_5494_dilations_0, groups = var_5494_groups_0, pad = var_5494_pad_0, pad_type = var_5494_pad_type_0, strides = var_5494_strides_0, weight = layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_531_cast_fp16)[name = string("op_5494_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_439_cast_fp16 = add(x = x_435_cast_fp16, y = var_5494_cast_fp16)[name = string("x_439_cast_fp16")];
            tensor<int32, [1]> var_5510_axes_0 = const()[name = string("op_5510_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_19_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_19_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5510_cast_fp16 = layer_norm(axes = var_5510_axes_0, epsilon = layers_19_norm_feed_forward2_eps_scaled_to_fp16, x = x_439_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<fp16, [1024]> input_533_gamma_0_to_fp16 = const()[name = string("input_533_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255908032)))];
            tensor<fp16, [1024]> input_533_beta_0_to_fp16 = const()[name = string("input_533_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255910144)))];
            fp16 input_533_epsilon_0_to_fp16 = const()[name = string("input_533_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_533_cast_fp16 = batch_norm(beta = input_533_beta_0_to_fp16, epsilon = input_533_epsilon_0_to_fp16, gamma = input_533_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5510_cast_fp16)[name = string("input_533_cast_fp16")];
            string input_535_pad_type_0 = const()[name = string("input_535_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_535_strides_0 = const()[name = string("input_535_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_535_pad_0 = const()[name = string("input_535_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_535_dilations_0 = const()[name = string("input_535_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_535_groups_0 = const()[name = string("input_535_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255912256))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258009472))))[name = string("layers_19_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_535_cast_fp16 = conv(dilations = input_535_dilations_0, groups = input_535_groups_0, pad = input_535_pad_0, pad_type = input_535_pad_type_0, strides = input_535_strides_0, weight = layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_533_cast_fp16)[name = string("input_535_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_537_cast_fp16 = silu(x = input_535_cast_fp16)[name = string("input_537_cast_fp16")];
            string var_5527_pad_type_0 = const()[name = string("op_5527_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5527_strides_0 = const()[name = string("op_5527_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5527_pad_0 = const()[name = string("op_5527_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5527_dilations_0 = const()[name = string("op_5527_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5527_groups_0 = const()[name = string("op_5527_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_5528_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258017728))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260114944))))[name = string("op_5528_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5528_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_5527_dilations_0, groups = var_5527_groups_0, pad = var_5527_pad_0, pad_type = var_5527_pad_type_0, strides = var_5527_strides_0, weight = op_5528_weight_0_to_fp16_quantized, x = input_537_cast_fp16)[name = string("op_5528_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_441_cast_fp16 = add(x = x_439_cast_fp16, y = var_5528_cast_fp16)[name = string("x_441_cast_fp16")];
            tensor<int32, [1]> var_5544_axes_0 = const()[name = string("op_5544_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_19_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_19_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5544_cast_fp16 = layer_norm(axes = var_5544_axes_0, epsilon = layers_19_norm_out_eps_scaled_to_fp16, x = x_441_cast_fp16)[name = string("op_5544_cast_fp16")];
            tensor<fp16, [1024]> x_443_gamma_0_to_fp16 = const()[name = string("x_443_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260117056)))];
            tensor<fp16, [1024]> x_443_beta_0_to_fp16 = const()[name = string("x_443_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260119168)))];
            fp16 x_443_epsilon_0_to_fp16 = const()[name = string("x_443_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_443_cast_fp16 = batch_norm(beta = x_443_beta_0_to_fp16, epsilon = x_443_epsilon_0_to_fp16, gamma = x_443_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5544_cast_fp16)[name = string("x_443_cast_fp16")];
            int32 var_5563 = const()[name = string("op_5563"), val = int32(1)];
            tensor<int32, [1]> var_5590_axes_0 = const()[name = string("op_5590_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_20_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_20_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5590_cast_fp16 = layer_norm(axes = var_5590_axes_0, epsilon = layers_20_norm_feed_forward1_eps_scaled_to_fp16, x = x_443_cast_fp16)[name = string("op_5590_cast_fp16")];
            tensor<fp16, [1024]> input_539_gamma_0_to_fp16 = const()[name = string("input_539_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260121280)))];
            tensor<fp16, [1024]> input_539_beta_0_to_fp16 = const()[name = string("input_539_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260123392)))];
            fp16 input_539_epsilon_0_to_fp16 = const()[name = string("input_539_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_539_cast_fp16 = batch_norm(beta = input_539_beta_0_to_fp16, epsilon = input_539_epsilon_0_to_fp16, gamma = input_539_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5590_cast_fp16)[name = string("input_539_cast_fp16")];
            string input_541_pad_type_0 = const()[name = string("input_541_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_541_strides_0 = const()[name = string("input_541_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_541_pad_0 = const()[name = string("input_541_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_541_dilations_0 = const()[name = string("input_541_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_541_groups_0 = const()[name = string("input_541_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260125504))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262222720))))[name = string("layers_20_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_541_cast_fp16 = conv(dilations = input_541_dilations_0, groups = input_541_groups_0, pad = input_541_pad_0, pad_type = input_541_pad_type_0, strides = input_541_strides_0, weight = layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_539_cast_fp16)[name = string("input_541_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_543_cast_fp16 = silu(x = input_541_cast_fp16)[name = string("input_543_cast_fp16")];
            string var_5607_pad_type_0 = const()[name = string("op_5607_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5607_strides_0 = const()[name = string("op_5607_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5607_pad_0 = const()[name = string("op_5607_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5607_dilations_0 = const()[name = string("op_5607_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5607_groups_0 = const()[name = string("op_5607_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_5608_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262230976))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264328192))))[name = string("op_5608_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5608_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_5607_dilations_0, groups = var_5607_groups_0, pad = var_5607_pad_0, pad_type = var_5607_pad_type_0, strides = var_5607_strides_0, weight = op_5608_weight_0_to_fp16_quantized, x = input_543_cast_fp16)[name = string("op_5608_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_445_cast_fp16 = add(x = x_443_cast_fp16, y = var_5608_cast_fp16)[name = string("x_445_cast_fp16")];
            tensor<int32, [1]> var_5624_axes_0 = const()[name = string("op_5624_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_20_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_20_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5624_cast_fp16 = layer_norm(axes = var_5624_axes_0, epsilon = layers_20_norm_self_att_eps_scaled_to_fp16, x = x_445_cast_fp16)[name = string("op_5624_cast_fp16")];
            tensor<fp16, [1024]> x_447_gamma_0_to_fp16 = const()[name = string("x_447_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264330304)))];
            tensor<fp16, [1024]> x_447_beta_0_to_fp16 = const()[name = string("x_447_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264332416)))];
            fp16 x_447_epsilon_0_to_fp16 = const()[name = string("x_447_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_447_cast_fp16 = batch_norm(beta = x_447_beta_0_to_fp16, epsilon = x_447_epsilon_0_to_fp16, gamma = x_447_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5624_cast_fp16)[name = string("x_447_cast_fp16")];
            string q_41_pad_type_0 = const()[name = string("q_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_41_strides_0 = const()[name = string("q_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_41_pad_0 = const()[name = string("q_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_41_dilations_0 = const()[name = string("q_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_41_groups_0 = const()[name = string("q_41_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264334528))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264858880))))[name = string("layers_20_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_41_cast_fp16 = conv(dilations = q_41_dilations_0, groups = q_41_groups_0, pad = q_41_pad_0, pad_type = q_41_pad_type_0, strides = q_41_strides_0, weight = layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = x_447_cast_fp16)[name = string("q_41_cast_fp16")];
            string k_41_pad_type_0 = const()[name = string("k_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_41_strides_0 = const()[name = string("k_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_41_pad_0 = const()[name = string("k_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_41_dilations_0 = const()[name = string("k_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_41_groups_0 = const()[name = string("k_41_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264860992))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265385344))))[name = string("layers_20_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_41_cast_fp16 = conv(dilations = k_41_dilations_0, groups = k_41_groups_0, pad = k_41_pad_0, pad_type = k_41_pad_type_0, strides = k_41_strides_0, weight = layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = x_447_cast_fp16)[name = string("k_41_cast_fp16")];
            string v_41_pad_type_0 = const()[name = string("v_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_41_strides_0 = const()[name = string("v_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_41_pad_0 = const()[name = string("v_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_41_dilations_0 = const()[name = string("v_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_41_groups_0 = const()[name = string("v_41_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265387456))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911808))))[name = string("layers_20_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_41_cast_fp16 = conv(dilations = v_41_dilations_0, groups = v_41_groups_0, pad = v_41_pad_0, pad_type = v_41_pad_type_0, strides = v_41_strides_0, weight = layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = x_447_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_41_to_fp16 = const()[name = string("bv_all_41_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265913920)))];
            tensor<fp16, [1, 1024, 1, 188]> var_5657_cast_fp16 = add(x = q_41_cast_fp16, y = bv_all_41_to_fp16)[name = string("op_5657_cast_fp16")];
            tensor<int32, [3]> var_5658 = const()[name = string("op_5658"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_41_cast_fp16 = reshape(shape = var_5658, x = var_5657_cast_fp16)[name = string("qb_41_cast_fp16")];
            bool bd_all_81_transpose_x_0 = const()[name = string("bd_all_81_transpose_x_0"), val = bool(false)];
            bool bd_all_81_transpose_y_0 = const()[name = string("bd_all_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_20_self_attn_pos_proj_to_fp16 = const()[name = string("layers_20_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265916032)))];
            tensor<fp16, [8, 375, 188]> bd_all_81_cast_fp16 = matmul(transpose_x = bd_all_81_transpose_x_0, transpose_y = bd_all_81_transpose_y_0, x = layers_20_self_attn_pos_proj_to_fp16, y = qb_41_cast_fp16)[name = string("bd_all_81_cast_fp16")];
            tensor<int32, [3]> x_449_perm_0 = const()[name = string("x_449_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")];
            fp16 const_90_to_fp16 = const()[name = string("const_90_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_449_cast_fp16 = transpose(perm = x_449_perm_0, x = bd_all_81_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [8, 188, 376]> x_451_cast_fp16 = pad(constant_val = const_90_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")];
            tensor<int32, [3]> var_5665 = const()[name = string("op_5665"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_453_cast_fp16 = reshape(shape = var_5665, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")];
            tensor<int32, [3]> var_5668_begin_0 = const()[name = string("op_5668_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_5668_end_0 = const()[name = string("op_5668_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_5668_end_mask_0 = const()[name = string("op_5668_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_5668_cast_fp16 = slice_by_index(begin = var_5668_begin_0, end = var_5668_end_0, end_mask = var_5668_end_mask_0, x = x_453_cast_fp16)[name = string("op_5668_cast_fp16")];
            tensor<int32, [3]> var_5669 = const()[name = string("op_5669"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_455_cast_fp16 = reshape(shape = var_5669, x = var_5668_cast_fp16)[name = string("x_455_cast_fp16")];
            tensor<int32, [3]> bd_all_83_begin_0 = const()[name = string("bd_all_83_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_83_end_0 = const()[name = string("bd_all_83_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_83_end_mask_0 = const()[name = string("bd_all_83_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_83_cast_fp16 = slice_by_index(begin = bd_all_83_begin_0, end = bd_all_83_end_0, end_mask = bd_all_83_end_mask_0, x = x_455_cast_fp16)[name = string("bd_all_83_cast_fp16")];
            tensor<int32, [4]> var_5674 = const()[name = string("op_5674"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5675_cast_fp16 = reshape(shape = var_5674, x = q_41_cast_fp16)[name = string("op_5675_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_5677_to_fp16 = const()[name = string("op_5677_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266684096)))];
            tensor<fp16, [8, 128, 1, 188]> qh_41_cast_fp16 = add(x = var_5675_cast_fp16, y = var_5677_to_fp16)[name = string("qh_41_cast_fp16")];
            tensor<int32, [4]> var_5679 = const()[name = string("op_5679"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5680_cast_fp16 = reshape(shape = var_5679, x = k_41_cast_fp16)[name = string("op_5680_cast_fp16")];
            tensor<int32, [4]> var_5681 = const()[name = string("op_5681"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_5683 = const()[name = string("op_5683"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5684_cast_fp16 = reshape(shape = var_5683, x = v_41_cast_fp16)[name = string("op_5684_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5687_perm_0 = const()[name = string("op_5687_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_41_axes_0 = const()[name = string("bd_41_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_5687_cast_fp16 = transpose(perm = var_5687_perm_0, x = bd_all_83_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [8, 188, 1, 188]> bd_41_cast_fp16 = expand_dims(axes = bd_41_axes_0, x = var_5687_cast_fp16)[name = string("bd_41_cast_fp16")];
            string ac_41_equation_0 = const()[name = string("ac_41_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_5682_cast_fp16 = transpose(perm = var_5681, x = var_5680_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [8, 188, 1, 188]> ac_41_cast_fp16 = einsum(equation = ac_41_equation_0, values = (var_5682_cast_fp16, qh_41_cast_fp16))[name = string("ac_41_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_5703_cast_fp16 = add(x = ac_41_cast_fp16, y = bd_41_cast_fp16)[name = string("op_5703_cast_fp16")];
            fp16 var_5704_to_fp16 = const()[name = string("op_5704_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_81_cast_fp16 = mul(x = var_5703_cast_fp16, y = var_5704_to_fp16)[name = string("scores_81_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = key_bias)[name = string("scores_83_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_5707_cast_fp16 = softmax(axis = var_5563, x = scores_83_cast_fp16)[name = string("op_5707_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_204 = const()[name = string("concat_204"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = var_5707_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [8, 188, 188]> reshape_60_cast_fp16 = reshape(shape = concat_204, x = transpose_40_cast_fp16)[name = string("reshape_60_cast_fp16")];
            tensor<int32, [3]> concat_205 = const()[name = string("concat_205"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = var_5684_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [8, 188, 128]> reshape_61_cast_fp16 = reshape(shape = concat_205, x = transpose_68_cast_fp16)[name = string("reshape_61_cast_fp16")];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = reshape_60_cast_fp16, y = reshape_61_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<int32, [4]> concat_209 = const()[name = string("concat_209"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_62_cast_fp16 = reshape(shape = concat_209, x = matmul_20_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [4]> ctx_41_perm_0 = const()[name = string("ctx_41_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_5710 = const()[name = string("op_5710"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_41_cast_fp16 = transpose(perm = ctx_41_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 1024, 1, 188]> input_545_cast_fp16 = reshape(shape = var_5710, x = ctx_41_cast_fp16)[name = string("input_545_cast_fp16")];
            string var_5717_pad_type_0 = const()[name = string("op_5717_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5717_strides_0 = const()[name = string("op_5717_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5717_pad_0 = const()[name = string("op_5717_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5717_dilations_0 = const()[name = string("op_5717_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5717_groups_0 = const()[name = string("op_5717_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266686208))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267210560))))[name = string("layers_20_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5717_cast_fp16 = conv(dilations = var_5717_dilations_0, groups = var_5717_groups_0, pad = var_5717_pad_0, pad_type = var_5717_pad_type_0, strides = var_5717_strides_0, weight = layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_545_cast_fp16)[name = string("op_5717_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_457_cast_fp16 = add(x = x_445_cast_fp16, y = var_5717_cast_fp16)[name = string("x_457_cast_fp16")];
            tensor<int32, [1]> var_5733_axes_0 = const()[name = string("op_5733_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_20_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_20_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5733_cast_fp16 = layer_norm(axes = var_5733_axes_0, epsilon = layers_20_norm_conv_eps_scaled_to_fp16, x = x_457_cast_fp16)[name = string("op_5733_cast_fp16")];
            tensor<fp16, [1024]> input_547_gamma_0_to_fp16 = const()[name = string("input_547_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267212672)))];
            tensor<fp16, [1024]> input_547_beta_0_to_fp16 = const()[name = string("input_547_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267214784)))];
            fp16 input_547_epsilon_0_to_fp16 = const()[name = string("input_547_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_547_cast_fp16 = batch_norm(beta = input_547_beta_0_to_fp16, epsilon = input_547_epsilon_0_to_fp16, gamma = input_547_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5733_cast_fp16)[name = string("input_547_cast_fp16")];
            string input_549_pad_type_0 = const()[name = string("input_549_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_549_strides_0 = const()[name = string("input_549_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_549_pad_0 = const()[name = string("input_549_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_549_dilations_0 = const()[name = string("input_549_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_549_groups_0 = const()[name = string("input_549_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267216896))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268265536))))[name = string("layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_549_cast_fp16 = conv(dilations = input_549_dilations_0, groups = input_549_groups_0, pad = input_549_pad_0, pad_type = input_549_pad_type_0, strides = input_549_strides_0, weight = layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_547_cast_fp16)[name = string("input_549_cast_fp16")];
            int32 x_459_split_num_splits_0 = const()[name = string("x_459_split_num_splits_0"), val = int32(2)];
            int32 x_459_split_axis_0 = const()[name = string("x_459_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_459_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_459_split_cast_fp16_1 = split(axis = x_459_split_axis_0, num_splits = x_459_split_num_splits_0, x = input_549_cast_fp16)[name = string("x_459_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_459_split_1_sigmoid_cast_fp16 = sigmoid(x = x_459_split_cast_fp16_1)[name = string("x_459_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_459_cast_fp16 = mul(x = x_459_split_cast_fp16_0, y = x_459_split_1_sigmoid_cast_fp16)[name = string("x_459_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_551_cast_fp16 = mul(x = x_459_cast_fp16, y = pad_mask)[name = string("input_551_cast_fp16")];
            string input_553_pad_type_0 = const()[name = string("input_553_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_553_pad_0 = const()[name = string("input_553_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_553_groups_0 = const()[name = string("input_553_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_553_strides_0 = const()[name = string("input_553_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_553_dilations_0 = const()[name = string("input_553_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_143_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268269696))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268274368))))[name = string("const_143_to_fp16_quantized")];
            tensor<fp16, [1024]> const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268276480)))];
            tensor<fp16, [1, 1024, 1, 188]> input_555_cast_fp16 = conv(bias = const_144_to_fp16, dilations = input_553_dilations_0, groups = input_553_groups_0, pad = input_553_pad_0, pad_type = input_553_pad_type_0, strides = input_553_strides_0, weight = const_143_to_fp16_quantized, x = input_551_cast_fp16)[name = string("input_555_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_557_cast_fp16 = silu(x = input_555_cast_fp16)[name = string("input_557_cast_fp16")];
            string var_5765_pad_type_0 = const()[name = string("op_5765_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5765_strides_0 = const()[name = string("op_5765_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5765_pad_0 = const()[name = string("op_5765_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5765_dilations_0 = const()[name = string("op_5765_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5765_groups_0 = const()[name = string("op_5765_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268278592))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268802944))))[name = string("layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5765_cast_fp16 = conv(dilations = var_5765_dilations_0, groups = var_5765_groups_0, pad = var_5765_pad_0, pad_type = var_5765_pad_type_0, strides = var_5765_strides_0, weight = layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_557_cast_fp16)[name = string("op_5765_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_461_cast_fp16 = add(x = x_457_cast_fp16, y = var_5765_cast_fp16)[name = string("x_461_cast_fp16")];
            tensor<int32, [1]> var_5781_axes_0 = const()[name = string("op_5781_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_20_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_20_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5781_cast_fp16 = layer_norm(axes = var_5781_axes_0, epsilon = layers_20_norm_feed_forward2_eps_scaled_to_fp16, x = x_461_cast_fp16)[name = string("op_5781_cast_fp16")];
            tensor<fp16, [1024]> input_559_gamma_0_to_fp16 = const()[name = string("input_559_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268805056)))];
            tensor<fp16, [1024]> input_559_beta_0_to_fp16 = const()[name = string("input_559_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268807168)))];
            fp16 input_559_epsilon_0_to_fp16 = const()[name = string("input_559_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_559_cast_fp16 = batch_norm(beta = input_559_beta_0_to_fp16, epsilon = input_559_epsilon_0_to_fp16, gamma = input_559_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5781_cast_fp16)[name = string("input_559_cast_fp16")];
            string input_561_pad_type_0 = const()[name = string("input_561_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_561_strides_0 = const()[name = string("input_561_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_561_pad_0 = const()[name = string("input_561_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_561_dilations_0 = const()[name = string("input_561_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_561_groups_0 = const()[name = string("input_561_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268809280))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270906496))))[name = string("layers_20_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_561_cast_fp16 = conv(dilations = input_561_dilations_0, groups = input_561_groups_0, pad = input_561_pad_0, pad_type = input_561_pad_type_0, strides = input_561_strides_0, weight = layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_559_cast_fp16)[name = string("input_561_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_563_cast_fp16 = silu(x = input_561_cast_fp16)[name = string("input_563_cast_fp16")];
            string var_5798_pad_type_0 = const()[name = string("op_5798_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5798_strides_0 = const()[name = string("op_5798_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5798_pad_0 = const()[name = string("op_5798_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5798_dilations_0 = const()[name = string("op_5798_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5798_groups_0 = const()[name = string("op_5798_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_5799_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270914752))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273011968))))[name = string("op_5799_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5799_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_5798_dilations_0, groups = var_5798_groups_0, pad = var_5798_pad_0, pad_type = var_5798_pad_type_0, strides = var_5798_strides_0, weight = op_5799_weight_0_to_fp16_quantized, x = input_563_cast_fp16)[name = string("op_5799_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_463_cast_fp16 = add(x = x_461_cast_fp16, y = var_5799_cast_fp16)[name = string("x_463_cast_fp16")];
            tensor<int32, [1]> var_5815_axes_0 = const()[name = string("op_5815_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_20_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_20_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5815_cast_fp16 = layer_norm(axes = var_5815_axes_0, epsilon = layers_20_norm_out_eps_scaled_to_fp16, x = x_463_cast_fp16)[name = string("op_5815_cast_fp16")];
            tensor<fp16, [1024]> x_465_gamma_0_to_fp16 = const()[name = string("x_465_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273014080)))];
            tensor<fp16, [1024]> x_465_beta_0_to_fp16 = const()[name = string("x_465_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273016192)))];
            fp16 x_465_epsilon_0_to_fp16 = const()[name = string("x_465_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_465_cast_fp16 = batch_norm(beta = x_465_beta_0_to_fp16, epsilon = x_465_epsilon_0_to_fp16, gamma = x_465_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5815_cast_fp16)[name = string("x_465_cast_fp16")];
            int32 var_5834 = const()[name = string("op_5834"), val = int32(1)];
            tensor<int32, [1]> var_5861_axes_0 = const()[name = string("op_5861_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_21_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_21_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5861_cast_fp16 = layer_norm(axes = var_5861_axes_0, epsilon = layers_21_norm_feed_forward1_eps_scaled_to_fp16, x = x_465_cast_fp16)[name = string("op_5861_cast_fp16")];
            tensor<fp16, [1024]> input_565_gamma_0_to_fp16 = const()[name = string("input_565_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273018304)))];
            tensor<fp16, [1024]> input_565_beta_0_to_fp16 = const()[name = string("input_565_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273020416)))];
            fp16 input_565_epsilon_0_to_fp16 = const()[name = string("input_565_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_565_cast_fp16 = batch_norm(beta = input_565_beta_0_to_fp16, epsilon = input_565_epsilon_0_to_fp16, gamma = input_565_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5861_cast_fp16)[name = string("input_565_cast_fp16")];
            string input_567_pad_type_0 = const()[name = string("input_567_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_567_strides_0 = const()[name = string("input_567_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_567_pad_0 = const()[name = string("input_567_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_567_dilations_0 = const()[name = string("input_567_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_567_groups_0 = const()[name = string("input_567_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273022528))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275119744))))[name = string("layers_21_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_567_cast_fp16 = conv(dilations = input_567_dilations_0, groups = input_567_groups_0, pad = input_567_pad_0, pad_type = input_567_pad_type_0, strides = input_567_strides_0, weight = layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_565_cast_fp16)[name = string("input_567_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_569_cast_fp16 = silu(x = input_567_cast_fp16)[name = string("input_569_cast_fp16")];
            string var_5878_pad_type_0 = const()[name = string("op_5878_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5878_strides_0 = const()[name = string("op_5878_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5878_pad_0 = const()[name = string("op_5878_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5878_dilations_0 = const()[name = string("op_5878_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5878_groups_0 = const()[name = string("op_5878_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_5879_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275128000))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277225216))))[name = string("op_5879_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5879_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_5878_dilations_0, groups = var_5878_groups_0, pad = var_5878_pad_0, pad_type = var_5878_pad_type_0, strides = var_5878_strides_0, weight = op_5879_weight_0_to_fp16_quantized, x = input_569_cast_fp16)[name = string("op_5879_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_467_cast_fp16 = add(x = x_465_cast_fp16, y = var_5879_cast_fp16)[name = string("x_467_cast_fp16")];
            tensor<int32, [1]> var_5895_axes_0 = const()[name = string("op_5895_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_21_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_21_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_5895_cast_fp16 = layer_norm(axes = var_5895_axes_0, epsilon = layers_21_norm_self_att_eps_scaled_to_fp16, x = x_467_cast_fp16)[name = string("op_5895_cast_fp16")];
            tensor<fp16, [1024]> x_469_gamma_0_to_fp16 = const()[name = string("x_469_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277227328)))];
            tensor<fp16, [1024]> x_469_beta_0_to_fp16 = const()[name = string("x_469_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277229440)))];
            fp16 x_469_epsilon_0_to_fp16 = const()[name = string("x_469_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_469_cast_fp16 = batch_norm(beta = x_469_beta_0_to_fp16, epsilon = x_469_epsilon_0_to_fp16, gamma = x_469_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_5895_cast_fp16)[name = string("x_469_cast_fp16")];
            string q_43_pad_type_0 = const()[name = string("q_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_43_strides_0 = const()[name = string("q_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_43_pad_0 = const()[name = string("q_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_43_dilations_0 = const()[name = string("q_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_43_groups_0 = const()[name = string("q_43_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277231552))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277755904))))[name = string("layers_21_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_43_cast_fp16 = conv(dilations = q_43_dilations_0, groups = q_43_groups_0, pad = q_43_pad_0, pad_type = q_43_pad_type_0, strides = q_43_strides_0, weight = layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = x_469_cast_fp16)[name = string("q_43_cast_fp16")];
            string k_43_pad_type_0 = const()[name = string("k_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_43_strides_0 = const()[name = string("k_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_43_pad_0 = const()[name = string("k_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_43_dilations_0 = const()[name = string("k_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_43_groups_0 = const()[name = string("k_43_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277758016))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278282368))))[name = string("layers_21_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_43_cast_fp16 = conv(dilations = k_43_dilations_0, groups = k_43_groups_0, pad = k_43_pad_0, pad_type = k_43_pad_type_0, strides = k_43_strides_0, weight = layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = x_469_cast_fp16)[name = string("k_43_cast_fp16")];
            string v_43_pad_type_0 = const()[name = string("v_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_43_strides_0 = const()[name = string("v_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_43_pad_0 = const()[name = string("v_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_43_dilations_0 = const()[name = string("v_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_43_groups_0 = const()[name = string("v_43_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278284480))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278808832))))[name = string("layers_21_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_43_cast_fp16 = conv(dilations = v_43_dilations_0, groups = v_43_groups_0, pad = v_43_pad_0, pad_type = v_43_pad_type_0, strides = v_43_strides_0, weight = layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = x_469_cast_fp16)[name = string("v_43_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_43_to_fp16 = const()[name = string("bv_all_43_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278810944)))];
            tensor<fp16, [1, 1024, 1, 188]> var_5928_cast_fp16 = add(x = q_43_cast_fp16, y = bv_all_43_to_fp16)[name = string("op_5928_cast_fp16")];
            tensor<int32, [3]> var_5929 = const()[name = string("op_5929"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_43_cast_fp16 = reshape(shape = var_5929, x = var_5928_cast_fp16)[name = string("qb_43_cast_fp16")];
            bool bd_all_85_transpose_x_0 = const()[name = string("bd_all_85_transpose_x_0"), val = bool(false)];
            bool bd_all_85_transpose_y_0 = const()[name = string("bd_all_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_21_self_attn_pos_proj_to_fp16 = const()[name = string("layers_21_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278813056)))];
            tensor<fp16, [8, 375, 188]> bd_all_85_cast_fp16 = matmul(transpose_x = bd_all_85_transpose_x_0, transpose_y = bd_all_85_transpose_y_0, x = layers_21_self_attn_pos_proj_to_fp16, y = qb_43_cast_fp16)[name = string("bd_all_85_cast_fp16")];
            tensor<int32, [3]> x_471_perm_0 = const()[name = string("x_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_473_pad_0 = const()[name = string("x_473_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_473_mode_0 = const()[name = string("x_473_mode_0"), val = string("constant")];
            fp16 const_94_to_fp16 = const()[name = string("const_94_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_471_cast_fp16 = transpose(perm = x_471_perm_0, x = bd_all_85_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [8, 188, 376]> x_473_cast_fp16 = pad(constant_val = const_94_to_fp16, mode = x_473_mode_0, pad = x_473_pad_0, x = x_471_cast_fp16)[name = string("x_473_cast_fp16")];
            tensor<int32, [3]> var_5936 = const()[name = string("op_5936"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_475_cast_fp16 = reshape(shape = var_5936, x = x_473_cast_fp16)[name = string("x_475_cast_fp16")];
            tensor<int32, [3]> var_5939_begin_0 = const()[name = string("op_5939_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_5939_end_0 = const()[name = string("op_5939_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_5939_end_mask_0 = const()[name = string("op_5939_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_5939_cast_fp16 = slice_by_index(begin = var_5939_begin_0, end = var_5939_end_0, end_mask = var_5939_end_mask_0, x = x_475_cast_fp16)[name = string("op_5939_cast_fp16")];
            tensor<int32, [3]> var_5940 = const()[name = string("op_5940"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_477_cast_fp16 = reshape(shape = var_5940, x = var_5939_cast_fp16)[name = string("x_477_cast_fp16")];
            tensor<int32, [3]> bd_all_87_begin_0 = const()[name = string("bd_all_87_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_87_end_0 = const()[name = string("bd_all_87_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_87_end_mask_0 = const()[name = string("bd_all_87_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_87_cast_fp16 = slice_by_index(begin = bd_all_87_begin_0, end = bd_all_87_end_0, end_mask = bd_all_87_end_mask_0, x = x_477_cast_fp16)[name = string("bd_all_87_cast_fp16")];
            tensor<int32, [4]> var_5945 = const()[name = string("op_5945"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5946_cast_fp16 = reshape(shape = var_5945, x = q_43_cast_fp16)[name = string("op_5946_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_5948_to_fp16 = const()[name = string("op_5948_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279581120)))];
            tensor<fp16, [8, 128, 1, 188]> qh_43_cast_fp16 = add(x = var_5946_cast_fp16, y = var_5948_to_fp16)[name = string("qh_43_cast_fp16")];
            tensor<int32, [4]> var_5950 = const()[name = string("op_5950"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5951_cast_fp16 = reshape(shape = var_5950, x = k_43_cast_fp16)[name = string("op_5951_cast_fp16")];
            tensor<int32, [4]> var_5952 = const()[name = string("op_5952"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_5954 = const()[name = string("op_5954"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_5955_cast_fp16 = reshape(shape = var_5954, x = v_43_cast_fp16)[name = string("op_5955_cast_fp16")];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5958_perm_0 = const()[name = string("op_5958_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_43_axes_0 = const()[name = string("bd_43_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_5958_cast_fp16 = transpose(perm = var_5958_perm_0, x = bd_all_87_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [8, 188, 1, 188]> bd_43_cast_fp16 = expand_dims(axes = bd_43_axes_0, x = var_5958_cast_fp16)[name = string("bd_43_cast_fp16")];
            string ac_43_equation_0 = const()[name = string("ac_43_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_5953_cast_fp16 = transpose(perm = var_5952, x = var_5951_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [8, 188, 1, 188]> ac_43_cast_fp16 = einsum(equation = ac_43_equation_0, values = (var_5953_cast_fp16, qh_43_cast_fp16))[name = string("ac_43_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_5974_cast_fp16 = add(x = ac_43_cast_fp16, y = bd_43_cast_fp16)[name = string("op_5974_cast_fp16")];
            fp16 var_5975_to_fp16 = const()[name = string("op_5975_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_85_cast_fp16 = mul(x = var_5974_cast_fp16, y = var_5975_to_fp16)[name = string("scores_85_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = key_bias)[name = string("scores_87_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_5978_cast_fp16 = softmax(axis = var_5834, x = scores_87_cast_fp16)[name = string("op_5978_cast_fp16")];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_214 = const()[name = string("concat_214"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = var_5978_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [8, 188, 188]> reshape_63_cast_fp16 = reshape(shape = concat_214, x = transpose_42_cast_fp16)[name = string("reshape_63_cast_fp16")];
            tensor<int32, [3]> concat_215 = const()[name = string("concat_215"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = var_5955_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [8, 188, 128]> reshape_64_cast_fp16 = reshape(shape = concat_215, x = transpose_69_cast_fp16)[name = string("reshape_64_cast_fp16")];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = reshape_63_cast_fp16, y = reshape_64_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<int32, [4]> concat_219 = const()[name = string("concat_219"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_65_cast_fp16 = reshape(shape = concat_219, x = matmul_21_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> ctx_43_perm_0 = const()[name = string("ctx_43_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_5981 = const()[name = string("op_5981"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_43_cast_fp16 = transpose(perm = ctx_43_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 1024, 1, 188]> input_571_cast_fp16 = reshape(shape = var_5981, x = ctx_43_cast_fp16)[name = string("input_571_cast_fp16")];
            string var_5988_pad_type_0 = const()[name = string("op_5988_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5988_strides_0 = const()[name = string("op_5988_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5988_pad_0 = const()[name = string("op_5988_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5988_dilations_0 = const()[name = string("op_5988_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5988_groups_0 = const()[name = string("op_5988_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279583232))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280107584))))[name = string("layers_21_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_5988_cast_fp16 = conv(dilations = var_5988_dilations_0, groups = var_5988_groups_0, pad = var_5988_pad_0, pad_type = var_5988_pad_type_0, strides = var_5988_strides_0, weight = layers_21_self_attn_linear_out_weight_to_fp16_quantized, x = input_571_cast_fp16)[name = string("op_5988_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_479_cast_fp16 = add(x = x_467_cast_fp16, y = var_5988_cast_fp16)[name = string("x_479_cast_fp16")];
            tensor<int32, [1]> var_6004_axes_0 = const()[name = string("op_6004_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_21_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_21_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6004_cast_fp16 = layer_norm(axes = var_6004_axes_0, epsilon = layers_21_norm_conv_eps_scaled_to_fp16, x = x_479_cast_fp16)[name = string("op_6004_cast_fp16")];
            tensor<fp16, [1024]> input_573_gamma_0_to_fp16 = const()[name = string("input_573_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280109696)))];
            tensor<fp16, [1024]> input_573_beta_0_to_fp16 = const()[name = string("input_573_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280111808)))];
            fp16 input_573_epsilon_0_to_fp16 = const()[name = string("input_573_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_573_cast_fp16 = batch_norm(beta = input_573_beta_0_to_fp16, epsilon = input_573_epsilon_0_to_fp16, gamma = input_573_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6004_cast_fp16)[name = string("input_573_cast_fp16")];
            string input_575_pad_type_0 = const()[name = string("input_575_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_575_strides_0 = const()[name = string("input_575_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_575_pad_0 = const()[name = string("input_575_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_575_dilations_0 = const()[name = string("input_575_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_575_groups_0 = const()[name = string("input_575_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280113920))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281162560))))[name = string("layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_575_cast_fp16 = conv(dilations = input_575_dilations_0, groups = input_575_groups_0, pad = input_575_pad_0, pad_type = input_575_pad_type_0, strides = input_575_strides_0, weight = layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_573_cast_fp16)[name = string("input_575_cast_fp16")];
            int32 x_481_split_num_splits_0 = const()[name = string("x_481_split_num_splits_0"), val = int32(2)];
            int32 x_481_split_axis_0 = const()[name = string("x_481_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_481_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_481_split_cast_fp16_1 = split(axis = x_481_split_axis_0, num_splits = x_481_split_num_splits_0, x = input_575_cast_fp16)[name = string("x_481_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_481_split_1_sigmoid_cast_fp16 = sigmoid(x = x_481_split_cast_fp16_1)[name = string("x_481_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_481_cast_fp16 = mul(x = x_481_split_cast_fp16_0, y = x_481_split_1_sigmoid_cast_fp16)[name = string("x_481_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_577_cast_fp16 = mul(x = x_481_cast_fp16, y = pad_mask)[name = string("input_577_cast_fp16")];
            string input_579_pad_type_0 = const()[name = string("input_579_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_579_pad_0 = const()[name = string("input_579_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_579_groups_0 = const()[name = string("input_579_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_579_strides_0 = const()[name = string("input_579_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_579_dilations_0 = const()[name = string("input_579_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_145_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281166720))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281171392))))[name = string("const_145_to_fp16_quantized")];
            tensor<fp16, [1024]> const_146_to_fp16 = const()[name = string("const_146_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281173504)))];
            tensor<fp16, [1, 1024, 1, 188]> input_581_cast_fp16 = conv(bias = const_146_to_fp16, dilations = input_579_dilations_0, groups = input_579_groups_0, pad = input_579_pad_0, pad_type = input_579_pad_type_0, strides = input_579_strides_0, weight = const_145_to_fp16_quantized, x = input_577_cast_fp16)[name = string("input_581_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_583_cast_fp16 = silu(x = input_581_cast_fp16)[name = string("input_583_cast_fp16")];
            string var_6036_pad_type_0 = const()[name = string("op_6036_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6036_strides_0 = const()[name = string("op_6036_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6036_pad_0 = const()[name = string("op_6036_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6036_dilations_0 = const()[name = string("op_6036_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6036_groups_0 = const()[name = string("op_6036_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281175616))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281699968))))[name = string("layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6036_cast_fp16 = conv(dilations = var_6036_dilations_0, groups = var_6036_groups_0, pad = var_6036_pad_0, pad_type = var_6036_pad_type_0, strides = var_6036_strides_0, weight = layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_583_cast_fp16)[name = string("op_6036_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_483_cast_fp16 = add(x = x_479_cast_fp16, y = var_6036_cast_fp16)[name = string("x_483_cast_fp16")];
            tensor<int32, [1]> var_6052_axes_0 = const()[name = string("op_6052_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_21_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_21_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6052_cast_fp16 = layer_norm(axes = var_6052_axes_0, epsilon = layers_21_norm_feed_forward2_eps_scaled_to_fp16, x = x_483_cast_fp16)[name = string("op_6052_cast_fp16")];
            tensor<fp16, [1024]> input_585_gamma_0_to_fp16 = const()[name = string("input_585_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281702080)))];
            tensor<fp16, [1024]> input_585_beta_0_to_fp16 = const()[name = string("input_585_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281704192)))];
            fp16 input_585_epsilon_0_to_fp16 = const()[name = string("input_585_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_585_cast_fp16 = batch_norm(beta = input_585_beta_0_to_fp16, epsilon = input_585_epsilon_0_to_fp16, gamma = input_585_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6052_cast_fp16)[name = string("input_585_cast_fp16")];
            string input_587_pad_type_0 = const()[name = string("input_587_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_587_strides_0 = const()[name = string("input_587_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_587_pad_0 = const()[name = string("input_587_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_587_dilations_0 = const()[name = string("input_587_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_587_groups_0 = const()[name = string("input_587_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_21_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281706304))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283803520))))[name = string("layers_21_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_587_cast_fp16 = conv(dilations = input_587_dilations_0, groups = input_587_groups_0, pad = input_587_pad_0, pad_type = input_587_pad_type_0, strides = input_587_strides_0, weight = layers_21_feed_forward2_linear1_weight_to_fp16_quantized, x = input_585_cast_fp16)[name = string("input_587_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_589_cast_fp16 = silu(x = input_587_cast_fp16)[name = string("input_589_cast_fp16")];
            string var_6069_pad_type_0 = const()[name = string("op_6069_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6069_strides_0 = const()[name = string("op_6069_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6069_pad_0 = const()[name = string("op_6069_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6069_dilations_0 = const()[name = string("op_6069_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6069_groups_0 = const()[name = string("op_6069_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_6070_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283811776))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285908992))))[name = string("op_6070_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6070_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_6069_dilations_0, groups = var_6069_groups_0, pad = var_6069_pad_0, pad_type = var_6069_pad_type_0, strides = var_6069_strides_0, weight = op_6070_weight_0_to_fp16_quantized, x = input_589_cast_fp16)[name = string("op_6070_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_485_cast_fp16 = add(x = x_483_cast_fp16, y = var_6070_cast_fp16)[name = string("x_485_cast_fp16")];
            tensor<int32, [1]> var_6086_axes_0 = const()[name = string("op_6086_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_21_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_21_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6086_cast_fp16 = layer_norm(axes = var_6086_axes_0, epsilon = layers_21_norm_out_eps_scaled_to_fp16, x = x_485_cast_fp16)[name = string("op_6086_cast_fp16")];
            tensor<fp16, [1024]> x_487_gamma_0_to_fp16 = const()[name = string("x_487_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285911104)))];
            tensor<fp16, [1024]> x_487_beta_0_to_fp16 = const()[name = string("x_487_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285913216)))];
            fp16 x_487_epsilon_0_to_fp16 = const()[name = string("x_487_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_487_cast_fp16 = batch_norm(beta = x_487_beta_0_to_fp16, epsilon = x_487_epsilon_0_to_fp16, gamma = x_487_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6086_cast_fp16)[name = string("x_487_cast_fp16")];
            int32 var_6105 = const()[name = string("op_6105"), val = int32(1)];
            tensor<int32, [1]> var_6132_axes_0 = const()[name = string("op_6132_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_22_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_22_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6132_cast_fp16 = layer_norm(axes = var_6132_axes_0, epsilon = layers_22_norm_feed_forward1_eps_scaled_to_fp16, x = x_487_cast_fp16)[name = string("op_6132_cast_fp16")];
            tensor<fp16, [1024]> input_591_gamma_0_to_fp16 = const()[name = string("input_591_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285915328)))];
            tensor<fp16, [1024]> input_591_beta_0_to_fp16 = const()[name = string("input_591_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285917440)))];
            fp16 input_591_epsilon_0_to_fp16 = const()[name = string("input_591_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_591_cast_fp16 = batch_norm(beta = input_591_beta_0_to_fp16, epsilon = input_591_epsilon_0_to_fp16, gamma = input_591_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6132_cast_fp16)[name = string("input_591_cast_fp16")];
            string input_593_pad_type_0 = const()[name = string("input_593_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_593_strides_0 = const()[name = string("input_593_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_593_pad_0 = const()[name = string("input_593_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_593_dilations_0 = const()[name = string("input_593_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_593_groups_0 = const()[name = string("input_593_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_22_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285919552))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288016768))))[name = string("layers_22_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_593_cast_fp16 = conv(dilations = input_593_dilations_0, groups = input_593_groups_0, pad = input_593_pad_0, pad_type = input_593_pad_type_0, strides = input_593_strides_0, weight = layers_22_feed_forward1_linear1_weight_to_fp16_quantized, x = input_591_cast_fp16)[name = string("input_593_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_595_cast_fp16 = silu(x = input_593_cast_fp16)[name = string("input_595_cast_fp16")];
            string var_6149_pad_type_0 = const()[name = string("op_6149_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6149_strides_0 = const()[name = string("op_6149_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6149_pad_0 = const()[name = string("op_6149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6149_dilations_0 = const()[name = string("op_6149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6149_groups_0 = const()[name = string("op_6149_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_6150_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288025024))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290122240))))[name = string("op_6150_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6150_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_6149_dilations_0, groups = var_6149_groups_0, pad = var_6149_pad_0, pad_type = var_6149_pad_type_0, strides = var_6149_strides_0, weight = op_6150_weight_0_to_fp16_quantized, x = input_595_cast_fp16)[name = string("op_6150_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_489_cast_fp16 = add(x = x_487_cast_fp16, y = var_6150_cast_fp16)[name = string("x_489_cast_fp16")];
            tensor<int32, [1]> var_6166_axes_0 = const()[name = string("op_6166_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_22_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_22_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6166_cast_fp16 = layer_norm(axes = var_6166_axes_0, epsilon = layers_22_norm_self_att_eps_scaled_to_fp16, x = x_489_cast_fp16)[name = string("op_6166_cast_fp16")];
            tensor<fp16, [1024]> x_491_gamma_0_to_fp16 = const()[name = string("x_491_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290124352)))];
            tensor<fp16, [1024]> x_491_beta_0_to_fp16 = const()[name = string("x_491_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290126464)))];
            fp16 x_491_epsilon_0_to_fp16 = const()[name = string("x_491_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_491_cast_fp16 = batch_norm(beta = x_491_beta_0_to_fp16, epsilon = x_491_epsilon_0_to_fp16, gamma = x_491_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6166_cast_fp16)[name = string("x_491_cast_fp16")];
            string q_45_pad_type_0 = const()[name = string("q_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_45_strides_0 = const()[name = string("q_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_45_pad_0 = const()[name = string("q_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_45_dilations_0 = const()[name = string("q_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_45_groups_0 = const()[name = string("q_45_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290128576))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290652928))))[name = string("layers_22_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_45_cast_fp16 = conv(dilations = q_45_dilations_0, groups = q_45_groups_0, pad = q_45_pad_0, pad_type = q_45_pad_type_0, strides = q_45_strides_0, weight = layers_22_self_attn_linear_q_weight_to_fp16_quantized, x = x_491_cast_fp16)[name = string("q_45_cast_fp16")];
            string k_45_pad_type_0 = const()[name = string("k_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_45_strides_0 = const()[name = string("k_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_45_pad_0 = const()[name = string("k_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_45_dilations_0 = const()[name = string("k_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_45_groups_0 = const()[name = string("k_45_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290655040))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291179392))))[name = string("layers_22_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_45_cast_fp16 = conv(dilations = k_45_dilations_0, groups = k_45_groups_0, pad = k_45_pad_0, pad_type = k_45_pad_type_0, strides = k_45_strides_0, weight = layers_22_self_attn_linear_k_weight_to_fp16_quantized, x = x_491_cast_fp16)[name = string("k_45_cast_fp16")];
            string v_45_pad_type_0 = const()[name = string("v_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_45_strides_0 = const()[name = string("v_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_45_pad_0 = const()[name = string("v_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_45_dilations_0 = const()[name = string("v_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_45_groups_0 = const()[name = string("v_45_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291181504))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291705856))))[name = string("layers_22_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_45_cast_fp16 = conv(dilations = v_45_dilations_0, groups = v_45_groups_0, pad = v_45_pad_0, pad_type = v_45_pad_type_0, strides = v_45_strides_0, weight = layers_22_self_attn_linear_v_weight_to_fp16_quantized, x = x_491_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_45_to_fp16 = const()[name = string("bv_all_45_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291707968)))];
            tensor<fp16, [1, 1024, 1, 188]> var_6199_cast_fp16 = add(x = q_45_cast_fp16, y = bv_all_45_to_fp16)[name = string("op_6199_cast_fp16")];
            tensor<int32, [3]> var_6200 = const()[name = string("op_6200"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_45_cast_fp16 = reshape(shape = var_6200, x = var_6199_cast_fp16)[name = string("qb_45_cast_fp16")];
            bool bd_all_89_transpose_x_0 = const()[name = string("bd_all_89_transpose_x_0"), val = bool(false)];
            bool bd_all_89_transpose_y_0 = const()[name = string("bd_all_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_22_self_attn_pos_proj_to_fp16 = const()[name = string("layers_22_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291710080)))];
            tensor<fp16, [8, 375, 188]> bd_all_89_cast_fp16 = matmul(transpose_x = bd_all_89_transpose_x_0, transpose_y = bd_all_89_transpose_y_0, x = layers_22_self_attn_pos_proj_to_fp16, y = qb_45_cast_fp16)[name = string("bd_all_89_cast_fp16")];
            tensor<int32, [3]> x_493_perm_0 = const()[name = string("x_493_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_495_pad_0 = const()[name = string("x_495_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_495_mode_0 = const()[name = string("x_495_mode_0"), val = string("constant")];
            fp16 const_98_to_fp16 = const()[name = string("const_98_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_493_cast_fp16 = transpose(perm = x_493_perm_0, x = bd_all_89_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [8, 188, 376]> x_495_cast_fp16 = pad(constant_val = const_98_to_fp16, mode = x_495_mode_0, pad = x_495_pad_0, x = x_493_cast_fp16)[name = string("x_495_cast_fp16")];
            tensor<int32, [3]> var_6207 = const()[name = string("op_6207"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_497_cast_fp16 = reshape(shape = var_6207, x = x_495_cast_fp16)[name = string("x_497_cast_fp16")];
            tensor<int32, [3]> var_6210_begin_0 = const()[name = string("op_6210_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_6210_end_0 = const()[name = string("op_6210_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_6210_end_mask_0 = const()[name = string("op_6210_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_6210_cast_fp16 = slice_by_index(begin = var_6210_begin_0, end = var_6210_end_0, end_mask = var_6210_end_mask_0, x = x_497_cast_fp16)[name = string("op_6210_cast_fp16")];
            tensor<int32, [3]> var_6211 = const()[name = string("op_6211"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_499_cast_fp16 = reshape(shape = var_6211, x = var_6210_cast_fp16)[name = string("x_499_cast_fp16")];
            tensor<int32, [3]> bd_all_91_begin_0 = const()[name = string("bd_all_91_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_91_end_0 = const()[name = string("bd_all_91_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_91_end_mask_0 = const()[name = string("bd_all_91_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_91_cast_fp16 = slice_by_index(begin = bd_all_91_begin_0, end = bd_all_91_end_0, end_mask = bd_all_91_end_mask_0, x = x_499_cast_fp16)[name = string("bd_all_91_cast_fp16")];
            tensor<int32, [4]> var_6216 = const()[name = string("op_6216"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_6217_cast_fp16 = reshape(shape = var_6216, x = q_45_cast_fp16)[name = string("op_6217_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_6219_to_fp16 = const()[name = string("op_6219_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292478144)))];
            tensor<fp16, [8, 128, 1, 188]> qh_45_cast_fp16 = add(x = var_6217_cast_fp16, y = var_6219_to_fp16)[name = string("qh_45_cast_fp16")];
            tensor<int32, [4]> var_6221 = const()[name = string("op_6221"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_6222_cast_fp16 = reshape(shape = var_6221, x = k_45_cast_fp16)[name = string("op_6222_cast_fp16")];
            tensor<int32, [4]> var_6223 = const()[name = string("op_6223"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_6225 = const()[name = string("op_6225"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_6226_cast_fp16 = reshape(shape = var_6225, x = v_45_cast_fp16)[name = string("op_6226_cast_fp16")];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6229_perm_0 = const()[name = string("op_6229_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_45_axes_0 = const()[name = string("bd_45_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_6229_cast_fp16 = transpose(perm = var_6229_perm_0, x = bd_all_91_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [8, 188, 1, 188]> bd_45_cast_fp16 = expand_dims(axes = bd_45_axes_0, x = var_6229_cast_fp16)[name = string("bd_45_cast_fp16")];
            string ac_45_equation_0 = const()[name = string("ac_45_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_6224_cast_fp16 = transpose(perm = var_6223, x = var_6222_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [8, 188, 1, 188]> ac_45_cast_fp16 = einsum(equation = ac_45_equation_0, values = (var_6224_cast_fp16, qh_45_cast_fp16))[name = string("ac_45_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_6245_cast_fp16 = add(x = ac_45_cast_fp16, y = bd_45_cast_fp16)[name = string("op_6245_cast_fp16")];
            fp16 var_6246_to_fp16 = const()[name = string("op_6246_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_89_cast_fp16 = mul(x = var_6245_cast_fp16, y = var_6246_to_fp16)[name = string("scores_89_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = key_bias)[name = string("scores_91_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_6249_cast_fp16 = softmax(axis = var_6105, x = scores_91_cast_fp16)[name = string("op_6249_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_224 = const()[name = string("concat_224"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = var_6249_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [8, 188, 188]> reshape_66_cast_fp16 = reshape(shape = concat_224, x = transpose_44_cast_fp16)[name = string("reshape_66_cast_fp16")];
            tensor<int32, [3]> concat_225 = const()[name = string("concat_225"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_70_cast_fp16 = transpose(perm = transpose_70_perm_0, x = var_6226_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [8, 188, 128]> reshape_67_cast_fp16 = reshape(shape = concat_225, x = transpose_70_cast_fp16)[name = string("reshape_67_cast_fp16")];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = reshape_66_cast_fp16, y = reshape_67_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<int32, [4]> concat_229 = const()[name = string("concat_229"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_68_cast_fp16 = reshape(shape = concat_229, x = matmul_22_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [4]> ctx_45_perm_0 = const()[name = string("ctx_45_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_6252 = const()[name = string("op_6252"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_45_cast_fp16 = transpose(perm = ctx_45_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 1024, 1, 188]> input_597_cast_fp16 = reshape(shape = var_6252, x = ctx_45_cast_fp16)[name = string("input_597_cast_fp16")];
            string var_6259_pad_type_0 = const()[name = string("op_6259_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6259_strides_0 = const()[name = string("op_6259_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6259_pad_0 = const()[name = string("op_6259_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6259_dilations_0 = const()[name = string("op_6259_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6259_groups_0 = const()[name = string("op_6259_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292480256))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293004608))))[name = string("layers_22_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6259_cast_fp16 = conv(dilations = var_6259_dilations_0, groups = var_6259_groups_0, pad = var_6259_pad_0, pad_type = var_6259_pad_type_0, strides = var_6259_strides_0, weight = layers_22_self_attn_linear_out_weight_to_fp16_quantized, x = input_597_cast_fp16)[name = string("op_6259_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_501_cast_fp16 = add(x = x_489_cast_fp16, y = var_6259_cast_fp16)[name = string("x_501_cast_fp16")];
            tensor<int32, [1]> var_6275_axes_0 = const()[name = string("op_6275_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_22_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_22_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6275_cast_fp16 = layer_norm(axes = var_6275_axes_0, epsilon = layers_22_norm_conv_eps_scaled_to_fp16, x = x_501_cast_fp16)[name = string("op_6275_cast_fp16")];
            tensor<fp16, [1024]> input_599_gamma_0_to_fp16 = const()[name = string("input_599_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293006720)))];
            tensor<fp16, [1024]> input_599_beta_0_to_fp16 = const()[name = string("input_599_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293008832)))];
            fp16 input_599_epsilon_0_to_fp16 = const()[name = string("input_599_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_599_cast_fp16 = batch_norm(beta = input_599_beta_0_to_fp16, epsilon = input_599_epsilon_0_to_fp16, gamma = input_599_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6275_cast_fp16)[name = string("input_599_cast_fp16")];
            string input_601_pad_type_0 = const()[name = string("input_601_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_601_strides_0 = const()[name = string("input_601_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_601_pad_0 = const()[name = string("input_601_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_601_dilations_0 = const()[name = string("input_601_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_601_groups_0 = const()[name = string("input_601_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293010944))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294059584))))[name = string("layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_601_cast_fp16 = conv(dilations = input_601_dilations_0, groups = input_601_groups_0, pad = input_601_pad_0, pad_type = input_601_pad_type_0, strides = input_601_strides_0, weight = layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_599_cast_fp16)[name = string("input_601_cast_fp16")];
            int32 x_503_split_num_splits_0 = const()[name = string("x_503_split_num_splits_0"), val = int32(2)];
            int32 x_503_split_axis_0 = const()[name = string("x_503_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_503_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_503_split_cast_fp16_1 = split(axis = x_503_split_axis_0, num_splits = x_503_split_num_splits_0, x = input_601_cast_fp16)[name = string("x_503_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_503_split_1_sigmoid_cast_fp16 = sigmoid(x = x_503_split_cast_fp16_1)[name = string("x_503_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_503_cast_fp16 = mul(x = x_503_split_cast_fp16_0, y = x_503_split_1_sigmoid_cast_fp16)[name = string("x_503_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_603_cast_fp16 = mul(x = x_503_cast_fp16, y = pad_mask)[name = string("input_603_cast_fp16")];
            string input_605_pad_type_0 = const()[name = string("input_605_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_605_pad_0 = const()[name = string("input_605_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_605_groups_0 = const()[name = string("input_605_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_605_strides_0 = const()[name = string("input_605_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_605_dilations_0 = const()[name = string("input_605_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_147_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294063744))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294068416))))[name = string("const_147_to_fp16_quantized")];
            tensor<fp16, [1024]> const_148_to_fp16 = const()[name = string("const_148_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294070528)))];
            tensor<fp16, [1, 1024, 1, 188]> input_607_cast_fp16 = conv(bias = const_148_to_fp16, dilations = input_605_dilations_0, groups = input_605_groups_0, pad = input_605_pad_0, pad_type = input_605_pad_type_0, strides = input_605_strides_0, weight = const_147_to_fp16_quantized, x = input_603_cast_fp16)[name = string("input_607_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_609_cast_fp16 = silu(x = input_607_cast_fp16)[name = string("input_609_cast_fp16")];
            string var_6307_pad_type_0 = const()[name = string("op_6307_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6307_strides_0 = const()[name = string("op_6307_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6307_pad_0 = const()[name = string("op_6307_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6307_dilations_0 = const()[name = string("op_6307_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6307_groups_0 = const()[name = string("op_6307_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294072640))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294596992))))[name = string("layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6307_cast_fp16 = conv(dilations = var_6307_dilations_0, groups = var_6307_groups_0, pad = var_6307_pad_0, pad_type = var_6307_pad_type_0, strides = var_6307_strides_0, weight = layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_609_cast_fp16)[name = string("op_6307_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_505_cast_fp16 = add(x = x_501_cast_fp16, y = var_6307_cast_fp16)[name = string("x_505_cast_fp16")];
            tensor<int32, [1]> var_6323_axes_0 = const()[name = string("op_6323_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_22_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_22_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6323_cast_fp16 = layer_norm(axes = var_6323_axes_0, epsilon = layers_22_norm_feed_forward2_eps_scaled_to_fp16, x = x_505_cast_fp16)[name = string("op_6323_cast_fp16")];
            tensor<fp16, [1024]> input_611_gamma_0_to_fp16 = const()[name = string("input_611_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294599104)))];
            tensor<fp16, [1024]> input_611_beta_0_to_fp16 = const()[name = string("input_611_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294601216)))];
            fp16 input_611_epsilon_0_to_fp16 = const()[name = string("input_611_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_611_cast_fp16 = batch_norm(beta = input_611_beta_0_to_fp16, epsilon = input_611_epsilon_0_to_fp16, gamma = input_611_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6323_cast_fp16)[name = string("input_611_cast_fp16")];
            string input_613_pad_type_0 = const()[name = string("input_613_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_613_strides_0 = const()[name = string("input_613_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_613_pad_0 = const()[name = string("input_613_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_613_dilations_0 = const()[name = string("input_613_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_613_groups_0 = const()[name = string("input_613_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_22_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294603328))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296700544))))[name = string("layers_22_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_613_cast_fp16 = conv(dilations = input_613_dilations_0, groups = input_613_groups_0, pad = input_613_pad_0, pad_type = input_613_pad_type_0, strides = input_613_strides_0, weight = layers_22_feed_forward2_linear1_weight_to_fp16_quantized, x = input_611_cast_fp16)[name = string("input_613_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_615_cast_fp16 = silu(x = input_613_cast_fp16)[name = string("input_615_cast_fp16")];
            string var_6340_pad_type_0 = const()[name = string("op_6340_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6340_strides_0 = const()[name = string("op_6340_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6340_pad_0 = const()[name = string("op_6340_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6340_dilations_0 = const()[name = string("op_6340_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6340_groups_0 = const()[name = string("op_6340_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_6341_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296708800))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298806016))))[name = string("op_6341_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6341_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_6340_dilations_0, groups = var_6340_groups_0, pad = var_6340_pad_0, pad_type = var_6340_pad_type_0, strides = var_6340_strides_0, weight = op_6341_weight_0_to_fp16_quantized, x = input_615_cast_fp16)[name = string("op_6341_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_507_cast_fp16 = add(x = x_505_cast_fp16, y = var_6341_cast_fp16)[name = string("x_507_cast_fp16")];
            tensor<int32, [1]> var_6357_axes_0 = const()[name = string("op_6357_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_22_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_22_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6357_cast_fp16 = layer_norm(axes = var_6357_axes_0, epsilon = layers_22_norm_out_eps_scaled_to_fp16, x = x_507_cast_fp16)[name = string("op_6357_cast_fp16")];
            tensor<fp16, [1024]> x_509_gamma_0_to_fp16 = const()[name = string("x_509_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298808128)))];
            tensor<fp16, [1024]> x_509_beta_0_to_fp16 = const()[name = string("x_509_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298810240)))];
            fp16 x_509_epsilon_0_to_fp16 = const()[name = string("x_509_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_509_cast_fp16 = batch_norm(beta = x_509_beta_0_to_fp16, epsilon = x_509_epsilon_0_to_fp16, gamma = x_509_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6357_cast_fp16)[name = string("x_509_cast_fp16")];
            int32 var_6376 = const()[name = string("op_6376"), val = int32(1)];
            tensor<int32, [1]> var_6403_axes_0 = const()[name = string("op_6403_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_23_norm_feed_forward1_eps_scaled_to_fp16 = const()[name = string("layers_23_norm_feed_forward1_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6403_cast_fp16 = layer_norm(axes = var_6403_axes_0, epsilon = layers_23_norm_feed_forward1_eps_scaled_to_fp16, x = x_509_cast_fp16)[name = string("op_6403_cast_fp16")];
            tensor<fp16, [1024]> input_617_gamma_0_to_fp16 = const()[name = string("input_617_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298812352)))];
            tensor<fp16, [1024]> input_617_beta_0_to_fp16 = const()[name = string("input_617_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298814464)))];
            fp16 input_617_epsilon_0_to_fp16 = const()[name = string("input_617_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_617_cast_fp16 = batch_norm(beta = input_617_beta_0_to_fp16, epsilon = input_617_epsilon_0_to_fp16, gamma = input_617_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6403_cast_fp16)[name = string("input_617_cast_fp16")];
            string input_619_pad_type_0 = const()[name = string("input_619_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_619_strides_0 = const()[name = string("input_619_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_619_pad_0 = const()[name = string("input_619_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_619_dilations_0 = const()[name = string("input_619_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_619_groups_0 = const()[name = string("input_619_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_23_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298816576))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300913792))))[name = string("layers_23_feed_forward1_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_619_cast_fp16 = conv(dilations = input_619_dilations_0, groups = input_619_groups_0, pad = input_619_pad_0, pad_type = input_619_pad_type_0, strides = input_619_strides_0, weight = layers_23_feed_forward1_linear1_weight_to_fp16_quantized, x = input_617_cast_fp16)[name = string("input_619_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_621_cast_fp16 = silu(x = input_619_cast_fp16)[name = string("input_621_cast_fp16")];
            string var_6420_pad_type_0 = const()[name = string("op_6420_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6420_strides_0 = const()[name = string("op_6420_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6420_pad_0 = const()[name = string("op_6420_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6420_dilations_0 = const()[name = string("op_6420_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6420_groups_0 = const()[name = string("op_6420_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_6421_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300922048))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303019264))))[name = string("op_6421_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6421_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_6420_dilations_0, groups = var_6420_groups_0, pad = var_6420_pad_0, pad_type = var_6420_pad_type_0, strides = var_6420_strides_0, weight = op_6421_weight_0_to_fp16_quantized, x = input_621_cast_fp16)[name = string("op_6421_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_511_cast_fp16 = add(x = x_509_cast_fp16, y = var_6421_cast_fp16)[name = string("x_511_cast_fp16")];
            tensor<int32, [1]> var_6437_axes_0 = const()[name = string("op_6437_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_23_norm_self_att_eps_scaled_to_fp16 = const()[name = string("layers_23_norm_self_att_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6437_cast_fp16 = layer_norm(axes = var_6437_axes_0, epsilon = layers_23_norm_self_att_eps_scaled_to_fp16, x = x_511_cast_fp16)[name = string("op_6437_cast_fp16")];
            tensor<fp16, [1024]> x_513_gamma_0_to_fp16 = const()[name = string("x_513_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303021376)))];
            tensor<fp16, [1024]> x_513_beta_0_to_fp16 = const()[name = string("x_513_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303023488)))];
            fp16 x_513_epsilon_0_to_fp16 = const()[name = string("x_513_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> x_513_cast_fp16 = batch_norm(beta = x_513_beta_0_to_fp16, epsilon = x_513_epsilon_0_to_fp16, gamma = x_513_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6437_cast_fp16)[name = string("x_513_cast_fp16")];
            string q_pad_type_0 = const()[name = string("q_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> q_strides_0 = const()[name = string("q_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> q_pad_0 = const()[name = string("q_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> q_dilations_0 = const()[name = string("q_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 q_groups_0 = const()[name = string("q_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303025600))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303549952))))[name = string("layers_23_self_attn_linear_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> q_cast_fp16 = conv(dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = layers_23_self_attn_linear_q_weight_to_fp16_quantized, x = x_513_cast_fp16)[name = string("q_cast_fp16")];
            string k_pad_type_0 = const()[name = string("k_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> k_strides_0 = const()[name = string("k_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> k_pad_0 = const()[name = string("k_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> k_dilations_0 = const()[name = string("k_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 k_groups_0 = const()[name = string("k_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303552064))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304076416))))[name = string("layers_23_self_attn_linear_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = layers_23_self_attn_linear_k_weight_to_fp16_quantized, x = x_513_cast_fp16)[name = string("k_cast_fp16")];
            string v_pad_type_0 = const()[name = string("v_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> v_strides_0 = const()[name = string("v_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_pad_0 = const()[name = string("v_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_dilations_0 = const()[name = string("v_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 v_groups_0 = const()[name = string("v_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304078528))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304602880))))[name = string("layers_23_self_attn_linear_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> v_cast_fp16 = conv(dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = layers_23_self_attn_linear_v_weight_to_fp16_quantized, x = x_513_cast_fp16)[name = string("v_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> bv_all_to_fp16 = const()[name = string("bv_all_to_fp16"), val = tensor<fp16, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304604992)))];
            tensor<fp16, [1, 1024, 1, 188]> var_6470_cast_fp16 = add(x = q_cast_fp16, y = bv_all_to_fp16)[name = string("op_6470_cast_fp16")];
            tensor<int32, [3]> var_6471 = const()[name = string("op_6471"), val = tensor<int32, [3]>([8, 128, 188])];
            tensor<fp16, [8, 128, 188]> qb_cast_fp16 = reshape(shape = var_6471, x = var_6470_cast_fp16)[name = string("qb_cast_fp16")];
            bool bd_all_93_transpose_x_0 = const()[name = string("bd_all_93_transpose_x_0"), val = bool(false)];
            bool bd_all_93_transpose_y_0 = const()[name = string("bd_all_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 375, 128]> layers_23_self_attn_pos_proj_to_fp16 = const()[name = string("layers_23_self_attn_pos_proj_to_fp16"), val = tensor<fp16, [8, 375, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304607104)))];
            tensor<fp16, [8, 375, 188]> bd_all_93_cast_fp16 = matmul(transpose_x = bd_all_93_transpose_x_0, transpose_y = bd_all_93_transpose_y_0, x = layers_23_self_attn_pos_proj_to_fp16, y = qb_cast_fp16)[name = string("bd_all_93_cast_fp16")];
            tensor<int32, [3]> x_515_perm_0 = const()[name = string("x_515_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [6]> x_517_pad_0 = const()[name = string("x_517_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
            string x_517_mode_0 = const()[name = string("x_517_mode_0"), val = string("constant")];
            fp16 const_102_to_fp16 = const()[name = string("const_102_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [8, 188, 375]> x_515_cast_fp16 = transpose(perm = x_515_perm_0, x = bd_all_93_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [8, 188, 376]> x_517_cast_fp16 = pad(constant_val = const_102_to_fp16, mode = x_517_mode_0, pad = x_517_pad_0, x = x_515_cast_fp16)[name = string("x_517_cast_fp16")];
            tensor<int32, [3]> var_6478 = const()[name = string("op_6478"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<fp16, [8, 376, 188]> x_519_cast_fp16 = reshape(shape = var_6478, x = x_517_cast_fp16)[name = string("x_519_cast_fp16")];
            tensor<int32, [3]> var_6481_begin_0 = const()[name = string("op_6481_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
            tensor<int32, [3]> var_6481_end_0 = const()[name = string("op_6481_end_0"), val = tensor<int32, [3]>([8, 376, 188])];
            tensor<bool, [3]> var_6481_end_mask_0 = const()[name = string("op_6481_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 375, 188]> var_6481_cast_fp16 = slice_by_index(begin = var_6481_begin_0, end = var_6481_end_0, end_mask = var_6481_end_mask_0, x = x_519_cast_fp16)[name = string("op_6481_cast_fp16")];
            tensor<int32, [3]> var_6482 = const()[name = string("op_6482"), val = tensor<int32, [3]>([8, 188, 375])];
            tensor<fp16, [8, 188, 375]> x_521_cast_fp16 = reshape(shape = var_6482, x = var_6481_cast_fp16)[name = string("x_521_cast_fp16")];
            tensor<int32, [3]> bd_all_begin_0 = const()[name = string("bd_all_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> bd_all_end_0 = const()[name = string("bd_all_end_0"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<bool, [3]> bd_all_end_mask_0 = const()[name = string("bd_all_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 188, 188]> bd_all_cast_fp16 = slice_by_index(begin = bd_all_begin_0, end = bd_all_end_0, end_mask = bd_all_end_mask_0, x = x_521_cast_fp16)[name = string("bd_all_cast_fp16")];
            tensor<int32, [4]> var_6487 = const()[name = string("op_6487"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_6488_cast_fp16 = reshape(shape = var_6487, x = q_cast_fp16)[name = string("op_6488_cast_fp16")];
            tensor<fp16, [8, 128, 1, 1]> var_6490_to_fp16 = const()[name = string("op_6490_to_fp16"), val = tensor<fp16, [8, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305375168)))];
            tensor<fp16, [8, 128, 1, 188]> qh_cast_fp16 = add(x = var_6488_cast_fp16, y = var_6490_to_fp16)[name = string("qh_cast_fp16")];
            tensor<int32, [4]> var_6492 = const()[name = string("op_6492"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_6493_cast_fp16 = reshape(shape = var_6492, x = k_cast_fp16)[name = string("op_6493_cast_fp16")];
            tensor<int32, [4]> var_6494 = const()[name = string("op_6494"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [4]> var_6496 = const()[name = string("op_6496"), val = tensor<int32, [4]>([8, 128, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> var_6497_cast_fp16 = reshape(shape = var_6496, x = v_cast_fp16)[name = string("op_6497_cast_fp16")];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6500_perm_0 = const()[name = string("op_6500_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> bd_axes_0 = const()[name = string("bd_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [8, 188, 188]> var_6500_cast_fp16 = transpose(perm = var_6500_perm_0, x = bd_all_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [8, 188, 1, 188]> bd_cast_fp16 = expand_dims(axes = bd_axes_0, x = var_6500_cast_fp16)[name = string("bd_cast_fp16")];
            string ac_equation_0 = const()[name = string("ac_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [8, 188, 1, 128]> var_6495_cast_fp16 = transpose(perm = var_6494, x = var_6493_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [8, 188, 1, 188]> ac_cast_fp16 = einsum(equation = ac_equation_0, values = (var_6495_cast_fp16, qh_cast_fp16))[name = string("ac_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_6516_cast_fp16 = add(x = ac_cast_fp16, y = bd_cast_fp16)[name = string("op_6516_cast_fp16")];
            fp16 var_6517_to_fp16 = const()[name = string("op_6517_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [8, 188, 1, 188]> scores_93_cast_fp16 = mul(x = var_6516_cast_fp16, y = var_6517_to_fp16)[name = string("scores_93_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> scores_cast_fp16 = add(x = scores_93_cast_fp16, y = key_bias)[name = string("scores_cast_fp16")];
            tensor<fp16, [8, 188, 1, 188]> var_6520_cast_fp16 = softmax(axis = var_6376, x = scores_cast_fp16)[name = string("op_6520_cast_fp16")];
            tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_234 = const()[name = string("concat_234"), val = tensor<int32, [3]>([8, 188, 188])];
            tensor<fp16, [8, 1, 188, 188]> transpose_46_cast_fp16 = transpose(perm = transpose_46_perm_0, x = var_6520_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [8, 188, 188]> reshape_69_cast_fp16 = reshape(shape = concat_234, x = transpose_46_cast_fp16)[name = string("reshape_69_cast_fp16")];
            tensor<int32, [3]> concat_235 = const()[name = string("concat_235"), val = tensor<int32, [3]>([8, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = var_6497_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [8, 188, 128]> reshape_70_cast_fp16 = reshape(shape = concat_235, x = transpose_71_cast_fp16)[name = string("reshape_70_cast_fp16")];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [8, 188, 128]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = reshape_69_cast_fp16, y = reshape_70_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<int32, [4]> concat_239 = const()[name = string("concat_239"), val = tensor<int32, [4]>([8, 1, 188, 128])];
            tensor<fp16, [8, 1, 188, 128]> reshape_71_cast_fp16 = reshape(shape = concat_239, x = matmul_23_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> ctx_perm_0 = const()[name = string("ctx_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_6523 = const()[name = string("op_6523"), val = tensor<int32, [4]>([1, 1024, 1, 188])];
            tensor<fp16, [8, 128, 1, 188]> ctx_cast_fp16 = transpose(perm = ctx_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 1024, 1, 188]> input_623_cast_fp16 = reshape(shape = var_6523, x = ctx_cast_fp16)[name = string("input_623_cast_fp16")];
            string var_6530_pad_type_0 = const()[name = string("op_6530_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6530_strides_0 = const()[name = string("op_6530_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6530_pad_0 = const()[name = string("op_6530_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6530_dilations_0 = const()[name = string("op_6530_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6530_groups_0 = const()[name = string("op_6530_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305377280))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305901632))))[name = string("layers_23_self_attn_linear_out_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6530_cast_fp16 = conv(dilations = var_6530_dilations_0, groups = var_6530_groups_0, pad = var_6530_pad_0, pad_type = var_6530_pad_type_0, strides = var_6530_strides_0, weight = layers_23_self_attn_linear_out_weight_to_fp16_quantized, x = input_623_cast_fp16)[name = string("op_6530_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_523_cast_fp16 = add(x = x_511_cast_fp16, y = var_6530_cast_fp16)[name = string("x_523_cast_fp16")];
            tensor<int32, [1]> var_6546_axes_0 = const()[name = string("op_6546_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_23_norm_conv_eps_scaled_to_fp16 = const()[name = string("layers_23_norm_conv_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6546_cast_fp16 = layer_norm(axes = var_6546_axes_0, epsilon = layers_23_norm_conv_eps_scaled_to_fp16, x = x_523_cast_fp16)[name = string("op_6546_cast_fp16")];
            tensor<fp16, [1024]> input_625_gamma_0_to_fp16 = const()[name = string("input_625_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305903744)))];
            tensor<fp16, [1024]> input_625_beta_0_to_fp16 = const()[name = string("input_625_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305905856)))];
            fp16 input_625_epsilon_0_to_fp16 = const()[name = string("input_625_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_625_cast_fp16 = batch_norm(beta = input_625_beta_0_to_fp16, epsilon = input_625_epsilon_0_to_fp16, gamma = input_625_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6546_cast_fp16)[name = string("input_625_cast_fp16")];
            string input_627_pad_type_0 = const()[name = string("input_627_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_627_strides_0 = const()[name = string("input_627_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_627_pad_0 = const()[name = string("input_627_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_627_dilations_0 = const()[name = string("input_627_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_627_groups_0 = const()[name = string("input_627_groups_0"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305907968))), scale = tensor<fp16, [2048, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306956608))))[name = string("layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 1, 188]> input_627_cast_fp16 = conv(dilations = input_627_dilations_0, groups = input_627_groups_0, pad = input_627_pad_0, pad_type = input_627_pad_type_0, strides = input_627_strides_0, weight = layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_625_cast_fp16)[name = string("input_627_cast_fp16")];
            int32 x_525_split_num_splits_0 = const()[name = string("x_525_split_num_splits_0"), val = int32(2)];
            int32 x_525_split_axis_0 = const()[name = string("x_525_split_axis_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 188]> x_525_split_cast_fp16_0, tensor<fp16, [1, 1024, 1, 188]> x_525_split_cast_fp16_1 = split(axis = x_525_split_axis_0, num_splits = x_525_split_num_splits_0, x = input_627_cast_fp16)[name = string("x_525_split_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_525_split_1_sigmoid_cast_fp16 = sigmoid(x = x_525_split_cast_fp16_1)[name = string("x_525_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_525_cast_fp16 = mul(x = x_525_split_cast_fp16_0, y = x_525_split_1_sigmoid_cast_fp16)[name = string("x_525_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_629_cast_fp16 = mul(x = x_525_cast_fp16, y = pad_mask)[name = string("input_629_cast_fp16")];
            string input_631_pad_type_0 = const()[name = string("input_631_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_631_pad_0 = const()[name = string("input_631_pad_0"), val = tensor<int32, [4]>([0, 0, 4, 4])];
            int32 input_631_groups_0 = const()[name = string("input_631_groups_0"), val = int32(1024)];
            tensor<int32, [2]> input_631_strides_0 = const()[name = string("input_631_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_631_dilations_0 = const()[name = string("input_631_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1024, 1, 1, 9]> const_149_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306960768))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306965440))))[name = string("const_149_to_fp16_quantized")];
            tensor<fp16, [1024]> const_150_to_fp16 = const()[name = string("const_150_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306967552)))];
            tensor<fp16, [1, 1024, 1, 188]> input_633_cast_fp16 = conv(bias = const_150_to_fp16, dilations = input_631_dilations_0, groups = input_631_groups_0, pad = input_631_pad_0, pad_type = input_631_pad_type_0, strides = input_631_strides_0, weight = const_149_to_fp16_quantized, x = input_629_cast_fp16)[name = string("input_633_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> input_635_cast_fp16 = silu(x = input_633_cast_fp16)[name = string("input_635_cast_fp16")];
            string var_6578_pad_type_0 = const()[name = string("op_6578_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6578_strides_0 = const()[name = string("op_6578_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6578_pad_0 = const()[name = string("op_6578_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6578_dilations_0 = const()[name = string("op_6578_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6578_groups_0 = const()[name = string("op_6578_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306969664))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307494016))))[name = string("layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6578_cast_fp16 = conv(dilations = var_6578_dilations_0, groups = var_6578_groups_0, pad = var_6578_pad_0, pad_type = var_6578_pad_type_0, strides = var_6578_strides_0, weight = layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_635_cast_fp16)[name = string("op_6578_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_527_cast_fp16 = add(x = x_523_cast_fp16, y = var_6578_cast_fp16)[name = string("x_527_cast_fp16")];
            tensor<int32, [1]> var_6594_axes_0 = const()[name = string("op_6594_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_23_norm_feed_forward2_eps_scaled_to_fp16 = const()[name = string("layers_23_norm_feed_forward2_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6594_cast_fp16 = layer_norm(axes = var_6594_axes_0, epsilon = layers_23_norm_feed_forward2_eps_scaled_to_fp16, x = x_527_cast_fp16)[name = string("op_6594_cast_fp16")];
            tensor<fp16, [1024]> input_637_gamma_0_to_fp16 = const()[name = string("input_637_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307496128)))];
            tensor<fp16, [1024]> input_637_beta_0_to_fp16 = const()[name = string("input_637_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307498240)))];
            fp16 input_637_epsilon_0_to_fp16 = const()[name = string("input_637_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_637_cast_fp16 = batch_norm(beta = input_637_beta_0_to_fp16, epsilon = input_637_epsilon_0_to_fp16, gamma = input_637_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6594_cast_fp16)[name = string("input_637_cast_fp16")];
            string input_639_pad_type_0 = const()[name = string("input_639_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_639_strides_0 = const()[name = string("input_639_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_639_pad_0 = const()[name = string("input_639_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_639_dilations_0 = const()[name = string("input_639_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_639_groups_0 = const()[name = string("input_639_groups_0"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_23_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307500352))), scale = tensor<fp16, [4096, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309597568))))[name = string("layers_23_feed_forward2_linear1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 4096, 1, 188]> input_639_cast_fp16 = conv(dilations = input_639_dilations_0, groups = input_639_groups_0, pad = input_639_pad_0, pad_type = input_639_pad_type_0, strides = input_639_strides_0, weight = layers_23_feed_forward2_linear1_weight_to_fp16_quantized, x = input_637_cast_fp16)[name = string("input_639_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 188]> input_641_cast_fp16 = silu(x = input_639_cast_fp16)[name = string("input_641_cast_fp16")];
            string var_6611_pad_type_0 = const()[name = string("op_6611_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6611_strides_0 = const()[name = string("op_6611_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6611_pad_0 = const()[name = string("op_6611_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6611_dilations_0 = const()[name = string("op_6611_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6611_groups_0 = const()[name = string("op_6611_groups_0"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> op_6612_weight_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309605824))), scale = tensor<fp16, [1024, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311703040))))[name = string("op_6612_weight_0_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 1, 188]> var_6612_cast_fp16 = conv(bias = input_19_mean_0_to_fp16, dilations = var_6611_dilations_0, groups = var_6611_groups_0, pad = var_6611_pad_0, pad_type = var_6611_pad_type_0, strides = var_6611_strides_0, weight = op_6612_weight_0_to_fp16_quantized, x = input_641_cast_fp16)[name = string("op_6612_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 188]> x_cast_fp16 = add(x = x_527_cast_fp16, y = var_6612_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [1]> var_6628_axes_0 = const()[name = string("op_6628_axes_0"), val = tensor<int32, [1]>([1])];
            fp16 layers_23_norm_out_eps_scaled_to_fp16 = const()[name = string("layers_23_norm_out_eps_scaled_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> var_6628_cast_fp16 = layer_norm(axes = var_6628_axes_0, epsilon = layers_23_norm_out_eps_scaled_to_fp16, x = x_cast_fp16)[name = string("op_6628_cast_fp16")];
            tensor<fp16, [1024]> input_gamma_0_to_fp16 = const()[name = string("input_gamma_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311705152)))];
            tensor<fp16, [1024]> input_beta_0_to_fp16 = const()[name = string("input_beta_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311707264)))];
            fp16 input_epsilon_0_to_fp16 = const()[name = string("input_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 1024, 1, 188]> input_cast_fp16 = batch_norm(beta = input_beta_0_to_fp16, epsilon = input_epsilon_0_to_fp16, gamma = input_gamma_0_to_fp16, mean = input_19_mean_0_to_fp16, variance = input_19_variance_0_to_fp16, x = var_6628_cast_fp16)[name = string("input_cast_fp16")];
            string var_6641_pad_type_0 = const()[name = string("op_6641_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_6641_strides_0 = const()[name = string("op_6641_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_6641_pad_0 = const()[name = string("op_6641_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_6641_dilations_0 = const()[name = string("op_6641_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_6641_groups_0 = const()[name = string("op_6641_groups_0"), val = int32(1)];
            tensor<fp16, [640, 1024, 1, 1]> joint_enc_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int4, [640, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311709376))), scale = tensor<fp16, [640, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312037120))))[name = string("joint_enc_weight_to_fp16_quantized")];
            tensor<fp16, [640]> joint_enc_bias_to_fp16 = const()[name = string("joint_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312038464)))];
            tensor<fp16, [1, 640, 1, 188]> enc_proj = conv(bias = joint_enc_bias_to_fp16, dilations = var_6641_dilations_0, groups = var_6641_groups_0, pad = var_6641_pad_0, pad_type = var_6641_pad_type_0, strides = var_6641_strides_0, weight = joint_enc_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("op_6641_cast_fp16")];
        } -> (enc_proj);
}